program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
{
    func len128<ios18>(tensor<int32, [1, 128]> attention_mask, tensor<int32, [1, 128]> input_ids) {
            fp16 var_15_to_fp16 = const()[name = string("op_15_to_fp16"), val = fp16(0x1p+0)];
            string keys_to_fp16_dtype_0 = const()[name = string("keys_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128]> attention_mask_to_fp16 = cast(dtype = keys_to_fp16_dtype_0, x = attention_mask)[name = string("cast_60")];
            tensor<fp16, [1, 128]> var_17_cast_fp16 = sub(x = var_15_to_fp16, y = attention_mask_to_fp16)[name = string("op_17_cast_fp16")];
            tensor<int32, [1]> var_24_axes_0 = const()[name = string("op_24_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 128]> var_24_cast_fp16 = expand_dims(axes = var_24_axes_0, x = var_17_cast_fp16)[name = string("op_24_cast_fp16")];
            tensor<int32, [1]> var_26_axes_0 = const()[name = string("op_26_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 128]> var_26_cast_fp16 = expand_dims(axes = var_26_axes_0, x = var_24_cast_fp16)[name = string("op_26_cast_fp16")];
            fp16 var_32_to_fp16 = const()[name = string("op_32_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 1, 1, 128]> key_bias_cast_fp16 = mul(x = var_26_cast_fp16, y = var_32_to_fp16)[name = string("key_bias_cast_fp16")];
            tensor<fp16, [1, 1, 128, 128]> outside_to_fp16 = const()[name = string("outside_to_fp16"), val = tensor<fp16, [1, 1, 128, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [1, 1, 128, 128]> attention_mask_cast_fp16 = add(x = key_bias_cast_fp16, y = outside_to_fp16)[name = string("attention_mask_cast_fp16")];
            int32 var_49 = const()[name = string("op_49"), val = int32(-1)];
            int32 input_1_batch_dims_0 = const()[name = string("input_1_batch_dims_0"), val = int32(0)];
            bool input_1_validate_indices_0 = const()[name = string("input_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [50368, 1024]> model_model_embeddings_tok_embeddings_weight_to_fp16 = const()[name = string("model_model_embeddings_tok_embeddings_weight_to_fp16"), val = tensor<fp16, [50368, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32896)))];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 128]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(50368)];
            tensor<int32, [1, 128]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 128]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 input_1_cast_fp16_axis_0 = const()[name = string("input_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 128, 1024]> input_1_cast_fp16 = gather(axis = input_1_cast_fp16_axis_0, batch_dims = input_1_batch_dims_0, indices = select_0, validate_indices = input_1_validate_indices_0, x = model_model_embeddings_tok_embeddings_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<int32, [1]> input_3_axes_0 = const()[name = string("input_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_embeddings_norm_weight_to_fp16 = const()[name = string("model_model_embeddings_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103186624)))];
            fp16 var_52_to_fp16 = const()[name = string("op_52_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 128, 1024]> input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, epsilon = var_52_to_fp16, gamma = model_model_embeddings_norm_weight_to_fp16, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_0_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_0_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103188736)))];
            tensor<fp16, [3072]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109480256)))];
            tensor<fp16, [1, 128, 3072]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_0_attn_Wqkv_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [5]> var_199 = const()[name = string("op_199"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_3_cast_fp16 = reshape(shape = var_199, x = linear_0_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> var_201_split_sizes_0 = const()[name = string("op_201_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_201_axis_0 = const()[name = string("op_201_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_201_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_201_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_201_cast_fp16_2 = split(axis = var_201_axis_0, split_sizes = var_201_split_sizes_0, x = qkv_3_cast_fp16)[name = string("op_201_cast_fp16")];
            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = string("squeeze_0_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = var_201_cast_fp16_0)[name = string("squeeze_0_cast_fp16")];
            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = string("squeeze_1_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = var_201_cast_fp16_1)[name = string("squeeze_1_cast_fp16")];
            tensor<int32, [1]> squeeze_2_axes_0 = const()[name = string("squeeze_2_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_2_cast_fp16 = squeeze(axes = squeeze_2_axes_0, x = var_201_cast_fp16_2)[name = string("squeeze_2_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_1_perm_0 = const()[name = string("value_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 1, 128, 64]> cos_7_to_fp16 = const()[name = string("cos_7_to_fp16"), val = tensor<fp16, [1, 1, 128, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109486464)))];
            tensor<fp16, [1, 16, 128, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = squeeze_0_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 16, 128, 64]> var_211_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_7_to_fp16)[name = string("op_211_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_223_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_223_cast_fp16")];
            bool var_225_interleave_0 = const()[name = string("op_225_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_225_cast_fp16 = concat(axis = var_49, interleave = var_225_interleave_0, values = (var_223_cast_fp16, x1_1_cast_fp16))[name = string("op_225_cast_fp16")];
            tensor<fp16, [1, 1, 128, 64]> sin_7_to_fp16 = const()[name = string("sin_7_to_fp16"), val = tensor<fp16, [1, 1, 128, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109502912)))];
            tensor<fp16, [1, 16, 128, 64]> var_226_cast_fp16 = mul(x = var_225_cast_fp16, y = sin_7_to_fp16)[name = string("op_226_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_1_cast_fp16 = add(x = var_211_cast_fp16, y = var_226_cast_fp16)[name = string("q_embed_1_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = squeeze_1_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 128, 64]> var_229_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_7_to_fp16)[name = string("op_229_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_241_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_241_cast_fp16")];
            bool var_243_interleave_0 = const()[name = string("op_243_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_243_cast_fp16 = concat(axis = var_49, interleave = var_243_interleave_0, values = (var_241_cast_fp16, x1_3_cast_fp16))[name = string("op_243_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_244_cast_fp16 = mul(x = var_243_cast_fp16, y = sin_7_to_fp16)[name = string("op_244_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_1_cast_fp16 = add(x = var_229_cast_fp16, y = var_244_cast_fp16)[name = string("k_embed_1_cast_fp16")];
            bool var_249_transpose_x_1 = const()[name = string("op_249_transpose_x_1"), val = bool(false)];
            bool var_249_transpose_y_1 = const()[name = string("op_249_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_249_cast_fp16 = matmul(transpose_x = var_249_transpose_x_1, transpose_y = var_249_transpose_y_1, x = q_embed_1_cast_fp16, y = k_embed_1_cast_fp16)[name = string("op_249_cast_fp16")];
            fp16 var_250_to_fp16 = const()[name = string("op_250_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_1_cast_fp16 = mul(x = var_249_cast_fp16, y = var_250_to_fp16)[name = string("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_5_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = key_bias_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_253_cast_fp16 = softmax(axis = var_49, x = input_5_cast_fp16)[name = string("op_253_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_1_cast_fp16 = transpose(perm = value_1_perm_0, x = squeeze_2_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = var_253_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_257_perm_0 = const()[name = string("op_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_259 = const()[name = string("op_259"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_257_cast_fp16 = transpose(perm = var_257_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 128, 1024]> var_260_cast_fp16 = reshape(shape = var_259, x = var_257_cast_fp16)[name = string("op_260_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_0_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_0_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109519360)))];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111616576)))];
            tensor<fp16, [1, 128, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_0_attn_Wo_weight_to_fp16, x = var_260_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_11_cast_fp16 = add(x = input_3_cast_fp16, y = linear_1_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<int32, [1]> input_13_axes_0 = const()[name = string("input_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_0_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_0_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111618688)))];
            tensor<fp16, [1, 128, 1024]> input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_0_mlp_norm_weight_to_fp16, x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_0_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_0_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111620800)))];
            tensor<fp16, [5248]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5248]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122368768)))];
            tensor<fp16, [1, 128, 5248]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_0_mlp_Wi_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [2]> var_272_split_sizes_0 = const()[name = string("op_272_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_272_axis_0 = const()[name = string("op_272_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_272_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_272_cast_fp16_1 = split(axis = var_272_axis_0, split_sizes = var_272_split_sizes_0, x = linear_2_cast_fp16)[name = string("op_272_cast_fp16")];
            string var_274_mode_0 = const()[name = string("op_274_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_274_cast_fp16 = gelu(mode = var_274_mode_0, x = var_272_cast_fp16_0)[name = string("op_274_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_17_cast_fp16 = mul(x = var_274_cast_fp16, y = var_272_cast_fp16_1)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_0_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_0_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122379328)))];
            tensor<fp16, [1, 128, 1024]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_0_mlp_Wo_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_21_cast_fp16 = add(x = input_11_cast_fp16, y = linear_3_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [1]> hidden_states_3_axes_0 = const()[name = string("hidden_states_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_1_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_1_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127753344)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_3_cast_fp16 = layer_norm(axes = hidden_states_3_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_1_attn_norm_weight_to_fp16, x = input_21_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_1_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_1_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127755456)))];
            tensor<fp16, [1, 128, 3072]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_1_attn_Wqkv_weight_to_fp16, x = hidden_states_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [5]> var_293 = const()[name = string("op_293"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_7_cast_fp16 = reshape(shape = var_293, x = linear_4_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> var_295_split_sizes_0 = const()[name = string("op_295_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_295_axis_0 = const()[name = string("op_295_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_295_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_295_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_295_cast_fp16_2 = split(axis = var_295_axis_0, split_sizes = var_295_split_sizes_0, x = qkv_7_cast_fp16)[name = string("op_295_cast_fp16")];
            tensor<int32, [1]> squeeze_3_axes_0 = const()[name = string("squeeze_3_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_3_cast_fp16 = squeeze(axes = squeeze_3_axes_0, x = var_295_cast_fp16_0)[name = string("squeeze_3_cast_fp16")];
            tensor<int32, [1]> squeeze_4_axes_0 = const()[name = string("squeeze_4_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_4_cast_fp16 = squeeze(axes = squeeze_4_axes_0, x = var_295_cast_fp16_1)[name = string("squeeze_4_cast_fp16")];
            tensor<int32, [1]> squeeze_5_axes_0 = const()[name = string("squeeze_5_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_5_cast_fp16 = squeeze(axes = squeeze_5_axes_0, x = var_295_cast_fp16_2)[name = string("squeeze_5_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 1, 128, 64]> cos_11_to_fp16 = const()[name = string("cos_11_to_fp16"), val = tensor<fp16, [1, 1, 128, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134046976)))];
            tensor<fp16, [1, 16, 128, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = squeeze_3_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 16, 128, 64]> var_305_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_11_to_fp16)[name = string("op_305_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_5_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_5_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_317_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_317_cast_fp16")];
            bool var_319_interleave_0 = const()[name = string("op_319_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_319_cast_fp16 = concat(axis = var_49, interleave = var_319_interleave_0, values = (var_317_cast_fp16, x1_5_cast_fp16))[name = string("op_319_cast_fp16")];
            tensor<fp16, [1, 1, 128, 64]> sin_11_to_fp16 = const()[name = string("sin_11_to_fp16"), val = tensor<fp16, [1, 1, 128, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134063424)))];
            tensor<fp16, [1, 16, 128, 64]> var_320_cast_fp16 = mul(x = var_319_cast_fp16, y = sin_11_to_fp16)[name = string("op_320_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_3_cast_fp16 = add(x = var_305_cast_fp16, y = var_320_cast_fp16)[name = string("q_embed_3_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = squeeze_4_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 16, 128, 64]> var_323_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_11_to_fp16)[name = string("op_323_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_5_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_5_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_335_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_335_cast_fp16")];
            bool var_337_interleave_0 = const()[name = string("op_337_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_337_cast_fp16 = concat(axis = var_49, interleave = var_337_interleave_0, values = (var_335_cast_fp16, x1_7_cast_fp16))[name = string("op_337_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_338_cast_fp16 = mul(x = var_337_cast_fp16, y = sin_11_to_fp16)[name = string("op_338_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_3_cast_fp16 = add(x = var_323_cast_fp16, y = var_338_cast_fp16)[name = string("k_embed_3_cast_fp16")];
            bool var_343_transpose_x_1 = const()[name = string("op_343_transpose_x_1"), val = bool(false)];
            bool var_343_transpose_y_1 = const()[name = string("op_343_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_343_cast_fp16 = matmul(transpose_x = var_343_transpose_x_1, transpose_y = var_343_transpose_y_1, x = q_embed_3_cast_fp16, y = k_embed_3_cast_fp16)[name = string("op_343_cast_fp16")];
            fp16 var_344_to_fp16 = const()[name = string("op_344_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_5_cast_fp16 = mul(x = var_343_cast_fp16, y = var_344_to_fp16)[name = string("attn_weights_5_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_23_cast_fp16 = add(x = attn_weights_5_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_347_cast_fp16 = softmax(axis = var_49, x = input_23_cast_fp16)[name = string("op_347_cast_fp16")];
            bool attn_output_7_transpose_x_0 = const()[name = string("attn_output_7_transpose_x_0"), val = bool(false)];
            bool attn_output_7_transpose_y_0 = const()[name = string("attn_output_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = squeeze_5_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = var_347_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<int32, [4]> var_351_perm_0 = const()[name = string("op_351_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_353 = const()[name = string("op_353"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_351_cast_fp16 = transpose(perm = var_351_perm_0, x = attn_output_7_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 128, 1024]> var_354_cast_fp16 = reshape(shape = var_353, x = var_351_cast_fp16)[name = string("op_354_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_1_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_1_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134079872)))];
            tensor<fp16, [1, 128, 1024]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_1_attn_Wo_weight_to_fp16, x = var_354_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_29_cast_fp16 = add(x = input_21_cast_fp16, y = linear_5_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_1_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_1_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136177088)))];
            tensor<fp16, [1, 128, 1024]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_1_mlp_norm_weight_to_fp16, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_1_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_1_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136179200)))];
            tensor<fp16, [1, 128, 5248]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_1_mlp_Wi_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [2]> var_366_split_sizes_0 = const()[name = string("op_366_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_366_axis_0 = const()[name = string("op_366_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_366_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_366_cast_fp16_1 = split(axis = var_366_axis_0, split_sizes = var_366_split_sizes_0, x = linear_6_cast_fp16)[name = string("op_366_cast_fp16")];
            string var_368_mode_0 = const()[name = string("op_368_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_368_cast_fp16 = gelu(mode = var_368_mode_0, x = var_366_cast_fp16_0)[name = string("op_368_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_35_cast_fp16 = mul(x = var_368_cast_fp16, y = var_366_cast_fp16_1)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_1_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_1_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146927168)))];
            tensor<fp16, [1, 128, 1024]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_1_mlp_Wo_weight_to_fp16, x = input_35_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_39_cast_fp16 = add(x = input_29_cast_fp16, y = linear_7_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<int32, [1]> hidden_states_5_axes_0 = const()[name = string("hidden_states_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_2_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_2_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152301184)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_5_cast_fp16 = layer_norm(axes = hidden_states_5_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_2_attn_norm_weight_to_fp16, x = input_39_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_2_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_2_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152303296)))];
            tensor<fp16, [1, 128, 3072]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_2_attn_Wqkv_weight_to_fp16, x = hidden_states_5_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [5]> var_387 = const()[name = string("op_387"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_11_cast_fp16 = reshape(shape = var_387, x = linear_8_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> var_389_split_sizes_0 = const()[name = string("op_389_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_389_axis_0 = const()[name = string("op_389_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_389_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_389_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_389_cast_fp16_2 = split(axis = var_389_axis_0, split_sizes = var_389_split_sizes_0, x = qkv_11_cast_fp16)[name = string("op_389_cast_fp16")];
            tensor<int32, [1]> squeeze_6_axes_0 = const()[name = string("squeeze_6_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_6_cast_fp16 = squeeze(axes = squeeze_6_axes_0, x = var_389_cast_fp16_0)[name = string("squeeze_6_cast_fp16")];
            tensor<int32, [1]> squeeze_7_axes_0 = const()[name = string("squeeze_7_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_7_cast_fp16 = squeeze(axes = squeeze_7_axes_0, x = var_389_cast_fp16_1)[name = string("squeeze_7_cast_fp16")];
            tensor<int32, [1]> squeeze_8_axes_0 = const()[name = string("squeeze_8_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_8_cast_fp16 = squeeze(axes = squeeze_8_axes_0, x = var_389_cast_fp16_2)[name = string("squeeze_8_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = squeeze_6_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 16, 128, 64]> var_399_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_11_to_fp16)[name = string("op_399_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_9_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_9_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_411_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_411_cast_fp16")];
            bool var_413_interleave_0 = const()[name = string("op_413_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_413_cast_fp16 = concat(axis = var_49, interleave = var_413_interleave_0, values = (var_411_cast_fp16, x1_9_cast_fp16))[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_414_cast_fp16 = mul(x = var_413_cast_fp16, y = sin_11_to_fp16)[name = string("op_414_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_5_cast_fp16 = add(x = var_399_cast_fp16, y = var_414_cast_fp16)[name = string("q_embed_5_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = squeeze_7_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 16, 128, 64]> var_417_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_11_to_fp16)[name = string("op_417_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_9_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_9_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_429_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_429_cast_fp16")];
            bool var_431_interleave_0 = const()[name = string("op_431_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_431_cast_fp16 = concat(axis = var_49, interleave = var_431_interleave_0, values = (var_429_cast_fp16, x1_11_cast_fp16))[name = string("op_431_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_432_cast_fp16 = mul(x = var_431_cast_fp16, y = sin_11_to_fp16)[name = string("op_432_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_5_cast_fp16 = add(x = var_417_cast_fp16, y = var_432_cast_fp16)[name = string("k_embed_5_cast_fp16")];
            bool var_437_transpose_x_1 = const()[name = string("op_437_transpose_x_1"), val = bool(false)];
            bool var_437_transpose_y_1 = const()[name = string("op_437_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_437_cast_fp16 = matmul(transpose_x = var_437_transpose_x_1, transpose_y = var_437_transpose_y_1, x = q_embed_5_cast_fp16, y = k_embed_5_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 var_438_to_fp16 = const()[name = string("op_438_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_9_cast_fp16 = mul(x = var_437_cast_fp16, y = var_438_to_fp16)[name = string("attn_weights_9_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_41_cast_fp16 = add(x = attn_weights_9_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_441_cast_fp16 = softmax(axis = var_49, x = input_41_cast_fp16)[name = string("op_441_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = squeeze_8_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = var_441_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_445_perm_0 = const()[name = string("op_445_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_447 = const()[name = string("op_447"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_445_cast_fp16 = transpose(perm = var_445_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 128, 1024]> var_448_cast_fp16 = reshape(shape = var_447, x = var_445_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_2_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_2_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158594816)))];
            tensor<fp16, [1, 128, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_2_attn_Wo_weight_to_fp16, x = var_448_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_47_cast_fp16 = add(x = input_39_cast_fp16, y = linear_9_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<int32, [1]> input_49_axes_0 = const()[name = string("input_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_2_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_2_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160692032)))];
            tensor<fp16, [1, 128, 1024]> input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_2_mlp_norm_weight_to_fp16, x = input_47_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_2_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_2_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160694144)))];
            tensor<fp16, [1, 128, 5248]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_2_mlp_Wi_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [2]> var_460_split_sizes_0 = const()[name = string("op_460_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_460_axis_0 = const()[name = string("op_460_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_460_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_460_cast_fp16_1 = split(axis = var_460_axis_0, split_sizes = var_460_split_sizes_0, x = linear_10_cast_fp16)[name = string("op_460_cast_fp16")];
            string var_462_mode_0 = const()[name = string("op_462_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_462_cast_fp16 = gelu(mode = var_462_mode_0, x = var_460_cast_fp16_0)[name = string("op_462_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_53_cast_fp16 = mul(x = var_462_cast_fp16, y = var_460_cast_fp16_1)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_2_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_2_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171442112)))];
            tensor<fp16, [1, 128, 1024]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_2_mlp_Wo_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_57_cast_fp16 = add(x = input_47_cast_fp16, y = linear_11_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<int32, [1]> hidden_states_7_axes_0 = const()[name = string("hidden_states_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_3_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_3_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176816128)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_7_cast_fp16 = layer_norm(axes = hidden_states_7_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_3_attn_norm_weight_to_fp16, x = input_57_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_3_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_3_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176818240)))];
            tensor<fp16, [1, 128, 3072]> linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_3_attn_Wqkv_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [5]> var_481 = const()[name = string("op_481"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_15_cast_fp16 = reshape(shape = var_481, x = linear_12_cast_fp16)[name = string("qkv_15_cast_fp16")];
            tensor<int32, [3]> var_483_split_sizes_0 = const()[name = string("op_483_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_483_axis_0 = const()[name = string("op_483_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_483_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_483_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_483_cast_fp16_2 = split(axis = var_483_axis_0, split_sizes = var_483_split_sizes_0, x = qkv_15_cast_fp16)[name = string("op_483_cast_fp16")];
            tensor<int32, [1]> squeeze_9_axes_0 = const()[name = string("squeeze_9_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_9_cast_fp16 = squeeze(axes = squeeze_9_axes_0, x = var_483_cast_fp16_0)[name = string("squeeze_9_cast_fp16")];
            tensor<int32, [1]> squeeze_10_axes_0 = const()[name = string("squeeze_10_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_10_cast_fp16 = squeeze(axes = squeeze_10_axes_0, x = var_483_cast_fp16_1)[name = string("squeeze_10_cast_fp16")];
            tensor<int32, [1]> squeeze_11_axes_0 = const()[name = string("squeeze_11_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_11_cast_fp16 = squeeze(axes = squeeze_11_axes_0, x = var_483_cast_fp16_2)[name = string("squeeze_11_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_7_perm_0 = const()[name = string("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = squeeze_9_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 128, 64]> var_493_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_7_to_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_13_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_13_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_505_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_505_cast_fp16")];
            bool var_507_interleave_0 = const()[name = string("op_507_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_507_cast_fp16 = concat(axis = var_49, interleave = var_507_interleave_0, values = (var_505_cast_fp16, x1_13_cast_fp16))[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_508_cast_fp16 = mul(x = var_507_cast_fp16, y = sin_7_to_fp16)[name = string("op_508_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_7_cast_fp16 = add(x = var_493_cast_fp16, y = var_508_cast_fp16)[name = string("q_embed_7_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = squeeze_10_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 16, 128, 64]> var_511_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_7_to_fp16)[name = string("op_511_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_13_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_13_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_523_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_523_cast_fp16")];
            bool var_525_interleave_0 = const()[name = string("op_525_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_525_cast_fp16 = concat(axis = var_49, interleave = var_525_interleave_0, values = (var_523_cast_fp16, x1_15_cast_fp16))[name = string("op_525_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_526_cast_fp16 = mul(x = var_525_cast_fp16, y = sin_7_to_fp16)[name = string("op_526_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_7_cast_fp16 = add(x = var_511_cast_fp16, y = var_526_cast_fp16)[name = string("k_embed_7_cast_fp16")];
            bool var_531_transpose_x_1 = const()[name = string("op_531_transpose_x_1"), val = bool(false)];
            bool var_531_transpose_y_1 = const()[name = string("op_531_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_531_cast_fp16 = matmul(transpose_x = var_531_transpose_x_1, transpose_y = var_531_transpose_y_1, x = q_embed_7_cast_fp16, y = k_embed_7_cast_fp16)[name = string("op_531_cast_fp16")];
            fp16 var_532_to_fp16 = const()[name = string("op_532_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_13_cast_fp16 = mul(x = var_531_cast_fp16, y = var_532_to_fp16)[name = string("attn_weights_13_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_59_cast_fp16 = add(x = attn_weights_13_cast_fp16, y = key_bias_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_535_cast_fp16 = softmax(axis = var_49, x = input_59_cast_fp16)[name = string("op_535_cast_fp16")];
            bool attn_output_19_transpose_x_0 = const()[name = string("attn_output_19_transpose_x_0"), val = bool(false)];
            bool attn_output_19_transpose_y_0 = const()[name = string("attn_output_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = squeeze_11_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = var_535_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_539_perm_0 = const()[name = string("op_539_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_541 = const()[name = string("op_541"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_539_cast_fp16 = transpose(perm = var_539_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 128, 1024]> var_542_cast_fp16 = reshape(shape = var_541, x = var_539_cast_fp16)[name = string("op_542_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_3_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_3_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183109760)))];
            tensor<fp16, [1, 128, 1024]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_3_attn_Wo_weight_to_fp16, x = var_542_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_65_cast_fp16 = add(x = input_57_cast_fp16, y = linear_13_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_3_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_3_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185206976)))];
            tensor<fp16, [1, 128, 1024]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_3_mlp_norm_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_3_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_3_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185209088)))];
            tensor<fp16, [1, 128, 5248]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_3_mlp_Wi_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [2]> var_554_split_sizes_0 = const()[name = string("op_554_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_554_axis_0 = const()[name = string("op_554_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_554_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_554_cast_fp16_1 = split(axis = var_554_axis_0, split_sizes = var_554_split_sizes_0, x = linear_14_cast_fp16)[name = string("op_554_cast_fp16")];
            string var_556_mode_0 = const()[name = string("op_556_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_556_cast_fp16 = gelu(mode = var_556_mode_0, x = var_554_cast_fp16_0)[name = string("op_556_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_71_cast_fp16 = mul(x = var_556_cast_fp16, y = var_554_cast_fp16_1)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_3_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_3_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195957056)))];
            tensor<fp16, [1, 128, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_3_mlp_Wo_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_75_cast_fp16 = add(x = input_65_cast_fp16, y = linear_15_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<int32, [1]> hidden_states_9_axes_0 = const()[name = string("hidden_states_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_4_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_4_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201331072)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_9_cast_fp16 = layer_norm(axes = hidden_states_9_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_4_attn_norm_weight_to_fp16, x = input_75_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_4_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_4_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201333184)))];
            tensor<fp16, [1, 128, 3072]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_4_attn_Wqkv_weight_to_fp16, x = hidden_states_9_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [5]> var_575 = const()[name = string("op_575"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_19_cast_fp16 = reshape(shape = var_575, x = linear_16_cast_fp16)[name = string("qkv_19_cast_fp16")];
            tensor<int32, [3]> var_577_split_sizes_0 = const()[name = string("op_577_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_577_axis_0 = const()[name = string("op_577_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_577_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_577_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_577_cast_fp16_2 = split(axis = var_577_axis_0, split_sizes = var_577_split_sizes_0, x = qkv_19_cast_fp16)[name = string("op_577_cast_fp16")];
            tensor<int32, [1]> squeeze_12_axes_0 = const()[name = string("squeeze_12_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_12_cast_fp16 = squeeze(axes = squeeze_12_axes_0, x = var_577_cast_fp16_0)[name = string("squeeze_12_cast_fp16")];
            tensor<int32, [1]> squeeze_13_axes_0 = const()[name = string("squeeze_13_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_13_cast_fp16 = squeeze(axes = squeeze_13_axes_0, x = var_577_cast_fp16_1)[name = string("squeeze_13_cast_fp16")];
            tensor<int32, [1]> squeeze_14_axes_0 = const()[name = string("squeeze_14_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_14_cast_fp16 = squeeze(axes = squeeze_14_axes_0, x = var_577_cast_fp16_2)[name = string("squeeze_14_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = squeeze_12_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 128, 64]> var_587_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_11_to_fp16)[name = string("op_587_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_17_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_17_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_599_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_55_promoted_to_fp16)[name = string("op_599_cast_fp16")];
            bool var_601_interleave_0 = const()[name = string("op_601_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_601_cast_fp16 = concat(axis = var_49, interleave = var_601_interleave_0, values = (var_599_cast_fp16, x1_17_cast_fp16))[name = string("op_601_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_602_cast_fp16 = mul(x = var_601_cast_fp16, y = sin_11_to_fp16)[name = string("op_602_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_9_cast_fp16 = add(x = var_587_cast_fp16, y = var_602_cast_fp16)[name = string("q_embed_9_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = squeeze_13_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 128, 64]> var_605_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_11_to_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_17_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_17_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_58_promoted_to_fp16 = const()[name = string("const_58_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_617_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_58_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            bool var_619_interleave_0 = const()[name = string("op_619_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_619_cast_fp16 = concat(axis = var_49, interleave = var_619_interleave_0, values = (var_617_cast_fp16, x1_19_cast_fp16))[name = string("op_619_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_620_cast_fp16 = mul(x = var_619_cast_fp16, y = sin_11_to_fp16)[name = string("op_620_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_9_cast_fp16 = add(x = var_605_cast_fp16, y = var_620_cast_fp16)[name = string("k_embed_9_cast_fp16")];
            bool var_625_transpose_x_1 = const()[name = string("op_625_transpose_x_1"), val = bool(false)];
            bool var_625_transpose_y_1 = const()[name = string("op_625_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_625_cast_fp16 = matmul(transpose_x = var_625_transpose_x_1, transpose_y = var_625_transpose_y_1, x = q_embed_9_cast_fp16, y = k_embed_9_cast_fp16)[name = string("op_625_cast_fp16")];
            fp16 var_626_to_fp16 = const()[name = string("op_626_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_17_cast_fp16 = mul(x = var_625_cast_fp16, y = var_626_to_fp16)[name = string("attn_weights_17_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_77_cast_fp16 = add(x = attn_weights_17_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_629_cast_fp16 = softmax(axis = var_49, x = input_77_cast_fp16)[name = string("op_629_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = squeeze_14_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = var_629_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_633_perm_0 = const()[name = string("op_633_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_635 = const()[name = string("op_635"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_633_cast_fp16 = transpose(perm = var_633_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 128, 1024]> var_636_cast_fp16 = reshape(shape = var_635, x = var_633_cast_fp16)[name = string("op_636_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_4_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_4_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207624704)))];
            tensor<fp16, [1, 128, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_4_attn_Wo_weight_to_fp16, x = var_636_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_83_cast_fp16 = add(x = input_75_cast_fp16, y = linear_17_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<int32, [1]> input_85_axes_0 = const()[name = string("input_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_4_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_4_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209721920)))];
            tensor<fp16, [1, 128, 1024]> input_85_cast_fp16 = layer_norm(axes = input_85_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_4_mlp_norm_weight_to_fp16, x = input_83_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_4_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_4_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209724032)))];
            tensor<fp16, [1, 128, 5248]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_4_mlp_Wi_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [2]> var_648_split_sizes_0 = const()[name = string("op_648_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_648_axis_0 = const()[name = string("op_648_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_648_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_648_cast_fp16_1 = split(axis = var_648_axis_0, split_sizes = var_648_split_sizes_0, x = linear_18_cast_fp16)[name = string("op_648_cast_fp16")];
            string var_650_mode_0 = const()[name = string("op_650_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_650_cast_fp16 = gelu(mode = var_650_mode_0, x = var_648_cast_fp16_0)[name = string("op_650_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_89_cast_fp16 = mul(x = var_650_cast_fp16, y = var_648_cast_fp16_1)[name = string("input_89_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_4_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_4_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220472000)))];
            tensor<fp16, [1, 128, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_4_mlp_Wo_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_93_cast_fp16 = add(x = input_83_cast_fp16, y = linear_19_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<int32, [1]> hidden_states_11_axes_0 = const()[name = string("hidden_states_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_5_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_5_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225846016)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_11_cast_fp16 = layer_norm(axes = hidden_states_11_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_5_attn_norm_weight_to_fp16, x = input_93_cast_fp16)[name = string("hidden_states_11_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_5_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_5_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225848128)))];
            tensor<fp16, [1, 128, 3072]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_5_attn_Wqkv_weight_to_fp16, x = hidden_states_11_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [5]> var_669 = const()[name = string("op_669"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_23_cast_fp16 = reshape(shape = var_669, x = linear_20_cast_fp16)[name = string("qkv_23_cast_fp16")];
            tensor<int32, [3]> var_671_split_sizes_0 = const()[name = string("op_671_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_671_axis_0 = const()[name = string("op_671_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_671_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_671_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_671_cast_fp16_2 = split(axis = var_671_axis_0, split_sizes = var_671_split_sizes_0, x = qkv_23_cast_fp16)[name = string("op_671_cast_fp16")];
            tensor<int32, [1]> squeeze_15_axes_0 = const()[name = string("squeeze_15_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_15_cast_fp16 = squeeze(axes = squeeze_15_axes_0, x = var_671_cast_fp16_0)[name = string("squeeze_15_cast_fp16")];
            tensor<int32, [1]> squeeze_16_axes_0 = const()[name = string("squeeze_16_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_16_cast_fp16 = squeeze(axes = squeeze_16_axes_0, x = var_671_cast_fp16_1)[name = string("squeeze_16_cast_fp16")];
            tensor<int32, [1]> squeeze_17_axes_0 = const()[name = string("squeeze_17_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_17_cast_fp16 = squeeze(axes = squeeze_17_axes_0, x = var_671_cast_fp16_2)[name = string("squeeze_17_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = squeeze_15_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 16, 128, 64]> var_681_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_11_to_fp16)[name = string("op_681_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_21_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_21_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_63_promoted_to_fp16 = const()[name = string("const_63_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_693_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_63_promoted_to_fp16)[name = string("op_693_cast_fp16")];
            bool var_695_interleave_0 = const()[name = string("op_695_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_695_cast_fp16 = concat(axis = var_49, interleave = var_695_interleave_0, values = (var_693_cast_fp16, x1_21_cast_fp16))[name = string("op_695_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_696_cast_fp16 = mul(x = var_695_cast_fp16, y = sin_11_to_fp16)[name = string("op_696_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_11_cast_fp16 = add(x = var_681_cast_fp16, y = var_696_cast_fp16)[name = string("q_embed_11_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = squeeze_16_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 128, 64]> var_699_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_11_to_fp16)[name = string("op_699_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_21_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_21_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_66_promoted_to_fp16 = const()[name = string("const_66_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_711_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_66_promoted_to_fp16)[name = string("op_711_cast_fp16")];
            bool var_713_interleave_0 = const()[name = string("op_713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_713_cast_fp16 = concat(axis = var_49, interleave = var_713_interleave_0, values = (var_711_cast_fp16, x1_23_cast_fp16))[name = string("op_713_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_714_cast_fp16 = mul(x = var_713_cast_fp16, y = sin_11_to_fp16)[name = string("op_714_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_11_cast_fp16 = add(x = var_699_cast_fp16, y = var_714_cast_fp16)[name = string("k_embed_11_cast_fp16")];
            bool var_719_transpose_x_1 = const()[name = string("op_719_transpose_x_1"), val = bool(false)];
            bool var_719_transpose_y_1 = const()[name = string("op_719_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_719_cast_fp16 = matmul(transpose_x = var_719_transpose_x_1, transpose_y = var_719_transpose_y_1, x = q_embed_11_cast_fp16, y = k_embed_11_cast_fp16)[name = string("op_719_cast_fp16")];
            fp16 var_720_to_fp16 = const()[name = string("op_720_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_21_cast_fp16 = mul(x = var_719_cast_fp16, y = var_720_to_fp16)[name = string("attn_weights_21_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_95_cast_fp16 = add(x = attn_weights_21_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_723_cast_fp16 = softmax(axis = var_49, x = input_95_cast_fp16)[name = string("op_723_cast_fp16")];
            bool attn_output_31_transpose_x_0 = const()[name = string("attn_output_31_transpose_x_0"), val = bool(false)];
            bool attn_output_31_transpose_y_0 = const()[name = string("attn_output_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = squeeze_17_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = var_723_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [4]> var_727_perm_0 = const()[name = string("op_727_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_729 = const()[name = string("op_729"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_727_cast_fp16 = transpose(perm = var_727_perm_0, x = attn_output_31_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 128, 1024]> var_730_cast_fp16 = reshape(shape = var_729, x = var_727_cast_fp16)[name = string("op_730_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_5_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_5_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232139648)))];
            tensor<fp16, [1, 128, 1024]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_5_attn_Wo_weight_to_fp16, x = var_730_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_101_cast_fp16 = add(x = input_93_cast_fp16, y = linear_21_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<int32, [1]> input_103_axes_0 = const()[name = string("input_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_5_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_5_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234236864)))];
            tensor<fp16, [1, 128, 1024]> input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_5_mlp_norm_weight_to_fp16, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_5_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_5_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234238976)))];
            tensor<fp16, [1, 128, 5248]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_5_mlp_Wi_weight_to_fp16, x = input_103_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [2]> var_742_split_sizes_0 = const()[name = string("op_742_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_742_axis_0 = const()[name = string("op_742_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_742_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_742_cast_fp16_1 = split(axis = var_742_axis_0, split_sizes = var_742_split_sizes_0, x = linear_22_cast_fp16)[name = string("op_742_cast_fp16")];
            string var_744_mode_0 = const()[name = string("op_744_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_744_cast_fp16 = gelu(mode = var_744_mode_0, x = var_742_cast_fp16_0)[name = string("op_744_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_107_cast_fp16 = mul(x = var_744_cast_fp16, y = var_742_cast_fp16_1)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_5_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_5_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244986944)))];
            tensor<fp16, [1, 128, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_5_mlp_Wo_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_111_cast_fp16 = add(x = input_101_cast_fp16, y = linear_23_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<int32, [1]> hidden_states_13_axes_0 = const()[name = string("hidden_states_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_6_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_6_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250360960)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_13_cast_fp16 = layer_norm(axes = hidden_states_13_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_6_attn_norm_weight_to_fp16, x = input_111_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_6_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_6_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250363072)))];
            tensor<fp16, [1, 128, 3072]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_6_attn_Wqkv_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [5]> var_763 = const()[name = string("op_763"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_27_cast_fp16 = reshape(shape = var_763, x = linear_24_cast_fp16)[name = string("qkv_27_cast_fp16")];
            tensor<int32, [3]> var_765_split_sizes_0 = const()[name = string("op_765_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_765_axis_0 = const()[name = string("op_765_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_765_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_765_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_765_cast_fp16_2 = split(axis = var_765_axis_0, split_sizes = var_765_split_sizes_0, x = qkv_27_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<int32, [1]> squeeze_18_axes_0 = const()[name = string("squeeze_18_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_18_cast_fp16 = squeeze(axes = squeeze_18_axes_0, x = var_765_cast_fp16_0)[name = string("squeeze_18_cast_fp16")];
            tensor<int32, [1]> squeeze_19_axes_0 = const()[name = string("squeeze_19_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_19_cast_fp16 = squeeze(axes = squeeze_19_axes_0, x = var_765_cast_fp16_1)[name = string("squeeze_19_cast_fp16")];
            tensor<int32, [1]> squeeze_20_axes_0 = const()[name = string("squeeze_20_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_20_cast_fp16 = squeeze(axes = squeeze_20_axes_0, x = var_765_cast_fp16_2)[name = string("squeeze_20_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = squeeze_18_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 16, 128, 64]> var_775_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_7_to_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_25_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_25_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_71_promoted_to_fp16 = const()[name = string("const_71_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_787_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_71_promoted_to_fp16)[name = string("op_787_cast_fp16")];
            bool var_789_interleave_0 = const()[name = string("op_789_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_789_cast_fp16 = concat(axis = var_49, interleave = var_789_interleave_0, values = (var_787_cast_fp16, x1_25_cast_fp16))[name = string("op_789_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_790_cast_fp16 = mul(x = var_789_cast_fp16, y = sin_7_to_fp16)[name = string("op_790_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_13_cast_fp16 = add(x = var_775_cast_fp16, y = var_790_cast_fp16)[name = string("q_embed_13_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = squeeze_19_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 16, 128, 64]> var_793_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_7_to_fp16)[name = string("op_793_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_25_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_25_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_805_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_74_promoted_to_fp16)[name = string("op_805_cast_fp16")];
            bool var_807_interleave_0 = const()[name = string("op_807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_807_cast_fp16 = concat(axis = var_49, interleave = var_807_interleave_0, values = (var_805_cast_fp16, x1_27_cast_fp16))[name = string("op_807_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_808_cast_fp16 = mul(x = var_807_cast_fp16, y = sin_7_to_fp16)[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_13_cast_fp16 = add(x = var_793_cast_fp16, y = var_808_cast_fp16)[name = string("k_embed_13_cast_fp16")];
            bool var_813_transpose_x_1 = const()[name = string("op_813_transpose_x_1"), val = bool(false)];
            bool var_813_transpose_y_1 = const()[name = string("op_813_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_813_cast_fp16 = matmul(transpose_x = var_813_transpose_x_1, transpose_y = var_813_transpose_y_1, x = q_embed_13_cast_fp16, y = k_embed_13_cast_fp16)[name = string("op_813_cast_fp16")];
            fp16 var_814_to_fp16 = const()[name = string("op_814_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_25_cast_fp16 = mul(x = var_813_cast_fp16, y = var_814_to_fp16)[name = string("attn_weights_25_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_113_cast_fp16 = add(x = attn_weights_25_cast_fp16, y = key_bias_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_817_cast_fp16 = softmax(axis = var_49, x = input_113_cast_fp16)[name = string("op_817_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = squeeze_20_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = var_817_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_821_perm_0 = const()[name = string("op_821_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_823 = const()[name = string("op_823"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_821_cast_fp16 = transpose(perm = var_821_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 128, 1024]> var_824_cast_fp16 = reshape(shape = var_823, x = var_821_cast_fp16)[name = string("op_824_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_6_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_6_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256654592)))];
            tensor<fp16, [1, 128, 1024]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_6_attn_Wo_weight_to_fp16, x = var_824_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_119_cast_fp16 = add(x = input_111_cast_fp16, y = linear_25_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_6_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_6_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258751808)))];
            tensor<fp16, [1, 128, 1024]> input_121_cast_fp16 = layer_norm(axes = input_121_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_6_mlp_norm_weight_to_fp16, x = input_119_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_6_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_6_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258753920)))];
            tensor<fp16, [1, 128, 5248]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_6_mlp_Wi_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [2]> var_836_split_sizes_0 = const()[name = string("op_836_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_836_axis_0 = const()[name = string("op_836_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_836_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_836_cast_fp16_1 = split(axis = var_836_axis_0, split_sizes = var_836_split_sizes_0, x = linear_26_cast_fp16)[name = string("op_836_cast_fp16")];
            string var_838_mode_0 = const()[name = string("op_838_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_838_cast_fp16 = gelu(mode = var_838_mode_0, x = var_836_cast_fp16_0)[name = string("op_838_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_125_cast_fp16 = mul(x = var_838_cast_fp16, y = var_836_cast_fp16_1)[name = string("input_125_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_6_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_6_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269501888)))];
            tensor<fp16, [1, 128, 1024]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_6_mlp_Wo_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_129_cast_fp16 = add(x = input_119_cast_fp16, y = linear_27_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<int32, [1]> hidden_states_15_axes_0 = const()[name = string("hidden_states_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_7_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_7_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274875904)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_15_cast_fp16 = layer_norm(axes = hidden_states_15_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_7_attn_norm_weight_to_fp16, x = input_129_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_7_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_7_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274878016)))];
            tensor<fp16, [1, 128, 3072]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_7_attn_Wqkv_weight_to_fp16, x = hidden_states_15_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [5]> var_857 = const()[name = string("op_857"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_31_cast_fp16 = reshape(shape = var_857, x = linear_28_cast_fp16)[name = string("qkv_31_cast_fp16")];
            tensor<int32, [3]> var_859_split_sizes_0 = const()[name = string("op_859_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_859_axis_0 = const()[name = string("op_859_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_859_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_859_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_859_cast_fp16_2 = split(axis = var_859_axis_0, split_sizes = var_859_split_sizes_0, x = qkv_31_cast_fp16)[name = string("op_859_cast_fp16")];
            tensor<int32, [1]> squeeze_21_axes_0 = const()[name = string("squeeze_21_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_21_cast_fp16 = squeeze(axes = squeeze_21_axes_0, x = var_859_cast_fp16_0)[name = string("squeeze_21_cast_fp16")];
            tensor<int32, [1]> squeeze_22_axes_0 = const()[name = string("squeeze_22_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_22_cast_fp16 = squeeze(axes = squeeze_22_axes_0, x = var_859_cast_fp16_1)[name = string("squeeze_22_cast_fp16")];
            tensor<int32, [1]> squeeze_23_axes_0 = const()[name = string("squeeze_23_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_23_cast_fp16 = squeeze(axes = squeeze_23_axes_0, x = var_859_cast_fp16_2)[name = string("squeeze_23_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = squeeze_21_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 16, 128, 64]> var_869_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_11_to_fp16)[name = string("op_869_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_29_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_29_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_79_promoted_to_fp16 = const()[name = string("const_79_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_881_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_79_promoted_to_fp16)[name = string("op_881_cast_fp16")];
            bool var_883_interleave_0 = const()[name = string("op_883_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_883_cast_fp16 = concat(axis = var_49, interleave = var_883_interleave_0, values = (var_881_cast_fp16, x1_29_cast_fp16))[name = string("op_883_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_884_cast_fp16 = mul(x = var_883_cast_fp16, y = sin_11_to_fp16)[name = string("op_884_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_15_cast_fp16 = add(x = var_869_cast_fp16, y = var_884_cast_fp16)[name = string("q_embed_15_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = squeeze_22_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 16, 128, 64]> var_887_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_11_to_fp16)[name = string("op_887_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_29_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_29_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_899_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_899_cast_fp16")];
            bool var_901_interleave_0 = const()[name = string("op_901_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_901_cast_fp16 = concat(axis = var_49, interleave = var_901_interleave_0, values = (var_899_cast_fp16, x1_31_cast_fp16))[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_902_cast_fp16 = mul(x = var_901_cast_fp16, y = sin_11_to_fp16)[name = string("op_902_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_15_cast_fp16 = add(x = var_887_cast_fp16, y = var_902_cast_fp16)[name = string("k_embed_15_cast_fp16")];
            bool var_907_transpose_x_1 = const()[name = string("op_907_transpose_x_1"), val = bool(false)];
            bool var_907_transpose_y_1 = const()[name = string("op_907_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_907_cast_fp16 = matmul(transpose_x = var_907_transpose_x_1, transpose_y = var_907_transpose_y_1, x = q_embed_15_cast_fp16, y = k_embed_15_cast_fp16)[name = string("op_907_cast_fp16")];
            fp16 var_908_to_fp16 = const()[name = string("op_908_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_29_cast_fp16 = mul(x = var_907_cast_fp16, y = var_908_to_fp16)[name = string("attn_weights_29_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_131_cast_fp16 = add(x = attn_weights_29_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_911_cast_fp16 = softmax(axis = var_49, x = input_131_cast_fp16)[name = string("op_911_cast_fp16")];
            bool attn_output_43_transpose_x_0 = const()[name = string("attn_output_43_transpose_x_0"), val = bool(false)];
            bool attn_output_43_transpose_y_0 = const()[name = string("attn_output_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = squeeze_23_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = var_911_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_915_perm_0 = const()[name = string("op_915_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_917 = const()[name = string("op_917"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_915_cast_fp16 = transpose(perm = var_915_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 128, 1024]> var_918_cast_fp16 = reshape(shape = var_917, x = var_915_cast_fp16)[name = string("op_918_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_7_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_7_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281169536)))];
            tensor<fp16, [1, 128, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_7_attn_Wo_weight_to_fp16, x = var_918_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_137_cast_fp16 = add(x = input_129_cast_fp16, y = linear_29_cast_fp16)[name = string("input_137_cast_fp16")];
            tensor<int32, [1]> input_139_axes_0 = const()[name = string("input_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_7_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_7_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283266752)))];
            tensor<fp16, [1, 128, 1024]> input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_7_mlp_norm_weight_to_fp16, x = input_137_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_7_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_7_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283268864)))];
            tensor<fp16, [1, 128, 5248]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_7_mlp_Wi_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [2]> var_930_split_sizes_0 = const()[name = string("op_930_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_930_axis_0 = const()[name = string("op_930_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_930_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_930_cast_fp16_1 = split(axis = var_930_axis_0, split_sizes = var_930_split_sizes_0, x = linear_30_cast_fp16)[name = string("op_930_cast_fp16")];
            string var_932_mode_0 = const()[name = string("op_932_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_932_cast_fp16 = gelu(mode = var_932_mode_0, x = var_930_cast_fp16_0)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_143_cast_fp16 = mul(x = var_932_cast_fp16, y = var_930_cast_fp16_1)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_7_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_7_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294016832)))];
            tensor<fp16, [1, 128, 1024]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_7_mlp_Wo_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_147_cast_fp16 = add(x = input_137_cast_fp16, y = linear_31_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<int32, [1]> hidden_states_17_axes_0 = const()[name = string("hidden_states_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_8_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_8_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299390848)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_17_cast_fp16 = layer_norm(axes = hidden_states_17_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_8_attn_norm_weight_to_fp16, x = input_147_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_8_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_8_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299392960)))];
            tensor<fp16, [1, 128, 3072]> linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_8_attn_Wqkv_weight_to_fp16, x = hidden_states_17_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<int32, [5]> var_951 = const()[name = string("op_951"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_35_cast_fp16 = reshape(shape = var_951, x = linear_32_cast_fp16)[name = string("qkv_35_cast_fp16")];
            tensor<int32, [3]> var_953_split_sizes_0 = const()[name = string("op_953_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_953_axis_0 = const()[name = string("op_953_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_953_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_953_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_953_cast_fp16_2 = split(axis = var_953_axis_0, split_sizes = var_953_split_sizes_0, x = qkv_35_cast_fp16)[name = string("op_953_cast_fp16")];
            tensor<int32, [1]> squeeze_24_axes_0 = const()[name = string("squeeze_24_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_24_cast_fp16 = squeeze(axes = squeeze_24_axes_0, x = var_953_cast_fp16_0)[name = string("squeeze_24_cast_fp16")];
            tensor<int32, [1]> squeeze_25_axes_0 = const()[name = string("squeeze_25_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_25_cast_fp16 = squeeze(axes = squeeze_25_axes_0, x = var_953_cast_fp16_1)[name = string("squeeze_25_cast_fp16")];
            tensor<int32, [1]> squeeze_26_axes_0 = const()[name = string("squeeze_26_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_26_cast_fp16 = squeeze(axes = squeeze_26_axes_0, x = var_953_cast_fp16_2)[name = string("squeeze_26_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = squeeze_24_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 128, 64]> var_963_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_11_to_fp16)[name = string("op_963_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_33_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_33_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_975_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_975_cast_fp16")];
            bool var_977_interleave_0 = const()[name = string("op_977_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_977_cast_fp16 = concat(axis = var_49, interleave = var_977_interleave_0, values = (var_975_cast_fp16, x1_33_cast_fp16))[name = string("op_977_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_978_cast_fp16 = mul(x = var_977_cast_fp16, y = sin_11_to_fp16)[name = string("op_978_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_17_cast_fp16 = add(x = var_963_cast_fp16, y = var_978_cast_fp16)[name = string("q_embed_17_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = squeeze_25_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 16, 128, 64]> var_981_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_11_to_fp16)[name = string("op_981_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_33_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_33_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_993_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_993_cast_fp16")];
            bool var_995_interleave_0 = const()[name = string("op_995_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_995_cast_fp16 = concat(axis = var_49, interleave = var_995_interleave_0, values = (var_993_cast_fp16, x1_35_cast_fp16))[name = string("op_995_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_996_cast_fp16 = mul(x = var_995_cast_fp16, y = sin_11_to_fp16)[name = string("op_996_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_17_cast_fp16 = add(x = var_981_cast_fp16, y = var_996_cast_fp16)[name = string("k_embed_17_cast_fp16")];
            bool var_1001_transpose_x_1 = const()[name = string("op_1001_transpose_x_1"), val = bool(false)];
            bool var_1001_transpose_y_1 = const()[name = string("op_1001_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1001_cast_fp16 = matmul(transpose_x = var_1001_transpose_x_1, transpose_y = var_1001_transpose_y_1, x = q_embed_17_cast_fp16, y = k_embed_17_cast_fp16)[name = string("op_1001_cast_fp16")];
            fp16 var_1002_to_fp16 = const()[name = string("op_1002_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_33_cast_fp16 = mul(x = var_1001_cast_fp16, y = var_1002_to_fp16)[name = string("attn_weights_33_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_149_cast_fp16 = add(x = attn_weights_33_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1005_cast_fp16 = softmax(axis = var_49, x = input_149_cast_fp16)[name = string("op_1005_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = squeeze_26_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = var_1005_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_1009_perm_0 = const()[name = string("op_1009_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1011 = const()[name = string("op_1011"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1009_cast_fp16 = transpose(perm = var_1009_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 128, 1024]> var_1012_cast_fp16 = reshape(shape = var_1011, x = var_1009_cast_fp16)[name = string("op_1012_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_8_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_8_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305684480)))];
            tensor<fp16, [1, 128, 1024]> linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_8_attn_Wo_weight_to_fp16, x = var_1012_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_155_cast_fp16 = add(x = input_147_cast_fp16, y = linear_33_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<int32, [1]> input_157_axes_0 = const()[name = string("input_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_8_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_8_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307781696)))];
            tensor<fp16, [1, 128, 1024]> input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_8_mlp_norm_weight_to_fp16, x = input_155_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_8_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_8_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307783808)))];
            tensor<fp16, [1, 128, 5248]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_8_mlp_Wi_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<int32, [2]> var_1024_split_sizes_0 = const()[name = string("op_1024_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1024_axis_0 = const()[name = string("op_1024_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1024_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1024_cast_fp16_1 = split(axis = var_1024_axis_0, split_sizes = var_1024_split_sizes_0, x = linear_34_cast_fp16)[name = string("op_1024_cast_fp16")];
            string var_1026_mode_0 = const()[name = string("op_1026_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1026_cast_fp16 = gelu(mode = var_1026_mode_0, x = var_1024_cast_fp16_0)[name = string("op_1026_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_161_cast_fp16 = mul(x = var_1026_cast_fp16, y = var_1024_cast_fp16_1)[name = string("input_161_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_8_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_8_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318531776)))];
            tensor<fp16, [1, 128, 1024]> linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_8_mlp_Wo_weight_to_fp16, x = input_161_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_165_cast_fp16 = add(x = input_155_cast_fp16, y = linear_35_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<int32, [1]> hidden_states_19_axes_0 = const()[name = string("hidden_states_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_9_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_9_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323905792)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_19_cast_fp16 = layer_norm(axes = hidden_states_19_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_9_attn_norm_weight_to_fp16, x = input_165_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_9_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_9_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323907904)))];
            tensor<fp16, [1, 128, 3072]> linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_9_attn_Wqkv_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [5]> var_1045 = const()[name = string("op_1045"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_39_cast_fp16 = reshape(shape = var_1045, x = linear_36_cast_fp16)[name = string("qkv_39_cast_fp16")];
            tensor<int32, [3]> var_1047_split_sizes_0 = const()[name = string("op_1047_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1047_axis_0 = const()[name = string("op_1047_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1047_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1047_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1047_cast_fp16_2 = split(axis = var_1047_axis_0, split_sizes = var_1047_split_sizes_0, x = qkv_39_cast_fp16)[name = string("op_1047_cast_fp16")];
            tensor<int32, [1]> squeeze_27_axes_0 = const()[name = string("squeeze_27_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_27_cast_fp16 = squeeze(axes = squeeze_27_axes_0, x = var_1047_cast_fp16_0)[name = string("squeeze_27_cast_fp16")];
            tensor<int32, [1]> squeeze_28_axes_0 = const()[name = string("squeeze_28_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_28_cast_fp16 = squeeze(axes = squeeze_28_axes_0, x = var_1047_cast_fp16_1)[name = string("squeeze_28_cast_fp16")];
            tensor<int32, [1]> squeeze_29_axes_0 = const()[name = string("squeeze_29_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_29_cast_fp16 = squeeze(axes = squeeze_29_axes_0, x = var_1047_cast_fp16_2)[name = string("squeeze_29_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = squeeze_27_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 128, 64]> var_1057_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_7_to_fp16)[name = string("op_1057_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_37_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_37_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1069_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_95_promoted_to_fp16)[name = string("op_1069_cast_fp16")];
            bool var_1071_interleave_0 = const()[name = string("op_1071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1071_cast_fp16 = concat(axis = var_49, interleave = var_1071_interleave_0, values = (var_1069_cast_fp16, x1_37_cast_fp16))[name = string("op_1071_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1072_cast_fp16 = mul(x = var_1071_cast_fp16, y = sin_7_to_fp16)[name = string("op_1072_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_19_cast_fp16 = add(x = var_1057_cast_fp16, y = var_1072_cast_fp16)[name = string("q_embed_19_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = squeeze_28_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 128, 64]> var_1075_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_7_to_fp16)[name = string("op_1075_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_37_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_37_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_98_promoted_to_fp16 = const()[name = string("const_98_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1087_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_98_promoted_to_fp16)[name = string("op_1087_cast_fp16")];
            bool var_1089_interleave_0 = const()[name = string("op_1089_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1089_cast_fp16 = concat(axis = var_49, interleave = var_1089_interleave_0, values = (var_1087_cast_fp16, x1_39_cast_fp16))[name = string("op_1089_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1090_cast_fp16 = mul(x = var_1089_cast_fp16, y = sin_7_to_fp16)[name = string("op_1090_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_19_cast_fp16 = add(x = var_1075_cast_fp16, y = var_1090_cast_fp16)[name = string("k_embed_19_cast_fp16")];
            bool var_1095_transpose_x_1 = const()[name = string("op_1095_transpose_x_1"), val = bool(false)];
            bool var_1095_transpose_y_1 = const()[name = string("op_1095_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1095_cast_fp16 = matmul(transpose_x = var_1095_transpose_x_1, transpose_y = var_1095_transpose_y_1, x = q_embed_19_cast_fp16, y = k_embed_19_cast_fp16)[name = string("op_1095_cast_fp16")];
            fp16 var_1096_to_fp16 = const()[name = string("op_1096_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_37_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_1096_to_fp16)[name = string("attn_weights_37_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_167_cast_fp16 = add(x = attn_weights_37_cast_fp16, y = key_bias_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1099_cast_fp16 = softmax(axis = var_49, x = input_167_cast_fp16)[name = string("op_1099_cast_fp16")];
            bool attn_output_55_transpose_x_0 = const()[name = string("attn_output_55_transpose_x_0"), val = bool(false)];
            bool attn_output_55_transpose_y_0 = const()[name = string("attn_output_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = squeeze_29_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_55_cast_fp16 = matmul(transpose_x = attn_output_55_transpose_x_0, transpose_y = attn_output_55_transpose_y_0, x = var_1099_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [4]> var_1103_perm_0 = const()[name = string("op_1103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1105 = const()[name = string("op_1105"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1103_cast_fp16 = transpose(perm = var_1103_perm_0, x = attn_output_55_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 128, 1024]> var_1106_cast_fp16 = reshape(shape = var_1105, x = var_1103_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_9_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_9_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330199424)))];
            tensor<fp16, [1, 128, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_9_attn_Wo_weight_to_fp16, x = var_1106_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_173_cast_fp16 = add(x = input_165_cast_fp16, y = linear_37_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<int32, [1]> input_175_axes_0 = const()[name = string("input_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_9_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_9_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332296640)))];
            tensor<fp16, [1, 128, 1024]> input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_9_mlp_norm_weight_to_fp16, x = input_173_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_9_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_9_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332298752)))];
            tensor<fp16, [1, 128, 5248]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_9_mlp_Wi_weight_to_fp16, x = input_175_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1118_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = linear_38_cast_fp16)[name = string("op_1118_cast_fp16")];
            string var_1120_mode_0 = const()[name = string("op_1120_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1120_cast_fp16 = gelu(mode = var_1120_mode_0, x = var_1118_cast_fp16_0)[name = string("op_1120_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_179_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1118_cast_fp16_1)[name = string("input_179_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_9_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_9_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343046720)))];
            tensor<fp16, [1, 128, 1024]> linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_9_mlp_Wo_weight_to_fp16, x = input_179_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_183_cast_fp16 = add(x = input_173_cast_fp16, y = linear_39_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<int32, [1]> hidden_states_21_axes_0 = const()[name = string("hidden_states_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_10_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_10_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348420736)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_21_cast_fp16 = layer_norm(axes = hidden_states_21_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_10_attn_norm_weight_to_fp16, x = input_183_cast_fp16)[name = string("hidden_states_21_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_10_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_10_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348422848)))];
            tensor<fp16, [1, 128, 3072]> linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_10_attn_Wqkv_weight_to_fp16, x = hidden_states_21_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<int32, [5]> var_1139 = const()[name = string("op_1139"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_43_cast_fp16 = reshape(shape = var_1139, x = linear_40_cast_fp16)[name = string("qkv_43_cast_fp16")];
            tensor<int32, [3]> var_1141_split_sizes_0 = const()[name = string("op_1141_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1141_axis_0 = const()[name = string("op_1141_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1141_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1141_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1141_cast_fp16_2 = split(axis = var_1141_axis_0, split_sizes = var_1141_split_sizes_0, x = qkv_43_cast_fp16)[name = string("op_1141_cast_fp16")];
            tensor<int32, [1]> squeeze_30_axes_0 = const()[name = string("squeeze_30_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_30_cast_fp16 = squeeze(axes = squeeze_30_axes_0, x = var_1141_cast_fp16_0)[name = string("squeeze_30_cast_fp16")];
            tensor<int32, [1]> squeeze_31_axes_0 = const()[name = string("squeeze_31_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_31_cast_fp16 = squeeze(axes = squeeze_31_axes_0, x = var_1141_cast_fp16_1)[name = string("squeeze_31_cast_fp16")];
            tensor<int32, [1]> squeeze_32_axes_0 = const()[name = string("squeeze_32_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_32_cast_fp16 = squeeze(axes = squeeze_32_axes_0, x = var_1141_cast_fp16_2)[name = string("squeeze_32_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = squeeze_30_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 16, 128, 64]> var_1151_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_11_to_fp16)[name = string("op_1151_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_41_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_41_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_103_promoted_to_fp16 = const()[name = string("const_103_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1163_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_103_promoted_to_fp16)[name = string("op_1163_cast_fp16")];
            bool var_1165_interleave_0 = const()[name = string("op_1165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1165_cast_fp16 = concat(axis = var_49, interleave = var_1165_interleave_0, values = (var_1163_cast_fp16, x1_41_cast_fp16))[name = string("op_1165_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1166_cast_fp16 = mul(x = var_1165_cast_fp16, y = sin_11_to_fp16)[name = string("op_1166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_21_cast_fp16 = add(x = var_1151_cast_fp16, y = var_1166_cast_fp16)[name = string("q_embed_21_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = squeeze_31_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 128, 64]> var_1169_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_11_to_fp16)[name = string("op_1169_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_41_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_41_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_106_promoted_to_fp16 = const()[name = string("const_106_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1181_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_106_promoted_to_fp16)[name = string("op_1181_cast_fp16")];
            bool var_1183_interleave_0 = const()[name = string("op_1183_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1183_cast_fp16 = concat(axis = var_49, interleave = var_1183_interleave_0, values = (var_1181_cast_fp16, x1_43_cast_fp16))[name = string("op_1183_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1184_cast_fp16 = mul(x = var_1183_cast_fp16, y = sin_11_to_fp16)[name = string("op_1184_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_21_cast_fp16 = add(x = var_1169_cast_fp16, y = var_1184_cast_fp16)[name = string("k_embed_21_cast_fp16")];
            bool var_1189_transpose_x_1 = const()[name = string("op_1189_transpose_x_1"), val = bool(false)];
            bool var_1189_transpose_y_1 = const()[name = string("op_1189_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1189_cast_fp16 = matmul(transpose_x = var_1189_transpose_x_1, transpose_y = var_1189_transpose_y_1, x = q_embed_21_cast_fp16, y = k_embed_21_cast_fp16)[name = string("op_1189_cast_fp16")];
            fp16 var_1190_to_fp16 = const()[name = string("op_1190_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_41_cast_fp16 = mul(x = var_1189_cast_fp16, y = var_1190_to_fp16)[name = string("attn_weights_41_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_185_cast_fp16 = add(x = attn_weights_41_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1193_cast_fp16 = softmax(axis = var_49, x = input_185_cast_fp16)[name = string("op_1193_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = squeeze_32_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = var_1193_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_1197_perm_0 = const()[name = string("op_1197_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1199 = const()[name = string("op_1199"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1197_cast_fp16 = transpose(perm = var_1197_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 128, 1024]> var_1200_cast_fp16 = reshape(shape = var_1199, x = var_1197_cast_fp16)[name = string("op_1200_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_10_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_10_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354714368)))];
            tensor<fp16, [1, 128, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_10_attn_Wo_weight_to_fp16, x = var_1200_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_191_cast_fp16 = add(x = input_183_cast_fp16, y = linear_41_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<int32, [1]> input_193_axes_0 = const()[name = string("input_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_10_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_10_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356811584)))];
            tensor<fp16, [1, 128, 1024]> input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_10_mlp_norm_weight_to_fp16, x = input_191_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_10_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_10_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356813696)))];
            tensor<fp16, [1, 128, 5248]> linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_10_mlp_Wi_weight_to_fp16, x = input_193_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [2]> var_1212_split_sizes_0 = const()[name = string("op_1212_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1212_axis_0 = const()[name = string("op_1212_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1212_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1212_cast_fp16_1 = split(axis = var_1212_axis_0, split_sizes = var_1212_split_sizes_0, x = linear_42_cast_fp16)[name = string("op_1212_cast_fp16")];
            string var_1214_mode_0 = const()[name = string("op_1214_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1214_cast_fp16 = gelu(mode = var_1214_mode_0, x = var_1212_cast_fp16_0)[name = string("op_1214_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_197_cast_fp16 = mul(x = var_1214_cast_fp16, y = var_1212_cast_fp16_1)[name = string("input_197_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_10_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_10_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367561664)))];
            tensor<fp16, [1, 128, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_10_mlp_Wo_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_201_cast_fp16 = add(x = input_191_cast_fp16, y = linear_43_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<int32, [1]> hidden_states_23_axes_0 = const()[name = string("hidden_states_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_11_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_11_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372935680)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_23_cast_fp16 = layer_norm(axes = hidden_states_23_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_11_attn_norm_weight_to_fp16, x = input_201_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_11_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_11_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372937792)))];
            tensor<fp16, [1, 128, 3072]> linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_11_attn_Wqkv_weight_to_fp16, x = hidden_states_23_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [5]> var_1233 = const()[name = string("op_1233"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_47_cast_fp16 = reshape(shape = var_1233, x = linear_44_cast_fp16)[name = string("qkv_47_cast_fp16")];
            tensor<int32, [3]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1235_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1235_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1235_cast_fp16_2 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = qkv_47_cast_fp16)[name = string("op_1235_cast_fp16")];
            tensor<int32, [1]> squeeze_33_axes_0 = const()[name = string("squeeze_33_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_33_cast_fp16 = squeeze(axes = squeeze_33_axes_0, x = var_1235_cast_fp16_0)[name = string("squeeze_33_cast_fp16")];
            tensor<int32, [1]> squeeze_34_axes_0 = const()[name = string("squeeze_34_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_34_cast_fp16 = squeeze(axes = squeeze_34_axes_0, x = var_1235_cast_fp16_1)[name = string("squeeze_34_cast_fp16")];
            tensor<int32, [1]> squeeze_35_axes_0 = const()[name = string("squeeze_35_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_35_cast_fp16 = squeeze(axes = squeeze_35_axes_0, x = var_1235_cast_fp16_2)[name = string("squeeze_35_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = squeeze_33_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 16, 128, 64]> var_1245_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_11_to_fp16)[name = string("op_1245_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_45_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_45_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_111_promoted_to_fp16 = const()[name = string("const_111_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1257_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_111_promoted_to_fp16)[name = string("op_1257_cast_fp16")];
            bool var_1259_interleave_0 = const()[name = string("op_1259_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1259_cast_fp16 = concat(axis = var_49, interleave = var_1259_interleave_0, values = (var_1257_cast_fp16, x1_45_cast_fp16))[name = string("op_1259_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1260_cast_fp16 = mul(x = var_1259_cast_fp16, y = sin_11_to_fp16)[name = string("op_1260_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_23_cast_fp16 = add(x = var_1245_cast_fp16, y = var_1260_cast_fp16)[name = string("q_embed_23_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = squeeze_34_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 16, 128, 64]> var_1263_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_11_to_fp16)[name = string("op_1263_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_45_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_45_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1275_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_114_promoted_to_fp16)[name = string("op_1275_cast_fp16")];
            bool var_1277_interleave_0 = const()[name = string("op_1277_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1277_cast_fp16 = concat(axis = var_49, interleave = var_1277_interleave_0, values = (var_1275_cast_fp16, x1_47_cast_fp16))[name = string("op_1277_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1278_cast_fp16 = mul(x = var_1277_cast_fp16, y = sin_11_to_fp16)[name = string("op_1278_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_23_cast_fp16 = add(x = var_1263_cast_fp16, y = var_1278_cast_fp16)[name = string("k_embed_23_cast_fp16")];
            bool var_1283_transpose_x_1 = const()[name = string("op_1283_transpose_x_1"), val = bool(false)];
            bool var_1283_transpose_y_1 = const()[name = string("op_1283_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1283_cast_fp16 = matmul(transpose_x = var_1283_transpose_x_1, transpose_y = var_1283_transpose_y_1, x = q_embed_23_cast_fp16, y = k_embed_23_cast_fp16)[name = string("op_1283_cast_fp16")];
            fp16 var_1284_to_fp16 = const()[name = string("op_1284_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_45_cast_fp16 = mul(x = var_1283_cast_fp16, y = var_1284_to_fp16)[name = string("attn_weights_45_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_203_cast_fp16 = add(x = attn_weights_45_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1287_cast_fp16 = softmax(axis = var_49, x = input_203_cast_fp16)[name = string("op_1287_cast_fp16")];
            bool attn_output_67_transpose_x_0 = const()[name = string("attn_output_67_transpose_x_0"), val = bool(false)];
            bool attn_output_67_transpose_y_0 = const()[name = string("attn_output_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = squeeze_35_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_67_cast_fp16 = matmul(transpose_x = attn_output_67_transpose_x_0, transpose_y = attn_output_67_transpose_y_0, x = var_1287_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_1291_perm_0 = const()[name = string("op_1291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1293 = const()[name = string("op_1293"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1291_cast_fp16 = transpose(perm = var_1291_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 128, 1024]> var_1294_cast_fp16 = reshape(shape = var_1293, x = var_1291_cast_fp16)[name = string("op_1294_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_11_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_11_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379229312)))];
            tensor<fp16, [1, 128, 1024]> linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_11_attn_Wo_weight_to_fp16, x = var_1294_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_209_cast_fp16 = add(x = input_201_cast_fp16, y = linear_45_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_11_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_11_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381326528)))];
            tensor<fp16, [1, 128, 1024]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_11_mlp_norm_weight_to_fp16, x = input_209_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_11_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_11_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381328640)))];
            tensor<fp16, [1, 128, 5248]> linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_11_mlp_Wi_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [2]> var_1306_split_sizes_0 = const()[name = string("op_1306_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1306_axis_0 = const()[name = string("op_1306_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1306_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1306_cast_fp16_1 = split(axis = var_1306_axis_0, split_sizes = var_1306_split_sizes_0, x = linear_46_cast_fp16)[name = string("op_1306_cast_fp16")];
            string var_1308_mode_0 = const()[name = string("op_1308_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1308_cast_fp16 = gelu(mode = var_1308_mode_0, x = var_1306_cast_fp16_0)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_215_cast_fp16 = mul(x = var_1308_cast_fp16, y = var_1306_cast_fp16_1)[name = string("input_215_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_11_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_11_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392076608)))];
            tensor<fp16, [1, 128, 1024]> linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_11_mlp_Wo_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_219_cast_fp16 = add(x = input_209_cast_fp16, y = linear_47_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<int32, [1]> hidden_states_25_axes_0 = const()[name = string("hidden_states_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_12_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_12_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397450624)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_25_cast_fp16 = layer_norm(axes = hidden_states_25_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_12_attn_norm_weight_to_fp16, x = input_219_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_12_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_12_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397452736)))];
            tensor<fp16, [1, 128, 3072]> linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_12_attn_Wqkv_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [5]> var_1327 = const()[name = string("op_1327"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_51_cast_fp16 = reshape(shape = var_1327, x = linear_48_cast_fp16)[name = string("qkv_51_cast_fp16")];
            tensor<int32, [3]> var_1329_split_sizes_0 = const()[name = string("op_1329_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1329_axis_0 = const()[name = string("op_1329_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1329_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1329_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1329_cast_fp16_2 = split(axis = var_1329_axis_0, split_sizes = var_1329_split_sizes_0, x = qkv_51_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<int32, [1]> squeeze_36_axes_0 = const()[name = string("squeeze_36_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_36_cast_fp16 = squeeze(axes = squeeze_36_axes_0, x = var_1329_cast_fp16_0)[name = string("squeeze_36_cast_fp16")];
            tensor<int32, [1]> squeeze_37_axes_0 = const()[name = string("squeeze_37_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_37_cast_fp16 = squeeze(axes = squeeze_37_axes_0, x = var_1329_cast_fp16_1)[name = string("squeeze_37_cast_fp16")];
            tensor<int32, [1]> squeeze_38_axes_0 = const()[name = string("squeeze_38_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_38_cast_fp16 = squeeze(axes = squeeze_38_axes_0, x = var_1329_cast_fp16_2)[name = string("squeeze_38_cast_fp16")];
            tensor<int32, [4]> q_49_perm_0 = const()[name = string("q_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_49_perm_0 = const()[name = string("k_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_49_cast_fp16 = transpose(perm = q_49_perm_0, x = squeeze_36_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 16, 128, 64]> var_1339_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_7_to_fp16)[name = string("op_1339_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_49_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_49_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_119_promoted_to_fp16 = const()[name = string("const_119_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1351_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_119_promoted_to_fp16)[name = string("op_1351_cast_fp16")];
            bool var_1353_interleave_0 = const()[name = string("op_1353_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1353_cast_fp16 = concat(axis = var_49, interleave = var_1353_interleave_0, values = (var_1351_cast_fp16, x1_49_cast_fp16))[name = string("op_1353_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1354_cast_fp16 = mul(x = var_1353_cast_fp16, y = sin_7_to_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_25_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1354_cast_fp16)[name = string("q_embed_25_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_49_cast_fp16 = transpose(perm = k_49_perm_0, x = squeeze_37_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 16, 128, 64]> var_1357_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_7_to_fp16)[name = string("op_1357_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_49_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_49_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1369_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_1369_cast_fp16")];
            bool var_1371_interleave_0 = const()[name = string("op_1371_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1371_cast_fp16 = concat(axis = var_49, interleave = var_1371_interleave_0, values = (var_1369_cast_fp16, x1_51_cast_fp16))[name = string("op_1371_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1372_cast_fp16 = mul(x = var_1371_cast_fp16, y = sin_7_to_fp16)[name = string("op_1372_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_25_cast_fp16 = add(x = var_1357_cast_fp16, y = var_1372_cast_fp16)[name = string("k_embed_25_cast_fp16")];
            bool var_1377_transpose_x_1 = const()[name = string("op_1377_transpose_x_1"), val = bool(false)];
            bool var_1377_transpose_y_1 = const()[name = string("op_1377_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1377_cast_fp16 = matmul(transpose_x = var_1377_transpose_x_1, transpose_y = var_1377_transpose_y_1, x = q_embed_25_cast_fp16, y = k_embed_25_cast_fp16)[name = string("op_1377_cast_fp16")];
            fp16 var_1378_to_fp16 = const()[name = string("op_1378_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_49_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1378_to_fp16)[name = string("attn_weights_49_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_221_cast_fp16 = add(x = attn_weights_49_cast_fp16, y = key_bias_cast_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1381_cast_fp16 = softmax(axis = var_49, x = input_221_cast_fp16)[name = string("op_1381_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = squeeze_38_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = var_1381_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_1385_perm_0 = const()[name = string("op_1385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1387 = const()[name = string("op_1387"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1385_cast_fp16 = transpose(perm = var_1385_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 128, 1024]> var_1388_cast_fp16 = reshape(shape = var_1387, x = var_1385_cast_fp16)[name = string("op_1388_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_12_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_12_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403744256)))];
            tensor<fp16, [1, 128, 1024]> linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_12_attn_Wo_weight_to_fp16, x = var_1388_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_227_cast_fp16 = add(x = input_219_cast_fp16, y = linear_49_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<int32, [1]> input_229_axes_0 = const()[name = string("input_229_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_12_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_12_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405841472)))];
            tensor<fp16, [1, 128, 1024]> input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_12_mlp_norm_weight_to_fp16, x = input_227_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_12_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_12_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405843584)))];
            tensor<fp16, [1, 128, 5248]> linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_12_mlp_Wi_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [2]> var_1400_split_sizes_0 = const()[name = string("op_1400_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1400_axis_0 = const()[name = string("op_1400_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1400_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1400_cast_fp16_1 = split(axis = var_1400_axis_0, split_sizes = var_1400_split_sizes_0, x = linear_50_cast_fp16)[name = string("op_1400_cast_fp16")];
            string var_1402_mode_0 = const()[name = string("op_1402_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1402_cast_fp16 = gelu(mode = var_1402_mode_0, x = var_1400_cast_fp16_0)[name = string("op_1402_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_233_cast_fp16 = mul(x = var_1402_cast_fp16, y = var_1400_cast_fp16_1)[name = string("input_233_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_12_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_12_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416591552)))];
            tensor<fp16, [1, 128, 1024]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_12_mlp_Wo_weight_to_fp16, x = input_233_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_237_cast_fp16 = add(x = input_227_cast_fp16, y = linear_51_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<int32, [1]> hidden_states_27_axes_0 = const()[name = string("hidden_states_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_13_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_13_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421965568)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_27_cast_fp16 = layer_norm(axes = hidden_states_27_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_13_attn_norm_weight_to_fp16, x = input_237_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_13_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_13_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421967680)))];
            tensor<fp16, [1, 128, 3072]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_13_attn_Wqkv_weight_to_fp16, x = hidden_states_27_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<int32, [5]> var_1421 = const()[name = string("op_1421"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_55_cast_fp16 = reshape(shape = var_1421, x = linear_52_cast_fp16)[name = string("qkv_55_cast_fp16")];
            tensor<int32, [3]> var_1423_split_sizes_0 = const()[name = string("op_1423_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1423_axis_0 = const()[name = string("op_1423_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1423_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1423_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1423_cast_fp16_2 = split(axis = var_1423_axis_0, split_sizes = var_1423_split_sizes_0, x = qkv_55_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [1]> squeeze_39_axes_0 = const()[name = string("squeeze_39_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_39_cast_fp16 = squeeze(axes = squeeze_39_axes_0, x = var_1423_cast_fp16_0)[name = string("squeeze_39_cast_fp16")];
            tensor<int32, [1]> squeeze_40_axes_0 = const()[name = string("squeeze_40_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_40_cast_fp16 = squeeze(axes = squeeze_40_axes_0, x = var_1423_cast_fp16_1)[name = string("squeeze_40_cast_fp16")];
            tensor<int32, [1]> squeeze_41_axes_0 = const()[name = string("squeeze_41_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_41_cast_fp16 = squeeze(axes = squeeze_41_axes_0, x = var_1423_cast_fp16_2)[name = string("squeeze_41_cast_fp16")];
            tensor<int32, [4]> q_53_perm_0 = const()[name = string("q_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_53_perm_0 = const()[name = string("k_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_53_cast_fp16 = transpose(perm = q_53_perm_0, x = squeeze_39_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 128, 64]> var_1433_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_11_to_fp16)[name = string("op_1433_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_53_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_53_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_127_promoted_to_fp16 = const()[name = string("const_127_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1445_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_127_promoted_to_fp16)[name = string("op_1445_cast_fp16")];
            bool var_1447_interleave_0 = const()[name = string("op_1447_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1447_cast_fp16 = concat(axis = var_49, interleave = var_1447_interleave_0, values = (var_1445_cast_fp16, x1_53_cast_fp16))[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1448_cast_fp16 = mul(x = var_1447_cast_fp16, y = sin_11_to_fp16)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_27_cast_fp16 = add(x = var_1433_cast_fp16, y = var_1448_cast_fp16)[name = string("q_embed_27_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_53_cast_fp16 = transpose(perm = k_53_perm_0, x = squeeze_40_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 16, 128, 64]> var_1451_cast_fp16 = mul(x = k_53_cast_fp16, y = cos_11_to_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_53_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_53_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1463_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_1463_cast_fp16")];
            bool var_1465_interleave_0 = const()[name = string("op_1465_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1465_cast_fp16 = concat(axis = var_49, interleave = var_1465_interleave_0, values = (var_1463_cast_fp16, x1_55_cast_fp16))[name = string("op_1465_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1466_cast_fp16 = mul(x = var_1465_cast_fp16, y = sin_11_to_fp16)[name = string("op_1466_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_27_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1466_cast_fp16)[name = string("k_embed_27_cast_fp16")];
            bool var_1471_transpose_x_1 = const()[name = string("op_1471_transpose_x_1"), val = bool(false)];
            bool var_1471_transpose_y_1 = const()[name = string("op_1471_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1471_cast_fp16 = matmul(transpose_x = var_1471_transpose_x_1, transpose_y = var_1471_transpose_y_1, x = q_embed_27_cast_fp16, y = k_embed_27_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp16 var_1472_to_fp16 = const()[name = string("op_1472_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_53_cast_fp16 = mul(x = var_1471_cast_fp16, y = var_1472_to_fp16)[name = string("attn_weights_53_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_239_cast_fp16 = add(x = attn_weights_53_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1475_cast_fp16 = softmax(axis = var_49, x = input_239_cast_fp16)[name = string("op_1475_cast_fp16")];
            bool attn_output_79_transpose_x_0 = const()[name = string("attn_output_79_transpose_x_0"), val = bool(false)];
            bool attn_output_79_transpose_y_0 = const()[name = string("attn_output_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = squeeze_41_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_79_cast_fp16 = matmul(transpose_x = attn_output_79_transpose_x_0, transpose_y = attn_output_79_transpose_y_0, x = var_1475_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [4]> var_1479_perm_0 = const()[name = string("op_1479_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1481 = const()[name = string("op_1481"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1479_cast_fp16 = transpose(perm = var_1479_perm_0, x = attn_output_79_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 128, 1024]> var_1482_cast_fp16 = reshape(shape = var_1481, x = var_1479_cast_fp16)[name = string("op_1482_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_13_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_13_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428259200)))];
            tensor<fp16, [1, 128, 1024]> linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_13_attn_Wo_weight_to_fp16, x = var_1482_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_245_cast_fp16 = add(x = input_237_cast_fp16, y = linear_53_cast_fp16)[name = string("input_245_cast_fp16")];
            tensor<int32, [1]> input_247_axes_0 = const()[name = string("input_247_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_13_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_13_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430356416)))];
            tensor<fp16, [1, 128, 1024]> input_247_cast_fp16 = layer_norm(axes = input_247_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_13_mlp_norm_weight_to_fp16, x = input_245_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_13_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_13_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430358528)))];
            tensor<fp16, [1, 128, 5248]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_13_mlp_Wi_weight_to_fp16, x = input_247_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<int32, [2]> var_1494_split_sizes_0 = const()[name = string("op_1494_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1494_axis_0 = const()[name = string("op_1494_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1494_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1494_cast_fp16_1 = split(axis = var_1494_axis_0, split_sizes = var_1494_split_sizes_0, x = linear_54_cast_fp16)[name = string("op_1494_cast_fp16")];
            string var_1496_mode_0 = const()[name = string("op_1496_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1496_cast_fp16 = gelu(mode = var_1496_mode_0, x = var_1494_cast_fp16_0)[name = string("op_1496_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_251_cast_fp16 = mul(x = var_1496_cast_fp16, y = var_1494_cast_fp16_1)[name = string("input_251_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_13_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_13_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441106496)))];
            tensor<fp16, [1, 128, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_13_mlp_Wo_weight_to_fp16, x = input_251_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_55_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<int32, [1]> hidden_states_29_axes_0 = const()[name = string("hidden_states_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_14_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_14_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446480512)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_29_cast_fp16 = layer_norm(axes = hidden_states_29_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_14_attn_norm_weight_to_fp16, x = input_255_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_14_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_14_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446482624)))];
            tensor<fp16, [1, 128, 3072]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_14_attn_Wqkv_weight_to_fp16, x = hidden_states_29_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [5]> var_1515 = const()[name = string("op_1515"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_59_cast_fp16 = reshape(shape = var_1515, x = linear_56_cast_fp16)[name = string("qkv_59_cast_fp16")];
            tensor<int32, [3]> var_1517_split_sizes_0 = const()[name = string("op_1517_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1517_axis_0 = const()[name = string("op_1517_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1517_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1517_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1517_cast_fp16_2 = split(axis = var_1517_axis_0, split_sizes = var_1517_split_sizes_0, x = qkv_59_cast_fp16)[name = string("op_1517_cast_fp16")];
            tensor<int32, [1]> squeeze_42_axes_0 = const()[name = string("squeeze_42_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_42_cast_fp16 = squeeze(axes = squeeze_42_axes_0, x = var_1517_cast_fp16_0)[name = string("squeeze_42_cast_fp16")];
            tensor<int32, [1]> squeeze_43_axes_0 = const()[name = string("squeeze_43_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_43_cast_fp16 = squeeze(axes = squeeze_43_axes_0, x = var_1517_cast_fp16_1)[name = string("squeeze_43_cast_fp16")];
            tensor<int32, [1]> squeeze_44_axes_0 = const()[name = string("squeeze_44_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_44_cast_fp16 = squeeze(axes = squeeze_44_axes_0, x = var_1517_cast_fp16_2)[name = string("squeeze_44_cast_fp16")];
            tensor<int32, [4]> q_57_perm_0 = const()[name = string("q_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_57_perm_0 = const()[name = string("k_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_57_cast_fp16 = transpose(perm = q_57_perm_0, x = squeeze_42_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 128, 64]> var_1527_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_11_to_fp16)[name = string("op_1527_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_57_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_57_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1539_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_135_promoted_to_fp16)[name = string("op_1539_cast_fp16")];
            bool var_1541_interleave_0 = const()[name = string("op_1541_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1541_cast_fp16 = concat(axis = var_49, interleave = var_1541_interleave_0, values = (var_1539_cast_fp16, x1_57_cast_fp16))[name = string("op_1541_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1542_cast_fp16 = mul(x = var_1541_cast_fp16, y = sin_11_to_fp16)[name = string("op_1542_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_29_cast_fp16 = add(x = var_1527_cast_fp16, y = var_1542_cast_fp16)[name = string("q_embed_29_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_57_cast_fp16 = transpose(perm = k_57_perm_0, x = squeeze_43_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 128, 64]> var_1545_cast_fp16 = mul(x = k_57_cast_fp16, y = cos_11_to_fp16)[name = string("op_1545_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_57_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_57_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_138_promoted_to_fp16 = const()[name = string("const_138_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1557_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_138_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            bool var_1559_interleave_0 = const()[name = string("op_1559_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1559_cast_fp16 = concat(axis = var_49, interleave = var_1559_interleave_0, values = (var_1557_cast_fp16, x1_59_cast_fp16))[name = string("op_1559_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1560_cast_fp16 = mul(x = var_1559_cast_fp16, y = sin_11_to_fp16)[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_29_cast_fp16 = add(x = var_1545_cast_fp16, y = var_1560_cast_fp16)[name = string("k_embed_29_cast_fp16")];
            bool var_1565_transpose_x_1 = const()[name = string("op_1565_transpose_x_1"), val = bool(false)];
            bool var_1565_transpose_y_1 = const()[name = string("op_1565_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1565_cast_fp16 = matmul(transpose_x = var_1565_transpose_x_1, transpose_y = var_1565_transpose_y_1, x = q_embed_29_cast_fp16, y = k_embed_29_cast_fp16)[name = string("op_1565_cast_fp16")];
            fp16 var_1566_to_fp16 = const()[name = string("op_1566_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_57_cast_fp16 = mul(x = var_1565_cast_fp16, y = var_1566_to_fp16)[name = string("attn_weights_57_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_257_cast_fp16 = add(x = attn_weights_57_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1569_cast_fp16 = softmax(axis = var_49, x = input_257_cast_fp16)[name = string("op_1569_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = squeeze_44_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = var_1569_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_1573_perm_0 = const()[name = string("op_1573_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1575 = const()[name = string("op_1575"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1573_cast_fp16 = transpose(perm = var_1573_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 128, 1024]> var_1576_cast_fp16 = reshape(shape = var_1575, x = var_1573_cast_fp16)[name = string("op_1576_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_14_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_14_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452774144)))];
            tensor<fp16, [1, 128, 1024]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_14_attn_Wo_weight_to_fp16, x = var_1576_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_263_cast_fp16 = add(x = input_255_cast_fp16, y = linear_57_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<int32, [1]> input_265_axes_0 = const()[name = string("input_265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_14_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_14_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454871360)))];
            tensor<fp16, [1, 128, 1024]> input_265_cast_fp16 = layer_norm(axes = input_265_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_14_mlp_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_14_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_14_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454873472)))];
            tensor<fp16, [1, 128, 5248]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_14_mlp_Wi_weight_to_fp16, x = input_265_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [2]> var_1588_split_sizes_0 = const()[name = string("op_1588_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1588_axis_0 = const()[name = string("op_1588_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1588_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1588_cast_fp16_1 = split(axis = var_1588_axis_0, split_sizes = var_1588_split_sizes_0, x = linear_58_cast_fp16)[name = string("op_1588_cast_fp16")];
            string var_1590_mode_0 = const()[name = string("op_1590_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1590_cast_fp16 = gelu(mode = var_1590_mode_0, x = var_1588_cast_fp16_0)[name = string("op_1590_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_269_cast_fp16 = mul(x = var_1590_cast_fp16, y = var_1588_cast_fp16_1)[name = string("input_269_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_14_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_14_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465621440)))];
            tensor<fp16, [1, 128, 1024]> linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_14_mlp_Wo_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_273_cast_fp16 = add(x = input_263_cast_fp16, y = linear_59_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<int32, [1]> hidden_states_31_axes_0 = const()[name = string("hidden_states_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_15_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_15_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470995456)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_31_cast_fp16 = layer_norm(axes = hidden_states_31_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_15_attn_norm_weight_to_fp16, x = input_273_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_15_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_15_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470997568)))];
            tensor<fp16, [1, 128, 3072]> linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_15_attn_Wqkv_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<int32, [5]> var_1609 = const()[name = string("op_1609"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_63_cast_fp16 = reshape(shape = var_1609, x = linear_60_cast_fp16)[name = string("qkv_63_cast_fp16")];
            tensor<int32, [3]> var_1611_split_sizes_0 = const()[name = string("op_1611_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1611_axis_0 = const()[name = string("op_1611_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1611_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1611_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1611_cast_fp16_2 = split(axis = var_1611_axis_0, split_sizes = var_1611_split_sizes_0, x = qkv_63_cast_fp16)[name = string("op_1611_cast_fp16")];
            tensor<int32, [1]> squeeze_45_axes_0 = const()[name = string("squeeze_45_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_45_cast_fp16 = squeeze(axes = squeeze_45_axes_0, x = var_1611_cast_fp16_0)[name = string("squeeze_45_cast_fp16")];
            tensor<int32, [1]> squeeze_46_axes_0 = const()[name = string("squeeze_46_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_46_cast_fp16 = squeeze(axes = squeeze_46_axes_0, x = var_1611_cast_fp16_1)[name = string("squeeze_46_cast_fp16")];
            tensor<int32, [1]> squeeze_47_axes_0 = const()[name = string("squeeze_47_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_47_cast_fp16 = squeeze(axes = squeeze_47_axes_0, x = var_1611_cast_fp16_2)[name = string("squeeze_47_cast_fp16")];
            tensor<int32, [4]> q_61_perm_0 = const()[name = string("q_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_61_perm_0 = const()[name = string("k_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_61_cast_fp16 = transpose(perm = q_61_perm_0, x = squeeze_45_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 16, 128, 64]> var_1621_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_7_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_61_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_61_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_143_promoted_to_fp16 = const()[name = string("const_143_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1633_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_143_promoted_to_fp16)[name = string("op_1633_cast_fp16")];
            bool var_1635_interleave_0 = const()[name = string("op_1635_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1635_cast_fp16 = concat(axis = var_49, interleave = var_1635_interleave_0, values = (var_1633_cast_fp16, x1_61_cast_fp16))[name = string("op_1635_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1636_cast_fp16 = mul(x = var_1635_cast_fp16, y = sin_7_to_fp16)[name = string("op_1636_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_31_cast_fp16 = add(x = var_1621_cast_fp16, y = var_1636_cast_fp16)[name = string("q_embed_31_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_61_cast_fp16 = transpose(perm = k_61_perm_0, x = squeeze_46_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 128, 64]> var_1639_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_7_to_fp16)[name = string("op_1639_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_61_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_61_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_146_promoted_to_fp16 = const()[name = string("const_146_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1651_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_146_promoted_to_fp16)[name = string("op_1651_cast_fp16")];
            bool var_1653_interleave_0 = const()[name = string("op_1653_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1653_cast_fp16 = concat(axis = var_49, interleave = var_1653_interleave_0, values = (var_1651_cast_fp16, x1_63_cast_fp16))[name = string("op_1653_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1654_cast_fp16 = mul(x = var_1653_cast_fp16, y = sin_7_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_31_cast_fp16 = add(x = var_1639_cast_fp16, y = var_1654_cast_fp16)[name = string("k_embed_31_cast_fp16")];
            bool var_1659_transpose_x_1 = const()[name = string("op_1659_transpose_x_1"), val = bool(false)];
            bool var_1659_transpose_y_1 = const()[name = string("op_1659_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1659_cast_fp16 = matmul(transpose_x = var_1659_transpose_x_1, transpose_y = var_1659_transpose_y_1, x = q_embed_31_cast_fp16, y = k_embed_31_cast_fp16)[name = string("op_1659_cast_fp16")];
            fp16 var_1660_to_fp16 = const()[name = string("op_1660_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_61_cast_fp16 = mul(x = var_1659_cast_fp16, y = var_1660_to_fp16)[name = string("attn_weights_61_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_275_cast_fp16 = add(x = attn_weights_61_cast_fp16, y = key_bias_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1663_cast_fp16 = softmax(axis = var_49, x = input_275_cast_fp16)[name = string("op_1663_cast_fp16")];
            bool attn_output_91_transpose_x_0 = const()[name = string("attn_output_91_transpose_x_0"), val = bool(false)];
            bool attn_output_91_transpose_y_0 = const()[name = string("attn_output_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = squeeze_47_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_91_cast_fp16 = matmul(transpose_x = attn_output_91_transpose_x_0, transpose_y = attn_output_91_transpose_y_0, x = var_1663_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_1667_perm_0 = const()[name = string("op_1667_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1669 = const()[name = string("op_1669"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1667_cast_fp16 = transpose(perm = var_1667_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 128, 1024]> var_1670_cast_fp16 = reshape(shape = var_1669, x = var_1667_cast_fp16)[name = string("op_1670_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_15_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_15_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477289088)))];
            tensor<fp16, [1, 128, 1024]> linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_15_attn_Wo_weight_to_fp16, x = var_1670_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_281_cast_fp16 = add(x = input_273_cast_fp16, y = linear_61_cast_fp16)[name = string("input_281_cast_fp16")];
            tensor<int32, [1]> input_283_axes_0 = const()[name = string("input_283_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_15_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_15_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(479386304)))];
            tensor<fp16, [1, 128, 1024]> input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_15_mlp_norm_weight_to_fp16, x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_15_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_15_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(479388416)))];
            tensor<fp16, [1, 128, 5248]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_15_mlp_Wi_weight_to_fp16, x = input_283_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<int32, [2]> var_1682_split_sizes_0 = const()[name = string("op_1682_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1682_axis_0 = const()[name = string("op_1682_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1682_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1682_cast_fp16_1 = split(axis = var_1682_axis_0, split_sizes = var_1682_split_sizes_0, x = linear_62_cast_fp16)[name = string("op_1682_cast_fp16")];
            string var_1684_mode_0 = const()[name = string("op_1684_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1684_cast_fp16 = gelu(mode = var_1684_mode_0, x = var_1682_cast_fp16_0)[name = string("op_1684_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_287_cast_fp16 = mul(x = var_1684_cast_fp16, y = var_1682_cast_fp16_1)[name = string("input_287_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_15_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_15_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490136384)))];
            tensor<fp16, [1, 128, 1024]> linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_15_mlp_Wo_weight_to_fp16, x = input_287_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_291_cast_fp16 = add(x = input_281_cast_fp16, y = linear_63_cast_fp16)[name = string("input_291_cast_fp16")];
            tensor<int32, [1]> hidden_states_33_axes_0 = const()[name = string("hidden_states_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_16_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_16_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495510400)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_33_cast_fp16 = layer_norm(axes = hidden_states_33_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_16_attn_norm_weight_to_fp16, x = input_291_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_16_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_16_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495512512)))];
            tensor<fp16, [1, 128, 3072]> linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_16_attn_Wqkv_weight_to_fp16, x = hidden_states_33_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [5]> var_1703 = const()[name = string("op_1703"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_67_cast_fp16 = reshape(shape = var_1703, x = linear_64_cast_fp16)[name = string("qkv_67_cast_fp16")];
            tensor<int32, [3]> var_1705_split_sizes_0 = const()[name = string("op_1705_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1705_axis_0 = const()[name = string("op_1705_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1705_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1705_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1705_cast_fp16_2 = split(axis = var_1705_axis_0, split_sizes = var_1705_split_sizes_0, x = qkv_67_cast_fp16)[name = string("op_1705_cast_fp16")];
            tensor<int32, [1]> squeeze_48_axes_0 = const()[name = string("squeeze_48_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_48_cast_fp16 = squeeze(axes = squeeze_48_axes_0, x = var_1705_cast_fp16_0)[name = string("squeeze_48_cast_fp16")];
            tensor<int32, [1]> squeeze_49_axes_0 = const()[name = string("squeeze_49_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_49_cast_fp16 = squeeze(axes = squeeze_49_axes_0, x = var_1705_cast_fp16_1)[name = string("squeeze_49_cast_fp16")];
            tensor<int32, [1]> squeeze_50_axes_0 = const()[name = string("squeeze_50_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_50_cast_fp16 = squeeze(axes = squeeze_50_axes_0, x = var_1705_cast_fp16_2)[name = string("squeeze_50_cast_fp16")];
            tensor<int32, [4]> q_65_perm_0 = const()[name = string("q_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_65_perm_0 = const()[name = string("k_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_65_cast_fp16 = transpose(perm = q_65_perm_0, x = squeeze_48_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 16, 128, 64]> var_1715_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_11_to_fp16)[name = string("op_1715_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_65_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_65_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_151_promoted_to_fp16 = const()[name = string("const_151_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1727_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_151_promoted_to_fp16)[name = string("op_1727_cast_fp16")];
            bool var_1729_interleave_0 = const()[name = string("op_1729_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1729_cast_fp16 = concat(axis = var_49, interleave = var_1729_interleave_0, values = (var_1727_cast_fp16, x1_65_cast_fp16))[name = string("op_1729_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1730_cast_fp16 = mul(x = var_1729_cast_fp16, y = sin_11_to_fp16)[name = string("op_1730_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_33_cast_fp16 = add(x = var_1715_cast_fp16, y = var_1730_cast_fp16)[name = string("q_embed_33_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_65_cast_fp16 = transpose(perm = k_65_perm_0, x = squeeze_49_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 16, 128, 64]> var_1733_cast_fp16 = mul(x = k_65_cast_fp16, y = cos_11_to_fp16)[name = string("op_1733_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_65_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_65_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_154_promoted_to_fp16 = const()[name = string("const_154_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1745_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_154_promoted_to_fp16)[name = string("op_1745_cast_fp16")];
            bool var_1747_interleave_0 = const()[name = string("op_1747_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1747_cast_fp16 = concat(axis = var_49, interleave = var_1747_interleave_0, values = (var_1745_cast_fp16, x1_67_cast_fp16))[name = string("op_1747_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1748_cast_fp16 = mul(x = var_1747_cast_fp16, y = sin_11_to_fp16)[name = string("op_1748_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_33_cast_fp16 = add(x = var_1733_cast_fp16, y = var_1748_cast_fp16)[name = string("k_embed_33_cast_fp16")];
            bool var_1753_transpose_x_1 = const()[name = string("op_1753_transpose_x_1"), val = bool(false)];
            bool var_1753_transpose_y_1 = const()[name = string("op_1753_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1753_cast_fp16 = matmul(transpose_x = var_1753_transpose_x_1, transpose_y = var_1753_transpose_y_1, x = q_embed_33_cast_fp16, y = k_embed_33_cast_fp16)[name = string("op_1753_cast_fp16")];
            fp16 var_1754_to_fp16 = const()[name = string("op_1754_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_65_cast_fp16 = mul(x = var_1753_cast_fp16, y = var_1754_to_fp16)[name = string("attn_weights_65_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_293_cast_fp16 = add(x = attn_weights_65_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_293_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1757_cast_fp16 = softmax(axis = var_49, x = input_293_cast_fp16)[name = string("op_1757_cast_fp16")];
            bool attn_output_97_transpose_x_0 = const()[name = string("attn_output_97_transpose_x_0"), val = bool(false)];
            bool attn_output_97_transpose_y_0 = const()[name = string("attn_output_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = squeeze_50_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_0, transpose_y = attn_output_97_transpose_y_0, x = var_1757_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_97_cast_fp16")];
            tensor<int32, [4]> var_1761_perm_0 = const()[name = string("op_1761_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1763 = const()[name = string("op_1763"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1761_cast_fp16 = transpose(perm = var_1761_perm_0, x = attn_output_97_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 128, 1024]> var_1764_cast_fp16 = reshape(shape = var_1763, x = var_1761_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_16_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_16_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501804032)))];
            tensor<fp16, [1, 128, 1024]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_16_attn_Wo_weight_to_fp16, x = var_1764_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_299_cast_fp16 = add(x = input_291_cast_fp16, y = linear_65_cast_fp16)[name = string("input_299_cast_fp16")];
            tensor<int32, [1]> input_301_axes_0 = const()[name = string("input_301_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_16_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_16_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503901248)))];
            tensor<fp16, [1, 128, 1024]> input_301_cast_fp16 = layer_norm(axes = input_301_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_16_mlp_norm_weight_to_fp16, x = input_299_cast_fp16)[name = string("input_301_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_16_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_16_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503903360)))];
            tensor<fp16, [1, 128, 5248]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_16_mlp_Wi_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [2]> var_1776_split_sizes_0 = const()[name = string("op_1776_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1776_axis_0 = const()[name = string("op_1776_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1776_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1776_cast_fp16_1 = split(axis = var_1776_axis_0, split_sizes = var_1776_split_sizes_0, x = linear_66_cast_fp16)[name = string("op_1776_cast_fp16")];
            string var_1778_mode_0 = const()[name = string("op_1778_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1778_cast_fp16 = gelu(mode = var_1778_mode_0, x = var_1776_cast_fp16_0)[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_305_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_1776_cast_fp16_1)[name = string("input_305_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_16_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_16_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514651328)))];
            tensor<fp16, [1, 128, 1024]> linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_16_mlp_Wo_weight_to_fp16, x = input_305_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_309_cast_fp16 = add(x = input_299_cast_fp16, y = linear_67_cast_fp16)[name = string("input_309_cast_fp16")];
            tensor<int32, [1]> hidden_states_35_axes_0 = const()[name = string("hidden_states_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_17_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_17_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520025344)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_35_cast_fp16 = layer_norm(axes = hidden_states_35_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_17_attn_norm_weight_to_fp16, x = input_309_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_17_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_17_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520027456)))];
            tensor<fp16, [1, 128, 3072]> linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_17_attn_Wqkv_weight_to_fp16, x = hidden_states_35_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [5]> var_1797 = const()[name = string("op_1797"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_71_cast_fp16 = reshape(shape = var_1797, x = linear_68_cast_fp16)[name = string("qkv_71_cast_fp16")];
            tensor<int32, [3]> var_1799_split_sizes_0 = const()[name = string("op_1799_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1799_axis_0 = const()[name = string("op_1799_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1799_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1799_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1799_cast_fp16_2 = split(axis = var_1799_axis_0, split_sizes = var_1799_split_sizes_0, x = qkv_71_cast_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> squeeze_51_axes_0 = const()[name = string("squeeze_51_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_51_cast_fp16 = squeeze(axes = squeeze_51_axes_0, x = var_1799_cast_fp16_0)[name = string("squeeze_51_cast_fp16")];
            tensor<int32, [1]> squeeze_52_axes_0 = const()[name = string("squeeze_52_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_52_cast_fp16 = squeeze(axes = squeeze_52_axes_0, x = var_1799_cast_fp16_1)[name = string("squeeze_52_cast_fp16")];
            tensor<int32, [1]> squeeze_53_axes_0 = const()[name = string("squeeze_53_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_53_cast_fp16 = squeeze(axes = squeeze_53_axes_0, x = var_1799_cast_fp16_2)[name = string("squeeze_53_cast_fp16")];
            tensor<int32, [4]> q_69_perm_0 = const()[name = string("q_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_69_perm_0 = const()[name = string("k_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_69_cast_fp16 = transpose(perm = q_69_perm_0, x = squeeze_51_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 16, 128, 64]> var_1809_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_11_to_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_69_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_69_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1821_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1821_cast_fp16")];
            bool var_1823_interleave_0 = const()[name = string("op_1823_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1823_cast_fp16 = concat(axis = var_49, interleave = var_1823_interleave_0, values = (var_1821_cast_fp16, x1_69_cast_fp16))[name = string("op_1823_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1824_cast_fp16 = mul(x = var_1823_cast_fp16, y = sin_11_to_fp16)[name = string("op_1824_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_35_cast_fp16 = add(x = var_1809_cast_fp16, y = var_1824_cast_fp16)[name = string("q_embed_35_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_69_cast_fp16 = transpose(perm = k_69_perm_0, x = squeeze_52_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 16, 128, 64]> var_1827_cast_fp16 = mul(x = k_69_cast_fp16, y = cos_11_to_fp16)[name = string("op_1827_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_69_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_69_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1839_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1839_cast_fp16")];
            bool var_1841_interleave_0 = const()[name = string("op_1841_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1841_cast_fp16 = concat(axis = var_49, interleave = var_1841_interleave_0, values = (var_1839_cast_fp16, x1_71_cast_fp16))[name = string("op_1841_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1842_cast_fp16 = mul(x = var_1841_cast_fp16, y = sin_11_to_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_35_cast_fp16 = add(x = var_1827_cast_fp16, y = var_1842_cast_fp16)[name = string("k_embed_35_cast_fp16")];
            bool var_1847_transpose_x_1 = const()[name = string("op_1847_transpose_x_1"), val = bool(false)];
            bool var_1847_transpose_y_1 = const()[name = string("op_1847_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1847_cast_fp16 = matmul(transpose_x = var_1847_transpose_x_1, transpose_y = var_1847_transpose_y_1, x = q_embed_35_cast_fp16, y = k_embed_35_cast_fp16)[name = string("op_1847_cast_fp16")];
            fp16 var_1848_to_fp16 = const()[name = string("op_1848_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_69_cast_fp16 = mul(x = var_1847_cast_fp16, y = var_1848_to_fp16)[name = string("attn_weights_69_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_311_cast_fp16 = add(x = attn_weights_69_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_311_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1851_cast_fp16 = softmax(axis = var_49, x = input_311_cast_fp16)[name = string("op_1851_cast_fp16")];
            bool attn_output_103_transpose_x_0 = const()[name = string("attn_output_103_transpose_x_0"), val = bool(false)];
            bool attn_output_103_transpose_y_0 = const()[name = string("attn_output_103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = squeeze_53_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_103_cast_fp16 = matmul(transpose_x = attn_output_103_transpose_x_0, transpose_y = attn_output_103_transpose_y_0, x = var_1851_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [4]> var_1855_perm_0 = const()[name = string("op_1855_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1857 = const()[name = string("op_1857"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1855_cast_fp16 = transpose(perm = var_1855_perm_0, x = attn_output_103_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 128, 1024]> var_1858_cast_fp16 = reshape(shape = var_1857, x = var_1855_cast_fp16)[name = string("op_1858_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_17_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_17_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526318976)))];
            tensor<fp16, [1, 128, 1024]> linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_17_attn_Wo_weight_to_fp16, x = var_1858_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_317_cast_fp16 = add(x = input_309_cast_fp16, y = linear_69_cast_fp16)[name = string("input_317_cast_fp16")];
            tensor<int32, [1]> input_319_axes_0 = const()[name = string("input_319_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_17_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_17_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528416192)))];
            tensor<fp16, [1, 128, 1024]> input_319_cast_fp16 = layer_norm(axes = input_319_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_17_mlp_norm_weight_to_fp16, x = input_317_cast_fp16)[name = string("input_319_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_17_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_17_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528418304)))];
            tensor<fp16, [1, 128, 5248]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_17_mlp_Wi_weight_to_fp16, x = input_319_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [2]> var_1870_split_sizes_0 = const()[name = string("op_1870_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1870_axis_0 = const()[name = string("op_1870_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1870_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1870_cast_fp16_1 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_70_cast_fp16)[name = string("op_1870_cast_fp16")];
            string var_1872_mode_0 = const()[name = string("op_1872_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1872_cast_fp16 = gelu(mode = var_1872_mode_0, x = var_1870_cast_fp16_0)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_323_cast_fp16 = mul(x = var_1872_cast_fp16, y = var_1870_cast_fp16_1)[name = string("input_323_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_17_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_17_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539166272)))];
            tensor<fp16, [1, 128, 1024]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_17_mlp_Wo_weight_to_fp16, x = input_323_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_327_cast_fp16 = add(x = input_317_cast_fp16, y = linear_71_cast_fp16)[name = string("input_327_cast_fp16")];
            tensor<int32, [1]> hidden_states_37_axes_0 = const()[name = string("hidden_states_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_18_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_18_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544540288)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_37_cast_fp16 = layer_norm(axes = hidden_states_37_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_18_attn_norm_weight_to_fp16, x = input_327_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_18_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_18_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544542400)))];
            tensor<fp16, [1, 128, 3072]> linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_18_attn_Wqkv_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [5]> var_1891 = const()[name = string("op_1891"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_75_cast_fp16 = reshape(shape = var_1891, x = linear_72_cast_fp16)[name = string("qkv_75_cast_fp16")];
            tensor<int32, [3]> var_1893_split_sizes_0 = const()[name = string("op_1893_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1893_axis_0 = const()[name = string("op_1893_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1893_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1893_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1893_cast_fp16_2 = split(axis = var_1893_axis_0, split_sizes = var_1893_split_sizes_0, x = qkv_75_cast_fp16)[name = string("op_1893_cast_fp16")];
            tensor<int32, [1]> squeeze_54_axes_0 = const()[name = string("squeeze_54_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_54_cast_fp16 = squeeze(axes = squeeze_54_axes_0, x = var_1893_cast_fp16_0)[name = string("squeeze_54_cast_fp16")];
            tensor<int32, [1]> squeeze_55_axes_0 = const()[name = string("squeeze_55_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_55_cast_fp16 = squeeze(axes = squeeze_55_axes_0, x = var_1893_cast_fp16_1)[name = string("squeeze_55_cast_fp16")];
            tensor<int32, [1]> squeeze_56_axes_0 = const()[name = string("squeeze_56_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_56_cast_fp16 = squeeze(axes = squeeze_56_axes_0, x = var_1893_cast_fp16_2)[name = string("squeeze_56_cast_fp16")];
            tensor<int32, [4]> q_73_perm_0 = const()[name = string("q_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_73_perm_0 = const()[name = string("k_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_73_cast_fp16 = transpose(perm = q_73_perm_0, x = squeeze_54_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 128, 64]> var_1903_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_7_to_fp16)[name = string("op_1903_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_73_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_73_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_167_promoted_to_fp16 = const()[name = string("const_167_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1915_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_167_promoted_to_fp16)[name = string("op_1915_cast_fp16")];
            bool var_1917_interleave_0 = const()[name = string("op_1917_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1917_cast_fp16 = concat(axis = var_49, interleave = var_1917_interleave_0, values = (var_1915_cast_fp16, x1_73_cast_fp16))[name = string("op_1917_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1918_cast_fp16 = mul(x = var_1917_cast_fp16, y = sin_7_to_fp16)[name = string("op_1918_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_37_cast_fp16 = add(x = var_1903_cast_fp16, y = var_1918_cast_fp16)[name = string("q_embed_37_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_73_cast_fp16 = transpose(perm = k_73_perm_0, x = squeeze_55_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 16, 128, 64]> var_1921_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_7_to_fp16)[name = string("op_1921_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_73_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_73_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_170_promoted_to_fp16 = const()[name = string("const_170_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_1933_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_170_promoted_to_fp16)[name = string("op_1933_cast_fp16")];
            bool var_1935_interleave_0 = const()[name = string("op_1935_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1935_cast_fp16 = concat(axis = var_49, interleave = var_1935_interleave_0, values = (var_1933_cast_fp16, x1_75_cast_fp16))[name = string("op_1935_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1936_cast_fp16 = mul(x = var_1935_cast_fp16, y = sin_7_to_fp16)[name = string("op_1936_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_37_cast_fp16 = add(x = var_1921_cast_fp16, y = var_1936_cast_fp16)[name = string("k_embed_37_cast_fp16")];
            bool var_1941_transpose_x_1 = const()[name = string("op_1941_transpose_x_1"), val = bool(false)];
            bool var_1941_transpose_y_1 = const()[name = string("op_1941_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_1941_cast_fp16 = matmul(transpose_x = var_1941_transpose_x_1, transpose_y = var_1941_transpose_y_1, x = q_embed_37_cast_fp16, y = k_embed_37_cast_fp16)[name = string("op_1941_cast_fp16")];
            fp16 var_1942_to_fp16 = const()[name = string("op_1942_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_73_cast_fp16 = mul(x = var_1941_cast_fp16, y = var_1942_to_fp16)[name = string("attn_weights_73_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_329_cast_fp16 = add(x = attn_weights_73_cast_fp16, y = key_bias_cast_fp16)[name = string("input_329_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1945_cast_fp16 = softmax(axis = var_49, x = input_329_cast_fp16)[name = string("op_1945_cast_fp16")];
            bool attn_output_109_transpose_x_0 = const()[name = string("attn_output_109_transpose_x_0"), val = bool(false)];
            bool attn_output_109_transpose_y_0 = const()[name = string("attn_output_109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = squeeze_56_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_109_cast_fp16 = matmul(transpose_x = attn_output_109_transpose_x_0, transpose_y = attn_output_109_transpose_y_0, x = var_1945_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_109_cast_fp16")];
            tensor<int32, [4]> var_1949_perm_0 = const()[name = string("op_1949_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1951 = const()[name = string("op_1951"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_1949_cast_fp16 = transpose(perm = var_1949_perm_0, x = attn_output_109_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 128, 1024]> var_1952_cast_fp16 = reshape(shape = var_1951, x = var_1949_cast_fp16)[name = string("op_1952_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_18_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_18_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550833920)))];
            tensor<fp16, [1, 128, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_18_attn_Wo_weight_to_fp16, x = var_1952_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_335_cast_fp16 = add(x = input_327_cast_fp16, y = linear_73_cast_fp16)[name = string("input_335_cast_fp16")];
            tensor<int32, [1]> input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_18_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_18_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552931136)))];
            tensor<fp16, [1, 128, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_18_mlp_norm_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_18_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_18_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552933248)))];
            tensor<fp16, [1, 128, 5248]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_18_mlp_Wi_weight_to_fp16, x = input_337_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [2]> var_1964_split_sizes_0 = const()[name = string("op_1964_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1964_axis_0 = const()[name = string("op_1964_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_1964_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_1964_cast_fp16_1 = split(axis = var_1964_axis_0, split_sizes = var_1964_split_sizes_0, x = linear_74_cast_fp16)[name = string("op_1964_cast_fp16")];
            string var_1966_mode_0 = const()[name = string("op_1966_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_1966_cast_fp16 = gelu(mode = var_1966_mode_0, x = var_1964_cast_fp16_0)[name = string("op_1966_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_341_cast_fp16 = mul(x = var_1966_cast_fp16, y = var_1964_cast_fp16_1)[name = string("input_341_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_18_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_18_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563681216)))];
            tensor<fp16, [1, 128, 1024]> linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_18_mlp_Wo_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_345_cast_fp16 = add(x = input_335_cast_fp16, y = linear_75_cast_fp16)[name = string("input_345_cast_fp16")];
            tensor<int32, [1]> hidden_states_39_axes_0 = const()[name = string("hidden_states_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_19_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_19_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569055232)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_39_cast_fp16 = layer_norm(axes = hidden_states_39_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_19_attn_norm_weight_to_fp16, x = input_345_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_19_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_19_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569057344)))];
            tensor<fp16, [1, 128, 3072]> linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_19_attn_Wqkv_weight_to_fp16, x = hidden_states_39_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<int32, [5]> var_1985 = const()[name = string("op_1985"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_79_cast_fp16 = reshape(shape = var_1985, x = linear_76_cast_fp16)[name = string("qkv_79_cast_fp16")];
            tensor<int32, [3]> var_1987_split_sizes_0 = const()[name = string("op_1987_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1987_axis_0 = const()[name = string("op_1987_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_1987_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_1987_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_1987_cast_fp16_2 = split(axis = var_1987_axis_0, split_sizes = var_1987_split_sizes_0, x = qkv_79_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [1]> squeeze_57_axes_0 = const()[name = string("squeeze_57_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_57_cast_fp16 = squeeze(axes = squeeze_57_axes_0, x = var_1987_cast_fp16_0)[name = string("squeeze_57_cast_fp16")];
            tensor<int32, [1]> squeeze_58_axes_0 = const()[name = string("squeeze_58_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_58_cast_fp16 = squeeze(axes = squeeze_58_axes_0, x = var_1987_cast_fp16_1)[name = string("squeeze_58_cast_fp16")];
            tensor<int32, [1]> squeeze_59_axes_0 = const()[name = string("squeeze_59_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_59_cast_fp16 = squeeze(axes = squeeze_59_axes_0, x = var_1987_cast_fp16_2)[name = string("squeeze_59_cast_fp16")];
            tensor<int32, [4]> q_77_perm_0 = const()[name = string("q_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_77_perm_0 = const()[name = string("k_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_77_cast_fp16 = transpose(perm = q_77_perm_0, x = squeeze_57_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 128, 64]> var_1997_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_11_to_fp16)[name = string("op_1997_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_77_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_77_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_175_promoted_to_fp16 = const()[name = string("const_175_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2009_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_175_promoted_to_fp16)[name = string("op_2009_cast_fp16")];
            bool var_2011_interleave_0 = const()[name = string("op_2011_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2011_cast_fp16 = concat(axis = var_49, interleave = var_2011_interleave_0, values = (var_2009_cast_fp16, x1_77_cast_fp16))[name = string("op_2011_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2012_cast_fp16 = mul(x = var_2011_cast_fp16, y = sin_11_to_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_39_cast_fp16 = add(x = var_1997_cast_fp16, y = var_2012_cast_fp16)[name = string("q_embed_39_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_77_cast_fp16 = transpose(perm = k_77_perm_0, x = squeeze_58_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 128, 64]> var_2015_cast_fp16 = mul(x = k_77_cast_fp16, y = cos_11_to_fp16)[name = string("op_2015_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_77_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_77_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_178_promoted_to_fp16 = const()[name = string("const_178_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2027_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_178_promoted_to_fp16)[name = string("op_2027_cast_fp16")];
            bool var_2029_interleave_0 = const()[name = string("op_2029_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2029_cast_fp16 = concat(axis = var_49, interleave = var_2029_interleave_0, values = (var_2027_cast_fp16, x1_79_cast_fp16))[name = string("op_2029_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2030_cast_fp16 = mul(x = var_2029_cast_fp16, y = sin_11_to_fp16)[name = string("op_2030_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_39_cast_fp16 = add(x = var_2015_cast_fp16, y = var_2030_cast_fp16)[name = string("k_embed_39_cast_fp16")];
            bool var_2035_transpose_x_1 = const()[name = string("op_2035_transpose_x_1"), val = bool(false)];
            bool var_2035_transpose_y_1 = const()[name = string("op_2035_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_2035_cast_fp16 = matmul(transpose_x = var_2035_transpose_x_1, transpose_y = var_2035_transpose_y_1, x = q_embed_39_cast_fp16, y = k_embed_39_cast_fp16)[name = string("op_2035_cast_fp16")];
            fp16 var_2036_to_fp16 = const()[name = string("op_2036_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_77_cast_fp16 = mul(x = var_2035_cast_fp16, y = var_2036_to_fp16)[name = string("attn_weights_77_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_347_cast_fp16 = add(x = attn_weights_77_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_347_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2039_cast_fp16 = softmax(axis = var_49, x = input_347_cast_fp16)[name = string("op_2039_cast_fp16")];
            bool attn_output_115_transpose_x_0 = const()[name = string("attn_output_115_transpose_x_0"), val = bool(false)];
            bool attn_output_115_transpose_y_0 = const()[name = string("attn_output_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = squeeze_59_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_115_cast_fp16 = matmul(transpose_x = attn_output_115_transpose_x_0, transpose_y = attn_output_115_transpose_y_0, x = var_2039_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_115_cast_fp16")];
            tensor<int32, [4]> var_2043_perm_0 = const()[name = string("op_2043_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2045 = const()[name = string("op_2045"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_2043_cast_fp16 = transpose(perm = var_2043_perm_0, x = attn_output_115_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 128, 1024]> var_2046_cast_fp16 = reshape(shape = var_2045, x = var_2043_cast_fp16)[name = string("op_2046_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_19_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_19_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575348864)))];
            tensor<fp16, [1, 128, 1024]> linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_19_attn_Wo_weight_to_fp16, x = var_2046_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_353_cast_fp16 = add(x = input_345_cast_fp16, y = linear_77_cast_fp16)[name = string("input_353_cast_fp16")];
            tensor<int32, [1]> input_355_axes_0 = const()[name = string("input_355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_19_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_19_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577446080)))];
            tensor<fp16, [1, 128, 1024]> input_355_cast_fp16 = layer_norm(axes = input_355_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_19_mlp_norm_weight_to_fp16, x = input_353_cast_fp16)[name = string("input_355_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_19_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_19_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577448192)))];
            tensor<fp16, [1, 128, 5248]> linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_19_mlp_Wi_weight_to_fp16, x = input_355_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [2]> var_2058_split_sizes_0 = const()[name = string("op_2058_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2058_axis_0 = const()[name = string("op_2058_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_2058_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_2058_cast_fp16_1 = split(axis = var_2058_axis_0, split_sizes = var_2058_split_sizes_0, x = linear_78_cast_fp16)[name = string("op_2058_cast_fp16")];
            string var_2060_mode_0 = const()[name = string("op_2060_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_2060_cast_fp16 = gelu(mode = var_2060_mode_0, x = var_2058_cast_fp16_0)[name = string("op_2060_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_359_cast_fp16 = mul(x = var_2060_cast_fp16, y = var_2058_cast_fp16_1)[name = string("input_359_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_19_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_19_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588196160)))];
            tensor<fp16, [1, 128, 1024]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_19_mlp_Wo_weight_to_fp16, x = input_359_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_363_cast_fp16 = add(x = input_353_cast_fp16, y = linear_79_cast_fp16)[name = string("input_363_cast_fp16")];
            tensor<int32, [1]> hidden_states_41_axes_0 = const()[name = string("hidden_states_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_20_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_20_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593570176)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_41_cast_fp16 = layer_norm(axes = hidden_states_41_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_20_attn_norm_weight_to_fp16, x = input_363_cast_fp16)[name = string("hidden_states_41_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_20_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_20_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593572288)))];
            tensor<fp16, [1, 128, 3072]> linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_20_attn_Wqkv_weight_to_fp16, x = hidden_states_41_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [5]> var_2079 = const()[name = string("op_2079"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_83_cast_fp16 = reshape(shape = var_2079, x = linear_80_cast_fp16)[name = string("qkv_83_cast_fp16")];
            tensor<int32, [3]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_2081_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_2081_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_2081_cast_fp16_2 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = qkv_83_cast_fp16)[name = string("op_2081_cast_fp16")];
            tensor<int32, [1]> squeeze_60_axes_0 = const()[name = string("squeeze_60_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_60_cast_fp16 = squeeze(axes = squeeze_60_axes_0, x = var_2081_cast_fp16_0)[name = string("squeeze_60_cast_fp16")];
            tensor<int32, [1]> squeeze_61_axes_0 = const()[name = string("squeeze_61_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_61_cast_fp16 = squeeze(axes = squeeze_61_axes_0, x = var_2081_cast_fp16_1)[name = string("squeeze_61_cast_fp16")];
            tensor<int32, [1]> squeeze_62_axes_0 = const()[name = string("squeeze_62_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_62_cast_fp16 = squeeze(axes = squeeze_62_axes_0, x = var_2081_cast_fp16_2)[name = string("squeeze_62_cast_fp16")];
            tensor<int32, [4]> q_81_perm_0 = const()[name = string("q_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_81_perm_0 = const()[name = string("k_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_81_cast_fp16 = transpose(perm = q_81_perm_0, x = squeeze_60_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 16, 128, 64]> var_2091_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_11_to_fp16)[name = string("op_2091_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_81_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_81_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_183_promoted_to_fp16 = const()[name = string("const_183_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2103_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_183_promoted_to_fp16)[name = string("op_2103_cast_fp16")];
            bool var_2105_interleave_0 = const()[name = string("op_2105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2105_cast_fp16 = concat(axis = var_49, interleave = var_2105_interleave_0, values = (var_2103_cast_fp16, x1_81_cast_fp16))[name = string("op_2105_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2106_cast_fp16 = mul(x = var_2105_cast_fp16, y = sin_11_to_fp16)[name = string("op_2106_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_41_cast_fp16 = add(x = var_2091_cast_fp16, y = var_2106_cast_fp16)[name = string("q_embed_41_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_81_cast_fp16 = transpose(perm = k_81_perm_0, x = squeeze_61_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 128, 64]> var_2109_cast_fp16 = mul(x = k_81_cast_fp16, y = cos_11_to_fp16)[name = string("op_2109_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_81_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_81_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_186_promoted_to_fp16 = const()[name = string("const_186_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2121_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_186_promoted_to_fp16)[name = string("op_2121_cast_fp16")];
            bool var_2123_interleave_0 = const()[name = string("op_2123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2123_cast_fp16 = concat(axis = var_49, interleave = var_2123_interleave_0, values = (var_2121_cast_fp16, x1_83_cast_fp16))[name = string("op_2123_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2124_cast_fp16 = mul(x = var_2123_cast_fp16, y = sin_11_to_fp16)[name = string("op_2124_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_41_cast_fp16 = add(x = var_2109_cast_fp16, y = var_2124_cast_fp16)[name = string("k_embed_41_cast_fp16")];
            bool var_2129_transpose_x_1 = const()[name = string("op_2129_transpose_x_1"), val = bool(false)];
            bool var_2129_transpose_y_1 = const()[name = string("op_2129_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_2129_cast_fp16 = matmul(transpose_x = var_2129_transpose_x_1, transpose_y = var_2129_transpose_y_1, x = q_embed_41_cast_fp16, y = k_embed_41_cast_fp16)[name = string("op_2129_cast_fp16")];
            fp16 var_2130_to_fp16 = const()[name = string("op_2130_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_81_cast_fp16 = mul(x = var_2129_cast_fp16, y = var_2130_to_fp16)[name = string("attn_weights_81_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_365_cast_fp16 = add(x = attn_weights_81_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_365_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2133_cast_fp16 = softmax(axis = var_49, x = input_365_cast_fp16)[name = string("op_2133_cast_fp16")];
            bool attn_output_121_transpose_x_0 = const()[name = string("attn_output_121_transpose_x_0"), val = bool(false)];
            bool attn_output_121_transpose_y_0 = const()[name = string("attn_output_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = squeeze_62_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_121_cast_fp16 = matmul(transpose_x = attn_output_121_transpose_x_0, transpose_y = attn_output_121_transpose_y_0, x = var_2133_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_121_cast_fp16")];
            tensor<int32, [4]> var_2137_perm_0 = const()[name = string("op_2137_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2139 = const()[name = string("op_2139"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_2137_cast_fp16 = transpose(perm = var_2137_perm_0, x = attn_output_121_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 128, 1024]> var_2140_cast_fp16 = reshape(shape = var_2139, x = var_2137_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_20_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_20_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599863808)))];
            tensor<fp16, [1, 128, 1024]> linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_20_attn_Wo_weight_to_fp16, x = var_2140_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_371_cast_fp16 = add(x = input_363_cast_fp16, y = linear_81_cast_fp16)[name = string("input_371_cast_fp16")];
            tensor<int32, [1]> input_373_axes_0 = const()[name = string("input_373_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_20_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_20_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601961024)))];
            tensor<fp16, [1, 128, 1024]> input_373_cast_fp16 = layer_norm(axes = input_373_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_20_mlp_norm_weight_to_fp16, x = input_371_cast_fp16)[name = string("input_373_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_20_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_20_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601963136)))];
            tensor<fp16, [1, 128, 5248]> linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_20_mlp_Wi_weight_to_fp16, x = input_373_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [2]> var_2152_split_sizes_0 = const()[name = string("op_2152_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2152_axis_0 = const()[name = string("op_2152_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_2152_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_2152_cast_fp16_1 = split(axis = var_2152_axis_0, split_sizes = var_2152_split_sizes_0, x = linear_82_cast_fp16)[name = string("op_2152_cast_fp16")];
            string var_2154_mode_0 = const()[name = string("op_2154_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_2154_cast_fp16 = gelu(mode = var_2154_mode_0, x = var_2152_cast_fp16_0)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_377_cast_fp16 = mul(x = var_2154_cast_fp16, y = var_2152_cast_fp16_1)[name = string("input_377_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_20_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_20_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612711104)))];
            tensor<fp16, [1, 128, 1024]> linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_20_mlp_Wo_weight_to_fp16, x = input_377_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_381_cast_fp16 = add(x = input_371_cast_fp16, y = linear_83_cast_fp16)[name = string("input_381_cast_fp16")];
            tensor<int32, [1]> hidden_states_43_axes_0 = const()[name = string("hidden_states_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_21_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_21_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618085120)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_43_cast_fp16 = layer_norm(axes = hidden_states_43_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_21_attn_norm_weight_to_fp16, x = input_381_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_21_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_21_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618087232)))];
            tensor<fp16, [1, 128, 3072]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_21_attn_Wqkv_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [5]> var_2173 = const()[name = string("op_2173"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_87_cast_fp16 = reshape(shape = var_2173, x = linear_84_cast_fp16)[name = string("qkv_87_cast_fp16")];
            tensor<int32, [3]> var_2175_split_sizes_0 = const()[name = string("op_2175_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2175_axis_0 = const()[name = string("op_2175_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_2175_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_2175_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_2175_cast_fp16_2 = split(axis = var_2175_axis_0, split_sizes = var_2175_split_sizes_0, x = qkv_87_cast_fp16)[name = string("op_2175_cast_fp16")];
            tensor<int32, [1]> squeeze_63_axes_0 = const()[name = string("squeeze_63_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_63_cast_fp16 = squeeze(axes = squeeze_63_axes_0, x = var_2175_cast_fp16_0)[name = string("squeeze_63_cast_fp16")];
            tensor<int32, [1]> squeeze_64_axes_0 = const()[name = string("squeeze_64_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_64_cast_fp16 = squeeze(axes = squeeze_64_axes_0, x = var_2175_cast_fp16_1)[name = string("squeeze_64_cast_fp16")];
            tensor<int32, [1]> squeeze_65_axes_0 = const()[name = string("squeeze_65_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_65_cast_fp16 = squeeze(axes = squeeze_65_axes_0, x = var_2175_cast_fp16_2)[name = string("squeeze_65_cast_fp16")];
            tensor<int32, [4]> q_85_perm_0 = const()[name = string("q_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_85_perm_0 = const()[name = string("k_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_85_cast_fp16 = transpose(perm = q_85_perm_0, x = squeeze_63_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 16, 128, 64]> var_2185_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_7_to_fp16)[name = string("op_2185_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_85_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_85_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_191_promoted_to_fp16 = const()[name = string("const_191_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2197_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_191_promoted_to_fp16)[name = string("op_2197_cast_fp16")];
            bool var_2199_interleave_0 = const()[name = string("op_2199_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2199_cast_fp16 = concat(axis = var_49, interleave = var_2199_interleave_0, values = (var_2197_cast_fp16, x1_85_cast_fp16))[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2200_cast_fp16 = mul(x = var_2199_cast_fp16, y = sin_7_to_fp16)[name = string("op_2200_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_43_cast_fp16 = add(x = var_2185_cast_fp16, y = var_2200_cast_fp16)[name = string("q_embed_43_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_85_cast_fp16 = transpose(perm = k_85_perm_0, x = squeeze_64_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 16, 128, 64]> var_2203_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_7_to_fp16)[name = string("op_2203_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_85_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_85_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_194_promoted_to_fp16 = const()[name = string("const_194_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2215_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_194_promoted_to_fp16)[name = string("op_2215_cast_fp16")];
            bool var_2217_interleave_0 = const()[name = string("op_2217_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2217_cast_fp16 = concat(axis = var_49, interleave = var_2217_interleave_0, values = (var_2215_cast_fp16, x1_87_cast_fp16))[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2218_cast_fp16 = mul(x = var_2217_cast_fp16, y = sin_7_to_fp16)[name = string("op_2218_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_43_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2218_cast_fp16)[name = string("k_embed_43_cast_fp16")];
            bool var_2223_transpose_x_1 = const()[name = string("op_2223_transpose_x_1"), val = bool(false)];
            bool var_2223_transpose_y_1 = const()[name = string("op_2223_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_2223_cast_fp16 = matmul(transpose_x = var_2223_transpose_x_1, transpose_y = var_2223_transpose_y_1, x = q_embed_43_cast_fp16, y = k_embed_43_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_85_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("attn_weights_85_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_383_cast_fp16 = add(x = attn_weights_85_cast_fp16, y = key_bias_cast_fp16)[name = string("input_383_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2227_cast_fp16 = softmax(axis = var_49, x = input_383_cast_fp16)[name = string("op_2227_cast_fp16")];
            bool attn_output_127_transpose_x_0 = const()[name = string("attn_output_127_transpose_x_0"), val = bool(false)];
            bool attn_output_127_transpose_y_0 = const()[name = string("attn_output_127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = squeeze_65_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_127_cast_fp16 = matmul(transpose_x = attn_output_127_transpose_x_0, transpose_y = attn_output_127_transpose_y_0, x = var_2227_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_127_cast_fp16")];
            tensor<int32, [4]> var_2231_perm_0 = const()[name = string("op_2231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_2231_cast_fp16 = transpose(perm = var_2231_perm_0, x = attn_output_127_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 128, 1024]> var_2234_cast_fp16 = reshape(shape = var_2233, x = var_2231_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_21_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_21_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624378752)))];
            tensor<fp16, [1, 128, 1024]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_21_attn_Wo_weight_to_fp16, x = var_2234_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_389_cast_fp16 = add(x = input_381_cast_fp16, y = linear_85_cast_fp16)[name = string("input_389_cast_fp16")];
            tensor<int32, [1]> input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_21_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_21_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626475968)))];
            tensor<fp16, [1, 128, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_21_mlp_norm_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_21_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_21_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626478080)))];
            tensor<fp16, [1, 128, 5248]> linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_21_mlp_Wi_weight_to_fp16, x = input_391_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [2]> var_2246_split_sizes_0 = const()[name = string("op_2246_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2246_axis_0 = const()[name = string("op_2246_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_2246_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_2246_cast_fp16_1 = split(axis = var_2246_axis_0, split_sizes = var_2246_split_sizes_0, x = linear_86_cast_fp16)[name = string("op_2246_cast_fp16")];
            string var_2248_mode_0 = const()[name = string("op_2248_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_2248_cast_fp16 = gelu(mode = var_2248_mode_0, x = var_2246_cast_fp16_0)[name = string("op_2248_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_395_cast_fp16 = mul(x = var_2248_cast_fp16, y = var_2246_cast_fp16_1)[name = string("input_395_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_21_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_21_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637226048)))];
            tensor<fp16, [1, 128, 1024]> linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_21_mlp_Wo_weight_to_fp16, x = input_395_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_399_cast_fp16 = add(x = input_389_cast_fp16, y = linear_87_cast_fp16)[name = string("input_399_cast_fp16")];
            tensor<int32, [1]> hidden_states_45_axes_0 = const()[name = string("hidden_states_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_22_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_22_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642600064)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_45_cast_fp16 = layer_norm(axes = hidden_states_45_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_22_attn_norm_weight_to_fp16, x = input_399_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_22_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_22_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642602176)))];
            tensor<fp16, [1, 128, 3072]> linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_22_attn_Wqkv_weight_to_fp16, x = hidden_states_45_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<int32, [5]> var_2267 = const()[name = string("op_2267"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_91_cast_fp16 = reshape(shape = var_2267, x = linear_88_cast_fp16)[name = string("qkv_91_cast_fp16")];
            tensor<int32, [3]> var_2269_split_sizes_0 = const()[name = string("op_2269_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2269_axis_0 = const()[name = string("op_2269_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_2269_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_2269_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_2269_cast_fp16_2 = split(axis = var_2269_axis_0, split_sizes = var_2269_split_sizes_0, x = qkv_91_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> squeeze_66_axes_0 = const()[name = string("squeeze_66_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_66_cast_fp16 = squeeze(axes = squeeze_66_axes_0, x = var_2269_cast_fp16_0)[name = string("squeeze_66_cast_fp16")];
            tensor<int32, [1]> squeeze_67_axes_0 = const()[name = string("squeeze_67_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_67_cast_fp16 = squeeze(axes = squeeze_67_axes_0, x = var_2269_cast_fp16_1)[name = string("squeeze_67_cast_fp16")];
            tensor<int32, [1]> squeeze_68_axes_0 = const()[name = string("squeeze_68_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_68_cast_fp16 = squeeze(axes = squeeze_68_axes_0, x = var_2269_cast_fp16_2)[name = string("squeeze_68_cast_fp16")];
            tensor<int32, [4]> q_89_perm_0 = const()[name = string("q_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_89_perm_0 = const()[name = string("k_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_89_cast_fp16 = transpose(perm = q_89_perm_0, x = squeeze_66_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 16, 128, 64]> var_2279_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_11_to_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_89_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_89_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_199_promoted_to_fp16 = const()[name = string("const_199_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2291_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_199_promoted_to_fp16)[name = string("op_2291_cast_fp16")];
            bool var_2293_interleave_0 = const()[name = string("op_2293_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2293_cast_fp16 = concat(axis = var_49, interleave = var_2293_interleave_0, values = (var_2291_cast_fp16, x1_89_cast_fp16))[name = string("op_2293_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2294_cast_fp16 = mul(x = var_2293_cast_fp16, y = sin_11_to_fp16)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_45_cast_fp16 = add(x = var_2279_cast_fp16, y = var_2294_cast_fp16)[name = string("q_embed_45_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_89_cast_fp16 = transpose(perm = k_89_perm_0, x = squeeze_67_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 16, 128, 64]> var_2297_cast_fp16 = mul(x = k_89_cast_fp16, y = cos_11_to_fp16)[name = string("op_2297_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_89_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_89_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_202_promoted_to_fp16 = const()[name = string("const_202_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2309_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_202_promoted_to_fp16)[name = string("op_2309_cast_fp16")];
            bool var_2311_interleave_0 = const()[name = string("op_2311_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2311_cast_fp16 = concat(axis = var_49, interleave = var_2311_interleave_0, values = (var_2309_cast_fp16, x1_91_cast_fp16))[name = string("op_2311_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2312_cast_fp16 = mul(x = var_2311_cast_fp16, y = sin_11_to_fp16)[name = string("op_2312_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_45_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2312_cast_fp16)[name = string("k_embed_45_cast_fp16")];
            bool var_2317_transpose_x_1 = const()[name = string("op_2317_transpose_x_1"), val = bool(false)];
            bool var_2317_transpose_y_1 = const()[name = string("op_2317_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_2317_cast_fp16 = matmul(transpose_x = var_2317_transpose_x_1, transpose_y = var_2317_transpose_y_1, x = q_embed_45_cast_fp16, y = k_embed_45_cast_fp16)[name = string("op_2317_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_89_cast_fp16 = mul(x = var_2317_cast_fp16, y = var_2318_to_fp16)[name = string("attn_weights_89_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_401_cast_fp16 = add(x = attn_weights_89_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_401_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2321_cast_fp16 = softmax(axis = var_49, x = input_401_cast_fp16)[name = string("op_2321_cast_fp16")];
            bool attn_output_133_transpose_x_0 = const()[name = string("attn_output_133_transpose_x_0"), val = bool(false)];
            bool attn_output_133_transpose_y_0 = const()[name = string("attn_output_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = squeeze_68_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_133_cast_fp16 = matmul(transpose_x = attn_output_133_transpose_x_0, transpose_y = attn_output_133_transpose_y_0, x = var_2321_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_133_cast_fp16")];
            tensor<int32, [4]> var_2325_perm_0 = const()[name = string("op_2325_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2327 = const()[name = string("op_2327"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_2325_cast_fp16 = transpose(perm = var_2325_perm_0, x = attn_output_133_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 128, 1024]> var_2328_cast_fp16 = reshape(shape = var_2327, x = var_2325_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_22_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_22_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648893696)))];
            tensor<fp16, [1, 128, 1024]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_22_attn_Wo_weight_to_fp16, x = var_2328_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_407_cast_fp16 = add(x = input_399_cast_fp16, y = linear_89_cast_fp16)[name = string("input_407_cast_fp16")];
            tensor<int32, [1]> input_409_axes_0 = const()[name = string("input_409_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_22_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_22_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650990912)))];
            tensor<fp16, [1, 128, 1024]> input_409_cast_fp16 = layer_norm(axes = input_409_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_22_mlp_norm_weight_to_fp16, x = input_407_cast_fp16)[name = string("input_409_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_22_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_22_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650993024)))];
            tensor<fp16, [1, 128, 5248]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_22_mlp_Wi_weight_to_fp16, x = input_409_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<int32, [2]> var_2340_split_sizes_0 = const()[name = string("op_2340_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2340_axis_0 = const()[name = string("op_2340_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_2340_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_90_cast_fp16)[name = string("op_2340_cast_fp16")];
            string var_2342_mode_0 = const()[name = string("op_2342_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_2342_cast_fp16 = gelu(mode = var_2342_mode_0, x = var_2340_cast_fp16_0)[name = string("op_2342_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_413_cast_fp16 = mul(x = var_2342_cast_fp16, y = var_2340_cast_fp16_1)[name = string("input_413_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_22_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_22_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661740992)))];
            tensor<fp16, [1, 128, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_22_mlp_Wo_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_417_cast_fp16 = add(x = input_407_cast_fp16, y = linear_91_cast_fp16)[name = string("input_417_cast_fp16")];
            tensor<int32, [1]> hidden_states_47_axes_0 = const()[name = string("hidden_states_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_23_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_23_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667115008)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_47_cast_fp16 = layer_norm(axes = hidden_states_47_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_23_attn_norm_weight_to_fp16, x = input_417_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_23_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_23_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667117120)))];
            tensor<fp16, [1, 128, 3072]> linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_23_attn_Wqkv_weight_to_fp16, x = hidden_states_47_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [5]> var_2361 = const()[name = string("op_2361"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_95_cast_fp16 = reshape(shape = var_2361, x = linear_92_cast_fp16)[name = string("qkv_95_cast_fp16")];
            tensor<int32, [3]> var_2363_split_sizes_0 = const()[name = string("op_2363_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2363_axis_0 = const()[name = string("op_2363_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_2363_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_2363_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_2363_cast_fp16_2 = split(axis = var_2363_axis_0, split_sizes = var_2363_split_sizes_0, x = qkv_95_cast_fp16)[name = string("op_2363_cast_fp16")];
            tensor<int32, [1]> squeeze_69_axes_0 = const()[name = string("squeeze_69_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_69_cast_fp16 = squeeze(axes = squeeze_69_axes_0, x = var_2363_cast_fp16_0)[name = string("squeeze_69_cast_fp16")];
            tensor<int32, [1]> squeeze_70_axes_0 = const()[name = string("squeeze_70_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_70_cast_fp16 = squeeze(axes = squeeze_70_axes_0, x = var_2363_cast_fp16_1)[name = string("squeeze_70_cast_fp16")];
            tensor<int32, [1]> squeeze_71_axes_0 = const()[name = string("squeeze_71_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_71_cast_fp16 = squeeze(axes = squeeze_71_axes_0, x = var_2363_cast_fp16_2)[name = string("squeeze_71_cast_fp16")];
            tensor<int32, [4]> q_93_perm_0 = const()[name = string("q_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_93_perm_0 = const()[name = string("k_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_93_cast_fp16 = transpose(perm = q_93_perm_0, x = squeeze_69_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 128, 64]> var_2373_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_11_to_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_93_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_93_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_207_promoted_to_fp16 = const()[name = string("const_207_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2385_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_207_promoted_to_fp16)[name = string("op_2385_cast_fp16")];
            bool var_2387_interleave_0 = const()[name = string("op_2387_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2387_cast_fp16 = concat(axis = var_49, interleave = var_2387_interleave_0, values = (var_2385_cast_fp16, x1_93_cast_fp16))[name = string("op_2387_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2388_cast_fp16 = mul(x = var_2387_cast_fp16, y = sin_11_to_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_47_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2388_cast_fp16)[name = string("q_embed_47_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_93_cast_fp16 = transpose(perm = k_93_perm_0, x = squeeze_70_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 16, 128, 64]> var_2391_cast_fp16 = mul(x = k_93_cast_fp16, y = cos_11_to_fp16)[name = string("op_2391_cast_fp16")];
            tensor<int32, [4]> x1_95_begin_0 = const()[name = string("x1_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_95_end_0 = const()[name = string("x1_95_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_95_end_mask_0 = const()[name = string("x1_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_95_cast_fp16 = slice_by_index(begin = x1_95_begin_0, end = x1_95_end_0, end_mask = x1_95_end_mask_0, x = k_93_cast_fp16)[name = string("x1_95_cast_fp16")];
            tensor<int32, [4]> x2_95_begin_0 = const()[name = string("x2_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_95_end_0 = const()[name = string("x2_95_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_95_end_mask_0 = const()[name = string("x2_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_95_cast_fp16 = slice_by_index(begin = x2_95_begin_0, end = x2_95_end_0, end_mask = x2_95_end_mask_0, x = k_93_cast_fp16)[name = string("x2_95_cast_fp16")];
            fp16 const_210_promoted_to_fp16 = const()[name = string("const_210_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2403_cast_fp16 = mul(x = x2_95_cast_fp16, y = const_210_promoted_to_fp16)[name = string("op_2403_cast_fp16")];
            bool var_2405_interleave_0 = const()[name = string("op_2405_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2405_cast_fp16 = concat(axis = var_49, interleave = var_2405_interleave_0, values = (var_2403_cast_fp16, x1_95_cast_fp16))[name = string("op_2405_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2406_cast_fp16 = mul(x = var_2405_cast_fp16, y = sin_11_to_fp16)[name = string("op_2406_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_47_cast_fp16 = add(x = var_2391_cast_fp16, y = var_2406_cast_fp16)[name = string("k_embed_47_cast_fp16")];
            bool var_2411_transpose_x_1 = const()[name = string("op_2411_transpose_x_1"), val = bool(false)];
            bool var_2411_transpose_y_1 = const()[name = string("op_2411_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_2411_cast_fp16 = matmul(transpose_x = var_2411_transpose_x_1, transpose_y = var_2411_transpose_y_1, x = q_embed_47_cast_fp16, y = k_embed_47_cast_fp16)[name = string("op_2411_cast_fp16")];
            fp16 var_2412_to_fp16 = const()[name = string("op_2412_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_93_cast_fp16 = mul(x = var_2411_cast_fp16, y = var_2412_to_fp16)[name = string("attn_weights_93_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_419_cast_fp16 = add(x = attn_weights_93_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_419_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2415_cast_fp16 = softmax(axis = var_49, x = input_419_cast_fp16)[name = string("op_2415_cast_fp16")];
            bool attn_output_139_transpose_x_0 = const()[name = string("attn_output_139_transpose_x_0"), val = bool(false)];
            bool attn_output_139_transpose_y_0 = const()[name = string("attn_output_139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = squeeze_71_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_139_cast_fp16 = matmul(transpose_x = attn_output_139_transpose_x_0, transpose_y = attn_output_139_transpose_y_0, x = var_2415_cast_fp16, y = value_47_cast_fp16)[name = string("attn_output_139_cast_fp16")];
            tensor<int32, [4]> var_2419_perm_0 = const()[name = string("op_2419_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2421 = const()[name = string("op_2421"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_2419_cast_fp16 = transpose(perm = var_2419_perm_0, x = attn_output_139_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 128, 1024]> var_2422_cast_fp16 = reshape(shape = var_2421, x = var_2419_cast_fp16)[name = string("op_2422_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_23_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_23_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673408640)))];
            tensor<fp16, [1, 128, 1024]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_23_attn_Wo_weight_to_fp16, x = var_2422_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_425_cast_fp16 = add(x = input_417_cast_fp16, y = linear_93_cast_fp16)[name = string("input_425_cast_fp16")];
            tensor<int32, [1]> input_427_axes_0 = const()[name = string("input_427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_23_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_23_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675505856)))];
            tensor<fp16, [1, 128, 1024]> input_427_cast_fp16 = layer_norm(axes = input_427_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_23_mlp_norm_weight_to_fp16, x = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_23_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_23_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675507968)))];
            tensor<fp16, [1, 128, 5248]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_23_mlp_Wi_weight_to_fp16, x = input_427_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<int32, [2]> var_2434_split_sizes_0 = const()[name = string("op_2434_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2434_axis_0 = const()[name = string("op_2434_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_2434_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_2434_cast_fp16_1 = split(axis = var_2434_axis_0, split_sizes = var_2434_split_sizes_0, x = linear_94_cast_fp16)[name = string("op_2434_cast_fp16")];
            string var_2436_mode_0 = const()[name = string("op_2436_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_2436_cast_fp16 = gelu(mode = var_2436_mode_0, x = var_2434_cast_fp16_0)[name = string("op_2436_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_431_cast_fp16 = mul(x = var_2436_cast_fp16, y = var_2434_cast_fp16_1)[name = string("input_431_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_23_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_23_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686255936)))];
            tensor<fp16, [1, 128, 1024]> linear_95_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_23_mlp_Wo_weight_to_fp16, x = input_431_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_435_cast_fp16 = add(x = input_425_cast_fp16, y = linear_95_cast_fp16)[name = string("input_435_cast_fp16")];
            tensor<int32, [1]> hidden_states_49_axes_0 = const()[name = string("hidden_states_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_24_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_24_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691629952)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_49_cast_fp16 = layer_norm(axes = hidden_states_49_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_24_attn_norm_weight_to_fp16, x = input_435_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_24_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_24_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691632064)))];
            tensor<fp16, [1, 128, 3072]> linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_24_attn_Wqkv_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<int32, [5]> var_2455 = const()[name = string("op_2455"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_99_cast_fp16 = reshape(shape = var_2455, x = linear_96_cast_fp16)[name = string("qkv_99_cast_fp16")];
            tensor<int32, [3]> var_2457_split_sizes_0 = const()[name = string("op_2457_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2457_axis_0 = const()[name = string("op_2457_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_2457_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_2457_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_2457_cast_fp16_2 = split(axis = var_2457_axis_0, split_sizes = var_2457_split_sizes_0, x = qkv_99_cast_fp16)[name = string("op_2457_cast_fp16")];
            tensor<int32, [1]> squeeze_72_axes_0 = const()[name = string("squeeze_72_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_72_cast_fp16 = squeeze(axes = squeeze_72_axes_0, x = var_2457_cast_fp16_0)[name = string("squeeze_72_cast_fp16")];
            tensor<int32, [1]> squeeze_73_axes_0 = const()[name = string("squeeze_73_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_73_cast_fp16 = squeeze(axes = squeeze_73_axes_0, x = var_2457_cast_fp16_1)[name = string("squeeze_73_cast_fp16")];
            tensor<int32, [1]> squeeze_74_axes_0 = const()[name = string("squeeze_74_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_74_cast_fp16 = squeeze(axes = squeeze_74_axes_0, x = var_2457_cast_fp16_2)[name = string("squeeze_74_cast_fp16")];
            tensor<int32, [4]> q_97_perm_0 = const()[name = string("q_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_97_perm_0 = const()[name = string("k_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_97_cast_fp16 = transpose(perm = q_97_perm_0, x = squeeze_72_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, 128, 64]> var_2467_cast_fp16 = mul(x = q_97_cast_fp16, y = cos_7_to_fp16)[name = string("op_2467_cast_fp16")];
            tensor<int32, [4]> x1_97_begin_0 = const()[name = string("x1_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_97_end_0 = const()[name = string("x1_97_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_97_end_mask_0 = const()[name = string("x1_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_97_cast_fp16 = slice_by_index(begin = x1_97_begin_0, end = x1_97_end_0, end_mask = x1_97_end_mask_0, x = q_97_cast_fp16)[name = string("x1_97_cast_fp16")];
            tensor<int32, [4]> x2_97_begin_0 = const()[name = string("x2_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_97_end_0 = const()[name = string("x2_97_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_97_end_mask_0 = const()[name = string("x2_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_97_cast_fp16 = slice_by_index(begin = x2_97_begin_0, end = x2_97_end_0, end_mask = x2_97_end_mask_0, x = q_97_cast_fp16)[name = string("x2_97_cast_fp16")];
            fp16 const_215_promoted_to_fp16 = const()[name = string("const_215_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2479_cast_fp16 = mul(x = x2_97_cast_fp16, y = const_215_promoted_to_fp16)[name = string("op_2479_cast_fp16")];
            bool var_2481_interleave_0 = const()[name = string("op_2481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2481_cast_fp16 = concat(axis = var_49, interleave = var_2481_interleave_0, values = (var_2479_cast_fp16, x1_97_cast_fp16))[name = string("op_2481_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2482_cast_fp16 = mul(x = var_2481_cast_fp16, y = sin_7_to_fp16)[name = string("op_2482_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_49_cast_fp16 = add(x = var_2467_cast_fp16, y = var_2482_cast_fp16)[name = string("q_embed_49_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_97_cast_fp16 = transpose(perm = k_97_perm_0, x = squeeze_73_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 16, 128, 64]> var_2485_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_7_to_fp16)[name = string("op_2485_cast_fp16")];
            tensor<int32, [4]> x1_99_begin_0 = const()[name = string("x1_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_99_end_0 = const()[name = string("x1_99_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_99_end_mask_0 = const()[name = string("x1_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_99_cast_fp16 = slice_by_index(begin = x1_99_begin_0, end = x1_99_end_0, end_mask = x1_99_end_mask_0, x = k_97_cast_fp16)[name = string("x1_99_cast_fp16")];
            tensor<int32, [4]> x2_99_begin_0 = const()[name = string("x2_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_99_end_0 = const()[name = string("x2_99_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_99_end_mask_0 = const()[name = string("x2_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_99_cast_fp16 = slice_by_index(begin = x2_99_begin_0, end = x2_99_end_0, end_mask = x2_99_end_mask_0, x = k_97_cast_fp16)[name = string("x2_99_cast_fp16")];
            fp16 const_218_promoted_to_fp16 = const()[name = string("const_218_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2497_cast_fp16 = mul(x = x2_99_cast_fp16, y = const_218_promoted_to_fp16)[name = string("op_2497_cast_fp16")];
            bool var_2499_interleave_0 = const()[name = string("op_2499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2499_cast_fp16 = concat(axis = var_49, interleave = var_2499_interleave_0, values = (var_2497_cast_fp16, x1_99_cast_fp16))[name = string("op_2499_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2500_cast_fp16 = mul(x = var_2499_cast_fp16, y = sin_7_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_49_cast_fp16 = add(x = var_2485_cast_fp16, y = var_2500_cast_fp16)[name = string("k_embed_49_cast_fp16")];
            bool var_2505_transpose_x_1 = const()[name = string("op_2505_transpose_x_1"), val = bool(false)];
            bool var_2505_transpose_y_1 = const()[name = string("op_2505_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_2505_cast_fp16 = matmul(transpose_x = var_2505_transpose_x_1, transpose_y = var_2505_transpose_y_1, x = q_embed_49_cast_fp16, y = k_embed_49_cast_fp16)[name = string("op_2505_cast_fp16")];
            fp16 var_2506_to_fp16 = const()[name = string("op_2506_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_97_cast_fp16 = mul(x = var_2505_cast_fp16, y = var_2506_to_fp16)[name = string("attn_weights_97_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_437_cast_fp16 = add(x = attn_weights_97_cast_fp16, y = key_bias_cast_fp16)[name = string("input_437_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2509_cast_fp16 = softmax(axis = var_49, x = input_437_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool attn_output_145_transpose_x_0 = const()[name = string("attn_output_145_transpose_x_0"), val = bool(false)];
            bool attn_output_145_transpose_y_0 = const()[name = string("attn_output_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = squeeze_74_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_145_cast_fp16 = matmul(transpose_x = attn_output_145_transpose_x_0, transpose_y = attn_output_145_transpose_y_0, x = var_2509_cast_fp16, y = value_49_cast_fp16)[name = string("attn_output_145_cast_fp16")];
            tensor<int32, [4]> var_2513_perm_0 = const()[name = string("op_2513_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2515 = const()[name = string("op_2515"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_2513_cast_fp16 = transpose(perm = var_2513_perm_0, x = attn_output_145_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 128, 1024]> var_2516_cast_fp16 = reshape(shape = var_2515, x = var_2513_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_24_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_24_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(697923584)))];
            tensor<fp16, [1, 128, 1024]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_24_attn_Wo_weight_to_fp16, x = var_2516_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_443_cast_fp16 = add(x = input_435_cast_fp16, y = linear_97_cast_fp16)[name = string("input_443_cast_fp16")];
            tensor<int32, [1]> input_445_axes_0 = const()[name = string("input_445_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_24_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_24_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700020800)))];
            tensor<fp16, [1, 128, 1024]> input_445_cast_fp16 = layer_norm(axes = input_445_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_24_mlp_norm_weight_to_fp16, x = input_443_cast_fp16)[name = string("input_445_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_24_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_24_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700022912)))];
            tensor<fp16, [1, 128, 5248]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_24_mlp_Wi_weight_to_fp16, x = input_445_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [2]> var_2528_split_sizes_0 = const()[name = string("op_2528_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2528_axis_0 = const()[name = string("op_2528_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_2528_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_2528_cast_fp16_1 = split(axis = var_2528_axis_0, split_sizes = var_2528_split_sizes_0, x = linear_98_cast_fp16)[name = string("op_2528_cast_fp16")];
            string var_2530_mode_0 = const()[name = string("op_2530_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_2530_cast_fp16 = gelu(mode = var_2530_mode_0, x = var_2528_cast_fp16_0)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_449_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2528_cast_fp16_1)[name = string("input_449_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_24_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_24_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710770880)))];
            tensor<fp16, [1, 128, 1024]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_24_mlp_Wo_weight_to_fp16, x = input_449_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_453_cast_fp16 = add(x = input_443_cast_fp16, y = linear_99_cast_fp16)[name = string("input_453_cast_fp16")];
            tensor<int32, [1]> hidden_states_51_axes_0 = const()[name = string("hidden_states_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_25_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_25_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(716144896)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_51_cast_fp16 = layer_norm(axes = hidden_states_51_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_25_attn_norm_weight_to_fp16, x = input_453_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_25_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_25_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(716147008)))];
            tensor<fp16, [1, 128, 3072]> linear_100_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_25_attn_Wqkv_weight_to_fp16, x = hidden_states_51_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [5]> var_2549 = const()[name = string("op_2549"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_103_cast_fp16 = reshape(shape = var_2549, x = linear_100_cast_fp16)[name = string("qkv_103_cast_fp16")];
            tensor<int32, [3]> var_2551_split_sizes_0 = const()[name = string("op_2551_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2551_axis_0 = const()[name = string("op_2551_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_2551_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_2551_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_2551_cast_fp16_2 = split(axis = var_2551_axis_0, split_sizes = var_2551_split_sizes_0, x = qkv_103_cast_fp16)[name = string("op_2551_cast_fp16")];
            tensor<int32, [1]> squeeze_75_axes_0 = const()[name = string("squeeze_75_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_75_cast_fp16 = squeeze(axes = squeeze_75_axes_0, x = var_2551_cast_fp16_0)[name = string("squeeze_75_cast_fp16")];
            tensor<int32, [1]> squeeze_76_axes_0 = const()[name = string("squeeze_76_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_76_cast_fp16 = squeeze(axes = squeeze_76_axes_0, x = var_2551_cast_fp16_1)[name = string("squeeze_76_cast_fp16")];
            tensor<int32, [1]> squeeze_77_axes_0 = const()[name = string("squeeze_77_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_77_cast_fp16 = squeeze(axes = squeeze_77_axes_0, x = var_2551_cast_fp16_2)[name = string("squeeze_77_cast_fp16")];
            tensor<int32, [4]> q_101_perm_0 = const()[name = string("q_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_101_perm_0 = const()[name = string("k_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_101_cast_fp16 = transpose(perm = q_101_perm_0, x = squeeze_75_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 16, 128, 64]> var_2561_cast_fp16 = mul(x = q_101_cast_fp16, y = cos_11_to_fp16)[name = string("op_2561_cast_fp16")];
            tensor<int32, [4]> x1_101_begin_0 = const()[name = string("x1_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_101_end_0 = const()[name = string("x1_101_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_101_end_mask_0 = const()[name = string("x1_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_101_cast_fp16 = slice_by_index(begin = x1_101_begin_0, end = x1_101_end_0, end_mask = x1_101_end_mask_0, x = q_101_cast_fp16)[name = string("x1_101_cast_fp16")];
            tensor<int32, [4]> x2_101_begin_0 = const()[name = string("x2_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_101_end_0 = const()[name = string("x2_101_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_101_end_mask_0 = const()[name = string("x2_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_101_cast_fp16 = slice_by_index(begin = x2_101_begin_0, end = x2_101_end_0, end_mask = x2_101_end_mask_0, x = q_101_cast_fp16)[name = string("x2_101_cast_fp16")];
            fp16 const_223_promoted_to_fp16 = const()[name = string("const_223_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2573_cast_fp16 = mul(x = x2_101_cast_fp16, y = const_223_promoted_to_fp16)[name = string("op_2573_cast_fp16")];
            bool var_2575_interleave_0 = const()[name = string("op_2575_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2575_cast_fp16 = concat(axis = var_49, interleave = var_2575_interleave_0, values = (var_2573_cast_fp16, x1_101_cast_fp16))[name = string("op_2575_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2576_cast_fp16 = mul(x = var_2575_cast_fp16, y = sin_11_to_fp16)[name = string("op_2576_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_51_cast_fp16 = add(x = var_2561_cast_fp16, y = var_2576_cast_fp16)[name = string("q_embed_51_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_101_cast_fp16 = transpose(perm = k_101_perm_0, x = squeeze_76_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 16, 128, 64]> var_2579_cast_fp16 = mul(x = k_101_cast_fp16, y = cos_11_to_fp16)[name = string("op_2579_cast_fp16")];
            tensor<int32, [4]> x1_103_begin_0 = const()[name = string("x1_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_103_end_0 = const()[name = string("x1_103_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_103_end_mask_0 = const()[name = string("x1_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_103_cast_fp16 = slice_by_index(begin = x1_103_begin_0, end = x1_103_end_0, end_mask = x1_103_end_mask_0, x = k_101_cast_fp16)[name = string("x1_103_cast_fp16")];
            tensor<int32, [4]> x2_103_begin_0 = const()[name = string("x2_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_103_end_0 = const()[name = string("x2_103_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_103_end_mask_0 = const()[name = string("x2_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_103_cast_fp16 = slice_by_index(begin = x2_103_begin_0, end = x2_103_end_0, end_mask = x2_103_end_mask_0, x = k_101_cast_fp16)[name = string("x2_103_cast_fp16")];
            fp16 const_226_promoted_to_fp16 = const()[name = string("const_226_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2591_cast_fp16 = mul(x = x2_103_cast_fp16, y = const_226_promoted_to_fp16)[name = string("op_2591_cast_fp16")];
            bool var_2593_interleave_0 = const()[name = string("op_2593_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2593_cast_fp16 = concat(axis = var_49, interleave = var_2593_interleave_0, values = (var_2591_cast_fp16, x1_103_cast_fp16))[name = string("op_2593_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2594_cast_fp16 = mul(x = var_2593_cast_fp16, y = sin_11_to_fp16)[name = string("op_2594_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_51_cast_fp16 = add(x = var_2579_cast_fp16, y = var_2594_cast_fp16)[name = string("k_embed_51_cast_fp16")];
            bool var_2599_transpose_x_1 = const()[name = string("op_2599_transpose_x_1"), val = bool(false)];
            bool var_2599_transpose_y_1 = const()[name = string("op_2599_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_2599_cast_fp16 = matmul(transpose_x = var_2599_transpose_x_1, transpose_y = var_2599_transpose_y_1, x = q_embed_51_cast_fp16, y = k_embed_51_cast_fp16)[name = string("op_2599_cast_fp16")];
            fp16 var_2600_to_fp16 = const()[name = string("op_2600_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_101_cast_fp16 = mul(x = var_2599_cast_fp16, y = var_2600_to_fp16)[name = string("attn_weights_101_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_455_cast_fp16 = add(x = attn_weights_101_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_455_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2603_cast_fp16 = softmax(axis = var_49, x = input_455_cast_fp16)[name = string("op_2603_cast_fp16")];
            bool attn_output_151_transpose_x_0 = const()[name = string("attn_output_151_transpose_x_0"), val = bool(false)];
            bool attn_output_151_transpose_y_0 = const()[name = string("attn_output_151_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = squeeze_77_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_151_cast_fp16 = matmul(transpose_x = attn_output_151_transpose_x_0, transpose_y = attn_output_151_transpose_y_0, x = var_2603_cast_fp16, y = value_51_cast_fp16)[name = string("attn_output_151_cast_fp16")];
            tensor<int32, [4]> var_2607_perm_0 = const()[name = string("op_2607_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2609 = const()[name = string("op_2609"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_2607_cast_fp16 = transpose(perm = var_2607_perm_0, x = attn_output_151_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 128, 1024]> var_2610_cast_fp16 = reshape(shape = var_2609, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_25_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_25_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(722438528)))];
            tensor<fp16, [1, 128, 1024]> linear_101_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_25_attn_Wo_weight_to_fp16, x = var_2610_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_461_cast_fp16 = add(x = input_453_cast_fp16, y = linear_101_cast_fp16)[name = string("input_461_cast_fp16")];
            tensor<int32, [1]> input_463_axes_0 = const()[name = string("input_463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_25_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_25_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724535744)))];
            tensor<fp16, [1, 128, 1024]> input_463_cast_fp16 = layer_norm(axes = input_463_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_25_mlp_norm_weight_to_fp16, x = input_461_cast_fp16)[name = string("input_463_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_25_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_25_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724537856)))];
            tensor<fp16, [1, 128, 5248]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_25_mlp_Wi_weight_to_fp16, x = input_463_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [2]> var_2622_split_sizes_0 = const()[name = string("op_2622_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2622_axis_0 = const()[name = string("op_2622_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_2622_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_2622_cast_fp16_1 = split(axis = var_2622_axis_0, split_sizes = var_2622_split_sizes_0, x = linear_102_cast_fp16)[name = string("op_2622_cast_fp16")];
            string var_2624_mode_0 = const()[name = string("op_2624_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_2624_cast_fp16 = gelu(mode = var_2624_mode_0, x = var_2622_cast_fp16_0)[name = string("op_2624_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_467_cast_fp16 = mul(x = var_2624_cast_fp16, y = var_2622_cast_fp16_1)[name = string("input_467_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_25_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_25_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(735285824)))];
            tensor<fp16, [1, 128, 1024]> linear_103_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_25_mlp_Wo_weight_to_fp16, x = input_467_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_471_cast_fp16 = add(x = input_461_cast_fp16, y = linear_103_cast_fp16)[name = string("input_471_cast_fp16")];
            tensor<int32, [1]> hidden_states_53_axes_0 = const()[name = string("hidden_states_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_26_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_26_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740659840)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_53_cast_fp16 = layer_norm(axes = hidden_states_53_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_26_attn_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_26_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_26_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740661952)))];
            tensor<fp16, [1, 128, 3072]> linear_104_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_26_attn_Wqkv_weight_to_fp16, x = hidden_states_53_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<int32, [5]> var_2643 = const()[name = string("op_2643"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_107_cast_fp16 = reshape(shape = var_2643, x = linear_104_cast_fp16)[name = string("qkv_107_cast_fp16")];
            tensor<int32, [3]> var_2645_split_sizes_0 = const()[name = string("op_2645_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2645_axis_0 = const()[name = string("op_2645_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_2645_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_2645_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_2645_cast_fp16_2 = split(axis = var_2645_axis_0, split_sizes = var_2645_split_sizes_0, x = qkv_107_cast_fp16)[name = string("op_2645_cast_fp16")];
            tensor<int32, [1]> squeeze_78_axes_0 = const()[name = string("squeeze_78_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_78_cast_fp16 = squeeze(axes = squeeze_78_axes_0, x = var_2645_cast_fp16_0)[name = string("squeeze_78_cast_fp16")];
            tensor<int32, [1]> squeeze_79_axes_0 = const()[name = string("squeeze_79_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_79_cast_fp16 = squeeze(axes = squeeze_79_axes_0, x = var_2645_cast_fp16_1)[name = string("squeeze_79_cast_fp16")];
            tensor<int32, [1]> squeeze_80_axes_0 = const()[name = string("squeeze_80_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_80_cast_fp16 = squeeze(axes = squeeze_80_axes_0, x = var_2645_cast_fp16_2)[name = string("squeeze_80_cast_fp16")];
            tensor<int32, [4]> q_105_perm_0 = const()[name = string("q_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_105_perm_0 = const()[name = string("k_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_105_cast_fp16 = transpose(perm = q_105_perm_0, x = squeeze_78_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 16, 128, 64]> var_2655_cast_fp16 = mul(x = q_105_cast_fp16, y = cos_11_to_fp16)[name = string("op_2655_cast_fp16")];
            tensor<int32, [4]> x1_105_begin_0 = const()[name = string("x1_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_105_end_0 = const()[name = string("x1_105_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_105_end_mask_0 = const()[name = string("x1_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_105_cast_fp16 = slice_by_index(begin = x1_105_begin_0, end = x1_105_end_0, end_mask = x1_105_end_mask_0, x = q_105_cast_fp16)[name = string("x1_105_cast_fp16")];
            tensor<int32, [4]> x2_105_begin_0 = const()[name = string("x2_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_105_end_0 = const()[name = string("x2_105_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_105_end_mask_0 = const()[name = string("x2_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_105_cast_fp16 = slice_by_index(begin = x2_105_begin_0, end = x2_105_end_0, end_mask = x2_105_end_mask_0, x = q_105_cast_fp16)[name = string("x2_105_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2667_cast_fp16 = mul(x = x2_105_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_2667_cast_fp16")];
            bool var_2669_interleave_0 = const()[name = string("op_2669_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2669_cast_fp16 = concat(axis = var_49, interleave = var_2669_interleave_0, values = (var_2667_cast_fp16, x1_105_cast_fp16))[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2670_cast_fp16 = mul(x = var_2669_cast_fp16, y = sin_11_to_fp16)[name = string("op_2670_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_53_cast_fp16 = add(x = var_2655_cast_fp16, y = var_2670_cast_fp16)[name = string("q_embed_53_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_105_cast_fp16 = transpose(perm = k_105_perm_0, x = squeeze_79_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 16, 128, 64]> var_2673_cast_fp16 = mul(x = k_105_cast_fp16, y = cos_11_to_fp16)[name = string("op_2673_cast_fp16")];
            tensor<int32, [4]> x1_107_begin_0 = const()[name = string("x1_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_107_end_0 = const()[name = string("x1_107_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_107_end_mask_0 = const()[name = string("x1_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_107_cast_fp16 = slice_by_index(begin = x1_107_begin_0, end = x1_107_end_0, end_mask = x1_107_end_mask_0, x = k_105_cast_fp16)[name = string("x1_107_cast_fp16")];
            tensor<int32, [4]> x2_107_begin_0 = const()[name = string("x2_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_107_end_0 = const()[name = string("x2_107_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_107_end_mask_0 = const()[name = string("x2_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_107_cast_fp16 = slice_by_index(begin = x2_107_begin_0, end = x2_107_end_0, end_mask = x2_107_end_mask_0, x = k_105_cast_fp16)[name = string("x2_107_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2685_cast_fp16 = mul(x = x2_107_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2685_cast_fp16")];
            bool var_2687_interleave_0 = const()[name = string("op_2687_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2687_cast_fp16 = concat(axis = var_49, interleave = var_2687_interleave_0, values = (var_2685_cast_fp16, x1_107_cast_fp16))[name = string("op_2687_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2688_cast_fp16 = mul(x = var_2687_cast_fp16, y = sin_11_to_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_53_cast_fp16 = add(x = var_2673_cast_fp16, y = var_2688_cast_fp16)[name = string("k_embed_53_cast_fp16")];
            bool var_2693_transpose_x_1 = const()[name = string("op_2693_transpose_x_1"), val = bool(false)];
            bool var_2693_transpose_y_1 = const()[name = string("op_2693_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_2693_cast_fp16 = matmul(transpose_x = var_2693_transpose_x_1, transpose_y = var_2693_transpose_y_1, x = q_embed_53_cast_fp16, y = k_embed_53_cast_fp16)[name = string("op_2693_cast_fp16")];
            fp16 var_2694_to_fp16 = const()[name = string("op_2694_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_105_cast_fp16 = mul(x = var_2693_cast_fp16, y = var_2694_to_fp16)[name = string("attn_weights_105_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_473_cast_fp16 = add(x = attn_weights_105_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_473_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2697_cast_fp16 = softmax(axis = var_49, x = input_473_cast_fp16)[name = string("op_2697_cast_fp16")];
            bool attn_output_157_transpose_x_0 = const()[name = string("attn_output_157_transpose_x_0"), val = bool(false)];
            bool attn_output_157_transpose_y_0 = const()[name = string("attn_output_157_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = squeeze_80_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_157_cast_fp16 = matmul(transpose_x = attn_output_157_transpose_x_0, transpose_y = attn_output_157_transpose_y_0, x = var_2697_cast_fp16, y = value_53_cast_fp16)[name = string("attn_output_157_cast_fp16")];
            tensor<int32, [4]> var_2701_perm_0 = const()[name = string("op_2701_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2703 = const()[name = string("op_2703"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_2701_cast_fp16 = transpose(perm = var_2701_perm_0, x = attn_output_157_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 128, 1024]> var_2704_cast_fp16 = reshape(shape = var_2703, x = var_2701_cast_fp16)[name = string("op_2704_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_26_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_26_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746953472)))];
            tensor<fp16, [1, 128, 1024]> linear_105_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_26_attn_Wo_weight_to_fp16, x = var_2704_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_479_cast_fp16 = add(x = input_471_cast_fp16, y = linear_105_cast_fp16)[name = string("input_479_cast_fp16")];
            tensor<int32, [1]> input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_26_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_26_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749050688)))];
            tensor<fp16, [1, 128, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_26_mlp_norm_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_26_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_26_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749052800)))];
            tensor<fp16, [1, 128, 5248]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_26_mlp_Wi_weight_to_fp16, x = input_481_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [2]> var_2716_split_sizes_0 = const()[name = string("op_2716_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2716_axis_0 = const()[name = string("op_2716_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_2716_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_2716_cast_fp16_1 = split(axis = var_2716_axis_0, split_sizes = var_2716_split_sizes_0, x = linear_106_cast_fp16)[name = string("op_2716_cast_fp16")];
            string var_2718_mode_0 = const()[name = string("op_2718_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_2718_cast_fp16 = gelu(mode = var_2718_mode_0, x = var_2716_cast_fp16_0)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_485_cast_fp16 = mul(x = var_2718_cast_fp16, y = var_2716_cast_fp16_1)[name = string("input_485_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_26_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_26_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(759800768)))];
            tensor<fp16, [1, 128, 1024]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_26_mlp_Wo_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_489_cast_fp16 = add(x = input_479_cast_fp16, y = linear_107_cast_fp16)[name = string("input_489_cast_fp16")];
            tensor<int32, [1]> hidden_states_axes_0 = const()[name = string("hidden_states_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_27_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_27_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765174784)))];
            tensor<fp16, [1, 128, 1024]> hidden_states_cast_fp16 = layer_norm(axes = hidden_states_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_27_attn_norm_weight_to_fp16, x = input_489_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_27_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_27_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765176896)))];
            tensor<fp16, [1, 128, 3072]> linear_108_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_27_attn_Wqkv_weight_to_fp16, x = hidden_states_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<int32, [5]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [5]>([1, 128, 3, -1, 64])];
            tensor<fp16, [1, 128, 3, 16, 64]> qkv_cast_fp16 = reshape(shape = var_2737, x = linear_108_cast_fp16)[name = string("qkv_cast_fp16")];
            tensor<int32, [3]> var_2739_split_sizes_0 = const()[name = string("op_2739_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2739_axis_0 = const()[name = string("op_2739_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 128, 1, 16, 64]> var_2739_cast_fp16_0, tensor<fp16, [1, 128, 1, 16, 64]> var_2739_cast_fp16_1, tensor<fp16, [1, 128, 1, 16, 64]> var_2739_cast_fp16_2 = split(axis = var_2739_axis_0, split_sizes = var_2739_split_sizes_0, x = qkv_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<int32, [1]> squeeze_81_axes_0 = const()[name = string("squeeze_81_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_81_cast_fp16 = squeeze(axes = squeeze_81_axes_0, x = var_2739_cast_fp16_0)[name = string("squeeze_81_cast_fp16")];
            tensor<int32, [1]> squeeze_82_axes_0 = const()[name = string("squeeze_82_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_82_cast_fp16 = squeeze(axes = squeeze_82_axes_0, x = var_2739_cast_fp16_1)[name = string("squeeze_82_cast_fp16")];
            tensor<int32, [1]> squeeze_83_axes_0 = const()[name = string("squeeze_83_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 128, 16, 64]> squeeze_83_cast_fp16 = squeeze(axes = squeeze_83_axes_0, x = var_2739_cast_fp16_2)[name = string("squeeze_83_cast_fp16")];
            tensor<int32, [4]> q_109_perm_0 = const()[name = string("q_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_109_perm_0 = const()[name = string("k_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 128, 64]> q_109_cast_fp16 = transpose(perm = q_109_perm_0, x = squeeze_81_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 16, 128, 64]> var_2749_cast_fp16 = mul(x = q_109_cast_fp16, y = cos_7_to_fp16)[name = string("op_2749_cast_fp16")];
            tensor<int32, [4]> x1_109_begin_0 = const()[name = string("x1_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_109_end_0 = const()[name = string("x1_109_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_109_end_mask_0 = const()[name = string("x1_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_109_cast_fp16 = slice_by_index(begin = x1_109_begin_0, end = x1_109_end_0, end_mask = x1_109_end_mask_0, x = q_109_cast_fp16)[name = string("x1_109_cast_fp16")];
            tensor<int32, [4]> x2_109_begin_0 = const()[name = string("x2_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_109_end_0 = const()[name = string("x2_109_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_109_end_mask_0 = const()[name = string("x2_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_109_cast_fp16 = slice_by_index(begin = x2_109_begin_0, end = x2_109_end_0, end_mask = x2_109_end_mask_0, x = q_109_cast_fp16)[name = string("x2_109_cast_fp16")];
            fp16 const_239_promoted_to_fp16 = const()[name = string("const_239_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2761_cast_fp16 = mul(x = x2_109_cast_fp16, y = const_239_promoted_to_fp16)[name = string("op_2761_cast_fp16")];
            bool var_2763_interleave_0 = const()[name = string("op_2763_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2763_cast_fp16 = concat(axis = var_49, interleave = var_2763_interleave_0, values = (var_2761_cast_fp16, x1_109_cast_fp16))[name = string("op_2763_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2764_cast_fp16 = mul(x = var_2763_cast_fp16, y = sin_7_to_fp16)[name = string("op_2764_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> q_embed_cast_fp16 = add(x = var_2749_cast_fp16, y = var_2764_cast_fp16)[name = string("q_embed_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_109_cast_fp16 = transpose(perm = k_109_perm_0, x = squeeze_82_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 16, 128, 64]> var_2767_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_7_to_fp16)[name = string("op_2767_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 16, 128, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 128, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_109_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 16, 128, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 128, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_109_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_242_promoted_to_fp16 = const()[name = string("const_242_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 128, 32]> var_2779_cast_fp16 = mul(x = x2_cast_fp16, y = const_242_promoted_to_fp16)[name = string("op_2779_cast_fp16")];
            bool var_2781_interleave_0 = const()[name = string("op_2781_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2781_cast_fp16 = concat(axis = var_49, interleave = var_2781_interleave_0, values = (var_2779_cast_fp16, x1_cast_fp16))[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2782_cast_fp16 = mul(x = var_2781_cast_fp16, y = sin_7_to_fp16)[name = string("op_2782_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> k_embed_cast_fp16 = add(x = var_2767_cast_fp16, y = var_2782_cast_fp16)[name = string("k_embed_cast_fp16")];
            bool var_2787_transpose_x_1 = const()[name = string("op_2787_transpose_x_1"), val = bool(false)];
            bool var_2787_transpose_y_1 = const()[name = string("op_2787_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 128]> var_2787_cast_fp16 = matmul(transpose_x = var_2787_transpose_x_1, transpose_y = var_2787_transpose_y_1, x = q_embed_cast_fp16, y = k_embed_cast_fp16)[name = string("op_2787_cast_fp16")];
            fp16 var_2788_to_fp16 = const()[name = string("op_2788_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 128, 128]> attn_weights_109_cast_fp16 = mul(x = var_2787_cast_fp16, y = var_2788_to_fp16)[name = string("attn_weights_109_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> input_491_cast_fp16 = add(x = attn_weights_109_cast_fp16, y = key_bias_cast_fp16)[name = string("input_491_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2791_cast_fp16 = softmax(axis = var_49, x = input_491_cast_fp16)[name = string("op_2791_cast_fp16")];
            bool attn_output_163_transpose_x_0 = const()[name = string("attn_output_163_transpose_x_0"), val = bool(false)];
            bool attn_output_163_transpose_y_0 = const()[name = string("attn_output_163_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> value_cast_fp16 = transpose(perm = value_perm_0, x = squeeze_83_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 16, 128, 64]> attn_output_163_cast_fp16 = matmul(transpose_x = attn_output_163_transpose_x_0, transpose_y = attn_output_163_transpose_y_0, x = var_2791_cast_fp16, y = value_cast_fp16)[name = string("attn_output_163_cast_fp16")];
            tensor<int32, [4]> var_2795_perm_0 = const()[name = string("op_2795_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2797 = const()[name = string("op_2797"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 64]> var_2795_cast_fp16 = transpose(perm = var_2795_perm_0, x = attn_output_163_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 128, 1024]> var_2798_cast_fp16 = reshape(shape = var_2797, x = var_2795_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_27_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_27_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(771468416)))];
            tensor<fp16, [1, 128, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_27_attn_Wo_weight_to_fp16, x = var_2798_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_497_cast_fp16 = add(x = input_489_cast_fp16, y = linear_109_cast_fp16)[name = string("input_497_cast_fp16")];
            tensor<int32, [1]> input_499_axes_0 = const()[name = string("input_499_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_27_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_27_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773565632)))];
            tensor<fp16, [1, 128, 1024]> input_499_cast_fp16 = layer_norm(axes = input_499_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_27_mlp_norm_weight_to_fp16, x = input_497_cast_fp16)[name = string("input_499_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_27_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_27_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773567744)))];
            tensor<fp16, [1, 128, 5248]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_27_mlp_Wi_weight_to_fp16, x = input_499_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [2]> var_2810_split_sizes_0 = const()[name = string("op_2810_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2810_axis_0 = const()[name = string("op_2810_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 128, 2624]> var_2810_cast_fp16_0, tensor<fp16, [1, 128, 2624]> var_2810_cast_fp16_1 = split(axis = var_2810_axis_0, split_sizes = var_2810_split_sizes_0, x = linear_110_cast_fp16)[name = string("op_2810_cast_fp16")];
            string var_2812_mode_0 = const()[name = string("op_2812_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 128, 2624]> var_2812_cast_fp16 = gelu(mode = var_2812_mode_0, x = var_2810_cast_fp16_0)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [1, 128, 2624]> input_503_cast_fp16 = mul(x = var_2812_cast_fp16, y = var_2810_cast_fp16_1)[name = string("input_503_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_27_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_27_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(784315712)))];
            tensor<fp16, [1, 128, 1024]> linear_111_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_27_mlp_Wo_weight_to_fp16, x = input_503_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> input_507_cast_fp16 = add(x = input_497_cast_fp16, y = linear_111_cast_fp16)[name = string("input_507_cast_fp16")];
            tensor<int32, [1]> hidden_axes_0 = const()[name = string("hidden_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_final_norm_weight_to_fp16 = const()[name = string("model_model_final_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789689728)))];
            tensor<fp16, [1, 128, 1024]> hidden_cast_fp16 = layer_norm(axes = hidden_axes_0, epsilon = var_52_to_fp16, gamma = model_model_final_norm_weight_to_fp16, x = input_507_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_2822_axes_0 = const()[name = string("op_2822_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 128, 1]> var_2822_cast_fp16 = expand_dims(axes = var_2822_axes_0, x = attention_mask_to_fp16)[name = string("op_2822_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2823_cast_fp16 = mul(x = hidden_cast_fp16, y = var_2822_cast_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_2828_axes_0 = const()[name = string("op_2828_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_2828_keep_dims_0 = const()[name = string("op_2828_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 1024]> var_2828_cast_fp16 = reduce_sum(axes = var_2828_axes_0, keep_dims = var_2828_keep_dims_0, x = var_2823_cast_fp16)[name = string("op_2828_cast_fp16")];
            tensor<int32, [1]> var_2833_axes_0 = const()[name = string("op_2833_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_2833_keep_dims_0 = const()[name = string("op_2833_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_2833_cast_fp16 = reduce_sum(axes = var_2833_axes_0, keep_dims = var_2833_keep_dims_0, x = attention_mask_to_fp16)[name = string("op_2833_cast_fp16")];
            tensor<fp16, [1, 1024]> input_509_cast_fp16 = real_div(x = var_2828_cast_fp16, y = var_2833_cast_fp16)[name = string("input_509_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_head_dense_weight_to_fp16 = const()[name = string("model_head_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789691840)))];
            tensor<fp16, [1, 1024]> linear_112_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_head_dense_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_112_cast_fp16")];
            string input_513_mode_0 = const()[name = string("input_513_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024]> input_513_cast_fp16 = gelu(mode = input_513_mode_0, x = linear_112_cast_fp16)[name = string("input_513_cast_fp16")];
            tensor<int32, [1]> input_515_axes_0 = const()[name = string("input_515_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_head_norm_weight_to_fp16 = const()[name = string("model_head_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791789056)))];
            fp16 var_2836_to_fp16 = const()[name = string("op_2836_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024]> input_515_cast_fp16 = layer_norm(axes = input_515_axes_0, epsilon = var_2836_to_fp16, gamma = model_head_norm_weight_to_fp16, x = input_513_cast_fp16)[name = string("input_515_cast_fp16")];
            tensor<fp16, [3, 1024]> model_classifier_weight_to_fp16 = const()[name = string("model_classifier_weight_to_fp16"), val = tensor<fp16, [3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791791168)))];
            tensor<fp16, [3]> model_classifier_bias_to_fp16 = const()[name = string("model_classifier_bias_to_fp16"), val = tensor<fp16, [3]>([-0x1.284p-12, -0x1.f2p-17, 0x1.41p-12])];
            tensor<fp16, [1, 3]> linear_113_cast_fp16 = linear(bias = model_classifier_bias_to_fp16, weight = model_classifier_weight_to_fp16, x = input_515_cast_fp16)[name = string("linear_113_cast_fp16")];
            string linear_113_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_113_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 3]> logits = cast(dtype = linear_113_cast_fp16_to_fp32_dtype_0, x = linear_113_cast_fp16)[name = string("cast_59")];
        } -> (logits);
    func len192<ios18>(tensor<int32, [1, 192]> attention_mask, tensor<int32, [1, 192]> input_ids) {
            fp16 var_15_to_fp16 = const()[name = string("op_15_to_fp16"), val = fp16(0x1p+0)];
            string keys_to_fp16_dtype_0 = const()[name = string("keys_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 192]> attention_mask_to_fp16 = cast(dtype = keys_to_fp16_dtype_0, x = attention_mask)[name = string("cast_60")];
            tensor<fp16, [1, 192]> var_17_cast_fp16 = sub(x = var_15_to_fp16, y = attention_mask_to_fp16)[name = string("op_17_cast_fp16")];
            tensor<int32, [1]> var_24_axes_0 = const()[name = string("op_24_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 192]> var_24_cast_fp16 = expand_dims(axes = var_24_axes_0, x = var_17_cast_fp16)[name = string("op_24_cast_fp16")];
            tensor<int32, [1]> var_26_axes_0 = const()[name = string("op_26_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 192]> var_26_cast_fp16 = expand_dims(axes = var_26_axes_0, x = var_24_cast_fp16)[name = string("op_26_cast_fp16")];
            fp16 var_32_to_fp16 = const()[name = string("op_32_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 1, 1, 192]> key_bias_cast_fp16 = mul(x = var_26_cast_fp16, y = var_32_to_fp16)[name = string("key_bias_cast_fp16")];
            tensor<fp16, [1, 1, 192, 192]> outside_to_fp16 = const()[name = string("outside_to_fp16"), val = tensor<fp16, [1, 1, 192, 192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791797376)))];
            tensor<fp16, [1, 1, 192, 192]> attention_mask_cast_fp16 = add(x = key_bias_cast_fp16, y = outside_to_fp16)[name = string("attention_mask_cast_fp16")];
            int32 var_49 = const()[name = string("op_49"), val = int32(-1)];
            int32 input_1_batch_dims_0 = const()[name = string("input_1_batch_dims_0"), val = int32(0)];
            bool input_1_validate_indices_0 = const()[name = string("input_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [50368, 1024]> model_model_embeddings_tok_embeddings_weight_to_fp16 = const()[name = string("model_model_embeddings_tok_embeddings_weight_to_fp16"), val = tensor<fp16, [50368, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32896)))];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 192]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(50368)];
            tensor<int32, [1, 192]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 192]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 input_1_cast_fp16_axis_0 = const()[name = string("input_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 192, 1024]> input_1_cast_fp16 = gather(axis = input_1_cast_fp16_axis_0, batch_dims = input_1_batch_dims_0, indices = select_0, validate_indices = input_1_validate_indices_0, x = model_model_embeddings_tok_embeddings_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<int32, [1]> input_3_axes_0 = const()[name = string("input_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_embeddings_norm_weight_to_fp16 = const()[name = string("model_model_embeddings_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103186624)))];
            fp16 var_52_to_fp16 = const()[name = string("op_52_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 192, 1024]> input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, epsilon = var_52_to_fp16, gamma = model_model_embeddings_norm_weight_to_fp16, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_0_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_0_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103188736)))];
            tensor<fp16, [3072]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109480256)))];
            tensor<fp16, [1, 192, 3072]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_0_attn_Wqkv_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [5]> var_199 = const()[name = string("op_199"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_3_cast_fp16 = reshape(shape = var_199, x = linear_0_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> var_201_split_sizes_0 = const()[name = string("op_201_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_201_axis_0 = const()[name = string("op_201_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_201_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_201_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_201_cast_fp16_2 = split(axis = var_201_axis_0, split_sizes = var_201_split_sizes_0, x = qkv_3_cast_fp16)[name = string("op_201_cast_fp16")];
            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = string("squeeze_0_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = var_201_cast_fp16_0)[name = string("squeeze_0_cast_fp16")];
            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = string("squeeze_1_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = var_201_cast_fp16_1)[name = string("squeeze_1_cast_fp16")];
            tensor<int32, [1]> squeeze_2_axes_0 = const()[name = string("squeeze_2_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_2_cast_fp16 = squeeze(axes = squeeze_2_axes_0, x = var_201_cast_fp16_2)[name = string("squeeze_2_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_1_perm_0 = const()[name = string("value_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 1, 192, 64]> cos_7_to_fp16 = const()[name = string("cos_7_to_fp16"), val = tensor<fp16, [1, 1, 192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791871168)))];
            tensor<fp16, [1, 16, 192, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = squeeze_0_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 16, 192, 64]> var_211_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_7_to_fp16)[name = string("op_211_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_223_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_223_cast_fp16")];
            bool var_225_interleave_0 = const()[name = string("op_225_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_225_cast_fp16 = concat(axis = var_49, interleave = var_225_interleave_0, values = (var_223_cast_fp16, x1_1_cast_fp16))[name = string("op_225_cast_fp16")];
            tensor<fp16, [1, 1, 192, 64]> sin_7_to_fp16 = const()[name = string("sin_7_to_fp16"), val = tensor<fp16, [1, 1, 192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791895808)))];
            tensor<fp16, [1, 16, 192, 64]> var_226_cast_fp16 = mul(x = var_225_cast_fp16, y = sin_7_to_fp16)[name = string("op_226_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_1_cast_fp16 = add(x = var_211_cast_fp16, y = var_226_cast_fp16)[name = string("q_embed_1_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = squeeze_1_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 192, 64]> var_229_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_7_to_fp16)[name = string("op_229_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_241_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_241_cast_fp16")];
            bool var_243_interleave_0 = const()[name = string("op_243_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_243_cast_fp16 = concat(axis = var_49, interleave = var_243_interleave_0, values = (var_241_cast_fp16, x1_3_cast_fp16))[name = string("op_243_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_244_cast_fp16 = mul(x = var_243_cast_fp16, y = sin_7_to_fp16)[name = string("op_244_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_1_cast_fp16 = add(x = var_229_cast_fp16, y = var_244_cast_fp16)[name = string("k_embed_1_cast_fp16")];
            bool var_249_transpose_x_1 = const()[name = string("op_249_transpose_x_1"), val = bool(false)];
            bool var_249_transpose_y_1 = const()[name = string("op_249_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_249_cast_fp16 = matmul(transpose_x = var_249_transpose_x_1, transpose_y = var_249_transpose_y_1, x = q_embed_1_cast_fp16, y = k_embed_1_cast_fp16)[name = string("op_249_cast_fp16")];
            fp16 var_250_to_fp16 = const()[name = string("op_250_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_1_cast_fp16 = mul(x = var_249_cast_fp16, y = var_250_to_fp16)[name = string("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_5_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = key_bias_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_253_cast_fp16 = softmax(axis = var_49, x = input_5_cast_fp16)[name = string("op_253_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_1_cast_fp16 = transpose(perm = value_1_perm_0, x = squeeze_2_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = var_253_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_257_perm_0 = const()[name = string("op_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_259 = const()[name = string("op_259"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_257_cast_fp16 = transpose(perm = var_257_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 192, 1024]> var_260_cast_fp16 = reshape(shape = var_259, x = var_257_cast_fp16)[name = string("op_260_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_0_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_0_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109519360)))];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111616576)))];
            tensor<fp16, [1, 192, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_0_attn_Wo_weight_to_fp16, x = var_260_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_11_cast_fp16 = add(x = input_3_cast_fp16, y = linear_1_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<int32, [1]> input_13_axes_0 = const()[name = string("input_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_0_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_0_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111618688)))];
            tensor<fp16, [1, 192, 1024]> input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_0_mlp_norm_weight_to_fp16, x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_0_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_0_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111620800)))];
            tensor<fp16, [5248]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5248]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122368768)))];
            tensor<fp16, [1, 192, 5248]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_0_mlp_Wi_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [2]> var_272_split_sizes_0 = const()[name = string("op_272_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_272_axis_0 = const()[name = string("op_272_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_272_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_272_cast_fp16_1 = split(axis = var_272_axis_0, split_sizes = var_272_split_sizes_0, x = linear_2_cast_fp16)[name = string("op_272_cast_fp16")];
            string var_274_mode_0 = const()[name = string("op_274_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_274_cast_fp16 = gelu(mode = var_274_mode_0, x = var_272_cast_fp16_0)[name = string("op_274_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_17_cast_fp16 = mul(x = var_274_cast_fp16, y = var_272_cast_fp16_1)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_0_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_0_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122379328)))];
            tensor<fp16, [1, 192, 1024]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_0_mlp_Wo_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_21_cast_fp16 = add(x = input_11_cast_fp16, y = linear_3_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [1]> hidden_states_3_axes_0 = const()[name = string("hidden_states_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_1_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_1_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127753344)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_3_cast_fp16 = layer_norm(axes = hidden_states_3_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_1_attn_norm_weight_to_fp16, x = input_21_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_1_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_1_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127755456)))];
            tensor<fp16, [1, 192, 3072]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_1_attn_Wqkv_weight_to_fp16, x = hidden_states_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [5]> var_293 = const()[name = string("op_293"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_7_cast_fp16 = reshape(shape = var_293, x = linear_4_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> var_295_split_sizes_0 = const()[name = string("op_295_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_295_axis_0 = const()[name = string("op_295_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_295_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_295_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_295_cast_fp16_2 = split(axis = var_295_axis_0, split_sizes = var_295_split_sizes_0, x = qkv_7_cast_fp16)[name = string("op_295_cast_fp16")];
            tensor<int32, [1]> squeeze_3_axes_0 = const()[name = string("squeeze_3_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_3_cast_fp16 = squeeze(axes = squeeze_3_axes_0, x = var_295_cast_fp16_0)[name = string("squeeze_3_cast_fp16")];
            tensor<int32, [1]> squeeze_4_axes_0 = const()[name = string("squeeze_4_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_4_cast_fp16 = squeeze(axes = squeeze_4_axes_0, x = var_295_cast_fp16_1)[name = string("squeeze_4_cast_fp16")];
            tensor<int32, [1]> squeeze_5_axes_0 = const()[name = string("squeeze_5_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_5_cast_fp16 = squeeze(axes = squeeze_5_axes_0, x = var_295_cast_fp16_2)[name = string("squeeze_5_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 1, 192, 64]> cos_11_to_fp16 = const()[name = string("cos_11_to_fp16"), val = tensor<fp16, [1, 1, 192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791920448)))];
            tensor<fp16, [1, 16, 192, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = squeeze_3_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 16, 192, 64]> var_305_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_11_to_fp16)[name = string("op_305_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_5_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_5_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_317_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_317_cast_fp16")];
            bool var_319_interleave_0 = const()[name = string("op_319_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_319_cast_fp16 = concat(axis = var_49, interleave = var_319_interleave_0, values = (var_317_cast_fp16, x1_5_cast_fp16))[name = string("op_319_cast_fp16")];
            tensor<fp16, [1, 1, 192, 64]> sin_11_to_fp16 = const()[name = string("sin_11_to_fp16"), val = tensor<fp16, [1, 1, 192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791945088)))];
            tensor<fp16, [1, 16, 192, 64]> var_320_cast_fp16 = mul(x = var_319_cast_fp16, y = sin_11_to_fp16)[name = string("op_320_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_3_cast_fp16 = add(x = var_305_cast_fp16, y = var_320_cast_fp16)[name = string("q_embed_3_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = squeeze_4_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 16, 192, 64]> var_323_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_11_to_fp16)[name = string("op_323_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_5_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_5_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_335_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_335_cast_fp16")];
            bool var_337_interleave_0 = const()[name = string("op_337_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_337_cast_fp16 = concat(axis = var_49, interleave = var_337_interleave_0, values = (var_335_cast_fp16, x1_7_cast_fp16))[name = string("op_337_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_338_cast_fp16 = mul(x = var_337_cast_fp16, y = sin_11_to_fp16)[name = string("op_338_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_3_cast_fp16 = add(x = var_323_cast_fp16, y = var_338_cast_fp16)[name = string("k_embed_3_cast_fp16")];
            bool var_343_transpose_x_1 = const()[name = string("op_343_transpose_x_1"), val = bool(false)];
            bool var_343_transpose_y_1 = const()[name = string("op_343_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_343_cast_fp16 = matmul(transpose_x = var_343_transpose_x_1, transpose_y = var_343_transpose_y_1, x = q_embed_3_cast_fp16, y = k_embed_3_cast_fp16)[name = string("op_343_cast_fp16")];
            fp16 var_344_to_fp16 = const()[name = string("op_344_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_5_cast_fp16 = mul(x = var_343_cast_fp16, y = var_344_to_fp16)[name = string("attn_weights_5_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_23_cast_fp16 = add(x = attn_weights_5_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_347_cast_fp16 = softmax(axis = var_49, x = input_23_cast_fp16)[name = string("op_347_cast_fp16")];
            bool attn_output_7_transpose_x_0 = const()[name = string("attn_output_7_transpose_x_0"), val = bool(false)];
            bool attn_output_7_transpose_y_0 = const()[name = string("attn_output_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = squeeze_5_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = var_347_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<int32, [4]> var_351_perm_0 = const()[name = string("op_351_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_353 = const()[name = string("op_353"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_351_cast_fp16 = transpose(perm = var_351_perm_0, x = attn_output_7_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 192, 1024]> var_354_cast_fp16 = reshape(shape = var_353, x = var_351_cast_fp16)[name = string("op_354_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_1_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_1_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134079872)))];
            tensor<fp16, [1, 192, 1024]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_1_attn_Wo_weight_to_fp16, x = var_354_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_29_cast_fp16 = add(x = input_21_cast_fp16, y = linear_5_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_1_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_1_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136177088)))];
            tensor<fp16, [1, 192, 1024]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_1_mlp_norm_weight_to_fp16, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_1_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_1_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136179200)))];
            tensor<fp16, [1, 192, 5248]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_1_mlp_Wi_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [2]> var_366_split_sizes_0 = const()[name = string("op_366_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_366_axis_0 = const()[name = string("op_366_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_366_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_366_cast_fp16_1 = split(axis = var_366_axis_0, split_sizes = var_366_split_sizes_0, x = linear_6_cast_fp16)[name = string("op_366_cast_fp16")];
            string var_368_mode_0 = const()[name = string("op_368_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_368_cast_fp16 = gelu(mode = var_368_mode_0, x = var_366_cast_fp16_0)[name = string("op_368_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_35_cast_fp16 = mul(x = var_368_cast_fp16, y = var_366_cast_fp16_1)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_1_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_1_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146927168)))];
            tensor<fp16, [1, 192, 1024]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_1_mlp_Wo_weight_to_fp16, x = input_35_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_39_cast_fp16 = add(x = input_29_cast_fp16, y = linear_7_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<int32, [1]> hidden_states_5_axes_0 = const()[name = string("hidden_states_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_2_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_2_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152301184)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_5_cast_fp16 = layer_norm(axes = hidden_states_5_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_2_attn_norm_weight_to_fp16, x = input_39_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_2_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_2_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152303296)))];
            tensor<fp16, [1, 192, 3072]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_2_attn_Wqkv_weight_to_fp16, x = hidden_states_5_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [5]> var_387 = const()[name = string("op_387"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_11_cast_fp16 = reshape(shape = var_387, x = linear_8_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> var_389_split_sizes_0 = const()[name = string("op_389_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_389_axis_0 = const()[name = string("op_389_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_389_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_389_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_389_cast_fp16_2 = split(axis = var_389_axis_0, split_sizes = var_389_split_sizes_0, x = qkv_11_cast_fp16)[name = string("op_389_cast_fp16")];
            tensor<int32, [1]> squeeze_6_axes_0 = const()[name = string("squeeze_6_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_6_cast_fp16 = squeeze(axes = squeeze_6_axes_0, x = var_389_cast_fp16_0)[name = string("squeeze_6_cast_fp16")];
            tensor<int32, [1]> squeeze_7_axes_0 = const()[name = string("squeeze_7_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_7_cast_fp16 = squeeze(axes = squeeze_7_axes_0, x = var_389_cast_fp16_1)[name = string("squeeze_7_cast_fp16")];
            tensor<int32, [1]> squeeze_8_axes_0 = const()[name = string("squeeze_8_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_8_cast_fp16 = squeeze(axes = squeeze_8_axes_0, x = var_389_cast_fp16_2)[name = string("squeeze_8_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = squeeze_6_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 16, 192, 64]> var_399_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_11_to_fp16)[name = string("op_399_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_9_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_9_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_411_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_411_cast_fp16")];
            bool var_413_interleave_0 = const()[name = string("op_413_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_413_cast_fp16 = concat(axis = var_49, interleave = var_413_interleave_0, values = (var_411_cast_fp16, x1_9_cast_fp16))[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_414_cast_fp16 = mul(x = var_413_cast_fp16, y = sin_11_to_fp16)[name = string("op_414_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_5_cast_fp16 = add(x = var_399_cast_fp16, y = var_414_cast_fp16)[name = string("q_embed_5_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = squeeze_7_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 16, 192, 64]> var_417_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_11_to_fp16)[name = string("op_417_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_9_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_9_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_429_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_429_cast_fp16")];
            bool var_431_interleave_0 = const()[name = string("op_431_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_431_cast_fp16 = concat(axis = var_49, interleave = var_431_interleave_0, values = (var_429_cast_fp16, x1_11_cast_fp16))[name = string("op_431_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_432_cast_fp16 = mul(x = var_431_cast_fp16, y = sin_11_to_fp16)[name = string("op_432_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_5_cast_fp16 = add(x = var_417_cast_fp16, y = var_432_cast_fp16)[name = string("k_embed_5_cast_fp16")];
            bool var_437_transpose_x_1 = const()[name = string("op_437_transpose_x_1"), val = bool(false)];
            bool var_437_transpose_y_1 = const()[name = string("op_437_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_437_cast_fp16 = matmul(transpose_x = var_437_transpose_x_1, transpose_y = var_437_transpose_y_1, x = q_embed_5_cast_fp16, y = k_embed_5_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 var_438_to_fp16 = const()[name = string("op_438_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_9_cast_fp16 = mul(x = var_437_cast_fp16, y = var_438_to_fp16)[name = string("attn_weights_9_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_41_cast_fp16 = add(x = attn_weights_9_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_441_cast_fp16 = softmax(axis = var_49, x = input_41_cast_fp16)[name = string("op_441_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = squeeze_8_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = var_441_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_445_perm_0 = const()[name = string("op_445_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_447 = const()[name = string("op_447"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_445_cast_fp16 = transpose(perm = var_445_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 192, 1024]> var_448_cast_fp16 = reshape(shape = var_447, x = var_445_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_2_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_2_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158594816)))];
            tensor<fp16, [1, 192, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_2_attn_Wo_weight_to_fp16, x = var_448_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_47_cast_fp16 = add(x = input_39_cast_fp16, y = linear_9_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<int32, [1]> input_49_axes_0 = const()[name = string("input_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_2_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_2_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160692032)))];
            tensor<fp16, [1, 192, 1024]> input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_2_mlp_norm_weight_to_fp16, x = input_47_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_2_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_2_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160694144)))];
            tensor<fp16, [1, 192, 5248]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_2_mlp_Wi_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [2]> var_460_split_sizes_0 = const()[name = string("op_460_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_460_axis_0 = const()[name = string("op_460_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_460_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_460_cast_fp16_1 = split(axis = var_460_axis_0, split_sizes = var_460_split_sizes_0, x = linear_10_cast_fp16)[name = string("op_460_cast_fp16")];
            string var_462_mode_0 = const()[name = string("op_462_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_462_cast_fp16 = gelu(mode = var_462_mode_0, x = var_460_cast_fp16_0)[name = string("op_462_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_53_cast_fp16 = mul(x = var_462_cast_fp16, y = var_460_cast_fp16_1)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_2_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_2_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171442112)))];
            tensor<fp16, [1, 192, 1024]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_2_mlp_Wo_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_57_cast_fp16 = add(x = input_47_cast_fp16, y = linear_11_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<int32, [1]> hidden_states_7_axes_0 = const()[name = string("hidden_states_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_3_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_3_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176816128)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_7_cast_fp16 = layer_norm(axes = hidden_states_7_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_3_attn_norm_weight_to_fp16, x = input_57_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_3_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_3_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176818240)))];
            tensor<fp16, [1, 192, 3072]> linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_3_attn_Wqkv_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [5]> var_481 = const()[name = string("op_481"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_15_cast_fp16 = reshape(shape = var_481, x = linear_12_cast_fp16)[name = string("qkv_15_cast_fp16")];
            tensor<int32, [3]> var_483_split_sizes_0 = const()[name = string("op_483_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_483_axis_0 = const()[name = string("op_483_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_483_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_483_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_483_cast_fp16_2 = split(axis = var_483_axis_0, split_sizes = var_483_split_sizes_0, x = qkv_15_cast_fp16)[name = string("op_483_cast_fp16")];
            tensor<int32, [1]> squeeze_9_axes_0 = const()[name = string("squeeze_9_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_9_cast_fp16 = squeeze(axes = squeeze_9_axes_0, x = var_483_cast_fp16_0)[name = string("squeeze_9_cast_fp16")];
            tensor<int32, [1]> squeeze_10_axes_0 = const()[name = string("squeeze_10_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_10_cast_fp16 = squeeze(axes = squeeze_10_axes_0, x = var_483_cast_fp16_1)[name = string("squeeze_10_cast_fp16")];
            tensor<int32, [1]> squeeze_11_axes_0 = const()[name = string("squeeze_11_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_11_cast_fp16 = squeeze(axes = squeeze_11_axes_0, x = var_483_cast_fp16_2)[name = string("squeeze_11_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_7_perm_0 = const()[name = string("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = squeeze_9_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 192, 64]> var_493_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_7_to_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_13_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_13_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_505_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_505_cast_fp16")];
            bool var_507_interleave_0 = const()[name = string("op_507_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_507_cast_fp16 = concat(axis = var_49, interleave = var_507_interleave_0, values = (var_505_cast_fp16, x1_13_cast_fp16))[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_508_cast_fp16 = mul(x = var_507_cast_fp16, y = sin_7_to_fp16)[name = string("op_508_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_7_cast_fp16 = add(x = var_493_cast_fp16, y = var_508_cast_fp16)[name = string("q_embed_7_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = squeeze_10_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 16, 192, 64]> var_511_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_7_to_fp16)[name = string("op_511_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_13_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_13_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_523_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_523_cast_fp16")];
            bool var_525_interleave_0 = const()[name = string("op_525_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_525_cast_fp16 = concat(axis = var_49, interleave = var_525_interleave_0, values = (var_523_cast_fp16, x1_15_cast_fp16))[name = string("op_525_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_526_cast_fp16 = mul(x = var_525_cast_fp16, y = sin_7_to_fp16)[name = string("op_526_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_7_cast_fp16 = add(x = var_511_cast_fp16, y = var_526_cast_fp16)[name = string("k_embed_7_cast_fp16")];
            bool var_531_transpose_x_1 = const()[name = string("op_531_transpose_x_1"), val = bool(false)];
            bool var_531_transpose_y_1 = const()[name = string("op_531_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_531_cast_fp16 = matmul(transpose_x = var_531_transpose_x_1, transpose_y = var_531_transpose_y_1, x = q_embed_7_cast_fp16, y = k_embed_7_cast_fp16)[name = string("op_531_cast_fp16")];
            fp16 var_532_to_fp16 = const()[name = string("op_532_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_13_cast_fp16 = mul(x = var_531_cast_fp16, y = var_532_to_fp16)[name = string("attn_weights_13_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_59_cast_fp16 = add(x = attn_weights_13_cast_fp16, y = key_bias_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_535_cast_fp16 = softmax(axis = var_49, x = input_59_cast_fp16)[name = string("op_535_cast_fp16")];
            bool attn_output_19_transpose_x_0 = const()[name = string("attn_output_19_transpose_x_0"), val = bool(false)];
            bool attn_output_19_transpose_y_0 = const()[name = string("attn_output_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = squeeze_11_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = var_535_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_539_perm_0 = const()[name = string("op_539_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_541 = const()[name = string("op_541"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_539_cast_fp16 = transpose(perm = var_539_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 192, 1024]> var_542_cast_fp16 = reshape(shape = var_541, x = var_539_cast_fp16)[name = string("op_542_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_3_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_3_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183109760)))];
            tensor<fp16, [1, 192, 1024]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_3_attn_Wo_weight_to_fp16, x = var_542_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_65_cast_fp16 = add(x = input_57_cast_fp16, y = linear_13_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_3_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_3_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185206976)))];
            tensor<fp16, [1, 192, 1024]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_3_mlp_norm_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_3_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_3_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185209088)))];
            tensor<fp16, [1, 192, 5248]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_3_mlp_Wi_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [2]> var_554_split_sizes_0 = const()[name = string("op_554_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_554_axis_0 = const()[name = string("op_554_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_554_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_554_cast_fp16_1 = split(axis = var_554_axis_0, split_sizes = var_554_split_sizes_0, x = linear_14_cast_fp16)[name = string("op_554_cast_fp16")];
            string var_556_mode_0 = const()[name = string("op_556_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_556_cast_fp16 = gelu(mode = var_556_mode_0, x = var_554_cast_fp16_0)[name = string("op_556_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_71_cast_fp16 = mul(x = var_556_cast_fp16, y = var_554_cast_fp16_1)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_3_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_3_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195957056)))];
            tensor<fp16, [1, 192, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_3_mlp_Wo_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_75_cast_fp16 = add(x = input_65_cast_fp16, y = linear_15_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<int32, [1]> hidden_states_9_axes_0 = const()[name = string("hidden_states_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_4_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_4_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201331072)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_9_cast_fp16 = layer_norm(axes = hidden_states_9_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_4_attn_norm_weight_to_fp16, x = input_75_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_4_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_4_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201333184)))];
            tensor<fp16, [1, 192, 3072]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_4_attn_Wqkv_weight_to_fp16, x = hidden_states_9_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [5]> var_575 = const()[name = string("op_575"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_19_cast_fp16 = reshape(shape = var_575, x = linear_16_cast_fp16)[name = string("qkv_19_cast_fp16")];
            tensor<int32, [3]> var_577_split_sizes_0 = const()[name = string("op_577_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_577_axis_0 = const()[name = string("op_577_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_577_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_577_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_577_cast_fp16_2 = split(axis = var_577_axis_0, split_sizes = var_577_split_sizes_0, x = qkv_19_cast_fp16)[name = string("op_577_cast_fp16")];
            tensor<int32, [1]> squeeze_12_axes_0 = const()[name = string("squeeze_12_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_12_cast_fp16 = squeeze(axes = squeeze_12_axes_0, x = var_577_cast_fp16_0)[name = string("squeeze_12_cast_fp16")];
            tensor<int32, [1]> squeeze_13_axes_0 = const()[name = string("squeeze_13_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_13_cast_fp16 = squeeze(axes = squeeze_13_axes_0, x = var_577_cast_fp16_1)[name = string("squeeze_13_cast_fp16")];
            tensor<int32, [1]> squeeze_14_axes_0 = const()[name = string("squeeze_14_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_14_cast_fp16 = squeeze(axes = squeeze_14_axes_0, x = var_577_cast_fp16_2)[name = string("squeeze_14_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = squeeze_12_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 192, 64]> var_587_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_11_to_fp16)[name = string("op_587_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_17_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_17_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_599_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_55_promoted_to_fp16)[name = string("op_599_cast_fp16")];
            bool var_601_interleave_0 = const()[name = string("op_601_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_601_cast_fp16 = concat(axis = var_49, interleave = var_601_interleave_0, values = (var_599_cast_fp16, x1_17_cast_fp16))[name = string("op_601_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_602_cast_fp16 = mul(x = var_601_cast_fp16, y = sin_11_to_fp16)[name = string("op_602_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_9_cast_fp16 = add(x = var_587_cast_fp16, y = var_602_cast_fp16)[name = string("q_embed_9_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = squeeze_13_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 192, 64]> var_605_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_11_to_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_17_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_17_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_58_promoted_to_fp16 = const()[name = string("const_58_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_617_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_58_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            bool var_619_interleave_0 = const()[name = string("op_619_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_619_cast_fp16 = concat(axis = var_49, interleave = var_619_interleave_0, values = (var_617_cast_fp16, x1_19_cast_fp16))[name = string("op_619_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_620_cast_fp16 = mul(x = var_619_cast_fp16, y = sin_11_to_fp16)[name = string("op_620_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_9_cast_fp16 = add(x = var_605_cast_fp16, y = var_620_cast_fp16)[name = string("k_embed_9_cast_fp16")];
            bool var_625_transpose_x_1 = const()[name = string("op_625_transpose_x_1"), val = bool(false)];
            bool var_625_transpose_y_1 = const()[name = string("op_625_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_625_cast_fp16 = matmul(transpose_x = var_625_transpose_x_1, transpose_y = var_625_transpose_y_1, x = q_embed_9_cast_fp16, y = k_embed_9_cast_fp16)[name = string("op_625_cast_fp16")];
            fp16 var_626_to_fp16 = const()[name = string("op_626_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_17_cast_fp16 = mul(x = var_625_cast_fp16, y = var_626_to_fp16)[name = string("attn_weights_17_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_77_cast_fp16 = add(x = attn_weights_17_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_629_cast_fp16 = softmax(axis = var_49, x = input_77_cast_fp16)[name = string("op_629_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = squeeze_14_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = var_629_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_633_perm_0 = const()[name = string("op_633_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_635 = const()[name = string("op_635"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_633_cast_fp16 = transpose(perm = var_633_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 192, 1024]> var_636_cast_fp16 = reshape(shape = var_635, x = var_633_cast_fp16)[name = string("op_636_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_4_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_4_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207624704)))];
            tensor<fp16, [1, 192, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_4_attn_Wo_weight_to_fp16, x = var_636_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_83_cast_fp16 = add(x = input_75_cast_fp16, y = linear_17_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<int32, [1]> input_85_axes_0 = const()[name = string("input_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_4_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_4_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209721920)))];
            tensor<fp16, [1, 192, 1024]> input_85_cast_fp16 = layer_norm(axes = input_85_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_4_mlp_norm_weight_to_fp16, x = input_83_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_4_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_4_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209724032)))];
            tensor<fp16, [1, 192, 5248]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_4_mlp_Wi_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [2]> var_648_split_sizes_0 = const()[name = string("op_648_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_648_axis_0 = const()[name = string("op_648_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_648_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_648_cast_fp16_1 = split(axis = var_648_axis_0, split_sizes = var_648_split_sizes_0, x = linear_18_cast_fp16)[name = string("op_648_cast_fp16")];
            string var_650_mode_0 = const()[name = string("op_650_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_650_cast_fp16 = gelu(mode = var_650_mode_0, x = var_648_cast_fp16_0)[name = string("op_650_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_89_cast_fp16 = mul(x = var_650_cast_fp16, y = var_648_cast_fp16_1)[name = string("input_89_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_4_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_4_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220472000)))];
            tensor<fp16, [1, 192, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_4_mlp_Wo_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_93_cast_fp16 = add(x = input_83_cast_fp16, y = linear_19_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<int32, [1]> hidden_states_11_axes_0 = const()[name = string("hidden_states_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_5_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_5_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225846016)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_11_cast_fp16 = layer_norm(axes = hidden_states_11_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_5_attn_norm_weight_to_fp16, x = input_93_cast_fp16)[name = string("hidden_states_11_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_5_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_5_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225848128)))];
            tensor<fp16, [1, 192, 3072]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_5_attn_Wqkv_weight_to_fp16, x = hidden_states_11_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [5]> var_669 = const()[name = string("op_669"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_23_cast_fp16 = reshape(shape = var_669, x = linear_20_cast_fp16)[name = string("qkv_23_cast_fp16")];
            tensor<int32, [3]> var_671_split_sizes_0 = const()[name = string("op_671_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_671_axis_0 = const()[name = string("op_671_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_671_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_671_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_671_cast_fp16_2 = split(axis = var_671_axis_0, split_sizes = var_671_split_sizes_0, x = qkv_23_cast_fp16)[name = string("op_671_cast_fp16")];
            tensor<int32, [1]> squeeze_15_axes_0 = const()[name = string("squeeze_15_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_15_cast_fp16 = squeeze(axes = squeeze_15_axes_0, x = var_671_cast_fp16_0)[name = string("squeeze_15_cast_fp16")];
            tensor<int32, [1]> squeeze_16_axes_0 = const()[name = string("squeeze_16_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_16_cast_fp16 = squeeze(axes = squeeze_16_axes_0, x = var_671_cast_fp16_1)[name = string("squeeze_16_cast_fp16")];
            tensor<int32, [1]> squeeze_17_axes_0 = const()[name = string("squeeze_17_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_17_cast_fp16 = squeeze(axes = squeeze_17_axes_0, x = var_671_cast_fp16_2)[name = string("squeeze_17_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = squeeze_15_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 16, 192, 64]> var_681_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_11_to_fp16)[name = string("op_681_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_21_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_21_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_63_promoted_to_fp16 = const()[name = string("const_63_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_693_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_63_promoted_to_fp16)[name = string("op_693_cast_fp16")];
            bool var_695_interleave_0 = const()[name = string("op_695_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_695_cast_fp16 = concat(axis = var_49, interleave = var_695_interleave_0, values = (var_693_cast_fp16, x1_21_cast_fp16))[name = string("op_695_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_696_cast_fp16 = mul(x = var_695_cast_fp16, y = sin_11_to_fp16)[name = string("op_696_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_11_cast_fp16 = add(x = var_681_cast_fp16, y = var_696_cast_fp16)[name = string("q_embed_11_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = squeeze_16_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 192, 64]> var_699_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_11_to_fp16)[name = string("op_699_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_21_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_21_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_66_promoted_to_fp16 = const()[name = string("const_66_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_711_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_66_promoted_to_fp16)[name = string("op_711_cast_fp16")];
            bool var_713_interleave_0 = const()[name = string("op_713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_713_cast_fp16 = concat(axis = var_49, interleave = var_713_interleave_0, values = (var_711_cast_fp16, x1_23_cast_fp16))[name = string("op_713_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_714_cast_fp16 = mul(x = var_713_cast_fp16, y = sin_11_to_fp16)[name = string("op_714_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_11_cast_fp16 = add(x = var_699_cast_fp16, y = var_714_cast_fp16)[name = string("k_embed_11_cast_fp16")];
            bool var_719_transpose_x_1 = const()[name = string("op_719_transpose_x_1"), val = bool(false)];
            bool var_719_transpose_y_1 = const()[name = string("op_719_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_719_cast_fp16 = matmul(transpose_x = var_719_transpose_x_1, transpose_y = var_719_transpose_y_1, x = q_embed_11_cast_fp16, y = k_embed_11_cast_fp16)[name = string("op_719_cast_fp16")];
            fp16 var_720_to_fp16 = const()[name = string("op_720_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_21_cast_fp16 = mul(x = var_719_cast_fp16, y = var_720_to_fp16)[name = string("attn_weights_21_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_95_cast_fp16 = add(x = attn_weights_21_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_723_cast_fp16 = softmax(axis = var_49, x = input_95_cast_fp16)[name = string("op_723_cast_fp16")];
            bool attn_output_31_transpose_x_0 = const()[name = string("attn_output_31_transpose_x_0"), val = bool(false)];
            bool attn_output_31_transpose_y_0 = const()[name = string("attn_output_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = squeeze_17_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = var_723_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [4]> var_727_perm_0 = const()[name = string("op_727_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_729 = const()[name = string("op_729"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_727_cast_fp16 = transpose(perm = var_727_perm_0, x = attn_output_31_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 192, 1024]> var_730_cast_fp16 = reshape(shape = var_729, x = var_727_cast_fp16)[name = string("op_730_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_5_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_5_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232139648)))];
            tensor<fp16, [1, 192, 1024]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_5_attn_Wo_weight_to_fp16, x = var_730_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_101_cast_fp16 = add(x = input_93_cast_fp16, y = linear_21_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<int32, [1]> input_103_axes_0 = const()[name = string("input_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_5_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_5_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234236864)))];
            tensor<fp16, [1, 192, 1024]> input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_5_mlp_norm_weight_to_fp16, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_5_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_5_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234238976)))];
            tensor<fp16, [1, 192, 5248]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_5_mlp_Wi_weight_to_fp16, x = input_103_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [2]> var_742_split_sizes_0 = const()[name = string("op_742_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_742_axis_0 = const()[name = string("op_742_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_742_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_742_cast_fp16_1 = split(axis = var_742_axis_0, split_sizes = var_742_split_sizes_0, x = linear_22_cast_fp16)[name = string("op_742_cast_fp16")];
            string var_744_mode_0 = const()[name = string("op_744_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_744_cast_fp16 = gelu(mode = var_744_mode_0, x = var_742_cast_fp16_0)[name = string("op_744_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_107_cast_fp16 = mul(x = var_744_cast_fp16, y = var_742_cast_fp16_1)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_5_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_5_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244986944)))];
            tensor<fp16, [1, 192, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_5_mlp_Wo_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_111_cast_fp16 = add(x = input_101_cast_fp16, y = linear_23_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<int32, [1]> hidden_states_13_axes_0 = const()[name = string("hidden_states_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_6_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_6_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250360960)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_13_cast_fp16 = layer_norm(axes = hidden_states_13_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_6_attn_norm_weight_to_fp16, x = input_111_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_6_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_6_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250363072)))];
            tensor<fp16, [1, 192, 3072]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_6_attn_Wqkv_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [5]> var_763 = const()[name = string("op_763"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_27_cast_fp16 = reshape(shape = var_763, x = linear_24_cast_fp16)[name = string("qkv_27_cast_fp16")];
            tensor<int32, [3]> var_765_split_sizes_0 = const()[name = string("op_765_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_765_axis_0 = const()[name = string("op_765_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_765_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_765_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_765_cast_fp16_2 = split(axis = var_765_axis_0, split_sizes = var_765_split_sizes_0, x = qkv_27_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<int32, [1]> squeeze_18_axes_0 = const()[name = string("squeeze_18_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_18_cast_fp16 = squeeze(axes = squeeze_18_axes_0, x = var_765_cast_fp16_0)[name = string("squeeze_18_cast_fp16")];
            tensor<int32, [1]> squeeze_19_axes_0 = const()[name = string("squeeze_19_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_19_cast_fp16 = squeeze(axes = squeeze_19_axes_0, x = var_765_cast_fp16_1)[name = string("squeeze_19_cast_fp16")];
            tensor<int32, [1]> squeeze_20_axes_0 = const()[name = string("squeeze_20_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_20_cast_fp16 = squeeze(axes = squeeze_20_axes_0, x = var_765_cast_fp16_2)[name = string("squeeze_20_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = squeeze_18_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 16, 192, 64]> var_775_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_7_to_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_25_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_25_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_71_promoted_to_fp16 = const()[name = string("const_71_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_787_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_71_promoted_to_fp16)[name = string("op_787_cast_fp16")];
            bool var_789_interleave_0 = const()[name = string("op_789_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_789_cast_fp16 = concat(axis = var_49, interleave = var_789_interleave_0, values = (var_787_cast_fp16, x1_25_cast_fp16))[name = string("op_789_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_790_cast_fp16 = mul(x = var_789_cast_fp16, y = sin_7_to_fp16)[name = string("op_790_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_13_cast_fp16 = add(x = var_775_cast_fp16, y = var_790_cast_fp16)[name = string("q_embed_13_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = squeeze_19_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 16, 192, 64]> var_793_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_7_to_fp16)[name = string("op_793_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_25_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_25_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_805_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_74_promoted_to_fp16)[name = string("op_805_cast_fp16")];
            bool var_807_interleave_0 = const()[name = string("op_807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_807_cast_fp16 = concat(axis = var_49, interleave = var_807_interleave_0, values = (var_805_cast_fp16, x1_27_cast_fp16))[name = string("op_807_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_808_cast_fp16 = mul(x = var_807_cast_fp16, y = sin_7_to_fp16)[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_13_cast_fp16 = add(x = var_793_cast_fp16, y = var_808_cast_fp16)[name = string("k_embed_13_cast_fp16")];
            bool var_813_transpose_x_1 = const()[name = string("op_813_transpose_x_1"), val = bool(false)];
            bool var_813_transpose_y_1 = const()[name = string("op_813_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_813_cast_fp16 = matmul(transpose_x = var_813_transpose_x_1, transpose_y = var_813_transpose_y_1, x = q_embed_13_cast_fp16, y = k_embed_13_cast_fp16)[name = string("op_813_cast_fp16")];
            fp16 var_814_to_fp16 = const()[name = string("op_814_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_25_cast_fp16 = mul(x = var_813_cast_fp16, y = var_814_to_fp16)[name = string("attn_weights_25_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_113_cast_fp16 = add(x = attn_weights_25_cast_fp16, y = key_bias_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_817_cast_fp16 = softmax(axis = var_49, x = input_113_cast_fp16)[name = string("op_817_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = squeeze_20_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = var_817_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_821_perm_0 = const()[name = string("op_821_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_823 = const()[name = string("op_823"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_821_cast_fp16 = transpose(perm = var_821_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 192, 1024]> var_824_cast_fp16 = reshape(shape = var_823, x = var_821_cast_fp16)[name = string("op_824_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_6_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_6_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256654592)))];
            tensor<fp16, [1, 192, 1024]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_6_attn_Wo_weight_to_fp16, x = var_824_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_119_cast_fp16 = add(x = input_111_cast_fp16, y = linear_25_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_6_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_6_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258751808)))];
            tensor<fp16, [1, 192, 1024]> input_121_cast_fp16 = layer_norm(axes = input_121_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_6_mlp_norm_weight_to_fp16, x = input_119_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_6_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_6_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258753920)))];
            tensor<fp16, [1, 192, 5248]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_6_mlp_Wi_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [2]> var_836_split_sizes_0 = const()[name = string("op_836_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_836_axis_0 = const()[name = string("op_836_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_836_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_836_cast_fp16_1 = split(axis = var_836_axis_0, split_sizes = var_836_split_sizes_0, x = linear_26_cast_fp16)[name = string("op_836_cast_fp16")];
            string var_838_mode_0 = const()[name = string("op_838_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_838_cast_fp16 = gelu(mode = var_838_mode_0, x = var_836_cast_fp16_0)[name = string("op_838_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_125_cast_fp16 = mul(x = var_838_cast_fp16, y = var_836_cast_fp16_1)[name = string("input_125_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_6_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_6_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269501888)))];
            tensor<fp16, [1, 192, 1024]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_6_mlp_Wo_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_129_cast_fp16 = add(x = input_119_cast_fp16, y = linear_27_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<int32, [1]> hidden_states_15_axes_0 = const()[name = string("hidden_states_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_7_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_7_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274875904)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_15_cast_fp16 = layer_norm(axes = hidden_states_15_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_7_attn_norm_weight_to_fp16, x = input_129_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_7_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_7_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274878016)))];
            tensor<fp16, [1, 192, 3072]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_7_attn_Wqkv_weight_to_fp16, x = hidden_states_15_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [5]> var_857 = const()[name = string("op_857"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_31_cast_fp16 = reshape(shape = var_857, x = linear_28_cast_fp16)[name = string("qkv_31_cast_fp16")];
            tensor<int32, [3]> var_859_split_sizes_0 = const()[name = string("op_859_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_859_axis_0 = const()[name = string("op_859_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_859_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_859_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_859_cast_fp16_2 = split(axis = var_859_axis_0, split_sizes = var_859_split_sizes_0, x = qkv_31_cast_fp16)[name = string("op_859_cast_fp16")];
            tensor<int32, [1]> squeeze_21_axes_0 = const()[name = string("squeeze_21_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_21_cast_fp16 = squeeze(axes = squeeze_21_axes_0, x = var_859_cast_fp16_0)[name = string("squeeze_21_cast_fp16")];
            tensor<int32, [1]> squeeze_22_axes_0 = const()[name = string("squeeze_22_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_22_cast_fp16 = squeeze(axes = squeeze_22_axes_0, x = var_859_cast_fp16_1)[name = string("squeeze_22_cast_fp16")];
            tensor<int32, [1]> squeeze_23_axes_0 = const()[name = string("squeeze_23_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_23_cast_fp16 = squeeze(axes = squeeze_23_axes_0, x = var_859_cast_fp16_2)[name = string("squeeze_23_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = squeeze_21_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 16, 192, 64]> var_869_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_11_to_fp16)[name = string("op_869_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_29_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_29_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_79_promoted_to_fp16 = const()[name = string("const_79_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_881_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_79_promoted_to_fp16)[name = string("op_881_cast_fp16")];
            bool var_883_interleave_0 = const()[name = string("op_883_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_883_cast_fp16 = concat(axis = var_49, interleave = var_883_interleave_0, values = (var_881_cast_fp16, x1_29_cast_fp16))[name = string("op_883_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_884_cast_fp16 = mul(x = var_883_cast_fp16, y = sin_11_to_fp16)[name = string("op_884_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_15_cast_fp16 = add(x = var_869_cast_fp16, y = var_884_cast_fp16)[name = string("q_embed_15_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = squeeze_22_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 16, 192, 64]> var_887_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_11_to_fp16)[name = string("op_887_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_29_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_29_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_899_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_899_cast_fp16")];
            bool var_901_interleave_0 = const()[name = string("op_901_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_901_cast_fp16 = concat(axis = var_49, interleave = var_901_interleave_0, values = (var_899_cast_fp16, x1_31_cast_fp16))[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_902_cast_fp16 = mul(x = var_901_cast_fp16, y = sin_11_to_fp16)[name = string("op_902_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_15_cast_fp16 = add(x = var_887_cast_fp16, y = var_902_cast_fp16)[name = string("k_embed_15_cast_fp16")];
            bool var_907_transpose_x_1 = const()[name = string("op_907_transpose_x_1"), val = bool(false)];
            bool var_907_transpose_y_1 = const()[name = string("op_907_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_907_cast_fp16 = matmul(transpose_x = var_907_transpose_x_1, transpose_y = var_907_transpose_y_1, x = q_embed_15_cast_fp16, y = k_embed_15_cast_fp16)[name = string("op_907_cast_fp16")];
            fp16 var_908_to_fp16 = const()[name = string("op_908_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_29_cast_fp16 = mul(x = var_907_cast_fp16, y = var_908_to_fp16)[name = string("attn_weights_29_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_131_cast_fp16 = add(x = attn_weights_29_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_911_cast_fp16 = softmax(axis = var_49, x = input_131_cast_fp16)[name = string("op_911_cast_fp16")];
            bool attn_output_43_transpose_x_0 = const()[name = string("attn_output_43_transpose_x_0"), val = bool(false)];
            bool attn_output_43_transpose_y_0 = const()[name = string("attn_output_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = squeeze_23_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = var_911_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_915_perm_0 = const()[name = string("op_915_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_917 = const()[name = string("op_917"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_915_cast_fp16 = transpose(perm = var_915_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 192, 1024]> var_918_cast_fp16 = reshape(shape = var_917, x = var_915_cast_fp16)[name = string("op_918_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_7_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_7_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281169536)))];
            tensor<fp16, [1, 192, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_7_attn_Wo_weight_to_fp16, x = var_918_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_137_cast_fp16 = add(x = input_129_cast_fp16, y = linear_29_cast_fp16)[name = string("input_137_cast_fp16")];
            tensor<int32, [1]> input_139_axes_0 = const()[name = string("input_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_7_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_7_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283266752)))];
            tensor<fp16, [1, 192, 1024]> input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_7_mlp_norm_weight_to_fp16, x = input_137_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_7_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_7_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283268864)))];
            tensor<fp16, [1, 192, 5248]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_7_mlp_Wi_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [2]> var_930_split_sizes_0 = const()[name = string("op_930_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_930_axis_0 = const()[name = string("op_930_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_930_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_930_cast_fp16_1 = split(axis = var_930_axis_0, split_sizes = var_930_split_sizes_0, x = linear_30_cast_fp16)[name = string("op_930_cast_fp16")];
            string var_932_mode_0 = const()[name = string("op_932_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_932_cast_fp16 = gelu(mode = var_932_mode_0, x = var_930_cast_fp16_0)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_143_cast_fp16 = mul(x = var_932_cast_fp16, y = var_930_cast_fp16_1)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_7_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_7_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294016832)))];
            tensor<fp16, [1, 192, 1024]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_7_mlp_Wo_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_147_cast_fp16 = add(x = input_137_cast_fp16, y = linear_31_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<int32, [1]> hidden_states_17_axes_0 = const()[name = string("hidden_states_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_8_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_8_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299390848)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_17_cast_fp16 = layer_norm(axes = hidden_states_17_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_8_attn_norm_weight_to_fp16, x = input_147_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_8_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_8_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299392960)))];
            tensor<fp16, [1, 192, 3072]> linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_8_attn_Wqkv_weight_to_fp16, x = hidden_states_17_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<int32, [5]> var_951 = const()[name = string("op_951"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_35_cast_fp16 = reshape(shape = var_951, x = linear_32_cast_fp16)[name = string("qkv_35_cast_fp16")];
            tensor<int32, [3]> var_953_split_sizes_0 = const()[name = string("op_953_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_953_axis_0 = const()[name = string("op_953_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_953_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_953_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_953_cast_fp16_2 = split(axis = var_953_axis_0, split_sizes = var_953_split_sizes_0, x = qkv_35_cast_fp16)[name = string("op_953_cast_fp16")];
            tensor<int32, [1]> squeeze_24_axes_0 = const()[name = string("squeeze_24_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_24_cast_fp16 = squeeze(axes = squeeze_24_axes_0, x = var_953_cast_fp16_0)[name = string("squeeze_24_cast_fp16")];
            tensor<int32, [1]> squeeze_25_axes_0 = const()[name = string("squeeze_25_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_25_cast_fp16 = squeeze(axes = squeeze_25_axes_0, x = var_953_cast_fp16_1)[name = string("squeeze_25_cast_fp16")];
            tensor<int32, [1]> squeeze_26_axes_0 = const()[name = string("squeeze_26_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_26_cast_fp16 = squeeze(axes = squeeze_26_axes_0, x = var_953_cast_fp16_2)[name = string("squeeze_26_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = squeeze_24_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 192, 64]> var_963_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_11_to_fp16)[name = string("op_963_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_33_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_33_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_975_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_975_cast_fp16")];
            bool var_977_interleave_0 = const()[name = string("op_977_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_977_cast_fp16 = concat(axis = var_49, interleave = var_977_interleave_0, values = (var_975_cast_fp16, x1_33_cast_fp16))[name = string("op_977_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_978_cast_fp16 = mul(x = var_977_cast_fp16, y = sin_11_to_fp16)[name = string("op_978_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_17_cast_fp16 = add(x = var_963_cast_fp16, y = var_978_cast_fp16)[name = string("q_embed_17_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = squeeze_25_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 16, 192, 64]> var_981_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_11_to_fp16)[name = string("op_981_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_33_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_33_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_993_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_993_cast_fp16")];
            bool var_995_interleave_0 = const()[name = string("op_995_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_995_cast_fp16 = concat(axis = var_49, interleave = var_995_interleave_0, values = (var_993_cast_fp16, x1_35_cast_fp16))[name = string("op_995_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_996_cast_fp16 = mul(x = var_995_cast_fp16, y = sin_11_to_fp16)[name = string("op_996_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_17_cast_fp16 = add(x = var_981_cast_fp16, y = var_996_cast_fp16)[name = string("k_embed_17_cast_fp16")];
            bool var_1001_transpose_x_1 = const()[name = string("op_1001_transpose_x_1"), val = bool(false)];
            bool var_1001_transpose_y_1 = const()[name = string("op_1001_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1001_cast_fp16 = matmul(transpose_x = var_1001_transpose_x_1, transpose_y = var_1001_transpose_y_1, x = q_embed_17_cast_fp16, y = k_embed_17_cast_fp16)[name = string("op_1001_cast_fp16")];
            fp16 var_1002_to_fp16 = const()[name = string("op_1002_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_33_cast_fp16 = mul(x = var_1001_cast_fp16, y = var_1002_to_fp16)[name = string("attn_weights_33_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_149_cast_fp16 = add(x = attn_weights_33_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1005_cast_fp16 = softmax(axis = var_49, x = input_149_cast_fp16)[name = string("op_1005_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = squeeze_26_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = var_1005_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_1009_perm_0 = const()[name = string("op_1009_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1011 = const()[name = string("op_1011"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1009_cast_fp16 = transpose(perm = var_1009_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 192, 1024]> var_1012_cast_fp16 = reshape(shape = var_1011, x = var_1009_cast_fp16)[name = string("op_1012_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_8_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_8_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305684480)))];
            tensor<fp16, [1, 192, 1024]> linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_8_attn_Wo_weight_to_fp16, x = var_1012_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_155_cast_fp16 = add(x = input_147_cast_fp16, y = linear_33_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<int32, [1]> input_157_axes_0 = const()[name = string("input_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_8_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_8_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307781696)))];
            tensor<fp16, [1, 192, 1024]> input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_8_mlp_norm_weight_to_fp16, x = input_155_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_8_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_8_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307783808)))];
            tensor<fp16, [1, 192, 5248]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_8_mlp_Wi_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<int32, [2]> var_1024_split_sizes_0 = const()[name = string("op_1024_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1024_axis_0 = const()[name = string("op_1024_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1024_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1024_cast_fp16_1 = split(axis = var_1024_axis_0, split_sizes = var_1024_split_sizes_0, x = linear_34_cast_fp16)[name = string("op_1024_cast_fp16")];
            string var_1026_mode_0 = const()[name = string("op_1026_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1026_cast_fp16 = gelu(mode = var_1026_mode_0, x = var_1024_cast_fp16_0)[name = string("op_1026_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_161_cast_fp16 = mul(x = var_1026_cast_fp16, y = var_1024_cast_fp16_1)[name = string("input_161_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_8_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_8_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318531776)))];
            tensor<fp16, [1, 192, 1024]> linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_8_mlp_Wo_weight_to_fp16, x = input_161_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_165_cast_fp16 = add(x = input_155_cast_fp16, y = linear_35_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<int32, [1]> hidden_states_19_axes_0 = const()[name = string("hidden_states_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_9_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_9_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323905792)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_19_cast_fp16 = layer_norm(axes = hidden_states_19_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_9_attn_norm_weight_to_fp16, x = input_165_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_9_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_9_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323907904)))];
            tensor<fp16, [1, 192, 3072]> linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_9_attn_Wqkv_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [5]> var_1045 = const()[name = string("op_1045"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_39_cast_fp16 = reshape(shape = var_1045, x = linear_36_cast_fp16)[name = string("qkv_39_cast_fp16")];
            tensor<int32, [3]> var_1047_split_sizes_0 = const()[name = string("op_1047_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1047_axis_0 = const()[name = string("op_1047_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1047_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1047_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1047_cast_fp16_2 = split(axis = var_1047_axis_0, split_sizes = var_1047_split_sizes_0, x = qkv_39_cast_fp16)[name = string("op_1047_cast_fp16")];
            tensor<int32, [1]> squeeze_27_axes_0 = const()[name = string("squeeze_27_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_27_cast_fp16 = squeeze(axes = squeeze_27_axes_0, x = var_1047_cast_fp16_0)[name = string("squeeze_27_cast_fp16")];
            tensor<int32, [1]> squeeze_28_axes_0 = const()[name = string("squeeze_28_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_28_cast_fp16 = squeeze(axes = squeeze_28_axes_0, x = var_1047_cast_fp16_1)[name = string("squeeze_28_cast_fp16")];
            tensor<int32, [1]> squeeze_29_axes_0 = const()[name = string("squeeze_29_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_29_cast_fp16 = squeeze(axes = squeeze_29_axes_0, x = var_1047_cast_fp16_2)[name = string("squeeze_29_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = squeeze_27_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 192, 64]> var_1057_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_7_to_fp16)[name = string("op_1057_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_37_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_37_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1069_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_95_promoted_to_fp16)[name = string("op_1069_cast_fp16")];
            bool var_1071_interleave_0 = const()[name = string("op_1071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1071_cast_fp16 = concat(axis = var_49, interleave = var_1071_interleave_0, values = (var_1069_cast_fp16, x1_37_cast_fp16))[name = string("op_1071_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1072_cast_fp16 = mul(x = var_1071_cast_fp16, y = sin_7_to_fp16)[name = string("op_1072_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_19_cast_fp16 = add(x = var_1057_cast_fp16, y = var_1072_cast_fp16)[name = string("q_embed_19_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = squeeze_28_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 192, 64]> var_1075_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_7_to_fp16)[name = string("op_1075_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_37_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_37_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_98_promoted_to_fp16 = const()[name = string("const_98_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1087_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_98_promoted_to_fp16)[name = string("op_1087_cast_fp16")];
            bool var_1089_interleave_0 = const()[name = string("op_1089_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1089_cast_fp16 = concat(axis = var_49, interleave = var_1089_interleave_0, values = (var_1087_cast_fp16, x1_39_cast_fp16))[name = string("op_1089_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1090_cast_fp16 = mul(x = var_1089_cast_fp16, y = sin_7_to_fp16)[name = string("op_1090_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_19_cast_fp16 = add(x = var_1075_cast_fp16, y = var_1090_cast_fp16)[name = string("k_embed_19_cast_fp16")];
            bool var_1095_transpose_x_1 = const()[name = string("op_1095_transpose_x_1"), val = bool(false)];
            bool var_1095_transpose_y_1 = const()[name = string("op_1095_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1095_cast_fp16 = matmul(transpose_x = var_1095_transpose_x_1, transpose_y = var_1095_transpose_y_1, x = q_embed_19_cast_fp16, y = k_embed_19_cast_fp16)[name = string("op_1095_cast_fp16")];
            fp16 var_1096_to_fp16 = const()[name = string("op_1096_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_37_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_1096_to_fp16)[name = string("attn_weights_37_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_167_cast_fp16 = add(x = attn_weights_37_cast_fp16, y = key_bias_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1099_cast_fp16 = softmax(axis = var_49, x = input_167_cast_fp16)[name = string("op_1099_cast_fp16")];
            bool attn_output_55_transpose_x_0 = const()[name = string("attn_output_55_transpose_x_0"), val = bool(false)];
            bool attn_output_55_transpose_y_0 = const()[name = string("attn_output_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = squeeze_29_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_55_cast_fp16 = matmul(transpose_x = attn_output_55_transpose_x_0, transpose_y = attn_output_55_transpose_y_0, x = var_1099_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [4]> var_1103_perm_0 = const()[name = string("op_1103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1105 = const()[name = string("op_1105"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1103_cast_fp16 = transpose(perm = var_1103_perm_0, x = attn_output_55_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 192, 1024]> var_1106_cast_fp16 = reshape(shape = var_1105, x = var_1103_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_9_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_9_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330199424)))];
            tensor<fp16, [1, 192, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_9_attn_Wo_weight_to_fp16, x = var_1106_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_173_cast_fp16 = add(x = input_165_cast_fp16, y = linear_37_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<int32, [1]> input_175_axes_0 = const()[name = string("input_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_9_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_9_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332296640)))];
            tensor<fp16, [1, 192, 1024]> input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_9_mlp_norm_weight_to_fp16, x = input_173_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_9_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_9_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332298752)))];
            tensor<fp16, [1, 192, 5248]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_9_mlp_Wi_weight_to_fp16, x = input_175_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1118_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = linear_38_cast_fp16)[name = string("op_1118_cast_fp16")];
            string var_1120_mode_0 = const()[name = string("op_1120_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1120_cast_fp16 = gelu(mode = var_1120_mode_0, x = var_1118_cast_fp16_0)[name = string("op_1120_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_179_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1118_cast_fp16_1)[name = string("input_179_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_9_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_9_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343046720)))];
            tensor<fp16, [1, 192, 1024]> linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_9_mlp_Wo_weight_to_fp16, x = input_179_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_183_cast_fp16 = add(x = input_173_cast_fp16, y = linear_39_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<int32, [1]> hidden_states_21_axes_0 = const()[name = string("hidden_states_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_10_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_10_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348420736)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_21_cast_fp16 = layer_norm(axes = hidden_states_21_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_10_attn_norm_weight_to_fp16, x = input_183_cast_fp16)[name = string("hidden_states_21_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_10_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_10_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348422848)))];
            tensor<fp16, [1, 192, 3072]> linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_10_attn_Wqkv_weight_to_fp16, x = hidden_states_21_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<int32, [5]> var_1139 = const()[name = string("op_1139"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_43_cast_fp16 = reshape(shape = var_1139, x = linear_40_cast_fp16)[name = string("qkv_43_cast_fp16")];
            tensor<int32, [3]> var_1141_split_sizes_0 = const()[name = string("op_1141_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1141_axis_0 = const()[name = string("op_1141_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1141_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1141_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1141_cast_fp16_2 = split(axis = var_1141_axis_0, split_sizes = var_1141_split_sizes_0, x = qkv_43_cast_fp16)[name = string("op_1141_cast_fp16")];
            tensor<int32, [1]> squeeze_30_axes_0 = const()[name = string("squeeze_30_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_30_cast_fp16 = squeeze(axes = squeeze_30_axes_0, x = var_1141_cast_fp16_0)[name = string("squeeze_30_cast_fp16")];
            tensor<int32, [1]> squeeze_31_axes_0 = const()[name = string("squeeze_31_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_31_cast_fp16 = squeeze(axes = squeeze_31_axes_0, x = var_1141_cast_fp16_1)[name = string("squeeze_31_cast_fp16")];
            tensor<int32, [1]> squeeze_32_axes_0 = const()[name = string("squeeze_32_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_32_cast_fp16 = squeeze(axes = squeeze_32_axes_0, x = var_1141_cast_fp16_2)[name = string("squeeze_32_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = squeeze_30_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 16, 192, 64]> var_1151_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_11_to_fp16)[name = string("op_1151_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_41_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_41_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_103_promoted_to_fp16 = const()[name = string("const_103_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1163_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_103_promoted_to_fp16)[name = string("op_1163_cast_fp16")];
            bool var_1165_interleave_0 = const()[name = string("op_1165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1165_cast_fp16 = concat(axis = var_49, interleave = var_1165_interleave_0, values = (var_1163_cast_fp16, x1_41_cast_fp16))[name = string("op_1165_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1166_cast_fp16 = mul(x = var_1165_cast_fp16, y = sin_11_to_fp16)[name = string("op_1166_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_21_cast_fp16 = add(x = var_1151_cast_fp16, y = var_1166_cast_fp16)[name = string("q_embed_21_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = squeeze_31_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 192, 64]> var_1169_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_11_to_fp16)[name = string("op_1169_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_41_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_41_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_106_promoted_to_fp16 = const()[name = string("const_106_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1181_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_106_promoted_to_fp16)[name = string("op_1181_cast_fp16")];
            bool var_1183_interleave_0 = const()[name = string("op_1183_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1183_cast_fp16 = concat(axis = var_49, interleave = var_1183_interleave_0, values = (var_1181_cast_fp16, x1_43_cast_fp16))[name = string("op_1183_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1184_cast_fp16 = mul(x = var_1183_cast_fp16, y = sin_11_to_fp16)[name = string("op_1184_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_21_cast_fp16 = add(x = var_1169_cast_fp16, y = var_1184_cast_fp16)[name = string("k_embed_21_cast_fp16")];
            bool var_1189_transpose_x_1 = const()[name = string("op_1189_transpose_x_1"), val = bool(false)];
            bool var_1189_transpose_y_1 = const()[name = string("op_1189_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1189_cast_fp16 = matmul(transpose_x = var_1189_transpose_x_1, transpose_y = var_1189_transpose_y_1, x = q_embed_21_cast_fp16, y = k_embed_21_cast_fp16)[name = string("op_1189_cast_fp16")];
            fp16 var_1190_to_fp16 = const()[name = string("op_1190_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_41_cast_fp16 = mul(x = var_1189_cast_fp16, y = var_1190_to_fp16)[name = string("attn_weights_41_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_185_cast_fp16 = add(x = attn_weights_41_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1193_cast_fp16 = softmax(axis = var_49, x = input_185_cast_fp16)[name = string("op_1193_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = squeeze_32_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = var_1193_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_1197_perm_0 = const()[name = string("op_1197_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1199 = const()[name = string("op_1199"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1197_cast_fp16 = transpose(perm = var_1197_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 192, 1024]> var_1200_cast_fp16 = reshape(shape = var_1199, x = var_1197_cast_fp16)[name = string("op_1200_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_10_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_10_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354714368)))];
            tensor<fp16, [1, 192, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_10_attn_Wo_weight_to_fp16, x = var_1200_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_191_cast_fp16 = add(x = input_183_cast_fp16, y = linear_41_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<int32, [1]> input_193_axes_0 = const()[name = string("input_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_10_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_10_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356811584)))];
            tensor<fp16, [1, 192, 1024]> input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_10_mlp_norm_weight_to_fp16, x = input_191_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_10_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_10_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356813696)))];
            tensor<fp16, [1, 192, 5248]> linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_10_mlp_Wi_weight_to_fp16, x = input_193_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [2]> var_1212_split_sizes_0 = const()[name = string("op_1212_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1212_axis_0 = const()[name = string("op_1212_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1212_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1212_cast_fp16_1 = split(axis = var_1212_axis_0, split_sizes = var_1212_split_sizes_0, x = linear_42_cast_fp16)[name = string("op_1212_cast_fp16")];
            string var_1214_mode_0 = const()[name = string("op_1214_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1214_cast_fp16 = gelu(mode = var_1214_mode_0, x = var_1212_cast_fp16_0)[name = string("op_1214_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_197_cast_fp16 = mul(x = var_1214_cast_fp16, y = var_1212_cast_fp16_1)[name = string("input_197_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_10_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_10_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367561664)))];
            tensor<fp16, [1, 192, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_10_mlp_Wo_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_201_cast_fp16 = add(x = input_191_cast_fp16, y = linear_43_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<int32, [1]> hidden_states_23_axes_0 = const()[name = string("hidden_states_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_11_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_11_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372935680)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_23_cast_fp16 = layer_norm(axes = hidden_states_23_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_11_attn_norm_weight_to_fp16, x = input_201_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_11_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_11_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372937792)))];
            tensor<fp16, [1, 192, 3072]> linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_11_attn_Wqkv_weight_to_fp16, x = hidden_states_23_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [5]> var_1233 = const()[name = string("op_1233"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_47_cast_fp16 = reshape(shape = var_1233, x = linear_44_cast_fp16)[name = string("qkv_47_cast_fp16")];
            tensor<int32, [3]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1235_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1235_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1235_cast_fp16_2 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = qkv_47_cast_fp16)[name = string("op_1235_cast_fp16")];
            tensor<int32, [1]> squeeze_33_axes_0 = const()[name = string("squeeze_33_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_33_cast_fp16 = squeeze(axes = squeeze_33_axes_0, x = var_1235_cast_fp16_0)[name = string("squeeze_33_cast_fp16")];
            tensor<int32, [1]> squeeze_34_axes_0 = const()[name = string("squeeze_34_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_34_cast_fp16 = squeeze(axes = squeeze_34_axes_0, x = var_1235_cast_fp16_1)[name = string("squeeze_34_cast_fp16")];
            tensor<int32, [1]> squeeze_35_axes_0 = const()[name = string("squeeze_35_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_35_cast_fp16 = squeeze(axes = squeeze_35_axes_0, x = var_1235_cast_fp16_2)[name = string("squeeze_35_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = squeeze_33_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 16, 192, 64]> var_1245_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_11_to_fp16)[name = string("op_1245_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_45_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_45_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_111_promoted_to_fp16 = const()[name = string("const_111_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1257_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_111_promoted_to_fp16)[name = string("op_1257_cast_fp16")];
            bool var_1259_interleave_0 = const()[name = string("op_1259_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1259_cast_fp16 = concat(axis = var_49, interleave = var_1259_interleave_0, values = (var_1257_cast_fp16, x1_45_cast_fp16))[name = string("op_1259_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1260_cast_fp16 = mul(x = var_1259_cast_fp16, y = sin_11_to_fp16)[name = string("op_1260_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_23_cast_fp16 = add(x = var_1245_cast_fp16, y = var_1260_cast_fp16)[name = string("q_embed_23_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = squeeze_34_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 16, 192, 64]> var_1263_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_11_to_fp16)[name = string("op_1263_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_45_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_45_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1275_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_114_promoted_to_fp16)[name = string("op_1275_cast_fp16")];
            bool var_1277_interleave_0 = const()[name = string("op_1277_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1277_cast_fp16 = concat(axis = var_49, interleave = var_1277_interleave_0, values = (var_1275_cast_fp16, x1_47_cast_fp16))[name = string("op_1277_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1278_cast_fp16 = mul(x = var_1277_cast_fp16, y = sin_11_to_fp16)[name = string("op_1278_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_23_cast_fp16 = add(x = var_1263_cast_fp16, y = var_1278_cast_fp16)[name = string("k_embed_23_cast_fp16")];
            bool var_1283_transpose_x_1 = const()[name = string("op_1283_transpose_x_1"), val = bool(false)];
            bool var_1283_transpose_y_1 = const()[name = string("op_1283_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1283_cast_fp16 = matmul(transpose_x = var_1283_transpose_x_1, transpose_y = var_1283_transpose_y_1, x = q_embed_23_cast_fp16, y = k_embed_23_cast_fp16)[name = string("op_1283_cast_fp16")];
            fp16 var_1284_to_fp16 = const()[name = string("op_1284_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_45_cast_fp16 = mul(x = var_1283_cast_fp16, y = var_1284_to_fp16)[name = string("attn_weights_45_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_203_cast_fp16 = add(x = attn_weights_45_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1287_cast_fp16 = softmax(axis = var_49, x = input_203_cast_fp16)[name = string("op_1287_cast_fp16")];
            bool attn_output_67_transpose_x_0 = const()[name = string("attn_output_67_transpose_x_0"), val = bool(false)];
            bool attn_output_67_transpose_y_0 = const()[name = string("attn_output_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = squeeze_35_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_67_cast_fp16 = matmul(transpose_x = attn_output_67_transpose_x_0, transpose_y = attn_output_67_transpose_y_0, x = var_1287_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_1291_perm_0 = const()[name = string("op_1291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1293 = const()[name = string("op_1293"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1291_cast_fp16 = transpose(perm = var_1291_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 192, 1024]> var_1294_cast_fp16 = reshape(shape = var_1293, x = var_1291_cast_fp16)[name = string("op_1294_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_11_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_11_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379229312)))];
            tensor<fp16, [1, 192, 1024]> linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_11_attn_Wo_weight_to_fp16, x = var_1294_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_209_cast_fp16 = add(x = input_201_cast_fp16, y = linear_45_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_11_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_11_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381326528)))];
            tensor<fp16, [1, 192, 1024]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_11_mlp_norm_weight_to_fp16, x = input_209_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_11_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_11_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381328640)))];
            tensor<fp16, [1, 192, 5248]> linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_11_mlp_Wi_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [2]> var_1306_split_sizes_0 = const()[name = string("op_1306_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1306_axis_0 = const()[name = string("op_1306_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1306_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1306_cast_fp16_1 = split(axis = var_1306_axis_0, split_sizes = var_1306_split_sizes_0, x = linear_46_cast_fp16)[name = string("op_1306_cast_fp16")];
            string var_1308_mode_0 = const()[name = string("op_1308_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1308_cast_fp16 = gelu(mode = var_1308_mode_0, x = var_1306_cast_fp16_0)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_215_cast_fp16 = mul(x = var_1308_cast_fp16, y = var_1306_cast_fp16_1)[name = string("input_215_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_11_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_11_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392076608)))];
            tensor<fp16, [1, 192, 1024]> linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_11_mlp_Wo_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_219_cast_fp16 = add(x = input_209_cast_fp16, y = linear_47_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<int32, [1]> hidden_states_25_axes_0 = const()[name = string("hidden_states_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_12_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_12_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397450624)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_25_cast_fp16 = layer_norm(axes = hidden_states_25_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_12_attn_norm_weight_to_fp16, x = input_219_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_12_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_12_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397452736)))];
            tensor<fp16, [1, 192, 3072]> linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_12_attn_Wqkv_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [5]> var_1327 = const()[name = string("op_1327"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_51_cast_fp16 = reshape(shape = var_1327, x = linear_48_cast_fp16)[name = string("qkv_51_cast_fp16")];
            tensor<int32, [3]> var_1329_split_sizes_0 = const()[name = string("op_1329_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1329_axis_0 = const()[name = string("op_1329_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1329_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1329_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1329_cast_fp16_2 = split(axis = var_1329_axis_0, split_sizes = var_1329_split_sizes_0, x = qkv_51_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<int32, [1]> squeeze_36_axes_0 = const()[name = string("squeeze_36_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_36_cast_fp16 = squeeze(axes = squeeze_36_axes_0, x = var_1329_cast_fp16_0)[name = string("squeeze_36_cast_fp16")];
            tensor<int32, [1]> squeeze_37_axes_0 = const()[name = string("squeeze_37_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_37_cast_fp16 = squeeze(axes = squeeze_37_axes_0, x = var_1329_cast_fp16_1)[name = string("squeeze_37_cast_fp16")];
            tensor<int32, [1]> squeeze_38_axes_0 = const()[name = string("squeeze_38_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_38_cast_fp16 = squeeze(axes = squeeze_38_axes_0, x = var_1329_cast_fp16_2)[name = string("squeeze_38_cast_fp16")];
            tensor<int32, [4]> q_49_perm_0 = const()[name = string("q_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_49_perm_0 = const()[name = string("k_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_49_cast_fp16 = transpose(perm = q_49_perm_0, x = squeeze_36_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 16, 192, 64]> var_1339_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_7_to_fp16)[name = string("op_1339_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_49_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_49_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_119_promoted_to_fp16 = const()[name = string("const_119_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1351_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_119_promoted_to_fp16)[name = string("op_1351_cast_fp16")];
            bool var_1353_interleave_0 = const()[name = string("op_1353_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1353_cast_fp16 = concat(axis = var_49, interleave = var_1353_interleave_0, values = (var_1351_cast_fp16, x1_49_cast_fp16))[name = string("op_1353_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1354_cast_fp16 = mul(x = var_1353_cast_fp16, y = sin_7_to_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_25_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1354_cast_fp16)[name = string("q_embed_25_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_49_cast_fp16 = transpose(perm = k_49_perm_0, x = squeeze_37_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 16, 192, 64]> var_1357_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_7_to_fp16)[name = string("op_1357_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_49_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_49_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1369_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_1369_cast_fp16")];
            bool var_1371_interleave_0 = const()[name = string("op_1371_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1371_cast_fp16 = concat(axis = var_49, interleave = var_1371_interleave_0, values = (var_1369_cast_fp16, x1_51_cast_fp16))[name = string("op_1371_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1372_cast_fp16 = mul(x = var_1371_cast_fp16, y = sin_7_to_fp16)[name = string("op_1372_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_25_cast_fp16 = add(x = var_1357_cast_fp16, y = var_1372_cast_fp16)[name = string("k_embed_25_cast_fp16")];
            bool var_1377_transpose_x_1 = const()[name = string("op_1377_transpose_x_1"), val = bool(false)];
            bool var_1377_transpose_y_1 = const()[name = string("op_1377_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1377_cast_fp16 = matmul(transpose_x = var_1377_transpose_x_1, transpose_y = var_1377_transpose_y_1, x = q_embed_25_cast_fp16, y = k_embed_25_cast_fp16)[name = string("op_1377_cast_fp16")];
            fp16 var_1378_to_fp16 = const()[name = string("op_1378_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_49_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1378_to_fp16)[name = string("attn_weights_49_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_221_cast_fp16 = add(x = attn_weights_49_cast_fp16, y = key_bias_cast_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1381_cast_fp16 = softmax(axis = var_49, x = input_221_cast_fp16)[name = string("op_1381_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = squeeze_38_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = var_1381_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_1385_perm_0 = const()[name = string("op_1385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1387 = const()[name = string("op_1387"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1385_cast_fp16 = transpose(perm = var_1385_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 192, 1024]> var_1388_cast_fp16 = reshape(shape = var_1387, x = var_1385_cast_fp16)[name = string("op_1388_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_12_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_12_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403744256)))];
            tensor<fp16, [1, 192, 1024]> linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_12_attn_Wo_weight_to_fp16, x = var_1388_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_227_cast_fp16 = add(x = input_219_cast_fp16, y = linear_49_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<int32, [1]> input_229_axes_0 = const()[name = string("input_229_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_12_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_12_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405841472)))];
            tensor<fp16, [1, 192, 1024]> input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_12_mlp_norm_weight_to_fp16, x = input_227_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_12_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_12_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405843584)))];
            tensor<fp16, [1, 192, 5248]> linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_12_mlp_Wi_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [2]> var_1400_split_sizes_0 = const()[name = string("op_1400_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1400_axis_0 = const()[name = string("op_1400_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1400_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1400_cast_fp16_1 = split(axis = var_1400_axis_0, split_sizes = var_1400_split_sizes_0, x = linear_50_cast_fp16)[name = string("op_1400_cast_fp16")];
            string var_1402_mode_0 = const()[name = string("op_1402_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1402_cast_fp16 = gelu(mode = var_1402_mode_0, x = var_1400_cast_fp16_0)[name = string("op_1402_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_233_cast_fp16 = mul(x = var_1402_cast_fp16, y = var_1400_cast_fp16_1)[name = string("input_233_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_12_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_12_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416591552)))];
            tensor<fp16, [1, 192, 1024]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_12_mlp_Wo_weight_to_fp16, x = input_233_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_237_cast_fp16 = add(x = input_227_cast_fp16, y = linear_51_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<int32, [1]> hidden_states_27_axes_0 = const()[name = string("hidden_states_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_13_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_13_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421965568)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_27_cast_fp16 = layer_norm(axes = hidden_states_27_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_13_attn_norm_weight_to_fp16, x = input_237_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_13_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_13_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421967680)))];
            tensor<fp16, [1, 192, 3072]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_13_attn_Wqkv_weight_to_fp16, x = hidden_states_27_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<int32, [5]> var_1421 = const()[name = string("op_1421"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_55_cast_fp16 = reshape(shape = var_1421, x = linear_52_cast_fp16)[name = string("qkv_55_cast_fp16")];
            tensor<int32, [3]> var_1423_split_sizes_0 = const()[name = string("op_1423_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1423_axis_0 = const()[name = string("op_1423_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1423_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1423_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1423_cast_fp16_2 = split(axis = var_1423_axis_0, split_sizes = var_1423_split_sizes_0, x = qkv_55_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [1]> squeeze_39_axes_0 = const()[name = string("squeeze_39_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_39_cast_fp16 = squeeze(axes = squeeze_39_axes_0, x = var_1423_cast_fp16_0)[name = string("squeeze_39_cast_fp16")];
            tensor<int32, [1]> squeeze_40_axes_0 = const()[name = string("squeeze_40_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_40_cast_fp16 = squeeze(axes = squeeze_40_axes_0, x = var_1423_cast_fp16_1)[name = string("squeeze_40_cast_fp16")];
            tensor<int32, [1]> squeeze_41_axes_0 = const()[name = string("squeeze_41_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_41_cast_fp16 = squeeze(axes = squeeze_41_axes_0, x = var_1423_cast_fp16_2)[name = string("squeeze_41_cast_fp16")];
            tensor<int32, [4]> q_53_perm_0 = const()[name = string("q_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_53_perm_0 = const()[name = string("k_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_53_cast_fp16 = transpose(perm = q_53_perm_0, x = squeeze_39_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 192, 64]> var_1433_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_11_to_fp16)[name = string("op_1433_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_53_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_53_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_127_promoted_to_fp16 = const()[name = string("const_127_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1445_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_127_promoted_to_fp16)[name = string("op_1445_cast_fp16")];
            bool var_1447_interleave_0 = const()[name = string("op_1447_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1447_cast_fp16 = concat(axis = var_49, interleave = var_1447_interleave_0, values = (var_1445_cast_fp16, x1_53_cast_fp16))[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1448_cast_fp16 = mul(x = var_1447_cast_fp16, y = sin_11_to_fp16)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_27_cast_fp16 = add(x = var_1433_cast_fp16, y = var_1448_cast_fp16)[name = string("q_embed_27_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_53_cast_fp16 = transpose(perm = k_53_perm_0, x = squeeze_40_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 16, 192, 64]> var_1451_cast_fp16 = mul(x = k_53_cast_fp16, y = cos_11_to_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_53_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_53_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1463_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_1463_cast_fp16")];
            bool var_1465_interleave_0 = const()[name = string("op_1465_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1465_cast_fp16 = concat(axis = var_49, interleave = var_1465_interleave_0, values = (var_1463_cast_fp16, x1_55_cast_fp16))[name = string("op_1465_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1466_cast_fp16 = mul(x = var_1465_cast_fp16, y = sin_11_to_fp16)[name = string("op_1466_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_27_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1466_cast_fp16)[name = string("k_embed_27_cast_fp16")];
            bool var_1471_transpose_x_1 = const()[name = string("op_1471_transpose_x_1"), val = bool(false)];
            bool var_1471_transpose_y_1 = const()[name = string("op_1471_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1471_cast_fp16 = matmul(transpose_x = var_1471_transpose_x_1, transpose_y = var_1471_transpose_y_1, x = q_embed_27_cast_fp16, y = k_embed_27_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp16 var_1472_to_fp16 = const()[name = string("op_1472_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_53_cast_fp16 = mul(x = var_1471_cast_fp16, y = var_1472_to_fp16)[name = string("attn_weights_53_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_239_cast_fp16 = add(x = attn_weights_53_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1475_cast_fp16 = softmax(axis = var_49, x = input_239_cast_fp16)[name = string("op_1475_cast_fp16")];
            bool attn_output_79_transpose_x_0 = const()[name = string("attn_output_79_transpose_x_0"), val = bool(false)];
            bool attn_output_79_transpose_y_0 = const()[name = string("attn_output_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = squeeze_41_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_79_cast_fp16 = matmul(transpose_x = attn_output_79_transpose_x_0, transpose_y = attn_output_79_transpose_y_0, x = var_1475_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [4]> var_1479_perm_0 = const()[name = string("op_1479_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1481 = const()[name = string("op_1481"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1479_cast_fp16 = transpose(perm = var_1479_perm_0, x = attn_output_79_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 192, 1024]> var_1482_cast_fp16 = reshape(shape = var_1481, x = var_1479_cast_fp16)[name = string("op_1482_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_13_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_13_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428259200)))];
            tensor<fp16, [1, 192, 1024]> linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_13_attn_Wo_weight_to_fp16, x = var_1482_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_245_cast_fp16 = add(x = input_237_cast_fp16, y = linear_53_cast_fp16)[name = string("input_245_cast_fp16")];
            tensor<int32, [1]> input_247_axes_0 = const()[name = string("input_247_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_13_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_13_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430356416)))];
            tensor<fp16, [1, 192, 1024]> input_247_cast_fp16 = layer_norm(axes = input_247_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_13_mlp_norm_weight_to_fp16, x = input_245_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_13_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_13_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430358528)))];
            tensor<fp16, [1, 192, 5248]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_13_mlp_Wi_weight_to_fp16, x = input_247_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<int32, [2]> var_1494_split_sizes_0 = const()[name = string("op_1494_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1494_axis_0 = const()[name = string("op_1494_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1494_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1494_cast_fp16_1 = split(axis = var_1494_axis_0, split_sizes = var_1494_split_sizes_0, x = linear_54_cast_fp16)[name = string("op_1494_cast_fp16")];
            string var_1496_mode_0 = const()[name = string("op_1496_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1496_cast_fp16 = gelu(mode = var_1496_mode_0, x = var_1494_cast_fp16_0)[name = string("op_1496_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_251_cast_fp16 = mul(x = var_1496_cast_fp16, y = var_1494_cast_fp16_1)[name = string("input_251_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_13_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_13_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441106496)))];
            tensor<fp16, [1, 192, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_13_mlp_Wo_weight_to_fp16, x = input_251_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_55_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<int32, [1]> hidden_states_29_axes_0 = const()[name = string("hidden_states_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_14_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_14_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446480512)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_29_cast_fp16 = layer_norm(axes = hidden_states_29_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_14_attn_norm_weight_to_fp16, x = input_255_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_14_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_14_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446482624)))];
            tensor<fp16, [1, 192, 3072]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_14_attn_Wqkv_weight_to_fp16, x = hidden_states_29_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [5]> var_1515 = const()[name = string("op_1515"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_59_cast_fp16 = reshape(shape = var_1515, x = linear_56_cast_fp16)[name = string("qkv_59_cast_fp16")];
            tensor<int32, [3]> var_1517_split_sizes_0 = const()[name = string("op_1517_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1517_axis_0 = const()[name = string("op_1517_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1517_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1517_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1517_cast_fp16_2 = split(axis = var_1517_axis_0, split_sizes = var_1517_split_sizes_0, x = qkv_59_cast_fp16)[name = string("op_1517_cast_fp16")];
            tensor<int32, [1]> squeeze_42_axes_0 = const()[name = string("squeeze_42_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_42_cast_fp16 = squeeze(axes = squeeze_42_axes_0, x = var_1517_cast_fp16_0)[name = string("squeeze_42_cast_fp16")];
            tensor<int32, [1]> squeeze_43_axes_0 = const()[name = string("squeeze_43_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_43_cast_fp16 = squeeze(axes = squeeze_43_axes_0, x = var_1517_cast_fp16_1)[name = string("squeeze_43_cast_fp16")];
            tensor<int32, [1]> squeeze_44_axes_0 = const()[name = string("squeeze_44_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_44_cast_fp16 = squeeze(axes = squeeze_44_axes_0, x = var_1517_cast_fp16_2)[name = string("squeeze_44_cast_fp16")];
            tensor<int32, [4]> q_57_perm_0 = const()[name = string("q_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_57_perm_0 = const()[name = string("k_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_57_cast_fp16 = transpose(perm = q_57_perm_0, x = squeeze_42_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 192, 64]> var_1527_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_11_to_fp16)[name = string("op_1527_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_57_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_57_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1539_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_135_promoted_to_fp16)[name = string("op_1539_cast_fp16")];
            bool var_1541_interleave_0 = const()[name = string("op_1541_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1541_cast_fp16 = concat(axis = var_49, interleave = var_1541_interleave_0, values = (var_1539_cast_fp16, x1_57_cast_fp16))[name = string("op_1541_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1542_cast_fp16 = mul(x = var_1541_cast_fp16, y = sin_11_to_fp16)[name = string("op_1542_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_29_cast_fp16 = add(x = var_1527_cast_fp16, y = var_1542_cast_fp16)[name = string("q_embed_29_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_57_cast_fp16 = transpose(perm = k_57_perm_0, x = squeeze_43_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 192, 64]> var_1545_cast_fp16 = mul(x = k_57_cast_fp16, y = cos_11_to_fp16)[name = string("op_1545_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_57_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_57_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_138_promoted_to_fp16 = const()[name = string("const_138_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1557_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_138_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            bool var_1559_interleave_0 = const()[name = string("op_1559_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1559_cast_fp16 = concat(axis = var_49, interleave = var_1559_interleave_0, values = (var_1557_cast_fp16, x1_59_cast_fp16))[name = string("op_1559_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1560_cast_fp16 = mul(x = var_1559_cast_fp16, y = sin_11_to_fp16)[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_29_cast_fp16 = add(x = var_1545_cast_fp16, y = var_1560_cast_fp16)[name = string("k_embed_29_cast_fp16")];
            bool var_1565_transpose_x_1 = const()[name = string("op_1565_transpose_x_1"), val = bool(false)];
            bool var_1565_transpose_y_1 = const()[name = string("op_1565_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1565_cast_fp16 = matmul(transpose_x = var_1565_transpose_x_1, transpose_y = var_1565_transpose_y_1, x = q_embed_29_cast_fp16, y = k_embed_29_cast_fp16)[name = string("op_1565_cast_fp16")];
            fp16 var_1566_to_fp16 = const()[name = string("op_1566_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_57_cast_fp16 = mul(x = var_1565_cast_fp16, y = var_1566_to_fp16)[name = string("attn_weights_57_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_257_cast_fp16 = add(x = attn_weights_57_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1569_cast_fp16 = softmax(axis = var_49, x = input_257_cast_fp16)[name = string("op_1569_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = squeeze_44_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = var_1569_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_1573_perm_0 = const()[name = string("op_1573_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1575 = const()[name = string("op_1575"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1573_cast_fp16 = transpose(perm = var_1573_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 192, 1024]> var_1576_cast_fp16 = reshape(shape = var_1575, x = var_1573_cast_fp16)[name = string("op_1576_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_14_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_14_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452774144)))];
            tensor<fp16, [1, 192, 1024]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_14_attn_Wo_weight_to_fp16, x = var_1576_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_263_cast_fp16 = add(x = input_255_cast_fp16, y = linear_57_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<int32, [1]> input_265_axes_0 = const()[name = string("input_265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_14_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_14_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454871360)))];
            tensor<fp16, [1, 192, 1024]> input_265_cast_fp16 = layer_norm(axes = input_265_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_14_mlp_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_14_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_14_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454873472)))];
            tensor<fp16, [1, 192, 5248]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_14_mlp_Wi_weight_to_fp16, x = input_265_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [2]> var_1588_split_sizes_0 = const()[name = string("op_1588_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1588_axis_0 = const()[name = string("op_1588_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1588_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1588_cast_fp16_1 = split(axis = var_1588_axis_0, split_sizes = var_1588_split_sizes_0, x = linear_58_cast_fp16)[name = string("op_1588_cast_fp16")];
            string var_1590_mode_0 = const()[name = string("op_1590_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1590_cast_fp16 = gelu(mode = var_1590_mode_0, x = var_1588_cast_fp16_0)[name = string("op_1590_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_269_cast_fp16 = mul(x = var_1590_cast_fp16, y = var_1588_cast_fp16_1)[name = string("input_269_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_14_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_14_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465621440)))];
            tensor<fp16, [1, 192, 1024]> linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_14_mlp_Wo_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_273_cast_fp16 = add(x = input_263_cast_fp16, y = linear_59_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<int32, [1]> hidden_states_31_axes_0 = const()[name = string("hidden_states_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_15_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_15_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470995456)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_31_cast_fp16 = layer_norm(axes = hidden_states_31_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_15_attn_norm_weight_to_fp16, x = input_273_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_15_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_15_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470997568)))];
            tensor<fp16, [1, 192, 3072]> linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_15_attn_Wqkv_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<int32, [5]> var_1609 = const()[name = string("op_1609"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_63_cast_fp16 = reshape(shape = var_1609, x = linear_60_cast_fp16)[name = string("qkv_63_cast_fp16")];
            tensor<int32, [3]> var_1611_split_sizes_0 = const()[name = string("op_1611_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1611_axis_0 = const()[name = string("op_1611_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1611_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1611_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1611_cast_fp16_2 = split(axis = var_1611_axis_0, split_sizes = var_1611_split_sizes_0, x = qkv_63_cast_fp16)[name = string("op_1611_cast_fp16")];
            tensor<int32, [1]> squeeze_45_axes_0 = const()[name = string("squeeze_45_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_45_cast_fp16 = squeeze(axes = squeeze_45_axes_0, x = var_1611_cast_fp16_0)[name = string("squeeze_45_cast_fp16")];
            tensor<int32, [1]> squeeze_46_axes_0 = const()[name = string("squeeze_46_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_46_cast_fp16 = squeeze(axes = squeeze_46_axes_0, x = var_1611_cast_fp16_1)[name = string("squeeze_46_cast_fp16")];
            tensor<int32, [1]> squeeze_47_axes_0 = const()[name = string("squeeze_47_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_47_cast_fp16 = squeeze(axes = squeeze_47_axes_0, x = var_1611_cast_fp16_2)[name = string("squeeze_47_cast_fp16")];
            tensor<int32, [4]> q_61_perm_0 = const()[name = string("q_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_61_perm_0 = const()[name = string("k_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_61_cast_fp16 = transpose(perm = q_61_perm_0, x = squeeze_45_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 16, 192, 64]> var_1621_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_7_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_61_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_61_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_143_promoted_to_fp16 = const()[name = string("const_143_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1633_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_143_promoted_to_fp16)[name = string("op_1633_cast_fp16")];
            bool var_1635_interleave_0 = const()[name = string("op_1635_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1635_cast_fp16 = concat(axis = var_49, interleave = var_1635_interleave_0, values = (var_1633_cast_fp16, x1_61_cast_fp16))[name = string("op_1635_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1636_cast_fp16 = mul(x = var_1635_cast_fp16, y = sin_7_to_fp16)[name = string("op_1636_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_31_cast_fp16 = add(x = var_1621_cast_fp16, y = var_1636_cast_fp16)[name = string("q_embed_31_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_61_cast_fp16 = transpose(perm = k_61_perm_0, x = squeeze_46_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 192, 64]> var_1639_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_7_to_fp16)[name = string("op_1639_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_61_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_61_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_146_promoted_to_fp16 = const()[name = string("const_146_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1651_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_146_promoted_to_fp16)[name = string("op_1651_cast_fp16")];
            bool var_1653_interleave_0 = const()[name = string("op_1653_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1653_cast_fp16 = concat(axis = var_49, interleave = var_1653_interleave_0, values = (var_1651_cast_fp16, x1_63_cast_fp16))[name = string("op_1653_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1654_cast_fp16 = mul(x = var_1653_cast_fp16, y = sin_7_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_31_cast_fp16 = add(x = var_1639_cast_fp16, y = var_1654_cast_fp16)[name = string("k_embed_31_cast_fp16")];
            bool var_1659_transpose_x_1 = const()[name = string("op_1659_transpose_x_1"), val = bool(false)];
            bool var_1659_transpose_y_1 = const()[name = string("op_1659_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1659_cast_fp16 = matmul(transpose_x = var_1659_transpose_x_1, transpose_y = var_1659_transpose_y_1, x = q_embed_31_cast_fp16, y = k_embed_31_cast_fp16)[name = string("op_1659_cast_fp16")];
            fp16 var_1660_to_fp16 = const()[name = string("op_1660_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_61_cast_fp16 = mul(x = var_1659_cast_fp16, y = var_1660_to_fp16)[name = string("attn_weights_61_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_275_cast_fp16 = add(x = attn_weights_61_cast_fp16, y = key_bias_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1663_cast_fp16 = softmax(axis = var_49, x = input_275_cast_fp16)[name = string("op_1663_cast_fp16")];
            bool attn_output_91_transpose_x_0 = const()[name = string("attn_output_91_transpose_x_0"), val = bool(false)];
            bool attn_output_91_transpose_y_0 = const()[name = string("attn_output_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = squeeze_47_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_91_cast_fp16 = matmul(transpose_x = attn_output_91_transpose_x_0, transpose_y = attn_output_91_transpose_y_0, x = var_1663_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_1667_perm_0 = const()[name = string("op_1667_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1669 = const()[name = string("op_1669"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1667_cast_fp16 = transpose(perm = var_1667_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 192, 1024]> var_1670_cast_fp16 = reshape(shape = var_1669, x = var_1667_cast_fp16)[name = string("op_1670_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_15_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_15_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477289088)))];
            tensor<fp16, [1, 192, 1024]> linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_15_attn_Wo_weight_to_fp16, x = var_1670_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_281_cast_fp16 = add(x = input_273_cast_fp16, y = linear_61_cast_fp16)[name = string("input_281_cast_fp16")];
            tensor<int32, [1]> input_283_axes_0 = const()[name = string("input_283_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_15_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_15_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(479386304)))];
            tensor<fp16, [1, 192, 1024]> input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_15_mlp_norm_weight_to_fp16, x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_15_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_15_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(479388416)))];
            tensor<fp16, [1, 192, 5248]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_15_mlp_Wi_weight_to_fp16, x = input_283_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<int32, [2]> var_1682_split_sizes_0 = const()[name = string("op_1682_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1682_axis_0 = const()[name = string("op_1682_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1682_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1682_cast_fp16_1 = split(axis = var_1682_axis_0, split_sizes = var_1682_split_sizes_0, x = linear_62_cast_fp16)[name = string("op_1682_cast_fp16")];
            string var_1684_mode_0 = const()[name = string("op_1684_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1684_cast_fp16 = gelu(mode = var_1684_mode_0, x = var_1682_cast_fp16_0)[name = string("op_1684_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_287_cast_fp16 = mul(x = var_1684_cast_fp16, y = var_1682_cast_fp16_1)[name = string("input_287_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_15_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_15_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490136384)))];
            tensor<fp16, [1, 192, 1024]> linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_15_mlp_Wo_weight_to_fp16, x = input_287_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_291_cast_fp16 = add(x = input_281_cast_fp16, y = linear_63_cast_fp16)[name = string("input_291_cast_fp16")];
            tensor<int32, [1]> hidden_states_33_axes_0 = const()[name = string("hidden_states_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_16_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_16_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495510400)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_33_cast_fp16 = layer_norm(axes = hidden_states_33_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_16_attn_norm_weight_to_fp16, x = input_291_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_16_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_16_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495512512)))];
            tensor<fp16, [1, 192, 3072]> linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_16_attn_Wqkv_weight_to_fp16, x = hidden_states_33_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [5]> var_1703 = const()[name = string("op_1703"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_67_cast_fp16 = reshape(shape = var_1703, x = linear_64_cast_fp16)[name = string("qkv_67_cast_fp16")];
            tensor<int32, [3]> var_1705_split_sizes_0 = const()[name = string("op_1705_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1705_axis_0 = const()[name = string("op_1705_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1705_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1705_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1705_cast_fp16_2 = split(axis = var_1705_axis_0, split_sizes = var_1705_split_sizes_0, x = qkv_67_cast_fp16)[name = string("op_1705_cast_fp16")];
            tensor<int32, [1]> squeeze_48_axes_0 = const()[name = string("squeeze_48_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_48_cast_fp16 = squeeze(axes = squeeze_48_axes_0, x = var_1705_cast_fp16_0)[name = string("squeeze_48_cast_fp16")];
            tensor<int32, [1]> squeeze_49_axes_0 = const()[name = string("squeeze_49_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_49_cast_fp16 = squeeze(axes = squeeze_49_axes_0, x = var_1705_cast_fp16_1)[name = string("squeeze_49_cast_fp16")];
            tensor<int32, [1]> squeeze_50_axes_0 = const()[name = string("squeeze_50_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_50_cast_fp16 = squeeze(axes = squeeze_50_axes_0, x = var_1705_cast_fp16_2)[name = string("squeeze_50_cast_fp16")];
            tensor<int32, [4]> q_65_perm_0 = const()[name = string("q_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_65_perm_0 = const()[name = string("k_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_65_cast_fp16 = transpose(perm = q_65_perm_0, x = squeeze_48_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 16, 192, 64]> var_1715_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_11_to_fp16)[name = string("op_1715_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_65_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_65_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_151_promoted_to_fp16 = const()[name = string("const_151_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1727_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_151_promoted_to_fp16)[name = string("op_1727_cast_fp16")];
            bool var_1729_interleave_0 = const()[name = string("op_1729_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1729_cast_fp16 = concat(axis = var_49, interleave = var_1729_interleave_0, values = (var_1727_cast_fp16, x1_65_cast_fp16))[name = string("op_1729_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1730_cast_fp16 = mul(x = var_1729_cast_fp16, y = sin_11_to_fp16)[name = string("op_1730_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_33_cast_fp16 = add(x = var_1715_cast_fp16, y = var_1730_cast_fp16)[name = string("q_embed_33_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_65_cast_fp16 = transpose(perm = k_65_perm_0, x = squeeze_49_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 16, 192, 64]> var_1733_cast_fp16 = mul(x = k_65_cast_fp16, y = cos_11_to_fp16)[name = string("op_1733_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_65_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_65_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_154_promoted_to_fp16 = const()[name = string("const_154_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1745_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_154_promoted_to_fp16)[name = string("op_1745_cast_fp16")];
            bool var_1747_interleave_0 = const()[name = string("op_1747_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1747_cast_fp16 = concat(axis = var_49, interleave = var_1747_interleave_0, values = (var_1745_cast_fp16, x1_67_cast_fp16))[name = string("op_1747_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1748_cast_fp16 = mul(x = var_1747_cast_fp16, y = sin_11_to_fp16)[name = string("op_1748_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_33_cast_fp16 = add(x = var_1733_cast_fp16, y = var_1748_cast_fp16)[name = string("k_embed_33_cast_fp16")];
            bool var_1753_transpose_x_1 = const()[name = string("op_1753_transpose_x_1"), val = bool(false)];
            bool var_1753_transpose_y_1 = const()[name = string("op_1753_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1753_cast_fp16 = matmul(transpose_x = var_1753_transpose_x_1, transpose_y = var_1753_transpose_y_1, x = q_embed_33_cast_fp16, y = k_embed_33_cast_fp16)[name = string("op_1753_cast_fp16")];
            fp16 var_1754_to_fp16 = const()[name = string("op_1754_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_65_cast_fp16 = mul(x = var_1753_cast_fp16, y = var_1754_to_fp16)[name = string("attn_weights_65_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_293_cast_fp16 = add(x = attn_weights_65_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_293_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1757_cast_fp16 = softmax(axis = var_49, x = input_293_cast_fp16)[name = string("op_1757_cast_fp16")];
            bool attn_output_97_transpose_x_0 = const()[name = string("attn_output_97_transpose_x_0"), val = bool(false)];
            bool attn_output_97_transpose_y_0 = const()[name = string("attn_output_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = squeeze_50_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_0, transpose_y = attn_output_97_transpose_y_0, x = var_1757_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_97_cast_fp16")];
            tensor<int32, [4]> var_1761_perm_0 = const()[name = string("op_1761_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1763 = const()[name = string("op_1763"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1761_cast_fp16 = transpose(perm = var_1761_perm_0, x = attn_output_97_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 192, 1024]> var_1764_cast_fp16 = reshape(shape = var_1763, x = var_1761_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_16_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_16_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501804032)))];
            tensor<fp16, [1, 192, 1024]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_16_attn_Wo_weight_to_fp16, x = var_1764_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_299_cast_fp16 = add(x = input_291_cast_fp16, y = linear_65_cast_fp16)[name = string("input_299_cast_fp16")];
            tensor<int32, [1]> input_301_axes_0 = const()[name = string("input_301_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_16_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_16_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503901248)))];
            tensor<fp16, [1, 192, 1024]> input_301_cast_fp16 = layer_norm(axes = input_301_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_16_mlp_norm_weight_to_fp16, x = input_299_cast_fp16)[name = string("input_301_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_16_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_16_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503903360)))];
            tensor<fp16, [1, 192, 5248]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_16_mlp_Wi_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [2]> var_1776_split_sizes_0 = const()[name = string("op_1776_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1776_axis_0 = const()[name = string("op_1776_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1776_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1776_cast_fp16_1 = split(axis = var_1776_axis_0, split_sizes = var_1776_split_sizes_0, x = linear_66_cast_fp16)[name = string("op_1776_cast_fp16")];
            string var_1778_mode_0 = const()[name = string("op_1778_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1778_cast_fp16 = gelu(mode = var_1778_mode_0, x = var_1776_cast_fp16_0)[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_305_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_1776_cast_fp16_1)[name = string("input_305_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_16_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_16_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514651328)))];
            tensor<fp16, [1, 192, 1024]> linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_16_mlp_Wo_weight_to_fp16, x = input_305_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_309_cast_fp16 = add(x = input_299_cast_fp16, y = linear_67_cast_fp16)[name = string("input_309_cast_fp16")];
            tensor<int32, [1]> hidden_states_35_axes_0 = const()[name = string("hidden_states_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_17_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_17_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520025344)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_35_cast_fp16 = layer_norm(axes = hidden_states_35_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_17_attn_norm_weight_to_fp16, x = input_309_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_17_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_17_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520027456)))];
            tensor<fp16, [1, 192, 3072]> linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_17_attn_Wqkv_weight_to_fp16, x = hidden_states_35_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [5]> var_1797 = const()[name = string("op_1797"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_71_cast_fp16 = reshape(shape = var_1797, x = linear_68_cast_fp16)[name = string("qkv_71_cast_fp16")];
            tensor<int32, [3]> var_1799_split_sizes_0 = const()[name = string("op_1799_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1799_axis_0 = const()[name = string("op_1799_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1799_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1799_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1799_cast_fp16_2 = split(axis = var_1799_axis_0, split_sizes = var_1799_split_sizes_0, x = qkv_71_cast_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> squeeze_51_axes_0 = const()[name = string("squeeze_51_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_51_cast_fp16 = squeeze(axes = squeeze_51_axes_0, x = var_1799_cast_fp16_0)[name = string("squeeze_51_cast_fp16")];
            tensor<int32, [1]> squeeze_52_axes_0 = const()[name = string("squeeze_52_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_52_cast_fp16 = squeeze(axes = squeeze_52_axes_0, x = var_1799_cast_fp16_1)[name = string("squeeze_52_cast_fp16")];
            tensor<int32, [1]> squeeze_53_axes_0 = const()[name = string("squeeze_53_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_53_cast_fp16 = squeeze(axes = squeeze_53_axes_0, x = var_1799_cast_fp16_2)[name = string("squeeze_53_cast_fp16")];
            tensor<int32, [4]> q_69_perm_0 = const()[name = string("q_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_69_perm_0 = const()[name = string("k_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_69_cast_fp16 = transpose(perm = q_69_perm_0, x = squeeze_51_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 16, 192, 64]> var_1809_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_11_to_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_69_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_69_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1821_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1821_cast_fp16")];
            bool var_1823_interleave_0 = const()[name = string("op_1823_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1823_cast_fp16 = concat(axis = var_49, interleave = var_1823_interleave_0, values = (var_1821_cast_fp16, x1_69_cast_fp16))[name = string("op_1823_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1824_cast_fp16 = mul(x = var_1823_cast_fp16, y = sin_11_to_fp16)[name = string("op_1824_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_35_cast_fp16 = add(x = var_1809_cast_fp16, y = var_1824_cast_fp16)[name = string("q_embed_35_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_69_cast_fp16 = transpose(perm = k_69_perm_0, x = squeeze_52_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 16, 192, 64]> var_1827_cast_fp16 = mul(x = k_69_cast_fp16, y = cos_11_to_fp16)[name = string("op_1827_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_69_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_69_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1839_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1839_cast_fp16")];
            bool var_1841_interleave_0 = const()[name = string("op_1841_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1841_cast_fp16 = concat(axis = var_49, interleave = var_1841_interleave_0, values = (var_1839_cast_fp16, x1_71_cast_fp16))[name = string("op_1841_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1842_cast_fp16 = mul(x = var_1841_cast_fp16, y = sin_11_to_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_35_cast_fp16 = add(x = var_1827_cast_fp16, y = var_1842_cast_fp16)[name = string("k_embed_35_cast_fp16")];
            bool var_1847_transpose_x_1 = const()[name = string("op_1847_transpose_x_1"), val = bool(false)];
            bool var_1847_transpose_y_1 = const()[name = string("op_1847_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1847_cast_fp16 = matmul(transpose_x = var_1847_transpose_x_1, transpose_y = var_1847_transpose_y_1, x = q_embed_35_cast_fp16, y = k_embed_35_cast_fp16)[name = string("op_1847_cast_fp16")];
            fp16 var_1848_to_fp16 = const()[name = string("op_1848_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_69_cast_fp16 = mul(x = var_1847_cast_fp16, y = var_1848_to_fp16)[name = string("attn_weights_69_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_311_cast_fp16 = add(x = attn_weights_69_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_311_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1851_cast_fp16 = softmax(axis = var_49, x = input_311_cast_fp16)[name = string("op_1851_cast_fp16")];
            bool attn_output_103_transpose_x_0 = const()[name = string("attn_output_103_transpose_x_0"), val = bool(false)];
            bool attn_output_103_transpose_y_0 = const()[name = string("attn_output_103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = squeeze_53_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_103_cast_fp16 = matmul(transpose_x = attn_output_103_transpose_x_0, transpose_y = attn_output_103_transpose_y_0, x = var_1851_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [4]> var_1855_perm_0 = const()[name = string("op_1855_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1857 = const()[name = string("op_1857"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1855_cast_fp16 = transpose(perm = var_1855_perm_0, x = attn_output_103_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 192, 1024]> var_1858_cast_fp16 = reshape(shape = var_1857, x = var_1855_cast_fp16)[name = string("op_1858_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_17_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_17_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526318976)))];
            tensor<fp16, [1, 192, 1024]> linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_17_attn_Wo_weight_to_fp16, x = var_1858_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_317_cast_fp16 = add(x = input_309_cast_fp16, y = linear_69_cast_fp16)[name = string("input_317_cast_fp16")];
            tensor<int32, [1]> input_319_axes_0 = const()[name = string("input_319_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_17_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_17_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528416192)))];
            tensor<fp16, [1, 192, 1024]> input_319_cast_fp16 = layer_norm(axes = input_319_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_17_mlp_norm_weight_to_fp16, x = input_317_cast_fp16)[name = string("input_319_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_17_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_17_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528418304)))];
            tensor<fp16, [1, 192, 5248]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_17_mlp_Wi_weight_to_fp16, x = input_319_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [2]> var_1870_split_sizes_0 = const()[name = string("op_1870_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1870_axis_0 = const()[name = string("op_1870_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1870_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1870_cast_fp16_1 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_70_cast_fp16)[name = string("op_1870_cast_fp16")];
            string var_1872_mode_0 = const()[name = string("op_1872_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1872_cast_fp16 = gelu(mode = var_1872_mode_0, x = var_1870_cast_fp16_0)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_323_cast_fp16 = mul(x = var_1872_cast_fp16, y = var_1870_cast_fp16_1)[name = string("input_323_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_17_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_17_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539166272)))];
            tensor<fp16, [1, 192, 1024]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_17_mlp_Wo_weight_to_fp16, x = input_323_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_327_cast_fp16 = add(x = input_317_cast_fp16, y = linear_71_cast_fp16)[name = string("input_327_cast_fp16")];
            tensor<int32, [1]> hidden_states_37_axes_0 = const()[name = string("hidden_states_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_18_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_18_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544540288)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_37_cast_fp16 = layer_norm(axes = hidden_states_37_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_18_attn_norm_weight_to_fp16, x = input_327_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_18_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_18_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544542400)))];
            tensor<fp16, [1, 192, 3072]> linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_18_attn_Wqkv_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [5]> var_1891 = const()[name = string("op_1891"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_75_cast_fp16 = reshape(shape = var_1891, x = linear_72_cast_fp16)[name = string("qkv_75_cast_fp16")];
            tensor<int32, [3]> var_1893_split_sizes_0 = const()[name = string("op_1893_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1893_axis_0 = const()[name = string("op_1893_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1893_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1893_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1893_cast_fp16_2 = split(axis = var_1893_axis_0, split_sizes = var_1893_split_sizes_0, x = qkv_75_cast_fp16)[name = string("op_1893_cast_fp16")];
            tensor<int32, [1]> squeeze_54_axes_0 = const()[name = string("squeeze_54_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_54_cast_fp16 = squeeze(axes = squeeze_54_axes_0, x = var_1893_cast_fp16_0)[name = string("squeeze_54_cast_fp16")];
            tensor<int32, [1]> squeeze_55_axes_0 = const()[name = string("squeeze_55_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_55_cast_fp16 = squeeze(axes = squeeze_55_axes_0, x = var_1893_cast_fp16_1)[name = string("squeeze_55_cast_fp16")];
            tensor<int32, [1]> squeeze_56_axes_0 = const()[name = string("squeeze_56_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_56_cast_fp16 = squeeze(axes = squeeze_56_axes_0, x = var_1893_cast_fp16_2)[name = string("squeeze_56_cast_fp16")];
            tensor<int32, [4]> q_73_perm_0 = const()[name = string("q_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_73_perm_0 = const()[name = string("k_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_73_cast_fp16 = transpose(perm = q_73_perm_0, x = squeeze_54_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 192, 64]> var_1903_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_7_to_fp16)[name = string("op_1903_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_73_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_73_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_167_promoted_to_fp16 = const()[name = string("const_167_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1915_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_167_promoted_to_fp16)[name = string("op_1915_cast_fp16")];
            bool var_1917_interleave_0 = const()[name = string("op_1917_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1917_cast_fp16 = concat(axis = var_49, interleave = var_1917_interleave_0, values = (var_1915_cast_fp16, x1_73_cast_fp16))[name = string("op_1917_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1918_cast_fp16 = mul(x = var_1917_cast_fp16, y = sin_7_to_fp16)[name = string("op_1918_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_37_cast_fp16 = add(x = var_1903_cast_fp16, y = var_1918_cast_fp16)[name = string("q_embed_37_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_73_cast_fp16 = transpose(perm = k_73_perm_0, x = squeeze_55_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 16, 192, 64]> var_1921_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_7_to_fp16)[name = string("op_1921_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_73_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_73_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_170_promoted_to_fp16 = const()[name = string("const_170_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_1933_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_170_promoted_to_fp16)[name = string("op_1933_cast_fp16")];
            bool var_1935_interleave_0 = const()[name = string("op_1935_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_1935_cast_fp16 = concat(axis = var_49, interleave = var_1935_interleave_0, values = (var_1933_cast_fp16, x1_75_cast_fp16))[name = string("op_1935_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_1936_cast_fp16 = mul(x = var_1935_cast_fp16, y = sin_7_to_fp16)[name = string("op_1936_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_37_cast_fp16 = add(x = var_1921_cast_fp16, y = var_1936_cast_fp16)[name = string("k_embed_37_cast_fp16")];
            bool var_1941_transpose_x_1 = const()[name = string("op_1941_transpose_x_1"), val = bool(false)];
            bool var_1941_transpose_y_1 = const()[name = string("op_1941_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_1941_cast_fp16 = matmul(transpose_x = var_1941_transpose_x_1, transpose_y = var_1941_transpose_y_1, x = q_embed_37_cast_fp16, y = k_embed_37_cast_fp16)[name = string("op_1941_cast_fp16")];
            fp16 var_1942_to_fp16 = const()[name = string("op_1942_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_73_cast_fp16 = mul(x = var_1941_cast_fp16, y = var_1942_to_fp16)[name = string("attn_weights_73_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_329_cast_fp16 = add(x = attn_weights_73_cast_fp16, y = key_bias_cast_fp16)[name = string("input_329_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_1945_cast_fp16 = softmax(axis = var_49, x = input_329_cast_fp16)[name = string("op_1945_cast_fp16")];
            bool attn_output_109_transpose_x_0 = const()[name = string("attn_output_109_transpose_x_0"), val = bool(false)];
            bool attn_output_109_transpose_y_0 = const()[name = string("attn_output_109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = squeeze_56_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_109_cast_fp16 = matmul(transpose_x = attn_output_109_transpose_x_0, transpose_y = attn_output_109_transpose_y_0, x = var_1945_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_109_cast_fp16")];
            tensor<int32, [4]> var_1949_perm_0 = const()[name = string("op_1949_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1951 = const()[name = string("op_1951"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_1949_cast_fp16 = transpose(perm = var_1949_perm_0, x = attn_output_109_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 192, 1024]> var_1952_cast_fp16 = reshape(shape = var_1951, x = var_1949_cast_fp16)[name = string("op_1952_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_18_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_18_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550833920)))];
            tensor<fp16, [1, 192, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_18_attn_Wo_weight_to_fp16, x = var_1952_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_335_cast_fp16 = add(x = input_327_cast_fp16, y = linear_73_cast_fp16)[name = string("input_335_cast_fp16")];
            tensor<int32, [1]> input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_18_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_18_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552931136)))];
            tensor<fp16, [1, 192, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_18_mlp_norm_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_18_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_18_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552933248)))];
            tensor<fp16, [1, 192, 5248]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_18_mlp_Wi_weight_to_fp16, x = input_337_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [2]> var_1964_split_sizes_0 = const()[name = string("op_1964_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1964_axis_0 = const()[name = string("op_1964_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_1964_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_1964_cast_fp16_1 = split(axis = var_1964_axis_0, split_sizes = var_1964_split_sizes_0, x = linear_74_cast_fp16)[name = string("op_1964_cast_fp16")];
            string var_1966_mode_0 = const()[name = string("op_1966_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_1966_cast_fp16 = gelu(mode = var_1966_mode_0, x = var_1964_cast_fp16_0)[name = string("op_1966_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_341_cast_fp16 = mul(x = var_1966_cast_fp16, y = var_1964_cast_fp16_1)[name = string("input_341_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_18_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_18_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563681216)))];
            tensor<fp16, [1, 192, 1024]> linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_18_mlp_Wo_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_345_cast_fp16 = add(x = input_335_cast_fp16, y = linear_75_cast_fp16)[name = string("input_345_cast_fp16")];
            tensor<int32, [1]> hidden_states_39_axes_0 = const()[name = string("hidden_states_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_19_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_19_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569055232)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_39_cast_fp16 = layer_norm(axes = hidden_states_39_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_19_attn_norm_weight_to_fp16, x = input_345_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_19_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_19_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569057344)))];
            tensor<fp16, [1, 192, 3072]> linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_19_attn_Wqkv_weight_to_fp16, x = hidden_states_39_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<int32, [5]> var_1985 = const()[name = string("op_1985"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_79_cast_fp16 = reshape(shape = var_1985, x = linear_76_cast_fp16)[name = string("qkv_79_cast_fp16")];
            tensor<int32, [3]> var_1987_split_sizes_0 = const()[name = string("op_1987_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1987_axis_0 = const()[name = string("op_1987_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_1987_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_1987_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_1987_cast_fp16_2 = split(axis = var_1987_axis_0, split_sizes = var_1987_split_sizes_0, x = qkv_79_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [1]> squeeze_57_axes_0 = const()[name = string("squeeze_57_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_57_cast_fp16 = squeeze(axes = squeeze_57_axes_0, x = var_1987_cast_fp16_0)[name = string("squeeze_57_cast_fp16")];
            tensor<int32, [1]> squeeze_58_axes_0 = const()[name = string("squeeze_58_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_58_cast_fp16 = squeeze(axes = squeeze_58_axes_0, x = var_1987_cast_fp16_1)[name = string("squeeze_58_cast_fp16")];
            tensor<int32, [1]> squeeze_59_axes_0 = const()[name = string("squeeze_59_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_59_cast_fp16 = squeeze(axes = squeeze_59_axes_0, x = var_1987_cast_fp16_2)[name = string("squeeze_59_cast_fp16")];
            tensor<int32, [4]> q_77_perm_0 = const()[name = string("q_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_77_perm_0 = const()[name = string("k_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_77_cast_fp16 = transpose(perm = q_77_perm_0, x = squeeze_57_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 192, 64]> var_1997_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_11_to_fp16)[name = string("op_1997_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_77_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_77_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_175_promoted_to_fp16 = const()[name = string("const_175_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2009_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_175_promoted_to_fp16)[name = string("op_2009_cast_fp16")];
            bool var_2011_interleave_0 = const()[name = string("op_2011_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2011_cast_fp16 = concat(axis = var_49, interleave = var_2011_interleave_0, values = (var_2009_cast_fp16, x1_77_cast_fp16))[name = string("op_2011_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2012_cast_fp16 = mul(x = var_2011_cast_fp16, y = sin_11_to_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_39_cast_fp16 = add(x = var_1997_cast_fp16, y = var_2012_cast_fp16)[name = string("q_embed_39_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_77_cast_fp16 = transpose(perm = k_77_perm_0, x = squeeze_58_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 192, 64]> var_2015_cast_fp16 = mul(x = k_77_cast_fp16, y = cos_11_to_fp16)[name = string("op_2015_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_77_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_77_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_178_promoted_to_fp16 = const()[name = string("const_178_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2027_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_178_promoted_to_fp16)[name = string("op_2027_cast_fp16")];
            bool var_2029_interleave_0 = const()[name = string("op_2029_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2029_cast_fp16 = concat(axis = var_49, interleave = var_2029_interleave_0, values = (var_2027_cast_fp16, x1_79_cast_fp16))[name = string("op_2029_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2030_cast_fp16 = mul(x = var_2029_cast_fp16, y = sin_11_to_fp16)[name = string("op_2030_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_39_cast_fp16 = add(x = var_2015_cast_fp16, y = var_2030_cast_fp16)[name = string("k_embed_39_cast_fp16")];
            bool var_2035_transpose_x_1 = const()[name = string("op_2035_transpose_x_1"), val = bool(false)];
            bool var_2035_transpose_y_1 = const()[name = string("op_2035_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_2035_cast_fp16 = matmul(transpose_x = var_2035_transpose_x_1, transpose_y = var_2035_transpose_y_1, x = q_embed_39_cast_fp16, y = k_embed_39_cast_fp16)[name = string("op_2035_cast_fp16")];
            fp16 var_2036_to_fp16 = const()[name = string("op_2036_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_77_cast_fp16 = mul(x = var_2035_cast_fp16, y = var_2036_to_fp16)[name = string("attn_weights_77_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_347_cast_fp16 = add(x = attn_weights_77_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_347_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_2039_cast_fp16 = softmax(axis = var_49, x = input_347_cast_fp16)[name = string("op_2039_cast_fp16")];
            bool attn_output_115_transpose_x_0 = const()[name = string("attn_output_115_transpose_x_0"), val = bool(false)];
            bool attn_output_115_transpose_y_0 = const()[name = string("attn_output_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = squeeze_59_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_115_cast_fp16 = matmul(transpose_x = attn_output_115_transpose_x_0, transpose_y = attn_output_115_transpose_y_0, x = var_2039_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_115_cast_fp16")];
            tensor<int32, [4]> var_2043_perm_0 = const()[name = string("op_2043_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2045 = const()[name = string("op_2045"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_2043_cast_fp16 = transpose(perm = var_2043_perm_0, x = attn_output_115_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 192, 1024]> var_2046_cast_fp16 = reshape(shape = var_2045, x = var_2043_cast_fp16)[name = string("op_2046_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_19_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_19_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575348864)))];
            tensor<fp16, [1, 192, 1024]> linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_19_attn_Wo_weight_to_fp16, x = var_2046_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_353_cast_fp16 = add(x = input_345_cast_fp16, y = linear_77_cast_fp16)[name = string("input_353_cast_fp16")];
            tensor<int32, [1]> input_355_axes_0 = const()[name = string("input_355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_19_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_19_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577446080)))];
            tensor<fp16, [1, 192, 1024]> input_355_cast_fp16 = layer_norm(axes = input_355_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_19_mlp_norm_weight_to_fp16, x = input_353_cast_fp16)[name = string("input_355_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_19_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_19_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577448192)))];
            tensor<fp16, [1, 192, 5248]> linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_19_mlp_Wi_weight_to_fp16, x = input_355_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [2]> var_2058_split_sizes_0 = const()[name = string("op_2058_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2058_axis_0 = const()[name = string("op_2058_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_2058_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_2058_cast_fp16_1 = split(axis = var_2058_axis_0, split_sizes = var_2058_split_sizes_0, x = linear_78_cast_fp16)[name = string("op_2058_cast_fp16")];
            string var_2060_mode_0 = const()[name = string("op_2060_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_2060_cast_fp16 = gelu(mode = var_2060_mode_0, x = var_2058_cast_fp16_0)[name = string("op_2060_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_359_cast_fp16 = mul(x = var_2060_cast_fp16, y = var_2058_cast_fp16_1)[name = string("input_359_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_19_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_19_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588196160)))];
            tensor<fp16, [1, 192, 1024]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_19_mlp_Wo_weight_to_fp16, x = input_359_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_363_cast_fp16 = add(x = input_353_cast_fp16, y = linear_79_cast_fp16)[name = string("input_363_cast_fp16")];
            tensor<int32, [1]> hidden_states_41_axes_0 = const()[name = string("hidden_states_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_20_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_20_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593570176)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_41_cast_fp16 = layer_norm(axes = hidden_states_41_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_20_attn_norm_weight_to_fp16, x = input_363_cast_fp16)[name = string("hidden_states_41_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_20_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_20_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593572288)))];
            tensor<fp16, [1, 192, 3072]> linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_20_attn_Wqkv_weight_to_fp16, x = hidden_states_41_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [5]> var_2079 = const()[name = string("op_2079"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_83_cast_fp16 = reshape(shape = var_2079, x = linear_80_cast_fp16)[name = string("qkv_83_cast_fp16")];
            tensor<int32, [3]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_2081_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_2081_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_2081_cast_fp16_2 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = qkv_83_cast_fp16)[name = string("op_2081_cast_fp16")];
            tensor<int32, [1]> squeeze_60_axes_0 = const()[name = string("squeeze_60_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_60_cast_fp16 = squeeze(axes = squeeze_60_axes_0, x = var_2081_cast_fp16_0)[name = string("squeeze_60_cast_fp16")];
            tensor<int32, [1]> squeeze_61_axes_0 = const()[name = string("squeeze_61_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_61_cast_fp16 = squeeze(axes = squeeze_61_axes_0, x = var_2081_cast_fp16_1)[name = string("squeeze_61_cast_fp16")];
            tensor<int32, [1]> squeeze_62_axes_0 = const()[name = string("squeeze_62_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_62_cast_fp16 = squeeze(axes = squeeze_62_axes_0, x = var_2081_cast_fp16_2)[name = string("squeeze_62_cast_fp16")];
            tensor<int32, [4]> q_81_perm_0 = const()[name = string("q_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_81_perm_0 = const()[name = string("k_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_81_cast_fp16 = transpose(perm = q_81_perm_0, x = squeeze_60_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 16, 192, 64]> var_2091_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_11_to_fp16)[name = string("op_2091_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_81_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_81_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_183_promoted_to_fp16 = const()[name = string("const_183_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2103_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_183_promoted_to_fp16)[name = string("op_2103_cast_fp16")];
            bool var_2105_interleave_0 = const()[name = string("op_2105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2105_cast_fp16 = concat(axis = var_49, interleave = var_2105_interleave_0, values = (var_2103_cast_fp16, x1_81_cast_fp16))[name = string("op_2105_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2106_cast_fp16 = mul(x = var_2105_cast_fp16, y = sin_11_to_fp16)[name = string("op_2106_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_41_cast_fp16 = add(x = var_2091_cast_fp16, y = var_2106_cast_fp16)[name = string("q_embed_41_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_81_cast_fp16 = transpose(perm = k_81_perm_0, x = squeeze_61_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 192, 64]> var_2109_cast_fp16 = mul(x = k_81_cast_fp16, y = cos_11_to_fp16)[name = string("op_2109_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_81_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_81_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_186_promoted_to_fp16 = const()[name = string("const_186_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2121_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_186_promoted_to_fp16)[name = string("op_2121_cast_fp16")];
            bool var_2123_interleave_0 = const()[name = string("op_2123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2123_cast_fp16 = concat(axis = var_49, interleave = var_2123_interleave_0, values = (var_2121_cast_fp16, x1_83_cast_fp16))[name = string("op_2123_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2124_cast_fp16 = mul(x = var_2123_cast_fp16, y = sin_11_to_fp16)[name = string("op_2124_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_41_cast_fp16 = add(x = var_2109_cast_fp16, y = var_2124_cast_fp16)[name = string("k_embed_41_cast_fp16")];
            bool var_2129_transpose_x_1 = const()[name = string("op_2129_transpose_x_1"), val = bool(false)];
            bool var_2129_transpose_y_1 = const()[name = string("op_2129_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_2129_cast_fp16 = matmul(transpose_x = var_2129_transpose_x_1, transpose_y = var_2129_transpose_y_1, x = q_embed_41_cast_fp16, y = k_embed_41_cast_fp16)[name = string("op_2129_cast_fp16")];
            fp16 var_2130_to_fp16 = const()[name = string("op_2130_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_81_cast_fp16 = mul(x = var_2129_cast_fp16, y = var_2130_to_fp16)[name = string("attn_weights_81_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_365_cast_fp16 = add(x = attn_weights_81_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_365_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_2133_cast_fp16 = softmax(axis = var_49, x = input_365_cast_fp16)[name = string("op_2133_cast_fp16")];
            bool attn_output_121_transpose_x_0 = const()[name = string("attn_output_121_transpose_x_0"), val = bool(false)];
            bool attn_output_121_transpose_y_0 = const()[name = string("attn_output_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = squeeze_62_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_121_cast_fp16 = matmul(transpose_x = attn_output_121_transpose_x_0, transpose_y = attn_output_121_transpose_y_0, x = var_2133_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_121_cast_fp16")];
            tensor<int32, [4]> var_2137_perm_0 = const()[name = string("op_2137_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2139 = const()[name = string("op_2139"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_2137_cast_fp16 = transpose(perm = var_2137_perm_0, x = attn_output_121_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 192, 1024]> var_2140_cast_fp16 = reshape(shape = var_2139, x = var_2137_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_20_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_20_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599863808)))];
            tensor<fp16, [1, 192, 1024]> linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_20_attn_Wo_weight_to_fp16, x = var_2140_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_371_cast_fp16 = add(x = input_363_cast_fp16, y = linear_81_cast_fp16)[name = string("input_371_cast_fp16")];
            tensor<int32, [1]> input_373_axes_0 = const()[name = string("input_373_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_20_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_20_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601961024)))];
            tensor<fp16, [1, 192, 1024]> input_373_cast_fp16 = layer_norm(axes = input_373_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_20_mlp_norm_weight_to_fp16, x = input_371_cast_fp16)[name = string("input_373_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_20_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_20_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601963136)))];
            tensor<fp16, [1, 192, 5248]> linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_20_mlp_Wi_weight_to_fp16, x = input_373_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [2]> var_2152_split_sizes_0 = const()[name = string("op_2152_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2152_axis_0 = const()[name = string("op_2152_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_2152_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_2152_cast_fp16_1 = split(axis = var_2152_axis_0, split_sizes = var_2152_split_sizes_0, x = linear_82_cast_fp16)[name = string("op_2152_cast_fp16")];
            string var_2154_mode_0 = const()[name = string("op_2154_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_2154_cast_fp16 = gelu(mode = var_2154_mode_0, x = var_2152_cast_fp16_0)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_377_cast_fp16 = mul(x = var_2154_cast_fp16, y = var_2152_cast_fp16_1)[name = string("input_377_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_20_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_20_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612711104)))];
            tensor<fp16, [1, 192, 1024]> linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_20_mlp_Wo_weight_to_fp16, x = input_377_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_381_cast_fp16 = add(x = input_371_cast_fp16, y = linear_83_cast_fp16)[name = string("input_381_cast_fp16")];
            tensor<int32, [1]> hidden_states_43_axes_0 = const()[name = string("hidden_states_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_21_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_21_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618085120)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_43_cast_fp16 = layer_norm(axes = hidden_states_43_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_21_attn_norm_weight_to_fp16, x = input_381_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_21_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_21_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618087232)))];
            tensor<fp16, [1, 192, 3072]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_21_attn_Wqkv_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [5]> var_2173 = const()[name = string("op_2173"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_87_cast_fp16 = reshape(shape = var_2173, x = linear_84_cast_fp16)[name = string("qkv_87_cast_fp16")];
            tensor<int32, [3]> var_2175_split_sizes_0 = const()[name = string("op_2175_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2175_axis_0 = const()[name = string("op_2175_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_2175_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_2175_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_2175_cast_fp16_2 = split(axis = var_2175_axis_0, split_sizes = var_2175_split_sizes_0, x = qkv_87_cast_fp16)[name = string("op_2175_cast_fp16")];
            tensor<int32, [1]> squeeze_63_axes_0 = const()[name = string("squeeze_63_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_63_cast_fp16 = squeeze(axes = squeeze_63_axes_0, x = var_2175_cast_fp16_0)[name = string("squeeze_63_cast_fp16")];
            tensor<int32, [1]> squeeze_64_axes_0 = const()[name = string("squeeze_64_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_64_cast_fp16 = squeeze(axes = squeeze_64_axes_0, x = var_2175_cast_fp16_1)[name = string("squeeze_64_cast_fp16")];
            tensor<int32, [1]> squeeze_65_axes_0 = const()[name = string("squeeze_65_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_65_cast_fp16 = squeeze(axes = squeeze_65_axes_0, x = var_2175_cast_fp16_2)[name = string("squeeze_65_cast_fp16")];
            tensor<int32, [4]> q_85_perm_0 = const()[name = string("q_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_85_perm_0 = const()[name = string("k_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_85_cast_fp16 = transpose(perm = q_85_perm_0, x = squeeze_63_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 16, 192, 64]> var_2185_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_7_to_fp16)[name = string("op_2185_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_85_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_85_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_191_promoted_to_fp16 = const()[name = string("const_191_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2197_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_191_promoted_to_fp16)[name = string("op_2197_cast_fp16")];
            bool var_2199_interleave_0 = const()[name = string("op_2199_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2199_cast_fp16 = concat(axis = var_49, interleave = var_2199_interleave_0, values = (var_2197_cast_fp16, x1_85_cast_fp16))[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2200_cast_fp16 = mul(x = var_2199_cast_fp16, y = sin_7_to_fp16)[name = string("op_2200_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_43_cast_fp16 = add(x = var_2185_cast_fp16, y = var_2200_cast_fp16)[name = string("q_embed_43_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_85_cast_fp16 = transpose(perm = k_85_perm_0, x = squeeze_64_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 16, 192, 64]> var_2203_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_7_to_fp16)[name = string("op_2203_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_85_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_85_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_194_promoted_to_fp16 = const()[name = string("const_194_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2215_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_194_promoted_to_fp16)[name = string("op_2215_cast_fp16")];
            bool var_2217_interleave_0 = const()[name = string("op_2217_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2217_cast_fp16 = concat(axis = var_49, interleave = var_2217_interleave_0, values = (var_2215_cast_fp16, x1_87_cast_fp16))[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2218_cast_fp16 = mul(x = var_2217_cast_fp16, y = sin_7_to_fp16)[name = string("op_2218_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_43_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2218_cast_fp16)[name = string("k_embed_43_cast_fp16")];
            bool var_2223_transpose_x_1 = const()[name = string("op_2223_transpose_x_1"), val = bool(false)];
            bool var_2223_transpose_y_1 = const()[name = string("op_2223_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_2223_cast_fp16 = matmul(transpose_x = var_2223_transpose_x_1, transpose_y = var_2223_transpose_y_1, x = q_embed_43_cast_fp16, y = k_embed_43_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_85_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("attn_weights_85_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_383_cast_fp16 = add(x = attn_weights_85_cast_fp16, y = key_bias_cast_fp16)[name = string("input_383_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_2227_cast_fp16 = softmax(axis = var_49, x = input_383_cast_fp16)[name = string("op_2227_cast_fp16")];
            bool attn_output_127_transpose_x_0 = const()[name = string("attn_output_127_transpose_x_0"), val = bool(false)];
            bool attn_output_127_transpose_y_0 = const()[name = string("attn_output_127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = squeeze_65_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_127_cast_fp16 = matmul(transpose_x = attn_output_127_transpose_x_0, transpose_y = attn_output_127_transpose_y_0, x = var_2227_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_127_cast_fp16")];
            tensor<int32, [4]> var_2231_perm_0 = const()[name = string("op_2231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_2231_cast_fp16 = transpose(perm = var_2231_perm_0, x = attn_output_127_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 192, 1024]> var_2234_cast_fp16 = reshape(shape = var_2233, x = var_2231_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_21_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_21_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624378752)))];
            tensor<fp16, [1, 192, 1024]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_21_attn_Wo_weight_to_fp16, x = var_2234_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_389_cast_fp16 = add(x = input_381_cast_fp16, y = linear_85_cast_fp16)[name = string("input_389_cast_fp16")];
            tensor<int32, [1]> input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_21_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_21_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626475968)))];
            tensor<fp16, [1, 192, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_21_mlp_norm_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_21_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_21_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626478080)))];
            tensor<fp16, [1, 192, 5248]> linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_21_mlp_Wi_weight_to_fp16, x = input_391_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [2]> var_2246_split_sizes_0 = const()[name = string("op_2246_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2246_axis_0 = const()[name = string("op_2246_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_2246_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_2246_cast_fp16_1 = split(axis = var_2246_axis_0, split_sizes = var_2246_split_sizes_0, x = linear_86_cast_fp16)[name = string("op_2246_cast_fp16")];
            string var_2248_mode_0 = const()[name = string("op_2248_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_2248_cast_fp16 = gelu(mode = var_2248_mode_0, x = var_2246_cast_fp16_0)[name = string("op_2248_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_395_cast_fp16 = mul(x = var_2248_cast_fp16, y = var_2246_cast_fp16_1)[name = string("input_395_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_21_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_21_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637226048)))];
            tensor<fp16, [1, 192, 1024]> linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_21_mlp_Wo_weight_to_fp16, x = input_395_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_399_cast_fp16 = add(x = input_389_cast_fp16, y = linear_87_cast_fp16)[name = string("input_399_cast_fp16")];
            tensor<int32, [1]> hidden_states_45_axes_0 = const()[name = string("hidden_states_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_22_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_22_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642600064)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_45_cast_fp16 = layer_norm(axes = hidden_states_45_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_22_attn_norm_weight_to_fp16, x = input_399_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_22_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_22_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642602176)))];
            tensor<fp16, [1, 192, 3072]> linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_22_attn_Wqkv_weight_to_fp16, x = hidden_states_45_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<int32, [5]> var_2267 = const()[name = string("op_2267"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_91_cast_fp16 = reshape(shape = var_2267, x = linear_88_cast_fp16)[name = string("qkv_91_cast_fp16")];
            tensor<int32, [3]> var_2269_split_sizes_0 = const()[name = string("op_2269_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2269_axis_0 = const()[name = string("op_2269_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_2269_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_2269_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_2269_cast_fp16_2 = split(axis = var_2269_axis_0, split_sizes = var_2269_split_sizes_0, x = qkv_91_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> squeeze_66_axes_0 = const()[name = string("squeeze_66_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_66_cast_fp16 = squeeze(axes = squeeze_66_axes_0, x = var_2269_cast_fp16_0)[name = string("squeeze_66_cast_fp16")];
            tensor<int32, [1]> squeeze_67_axes_0 = const()[name = string("squeeze_67_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_67_cast_fp16 = squeeze(axes = squeeze_67_axes_0, x = var_2269_cast_fp16_1)[name = string("squeeze_67_cast_fp16")];
            tensor<int32, [1]> squeeze_68_axes_0 = const()[name = string("squeeze_68_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_68_cast_fp16 = squeeze(axes = squeeze_68_axes_0, x = var_2269_cast_fp16_2)[name = string("squeeze_68_cast_fp16")];
            tensor<int32, [4]> q_89_perm_0 = const()[name = string("q_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_89_perm_0 = const()[name = string("k_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_89_cast_fp16 = transpose(perm = q_89_perm_0, x = squeeze_66_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 16, 192, 64]> var_2279_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_11_to_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_89_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_89_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_199_promoted_to_fp16 = const()[name = string("const_199_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2291_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_199_promoted_to_fp16)[name = string("op_2291_cast_fp16")];
            bool var_2293_interleave_0 = const()[name = string("op_2293_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2293_cast_fp16 = concat(axis = var_49, interleave = var_2293_interleave_0, values = (var_2291_cast_fp16, x1_89_cast_fp16))[name = string("op_2293_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2294_cast_fp16 = mul(x = var_2293_cast_fp16, y = sin_11_to_fp16)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_45_cast_fp16 = add(x = var_2279_cast_fp16, y = var_2294_cast_fp16)[name = string("q_embed_45_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_89_cast_fp16 = transpose(perm = k_89_perm_0, x = squeeze_67_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 16, 192, 64]> var_2297_cast_fp16 = mul(x = k_89_cast_fp16, y = cos_11_to_fp16)[name = string("op_2297_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_89_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_89_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_202_promoted_to_fp16 = const()[name = string("const_202_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2309_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_202_promoted_to_fp16)[name = string("op_2309_cast_fp16")];
            bool var_2311_interleave_0 = const()[name = string("op_2311_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2311_cast_fp16 = concat(axis = var_49, interleave = var_2311_interleave_0, values = (var_2309_cast_fp16, x1_91_cast_fp16))[name = string("op_2311_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2312_cast_fp16 = mul(x = var_2311_cast_fp16, y = sin_11_to_fp16)[name = string("op_2312_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_45_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2312_cast_fp16)[name = string("k_embed_45_cast_fp16")];
            bool var_2317_transpose_x_1 = const()[name = string("op_2317_transpose_x_1"), val = bool(false)];
            bool var_2317_transpose_y_1 = const()[name = string("op_2317_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_2317_cast_fp16 = matmul(transpose_x = var_2317_transpose_x_1, transpose_y = var_2317_transpose_y_1, x = q_embed_45_cast_fp16, y = k_embed_45_cast_fp16)[name = string("op_2317_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_89_cast_fp16 = mul(x = var_2317_cast_fp16, y = var_2318_to_fp16)[name = string("attn_weights_89_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_401_cast_fp16 = add(x = attn_weights_89_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_401_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_2321_cast_fp16 = softmax(axis = var_49, x = input_401_cast_fp16)[name = string("op_2321_cast_fp16")];
            bool attn_output_133_transpose_x_0 = const()[name = string("attn_output_133_transpose_x_0"), val = bool(false)];
            bool attn_output_133_transpose_y_0 = const()[name = string("attn_output_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = squeeze_68_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_133_cast_fp16 = matmul(transpose_x = attn_output_133_transpose_x_0, transpose_y = attn_output_133_transpose_y_0, x = var_2321_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_133_cast_fp16")];
            tensor<int32, [4]> var_2325_perm_0 = const()[name = string("op_2325_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2327 = const()[name = string("op_2327"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_2325_cast_fp16 = transpose(perm = var_2325_perm_0, x = attn_output_133_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 192, 1024]> var_2328_cast_fp16 = reshape(shape = var_2327, x = var_2325_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_22_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_22_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648893696)))];
            tensor<fp16, [1, 192, 1024]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_22_attn_Wo_weight_to_fp16, x = var_2328_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_407_cast_fp16 = add(x = input_399_cast_fp16, y = linear_89_cast_fp16)[name = string("input_407_cast_fp16")];
            tensor<int32, [1]> input_409_axes_0 = const()[name = string("input_409_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_22_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_22_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650990912)))];
            tensor<fp16, [1, 192, 1024]> input_409_cast_fp16 = layer_norm(axes = input_409_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_22_mlp_norm_weight_to_fp16, x = input_407_cast_fp16)[name = string("input_409_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_22_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_22_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650993024)))];
            tensor<fp16, [1, 192, 5248]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_22_mlp_Wi_weight_to_fp16, x = input_409_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<int32, [2]> var_2340_split_sizes_0 = const()[name = string("op_2340_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2340_axis_0 = const()[name = string("op_2340_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_2340_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_90_cast_fp16)[name = string("op_2340_cast_fp16")];
            string var_2342_mode_0 = const()[name = string("op_2342_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_2342_cast_fp16 = gelu(mode = var_2342_mode_0, x = var_2340_cast_fp16_0)[name = string("op_2342_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_413_cast_fp16 = mul(x = var_2342_cast_fp16, y = var_2340_cast_fp16_1)[name = string("input_413_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_22_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_22_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661740992)))];
            tensor<fp16, [1, 192, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_22_mlp_Wo_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_417_cast_fp16 = add(x = input_407_cast_fp16, y = linear_91_cast_fp16)[name = string("input_417_cast_fp16")];
            tensor<int32, [1]> hidden_states_47_axes_0 = const()[name = string("hidden_states_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_23_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_23_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667115008)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_47_cast_fp16 = layer_norm(axes = hidden_states_47_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_23_attn_norm_weight_to_fp16, x = input_417_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_23_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_23_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667117120)))];
            tensor<fp16, [1, 192, 3072]> linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_23_attn_Wqkv_weight_to_fp16, x = hidden_states_47_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [5]> var_2361 = const()[name = string("op_2361"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_95_cast_fp16 = reshape(shape = var_2361, x = linear_92_cast_fp16)[name = string("qkv_95_cast_fp16")];
            tensor<int32, [3]> var_2363_split_sizes_0 = const()[name = string("op_2363_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2363_axis_0 = const()[name = string("op_2363_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_2363_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_2363_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_2363_cast_fp16_2 = split(axis = var_2363_axis_0, split_sizes = var_2363_split_sizes_0, x = qkv_95_cast_fp16)[name = string("op_2363_cast_fp16")];
            tensor<int32, [1]> squeeze_69_axes_0 = const()[name = string("squeeze_69_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_69_cast_fp16 = squeeze(axes = squeeze_69_axes_0, x = var_2363_cast_fp16_0)[name = string("squeeze_69_cast_fp16")];
            tensor<int32, [1]> squeeze_70_axes_0 = const()[name = string("squeeze_70_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_70_cast_fp16 = squeeze(axes = squeeze_70_axes_0, x = var_2363_cast_fp16_1)[name = string("squeeze_70_cast_fp16")];
            tensor<int32, [1]> squeeze_71_axes_0 = const()[name = string("squeeze_71_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_71_cast_fp16 = squeeze(axes = squeeze_71_axes_0, x = var_2363_cast_fp16_2)[name = string("squeeze_71_cast_fp16")];
            tensor<int32, [4]> q_93_perm_0 = const()[name = string("q_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_93_perm_0 = const()[name = string("k_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_93_cast_fp16 = transpose(perm = q_93_perm_0, x = squeeze_69_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 192, 64]> var_2373_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_11_to_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_93_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_93_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_207_promoted_to_fp16 = const()[name = string("const_207_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2385_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_207_promoted_to_fp16)[name = string("op_2385_cast_fp16")];
            bool var_2387_interleave_0 = const()[name = string("op_2387_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2387_cast_fp16 = concat(axis = var_49, interleave = var_2387_interleave_0, values = (var_2385_cast_fp16, x1_93_cast_fp16))[name = string("op_2387_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2388_cast_fp16 = mul(x = var_2387_cast_fp16, y = sin_11_to_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_47_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2388_cast_fp16)[name = string("q_embed_47_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_93_cast_fp16 = transpose(perm = k_93_perm_0, x = squeeze_70_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 16, 192, 64]> var_2391_cast_fp16 = mul(x = k_93_cast_fp16, y = cos_11_to_fp16)[name = string("op_2391_cast_fp16")];
            tensor<int32, [4]> x1_95_begin_0 = const()[name = string("x1_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_95_end_0 = const()[name = string("x1_95_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_95_end_mask_0 = const()[name = string("x1_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_95_cast_fp16 = slice_by_index(begin = x1_95_begin_0, end = x1_95_end_0, end_mask = x1_95_end_mask_0, x = k_93_cast_fp16)[name = string("x1_95_cast_fp16")];
            tensor<int32, [4]> x2_95_begin_0 = const()[name = string("x2_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_95_end_0 = const()[name = string("x2_95_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_95_end_mask_0 = const()[name = string("x2_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_95_cast_fp16 = slice_by_index(begin = x2_95_begin_0, end = x2_95_end_0, end_mask = x2_95_end_mask_0, x = k_93_cast_fp16)[name = string("x2_95_cast_fp16")];
            fp16 const_210_promoted_to_fp16 = const()[name = string("const_210_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2403_cast_fp16 = mul(x = x2_95_cast_fp16, y = const_210_promoted_to_fp16)[name = string("op_2403_cast_fp16")];
            bool var_2405_interleave_0 = const()[name = string("op_2405_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2405_cast_fp16 = concat(axis = var_49, interleave = var_2405_interleave_0, values = (var_2403_cast_fp16, x1_95_cast_fp16))[name = string("op_2405_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2406_cast_fp16 = mul(x = var_2405_cast_fp16, y = sin_11_to_fp16)[name = string("op_2406_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_47_cast_fp16 = add(x = var_2391_cast_fp16, y = var_2406_cast_fp16)[name = string("k_embed_47_cast_fp16")];
            bool var_2411_transpose_x_1 = const()[name = string("op_2411_transpose_x_1"), val = bool(false)];
            bool var_2411_transpose_y_1 = const()[name = string("op_2411_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_2411_cast_fp16 = matmul(transpose_x = var_2411_transpose_x_1, transpose_y = var_2411_transpose_y_1, x = q_embed_47_cast_fp16, y = k_embed_47_cast_fp16)[name = string("op_2411_cast_fp16")];
            fp16 var_2412_to_fp16 = const()[name = string("op_2412_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_93_cast_fp16 = mul(x = var_2411_cast_fp16, y = var_2412_to_fp16)[name = string("attn_weights_93_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_419_cast_fp16 = add(x = attn_weights_93_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_419_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_2415_cast_fp16 = softmax(axis = var_49, x = input_419_cast_fp16)[name = string("op_2415_cast_fp16")];
            bool attn_output_139_transpose_x_0 = const()[name = string("attn_output_139_transpose_x_0"), val = bool(false)];
            bool attn_output_139_transpose_y_0 = const()[name = string("attn_output_139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = squeeze_71_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_139_cast_fp16 = matmul(transpose_x = attn_output_139_transpose_x_0, transpose_y = attn_output_139_transpose_y_0, x = var_2415_cast_fp16, y = value_47_cast_fp16)[name = string("attn_output_139_cast_fp16")];
            tensor<int32, [4]> var_2419_perm_0 = const()[name = string("op_2419_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2421 = const()[name = string("op_2421"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_2419_cast_fp16 = transpose(perm = var_2419_perm_0, x = attn_output_139_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 192, 1024]> var_2422_cast_fp16 = reshape(shape = var_2421, x = var_2419_cast_fp16)[name = string("op_2422_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_23_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_23_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673408640)))];
            tensor<fp16, [1, 192, 1024]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_23_attn_Wo_weight_to_fp16, x = var_2422_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_425_cast_fp16 = add(x = input_417_cast_fp16, y = linear_93_cast_fp16)[name = string("input_425_cast_fp16")];
            tensor<int32, [1]> input_427_axes_0 = const()[name = string("input_427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_23_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_23_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675505856)))];
            tensor<fp16, [1, 192, 1024]> input_427_cast_fp16 = layer_norm(axes = input_427_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_23_mlp_norm_weight_to_fp16, x = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_23_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_23_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675507968)))];
            tensor<fp16, [1, 192, 5248]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_23_mlp_Wi_weight_to_fp16, x = input_427_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<int32, [2]> var_2434_split_sizes_0 = const()[name = string("op_2434_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2434_axis_0 = const()[name = string("op_2434_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_2434_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_2434_cast_fp16_1 = split(axis = var_2434_axis_0, split_sizes = var_2434_split_sizes_0, x = linear_94_cast_fp16)[name = string("op_2434_cast_fp16")];
            string var_2436_mode_0 = const()[name = string("op_2436_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_2436_cast_fp16 = gelu(mode = var_2436_mode_0, x = var_2434_cast_fp16_0)[name = string("op_2436_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_431_cast_fp16 = mul(x = var_2436_cast_fp16, y = var_2434_cast_fp16_1)[name = string("input_431_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_23_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_23_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686255936)))];
            tensor<fp16, [1, 192, 1024]> linear_95_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_23_mlp_Wo_weight_to_fp16, x = input_431_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_435_cast_fp16 = add(x = input_425_cast_fp16, y = linear_95_cast_fp16)[name = string("input_435_cast_fp16")];
            tensor<int32, [1]> hidden_states_49_axes_0 = const()[name = string("hidden_states_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_24_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_24_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691629952)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_49_cast_fp16 = layer_norm(axes = hidden_states_49_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_24_attn_norm_weight_to_fp16, x = input_435_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_24_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_24_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691632064)))];
            tensor<fp16, [1, 192, 3072]> linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_24_attn_Wqkv_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<int32, [5]> var_2455 = const()[name = string("op_2455"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_99_cast_fp16 = reshape(shape = var_2455, x = linear_96_cast_fp16)[name = string("qkv_99_cast_fp16")];
            tensor<int32, [3]> var_2457_split_sizes_0 = const()[name = string("op_2457_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2457_axis_0 = const()[name = string("op_2457_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_2457_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_2457_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_2457_cast_fp16_2 = split(axis = var_2457_axis_0, split_sizes = var_2457_split_sizes_0, x = qkv_99_cast_fp16)[name = string("op_2457_cast_fp16")];
            tensor<int32, [1]> squeeze_72_axes_0 = const()[name = string("squeeze_72_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_72_cast_fp16 = squeeze(axes = squeeze_72_axes_0, x = var_2457_cast_fp16_0)[name = string("squeeze_72_cast_fp16")];
            tensor<int32, [1]> squeeze_73_axes_0 = const()[name = string("squeeze_73_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_73_cast_fp16 = squeeze(axes = squeeze_73_axes_0, x = var_2457_cast_fp16_1)[name = string("squeeze_73_cast_fp16")];
            tensor<int32, [1]> squeeze_74_axes_0 = const()[name = string("squeeze_74_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_74_cast_fp16 = squeeze(axes = squeeze_74_axes_0, x = var_2457_cast_fp16_2)[name = string("squeeze_74_cast_fp16")];
            tensor<int32, [4]> q_97_perm_0 = const()[name = string("q_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_97_perm_0 = const()[name = string("k_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_97_cast_fp16 = transpose(perm = q_97_perm_0, x = squeeze_72_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, 192, 64]> var_2467_cast_fp16 = mul(x = q_97_cast_fp16, y = cos_7_to_fp16)[name = string("op_2467_cast_fp16")];
            tensor<int32, [4]> x1_97_begin_0 = const()[name = string("x1_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_97_end_0 = const()[name = string("x1_97_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_97_end_mask_0 = const()[name = string("x1_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_97_cast_fp16 = slice_by_index(begin = x1_97_begin_0, end = x1_97_end_0, end_mask = x1_97_end_mask_0, x = q_97_cast_fp16)[name = string("x1_97_cast_fp16")];
            tensor<int32, [4]> x2_97_begin_0 = const()[name = string("x2_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_97_end_0 = const()[name = string("x2_97_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_97_end_mask_0 = const()[name = string("x2_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_97_cast_fp16 = slice_by_index(begin = x2_97_begin_0, end = x2_97_end_0, end_mask = x2_97_end_mask_0, x = q_97_cast_fp16)[name = string("x2_97_cast_fp16")];
            fp16 const_215_promoted_to_fp16 = const()[name = string("const_215_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2479_cast_fp16 = mul(x = x2_97_cast_fp16, y = const_215_promoted_to_fp16)[name = string("op_2479_cast_fp16")];
            bool var_2481_interleave_0 = const()[name = string("op_2481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2481_cast_fp16 = concat(axis = var_49, interleave = var_2481_interleave_0, values = (var_2479_cast_fp16, x1_97_cast_fp16))[name = string("op_2481_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2482_cast_fp16 = mul(x = var_2481_cast_fp16, y = sin_7_to_fp16)[name = string("op_2482_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_49_cast_fp16 = add(x = var_2467_cast_fp16, y = var_2482_cast_fp16)[name = string("q_embed_49_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_97_cast_fp16 = transpose(perm = k_97_perm_0, x = squeeze_73_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 16, 192, 64]> var_2485_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_7_to_fp16)[name = string("op_2485_cast_fp16")];
            tensor<int32, [4]> x1_99_begin_0 = const()[name = string("x1_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_99_end_0 = const()[name = string("x1_99_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_99_end_mask_0 = const()[name = string("x1_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_99_cast_fp16 = slice_by_index(begin = x1_99_begin_0, end = x1_99_end_0, end_mask = x1_99_end_mask_0, x = k_97_cast_fp16)[name = string("x1_99_cast_fp16")];
            tensor<int32, [4]> x2_99_begin_0 = const()[name = string("x2_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_99_end_0 = const()[name = string("x2_99_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_99_end_mask_0 = const()[name = string("x2_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_99_cast_fp16 = slice_by_index(begin = x2_99_begin_0, end = x2_99_end_0, end_mask = x2_99_end_mask_0, x = k_97_cast_fp16)[name = string("x2_99_cast_fp16")];
            fp16 const_218_promoted_to_fp16 = const()[name = string("const_218_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2497_cast_fp16 = mul(x = x2_99_cast_fp16, y = const_218_promoted_to_fp16)[name = string("op_2497_cast_fp16")];
            bool var_2499_interleave_0 = const()[name = string("op_2499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2499_cast_fp16 = concat(axis = var_49, interleave = var_2499_interleave_0, values = (var_2497_cast_fp16, x1_99_cast_fp16))[name = string("op_2499_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2500_cast_fp16 = mul(x = var_2499_cast_fp16, y = sin_7_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_49_cast_fp16 = add(x = var_2485_cast_fp16, y = var_2500_cast_fp16)[name = string("k_embed_49_cast_fp16")];
            bool var_2505_transpose_x_1 = const()[name = string("op_2505_transpose_x_1"), val = bool(false)];
            bool var_2505_transpose_y_1 = const()[name = string("op_2505_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_2505_cast_fp16 = matmul(transpose_x = var_2505_transpose_x_1, transpose_y = var_2505_transpose_y_1, x = q_embed_49_cast_fp16, y = k_embed_49_cast_fp16)[name = string("op_2505_cast_fp16")];
            fp16 var_2506_to_fp16 = const()[name = string("op_2506_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_97_cast_fp16 = mul(x = var_2505_cast_fp16, y = var_2506_to_fp16)[name = string("attn_weights_97_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_437_cast_fp16 = add(x = attn_weights_97_cast_fp16, y = key_bias_cast_fp16)[name = string("input_437_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_2509_cast_fp16 = softmax(axis = var_49, x = input_437_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool attn_output_145_transpose_x_0 = const()[name = string("attn_output_145_transpose_x_0"), val = bool(false)];
            bool attn_output_145_transpose_y_0 = const()[name = string("attn_output_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = squeeze_74_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_145_cast_fp16 = matmul(transpose_x = attn_output_145_transpose_x_0, transpose_y = attn_output_145_transpose_y_0, x = var_2509_cast_fp16, y = value_49_cast_fp16)[name = string("attn_output_145_cast_fp16")];
            tensor<int32, [4]> var_2513_perm_0 = const()[name = string("op_2513_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2515 = const()[name = string("op_2515"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_2513_cast_fp16 = transpose(perm = var_2513_perm_0, x = attn_output_145_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 192, 1024]> var_2516_cast_fp16 = reshape(shape = var_2515, x = var_2513_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_24_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_24_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(697923584)))];
            tensor<fp16, [1, 192, 1024]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_24_attn_Wo_weight_to_fp16, x = var_2516_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_443_cast_fp16 = add(x = input_435_cast_fp16, y = linear_97_cast_fp16)[name = string("input_443_cast_fp16")];
            tensor<int32, [1]> input_445_axes_0 = const()[name = string("input_445_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_24_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_24_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700020800)))];
            tensor<fp16, [1, 192, 1024]> input_445_cast_fp16 = layer_norm(axes = input_445_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_24_mlp_norm_weight_to_fp16, x = input_443_cast_fp16)[name = string("input_445_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_24_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_24_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700022912)))];
            tensor<fp16, [1, 192, 5248]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_24_mlp_Wi_weight_to_fp16, x = input_445_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [2]> var_2528_split_sizes_0 = const()[name = string("op_2528_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2528_axis_0 = const()[name = string("op_2528_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_2528_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_2528_cast_fp16_1 = split(axis = var_2528_axis_0, split_sizes = var_2528_split_sizes_0, x = linear_98_cast_fp16)[name = string("op_2528_cast_fp16")];
            string var_2530_mode_0 = const()[name = string("op_2530_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_2530_cast_fp16 = gelu(mode = var_2530_mode_0, x = var_2528_cast_fp16_0)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_449_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2528_cast_fp16_1)[name = string("input_449_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_24_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_24_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710770880)))];
            tensor<fp16, [1, 192, 1024]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_24_mlp_Wo_weight_to_fp16, x = input_449_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_453_cast_fp16 = add(x = input_443_cast_fp16, y = linear_99_cast_fp16)[name = string("input_453_cast_fp16")];
            tensor<int32, [1]> hidden_states_51_axes_0 = const()[name = string("hidden_states_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_25_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_25_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(716144896)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_51_cast_fp16 = layer_norm(axes = hidden_states_51_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_25_attn_norm_weight_to_fp16, x = input_453_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_25_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_25_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(716147008)))];
            tensor<fp16, [1, 192, 3072]> linear_100_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_25_attn_Wqkv_weight_to_fp16, x = hidden_states_51_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [5]> var_2549 = const()[name = string("op_2549"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_103_cast_fp16 = reshape(shape = var_2549, x = linear_100_cast_fp16)[name = string("qkv_103_cast_fp16")];
            tensor<int32, [3]> var_2551_split_sizes_0 = const()[name = string("op_2551_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2551_axis_0 = const()[name = string("op_2551_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_2551_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_2551_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_2551_cast_fp16_2 = split(axis = var_2551_axis_0, split_sizes = var_2551_split_sizes_0, x = qkv_103_cast_fp16)[name = string("op_2551_cast_fp16")];
            tensor<int32, [1]> squeeze_75_axes_0 = const()[name = string("squeeze_75_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_75_cast_fp16 = squeeze(axes = squeeze_75_axes_0, x = var_2551_cast_fp16_0)[name = string("squeeze_75_cast_fp16")];
            tensor<int32, [1]> squeeze_76_axes_0 = const()[name = string("squeeze_76_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_76_cast_fp16 = squeeze(axes = squeeze_76_axes_0, x = var_2551_cast_fp16_1)[name = string("squeeze_76_cast_fp16")];
            tensor<int32, [1]> squeeze_77_axes_0 = const()[name = string("squeeze_77_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_77_cast_fp16 = squeeze(axes = squeeze_77_axes_0, x = var_2551_cast_fp16_2)[name = string("squeeze_77_cast_fp16")];
            tensor<int32, [4]> q_101_perm_0 = const()[name = string("q_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_101_perm_0 = const()[name = string("k_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_101_cast_fp16 = transpose(perm = q_101_perm_0, x = squeeze_75_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 16, 192, 64]> var_2561_cast_fp16 = mul(x = q_101_cast_fp16, y = cos_11_to_fp16)[name = string("op_2561_cast_fp16")];
            tensor<int32, [4]> x1_101_begin_0 = const()[name = string("x1_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_101_end_0 = const()[name = string("x1_101_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_101_end_mask_0 = const()[name = string("x1_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_101_cast_fp16 = slice_by_index(begin = x1_101_begin_0, end = x1_101_end_0, end_mask = x1_101_end_mask_0, x = q_101_cast_fp16)[name = string("x1_101_cast_fp16")];
            tensor<int32, [4]> x2_101_begin_0 = const()[name = string("x2_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_101_end_0 = const()[name = string("x2_101_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_101_end_mask_0 = const()[name = string("x2_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_101_cast_fp16 = slice_by_index(begin = x2_101_begin_0, end = x2_101_end_0, end_mask = x2_101_end_mask_0, x = q_101_cast_fp16)[name = string("x2_101_cast_fp16")];
            fp16 const_223_promoted_to_fp16 = const()[name = string("const_223_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2573_cast_fp16 = mul(x = x2_101_cast_fp16, y = const_223_promoted_to_fp16)[name = string("op_2573_cast_fp16")];
            bool var_2575_interleave_0 = const()[name = string("op_2575_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2575_cast_fp16 = concat(axis = var_49, interleave = var_2575_interleave_0, values = (var_2573_cast_fp16, x1_101_cast_fp16))[name = string("op_2575_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2576_cast_fp16 = mul(x = var_2575_cast_fp16, y = sin_11_to_fp16)[name = string("op_2576_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_51_cast_fp16 = add(x = var_2561_cast_fp16, y = var_2576_cast_fp16)[name = string("q_embed_51_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_101_cast_fp16 = transpose(perm = k_101_perm_0, x = squeeze_76_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 16, 192, 64]> var_2579_cast_fp16 = mul(x = k_101_cast_fp16, y = cos_11_to_fp16)[name = string("op_2579_cast_fp16")];
            tensor<int32, [4]> x1_103_begin_0 = const()[name = string("x1_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_103_end_0 = const()[name = string("x1_103_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_103_end_mask_0 = const()[name = string("x1_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_103_cast_fp16 = slice_by_index(begin = x1_103_begin_0, end = x1_103_end_0, end_mask = x1_103_end_mask_0, x = k_101_cast_fp16)[name = string("x1_103_cast_fp16")];
            tensor<int32, [4]> x2_103_begin_0 = const()[name = string("x2_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_103_end_0 = const()[name = string("x2_103_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_103_end_mask_0 = const()[name = string("x2_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_103_cast_fp16 = slice_by_index(begin = x2_103_begin_0, end = x2_103_end_0, end_mask = x2_103_end_mask_0, x = k_101_cast_fp16)[name = string("x2_103_cast_fp16")];
            fp16 const_226_promoted_to_fp16 = const()[name = string("const_226_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2591_cast_fp16 = mul(x = x2_103_cast_fp16, y = const_226_promoted_to_fp16)[name = string("op_2591_cast_fp16")];
            bool var_2593_interleave_0 = const()[name = string("op_2593_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2593_cast_fp16 = concat(axis = var_49, interleave = var_2593_interleave_0, values = (var_2591_cast_fp16, x1_103_cast_fp16))[name = string("op_2593_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2594_cast_fp16 = mul(x = var_2593_cast_fp16, y = sin_11_to_fp16)[name = string("op_2594_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_51_cast_fp16 = add(x = var_2579_cast_fp16, y = var_2594_cast_fp16)[name = string("k_embed_51_cast_fp16")];
            bool var_2599_transpose_x_1 = const()[name = string("op_2599_transpose_x_1"), val = bool(false)];
            bool var_2599_transpose_y_1 = const()[name = string("op_2599_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_2599_cast_fp16 = matmul(transpose_x = var_2599_transpose_x_1, transpose_y = var_2599_transpose_y_1, x = q_embed_51_cast_fp16, y = k_embed_51_cast_fp16)[name = string("op_2599_cast_fp16")];
            fp16 var_2600_to_fp16 = const()[name = string("op_2600_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_101_cast_fp16 = mul(x = var_2599_cast_fp16, y = var_2600_to_fp16)[name = string("attn_weights_101_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_455_cast_fp16 = add(x = attn_weights_101_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_455_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_2603_cast_fp16 = softmax(axis = var_49, x = input_455_cast_fp16)[name = string("op_2603_cast_fp16")];
            bool attn_output_151_transpose_x_0 = const()[name = string("attn_output_151_transpose_x_0"), val = bool(false)];
            bool attn_output_151_transpose_y_0 = const()[name = string("attn_output_151_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = squeeze_77_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_151_cast_fp16 = matmul(transpose_x = attn_output_151_transpose_x_0, transpose_y = attn_output_151_transpose_y_0, x = var_2603_cast_fp16, y = value_51_cast_fp16)[name = string("attn_output_151_cast_fp16")];
            tensor<int32, [4]> var_2607_perm_0 = const()[name = string("op_2607_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2609 = const()[name = string("op_2609"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_2607_cast_fp16 = transpose(perm = var_2607_perm_0, x = attn_output_151_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 192, 1024]> var_2610_cast_fp16 = reshape(shape = var_2609, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_25_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_25_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(722438528)))];
            tensor<fp16, [1, 192, 1024]> linear_101_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_25_attn_Wo_weight_to_fp16, x = var_2610_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_461_cast_fp16 = add(x = input_453_cast_fp16, y = linear_101_cast_fp16)[name = string("input_461_cast_fp16")];
            tensor<int32, [1]> input_463_axes_0 = const()[name = string("input_463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_25_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_25_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724535744)))];
            tensor<fp16, [1, 192, 1024]> input_463_cast_fp16 = layer_norm(axes = input_463_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_25_mlp_norm_weight_to_fp16, x = input_461_cast_fp16)[name = string("input_463_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_25_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_25_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724537856)))];
            tensor<fp16, [1, 192, 5248]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_25_mlp_Wi_weight_to_fp16, x = input_463_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [2]> var_2622_split_sizes_0 = const()[name = string("op_2622_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2622_axis_0 = const()[name = string("op_2622_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_2622_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_2622_cast_fp16_1 = split(axis = var_2622_axis_0, split_sizes = var_2622_split_sizes_0, x = linear_102_cast_fp16)[name = string("op_2622_cast_fp16")];
            string var_2624_mode_0 = const()[name = string("op_2624_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_2624_cast_fp16 = gelu(mode = var_2624_mode_0, x = var_2622_cast_fp16_0)[name = string("op_2624_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_467_cast_fp16 = mul(x = var_2624_cast_fp16, y = var_2622_cast_fp16_1)[name = string("input_467_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_25_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_25_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(735285824)))];
            tensor<fp16, [1, 192, 1024]> linear_103_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_25_mlp_Wo_weight_to_fp16, x = input_467_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_471_cast_fp16 = add(x = input_461_cast_fp16, y = linear_103_cast_fp16)[name = string("input_471_cast_fp16")];
            tensor<int32, [1]> hidden_states_53_axes_0 = const()[name = string("hidden_states_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_26_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_26_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740659840)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_53_cast_fp16 = layer_norm(axes = hidden_states_53_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_26_attn_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_26_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_26_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740661952)))];
            tensor<fp16, [1, 192, 3072]> linear_104_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_26_attn_Wqkv_weight_to_fp16, x = hidden_states_53_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<int32, [5]> var_2643 = const()[name = string("op_2643"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_107_cast_fp16 = reshape(shape = var_2643, x = linear_104_cast_fp16)[name = string("qkv_107_cast_fp16")];
            tensor<int32, [3]> var_2645_split_sizes_0 = const()[name = string("op_2645_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2645_axis_0 = const()[name = string("op_2645_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_2645_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_2645_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_2645_cast_fp16_2 = split(axis = var_2645_axis_0, split_sizes = var_2645_split_sizes_0, x = qkv_107_cast_fp16)[name = string("op_2645_cast_fp16")];
            tensor<int32, [1]> squeeze_78_axes_0 = const()[name = string("squeeze_78_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_78_cast_fp16 = squeeze(axes = squeeze_78_axes_0, x = var_2645_cast_fp16_0)[name = string("squeeze_78_cast_fp16")];
            tensor<int32, [1]> squeeze_79_axes_0 = const()[name = string("squeeze_79_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_79_cast_fp16 = squeeze(axes = squeeze_79_axes_0, x = var_2645_cast_fp16_1)[name = string("squeeze_79_cast_fp16")];
            tensor<int32, [1]> squeeze_80_axes_0 = const()[name = string("squeeze_80_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_80_cast_fp16 = squeeze(axes = squeeze_80_axes_0, x = var_2645_cast_fp16_2)[name = string("squeeze_80_cast_fp16")];
            tensor<int32, [4]> q_105_perm_0 = const()[name = string("q_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_105_perm_0 = const()[name = string("k_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_105_cast_fp16 = transpose(perm = q_105_perm_0, x = squeeze_78_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 16, 192, 64]> var_2655_cast_fp16 = mul(x = q_105_cast_fp16, y = cos_11_to_fp16)[name = string("op_2655_cast_fp16")];
            tensor<int32, [4]> x1_105_begin_0 = const()[name = string("x1_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_105_end_0 = const()[name = string("x1_105_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_105_end_mask_0 = const()[name = string("x1_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_105_cast_fp16 = slice_by_index(begin = x1_105_begin_0, end = x1_105_end_0, end_mask = x1_105_end_mask_0, x = q_105_cast_fp16)[name = string("x1_105_cast_fp16")];
            tensor<int32, [4]> x2_105_begin_0 = const()[name = string("x2_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_105_end_0 = const()[name = string("x2_105_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_105_end_mask_0 = const()[name = string("x2_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_105_cast_fp16 = slice_by_index(begin = x2_105_begin_0, end = x2_105_end_0, end_mask = x2_105_end_mask_0, x = q_105_cast_fp16)[name = string("x2_105_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2667_cast_fp16 = mul(x = x2_105_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_2667_cast_fp16")];
            bool var_2669_interleave_0 = const()[name = string("op_2669_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2669_cast_fp16 = concat(axis = var_49, interleave = var_2669_interleave_0, values = (var_2667_cast_fp16, x1_105_cast_fp16))[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2670_cast_fp16 = mul(x = var_2669_cast_fp16, y = sin_11_to_fp16)[name = string("op_2670_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_53_cast_fp16 = add(x = var_2655_cast_fp16, y = var_2670_cast_fp16)[name = string("q_embed_53_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_105_cast_fp16 = transpose(perm = k_105_perm_0, x = squeeze_79_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 16, 192, 64]> var_2673_cast_fp16 = mul(x = k_105_cast_fp16, y = cos_11_to_fp16)[name = string("op_2673_cast_fp16")];
            tensor<int32, [4]> x1_107_begin_0 = const()[name = string("x1_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_107_end_0 = const()[name = string("x1_107_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_107_end_mask_0 = const()[name = string("x1_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_107_cast_fp16 = slice_by_index(begin = x1_107_begin_0, end = x1_107_end_0, end_mask = x1_107_end_mask_0, x = k_105_cast_fp16)[name = string("x1_107_cast_fp16")];
            tensor<int32, [4]> x2_107_begin_0 = const()[name = string("x2_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_107_end_0 = const()[name = string("x2_107_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_107_end_mask_0 = const()[name = string("x2_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_107_cast_fp16 = slice_by_index(begin = x2_107_begin_0, end = x2_107_end_0, end_mask = x2_107_end_mask_0, x = k_105_cast_fp16)[name = string("x2_107_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2685_cast_fp16 = mul(x = x2_107_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2685_cast_fp16")];
            bool var_2687_interleave_0 = const()[name = string("op_2687_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2687_cast_fp16 = concat(axis = var_49, interleave = var_2687_interleave_0, values = (var_2685_cast_fp16, x1_107_cast_fp16))[name = string("op_2687_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2688_cast_fp16 = mul(x = var_2687_cast_fp16, y = sin_11_to_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_53_cast_fp16 = add(x = var_2673_cast_fp16, y = var_2688_cast_fp16)[name = string("k_embed_53_cast_fp16")];
            bool var_2693_transpose_x_1 = const()[name = string("op_2693_transpose_x_1"), val = bool(false)];
            bool var_2693_transpose_y_1 = const()[name = string("op_2693_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_2693_cast_fp16 = matmul(transpose_x = var_2693_transpose_x_1, transpose_y = var_2693_transpose_y_1, x = q_embed_53_cast_fp16, y = k_embed_53_cast_fp16)[name = string("op_2693_cast_fp16")];
            fp16 var_2694_to_fp16 = const()[name = string("op_2694_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_105_cast_fp16 = mul(x = var_2693_cast_fp16, y = var_2694_to_fp16)[name = string("attn_weights_105_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_473_cast_fp16 = add(x = attn_weights_105_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_473_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_2697_cast_fp16 = softmax(axis = var_49, x = input_473_cast_fp16)[name = string("op_2697_cast_fp16")];
            bool attn_output_157_transpose_x_0 = const()[name = string("attn_output_157_transpose_x_0"), val = bool(false)];
            bool attn_output_157_transpose_y_0 = const()[name = string("attn_output_157_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = squeeze_80_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_157_cast_fp16 = matmul(transpose_x = attn_output_157_transpose_x_0, transpose_y = attn_output_157_transpose_y_0, x = var_2697_cast_fp16, y = value_53_cast_fp16)[name = string("attn_output_157_cast_fp16")];
            tensor<int32, [4]> var_2701_perm_0 = const()[name = string("op_2701_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2703 = const()[name = string("op_2703"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_2701_cast_fp16 = transpose(perm = var_2701_perm_0, x = attn_output_157_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 192, 1024]> var_2704_cast_fp16 = reshape(shape = var_2703, x = var_2701_cast_fp16)[name = string("op_2704_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_26_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_26_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746953472)))];
            tensor<fp16, [1, 192, 1024]> linear_105_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_26_attn_Wo_weight_to_fp16, x = var_2704_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_479_cast_fp16 = add(x = input_471_cast_fp16, y = linear_105_cast_fp16)[name = string("input_479_cast_fp16")];
            tensor<int32, [1]> input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_26_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_26_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749050688)))];
            tensor<fp16, [1, 192, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_26_mlp_norm_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_26_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_26_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749052800)))];
            tensor<fp16, [1, 192, 5248]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_26_mlp_Wi_weight_to_fp16, x = input_481_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [2]> var_2716_split_sizes_0 = const()[name = string("op_2716_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2716_axis_0 = const()[name = string("op_2716_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_2716_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_2716_cast_fp16_1 = split(axis = var_2716_axis_0, split_sizes = var_2716_split_sizes_0, x = linear_106_cast_fp16)[name = string("op_2716_cast_fp16")];
            string var_2718_mode_0 = const()[name = string("op_2718_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_2718_cast_fp16 = gelu(mode = var_2718_mode_0, x = var_2716_cast_fp16_0)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_485_cast_fp16 = mul(x = var_2718_cast_fp16, y = var_2716_cast_fp16_1)[name = string("input_485_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_26_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_26_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(759800768)))];
            tensor<fp16, [1, 192, 1024]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_26_mlp_Wo_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_489_cast_fp16 = add(x = input_479_cast_fp16, y = linear_107_cast_fp16)[name = string("input_489_cast_fp16")];
            tensor<int32, [1]> hidden_states_axes_0 = const()[name = string("hidden_states_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_27_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_27_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765174784)))];
            tensor<fp16, [1, 192, 1024]> hidden_states_cast_fp16 = layer_norm(axes = hidden_states_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_27_attn_norm_weight_to_fp16, x = input_489_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_27_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_27_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765176896)))];
            tensor<fp16, [1, 192, 3072]> linear_108_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_27_attn_Wqkv_weight_to_fp16, x = hidden_states_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<int32, [5]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [5]>([1, 192, 3, -1, 64])];
            tensor<fp16, [1, 192, 3, 16, 64]> qkv_cast_fp16 = reshape(shape = var_2737, x = linear_108_cast_fp16)[name = string("qkv_cast_fp16")];
            tensor<int32, [3]> var_2739_split_sizes_0 = const()[name = string("op_2739_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2739_axis_0 = const()[name = string("op_2739_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 192, 1, 16, 64]> var_2739_cast_fp16_0, tensor<fp16, [1, 192, 1, 16, 64]> var_2739_cast_fp16_1, tensor<fp16, [1, 192, 1, 16, 64]> var_2739_cast_fp16_2 = split(axis = var_2739_axis_0, split_sizes = var_2739_split_sizes_0, x = qkv_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<int32, [1]> squeeze_81_axes_0 = const()[name = string("squeeze_81_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_81_cast_fp16 = squeeze(axes = squeeze_81_axes_0, x = var_2739_cast_fp16_0)[name = string("squeeze_81_cast_fp16")];
            tensor<int32, [1]> squeeze_82_axes_0 = const()[name = string("squeeze_82_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_82_cast_fp16 = squeeze(axes = squeeze_82_axes_0, x = var_2739_cast_fp16_1)[name = string("squeeze_82_cast_fp16")];
            tensor<int32, [1]> squeeze_83_axes_0 = const()[name = string("squeeze_83_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 192, 16, 64]> squeeze_83_cast_fp16 = squeeze(axes = squeeze_83_axes_0, x = var_2739_cast_fp16_2)[name = string("squeeze_83_cast_fp16")];
            tensor<int32, [4]> q_109_perm_0 = const()[name = string("q_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_109_perm_0 = const()[name = string("k_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 192, 64]> q_109_cast_fp16 = transpose(perm = q_109_perm_0, x = squeeze_81_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 16, 192, 64]> var_2749_cast_fp16 = mul(x = q_109_cast_fp16, y = cos_7_to_fp16)[name = string("op_2749_cast_fp16")];
            tensor<int32, [4]> x1_109_begin_0 = const()[name = string("x1_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_109_end_0 = const()[name = string("x1_109_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_109_end_mask_0 = const()[name = string("x1_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_109_cast_fp16 = slice_by_index(begin = x1_109_begin_0, end = x1_109_end_0, end_mask = x1_109_end_mask_0, x = q_109_cast_fp16)[name = string("x1_109_cast_fp16")];
            tensor<int32, [4]> x2_109_begin_0 = const()[name = string("x2_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_109_end_0 = const()[name = string("x2_109_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_109_end_mask_0 = const()[name = string("x2_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_109_cast_fp16 = slice_by_index(begin = x2_109_begin_0, end = x2_109_end_0, end_mask = x2_109_end_mask_0, x = q_109_cast_fp16)[name = string("x2_109_cast_fp16")];
            fp16 const_239_promoted_to_fp16 = const()[name = string("const_239_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2761_cast_fp16 = mul(x = x2_109_cast_fp16, y = const_239_promoted_to_fp16)[name = string("op_2761_cast_fp16")];
            bool var_2763_interleave_0 = const()[name = string("op_2763_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2763_cast_fp16 = concat(axis = var_49, interleave = var_2763_interleave_0, values = (var_2761_cast_fp16, x1_109_cast_fp16))[name = string("op_2763_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2764_cast_fp16 = mul(x = var_2763_cast_fp16, y = sin_7_to_fp16)[name = string("op_2764_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> q_embed_cast_fp16 = add(x = var_2749_cast_fp16, y = var_2764_cast_fp16)[name = string("q_embed_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_109_cast_fp16 = transpose(perm = k_109_perm_0, x = squeeze_82_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 16, 192, 64]> var_2767_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_7_to_fp16)[name = string("op_2767_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 16, 192, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 192, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_109_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 16, 192, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 192, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_109_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_242_promoted_to_fp16 = const()[name = string("const_242_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 192, 32]> var_2779_cast_fp16 = mul(x = x2_cast_fp16, y = const_242_promoted_to_fp16)[name = string("op_2779_cast_fp16")];
            bool var_2781_interleave_0 = const()[name = string("op_2781_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> var_2781_cast_fp16 = concat(axis = var_49, interleave = var_2781_interleave_0, values = (var_2779_cast_fp16, x1_cast_fp16))[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> var_2782_cast_fp16 = mul(x = var_2781_cast_fp16, y = sin_7_to_fp16)[name = string("op_2782_cast_fp16")];
            tensor<fp16, [1, 16, 192, 64]> k_embed_cast_fp16 = add(x = var_2767_cast_fp16, y = var_2782_cast_fp16)[name = string("k_embed_cast_fp16")];
            bool var_2787_transpose_x_1 = const()[name = string("op_2787_transpose_x_1"), val = bool(false)];
            bool var_2787_transpose_y_1 = const()[name = string("op_2787_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 192, 192]> var_2787_cast_fp16 = matmul(transpose_x = var_2787_transpose_x_1, transpose_y = var_2787_transpose_y_1, x = q_embed_cast_fp16, y = k_embed_cast_fp16)[name = string("op_2787_cast_fp16")];
            fp16 var_2788_to_fp16 = const()[name = string("op_2788_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 192, 192]> attn_weights_109_cast_fp16 = mul(x = var_2787_cast_fp16, y = var_2788_to_fp16)[name = string("attn_weights_109_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> input_491_cast_fp16 = add(x = attn_weights_109_cast_fp16, y = key_bias_cast_fp16)[name = string("input_491_cast_fp16")];
            tensor<fp16, [1, 16, 192, 192]> var_2791_cast_fp16 = softmax(axis = var_49, x = input_491_cast_fp16)[name = string("op_2791_cast_fp16")];
            bool attn_output_163_transpose_x_0 = const()[name = string("attn_output_163_transpose_x_0"), val = bool(false)];
            bool attn_output_163_transpose_y_0 = const()[name = string("attn_output_163_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 192, 64]> value_cast_fp16 = transpose(perm = value_perm_0, x = squeeze_83_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 16, 192, 64]> attn_output_163_cast_fp16 = matmul(transpose_x = attn_output_163_transpose_x_0, transpose_y = attn_output_163_transpose_y_0, x = var_2791_cast_fp16, y = value_cast_fp16)[name = string("attn_output_163_cast_fp16")];
            tensor<int32, [4]> var_2795_perm_0 = const()[name = string("op_2795_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2797 = const()[name = string("op_2797"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, 16, 64]> var_2795_cast_fp16 = transpose(perm = var_2795_perm_0, x = attn_output_163_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 192, 1024]> var_2798_cast_fp16 = reshape(shape = var_2797, x = var_2795_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_27_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_27_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(771468416)))];
            tensor<fp16, [1, 192, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_27_attn_Wo_weight_to_fp16, x = var_2798_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_497_cast_fp16 = add(x = input_489_cast_fp16, y = linear_109_cast_fp16)[name = string("input_497_cast_fp16")];
            tensor<int32, [1]> input_499_axes_0 = const()[name = string("input_499_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_27_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_27_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773565632)))];
            tensor<fp16, [1, 192, 1024]> input_499_cast_fp16 = layer_norm(axes = input_499_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_27_mlp_norm_weight_to_fp16, x = input_497_cast_fp16)[name = string("input_499_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_27_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_27_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773567744)))];
            tensor<fp16, [1, 192, 5248]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_27_mlp_Wi_weight_to_fp16, x = input_499_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [2]> var_2810_split_sizes_0 = const()[name = string("op_2810_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2810_axis_0 = const()[name = string("op_2810_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 192, 2624]> var_2810_cast_fp16_0, tensor<fp16, [1, 192, 2624]> var_2810_cast_fp16_1 = split(axis = var_2810_axis_0, split_sizes = var_2810_split_sizes_0, x = linear_110_cast_fp16)[name = string("op_2810_cast_fp16")];
            string var_2812_mode_0 = const()[name = string("op_2812_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 192, 2624]> var_2812_cast_fp16 = gelu(mode = var_2812_mode_0, x = var_2810_cast_fp16_0)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [1, 192, 2624]> input_503_cast_fp16 = mul(x = var_2812_cast_fp16, y = var_2810_cast_fp16_1)[name = string("input_503_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_27_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_27_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(784315712)))];
            tensor<fp16, [1, 192, 1024]> linear_111_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_27_mlp_Wo_weight_to_fp16, x = input_503_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> input_507_cast_fp16 = add(x = input_497_cast_fp16, y = linear_111_cast_fp16)[name = string("input_507_cast_fp16")];
            tensor<int32, [1]> hidden_axes_0 = const()[name = string("hidden_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_final_norm_weight_to_fp16 = const()[name = string("model_model_final_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789689728)))];
            tensor<fp16, [1, 192, 1024]> hidden_cast_fp16 = layer_norm(axes = hidden_axes_0, epsilon = var_52_to_fp16, gamma = model_model_final_norm_weight_to_fp16, x = input_507_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_2822_axes_0 = const()[name = string("op_2822_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 192, 1]> var_2822_cast_fp16 = expand_dims(axes = var_2822_axes_0, x = attention_mask_to_fp16)[name = string("op_2822_cast_fp16")];
            tensor<fp16, [1, 192, 1024]> var_2823_cast_fp16 = mul(x = hidden_cast_fp16, y = var_2822_cast_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_2828_axes_0 = const()[name = string("op_2828_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_2828_keep_dims_0 = const()[name = string("op_2828_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 1024]> var_2828_cast_fp16 = reduce_sum(axes = var_2828_axes_0, keep_dims = var_2828_keep_dims_0, x = var_2823_cast_fp16)[name = string("op_2828_cast_fp16")];
            tensor<int32, [1]> var_2833_axes_0 = const()[name = string("op_2833_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_2833_keep_dims_0 = const()[name = string("op_2833_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_2833_cast_fp16 = reduce_sum(axes = var_2833_axes_0, keep_dims = var_2833_keep_dims_0, x = attention_mask_to_fp16)[name = string("op_2833_cast_fp16")];
            tensor<fp16, [1, 1024]> input_509_cast_fp16 = real_div(x = var_2828_cast_fp16, y = var_2833_cast_fp16)[name = string("input_509_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_head_dense_weight_to_fp16 = const()[name = string("model_head_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789691840)))];
            tensor<fp16, [1, 1024]> linear_112_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_head_dense_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_112_cast_fp16")];
            string input_513_mode_0 = const()[name = string("input_513_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024]> input_513_cast_fp16 = gelu(mode = input_513_mode_0, x = linear_112_cast_fp16)[name = string("input_513_cast_fp16")];
            tensor<int32, [1]> input_515_axes_0 = const()[name = string("input_515_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_head_norm_weight_to_fp16 = const()[name = string("model_head_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791789056)))];
            fp16 var_2836_to_fp16 = const()[name = string("op_2836_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024]> input_515_cast_fp16 = layer_norm(axes = input_515_axes_0, epsilon = var_2836_to_fp16, gamma = model_head_norm_weight_to_fp16, x = input_513_cast_fp16)[name = string("input_515_cast_fp16")];
            tensor<fp16, [3, 1024]> model_classifier_weight_to_fp16 = const()[name = string("model_classifier_weight_to_fp16"), val = tensor<fp16, [3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791791168)))];
            tensor<fp16, [3]> model_classifier_bias_to_fp16 = const()[name = string("model_classifier_bias_to_fp16"), val = tensor<fp16, [3]>([-0x1.284p-12, -0x1.f2p-17, 0x1.41p-12])];
            tensor<fp16, [1, 3]> linear_113_cast_fp16 = linear(bias = model_classifier_bias_to_fp16, weight = model_classifier_weight_to_fp16, x = input_515_cast_fp16)[name = string("linear_113_cast_fp16")];
            string linear_113_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_113_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 3]> logits = cast(dtype = linear_113_cast_fp16_to_fp32_dtype_0, x = linear_113_cast_fp16)[name = string("cast_59")];
        } -> (logits);
    func len224<ios18>(tensor<int32, [1, 224]> attention_mask, tensor<int32, [1, 224]> input_ids) {
            fp16 var_15_to_fp16 = const()[name = string("op_15_to_fp16"), val = fp16(0x1p+0)];
            string keys_to_fp16_dtype_0 = const()[name = string("keys_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 224]> attention_mask_to_fp16 = cast(dtype = keys_to_fp16_dtype_0, x = attention_mask)[name = string("cast_60")];
            tensor<fp16, [1, 224]> var_17_cast_fp16 = sub(x = var_15_to_fp16, y = attention_mask_to_fp16)[name = string("op_17_cast_fp16")];
            tensor<int32, [1]> var_24_axes_0 = const()[name = string("op_24_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 224]> var_24_cast_fp16 = expand_dims(axes = var_24_axes_0, x = var_17_cast_fp16)[name = string("op_24_cast_fp16")];
            tensor<int32, [1]> var_26_axes_0 = const()[name = string("op_26_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 224]> var_26_cast_fp16 = expand_dims(axes = var_26_axes_0, x = var_24_cast_fp16)[name = string("op_26_cast_fp16")];
            fp16 var_32_to_fp16 = const()[name = string("op_32_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 1, 1, 224]> key_bias_cast_fp16 = mul(x = var_26_cast_fp16, y = var_32_to_fp16)[name = string("key_bias_cast_fp16")];
            tensor<fp16, [1, 1, 224, 224]> outside_to_fp16 = const()[name = string("outside_to_fp16"), val = tensor<fp16, [1, 1, 224, 224]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791969728)))];
            tensor<fp16, [1, 1, 224, 224]> attention_mask_cast_fp16 = add(x = key_bias_cast_fp16, y = outside_to_fp16)[name = string("attention_mask_cast_fp16")];
            int32 var_49 = const()[name = string("op_49"), val = int32(-1)];
            int32 input_1_batch_dims_0 = const()[name = string("input_1_batch_dims_0"), val = int32(0)];
            bool input_1_validate_indices_0 = const()[name = string("input_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [50368, 1024]> model_model_embeddings_tok_embeddings_weight_to_fp16 = const()[name = string("model_model_embeddings_tok_embeddings_weight_to_fp16"), val = tensor<fp16, [50368, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32896)))];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 224]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(50368)];
            tensor<int32, [1, 224]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 224]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 input_1_cast_fp16_axis_0 = const()[name = string("input_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 224, 1024]> input_1_cast_fp16 = gather(axis = input_1_cast_fp16_axis_0, batch_dims = input_1_batch_dims_0, indices = select_0, validate_indices = input_1_validate_indices_0, x = model_model_embeddings_tok_embeddings_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<int32, [1]> input_3_axes_0 = const()[name = string("input_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_embeddings_norm_weight_to_fp16 = const()[name = string("model_model_embeddings_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103186624)))];
            fp16 var_52_to_fp16 = const()[name = string("op_52_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 224, 1024]> input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, epsilon = var_52_to_fp16, gamma = model_model_embeddings_norm_weight_to_fp16, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_0_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_0_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103188736)))];
            tensor<fp16, [3072]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109480256)))];
            tensor<fp16, [1, 224, 3072]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_0_attn_Wqkv_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [5]> var_199 = const()[name = string("op_199"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_3_cast_fp16 = reshape(shape = var_199, x = linear_0_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> var_201_split_sizes_0 = const()[name = string("op_201_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_201_axis_0 = const()[name = string("op_201_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_201_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_201_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_201_cast_fp16_2 = split(axis = var_201_axis_0, split_sizes = var_201_split_sizes_0, x = qkv_3_cast_fp16)[name = string("op_201_cast_fp16")];
            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = string("squeeze_0_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = var_201_cast_fp16_0)[name = string("squeeze_0_cast_fp16")];
            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = string("squeeze_1_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = var_201_cast_fp16_1)[name = string("squeeze_1_cast_fp16")];
            tensor<int32, [1]> squeeze_2_axes_0 = const()[name = string("squeeze_2_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_2_cast_fp16 = squeeze(axes = squeeze_2_axes_0, x = var_201_cast_fp16_2)[name = string("squeeze_2_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_1_perm_0 = const()[name = string("value_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 1, 224, 64]> cos_7_to_fp16 = const()[name = string("cos_7_to_fp16"), val = tensor<fp16, [1, 1, 224, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792070144)))];
            tensor<fp16, [1, 16, 224, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = squeeze_0_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 16, 224, 64]> var_211_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_7_to_fp16)[name = string("op_211_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_223_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_223_cast_fp16")];
            bool var_225_interleave_0 = const()[name = string("op_225_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_225_cast_fp16 = concat(axis = var_49, interleave = var_225_interleave_0, values = (var_223_cast_fp16, x1_1_cast_fp16))[name = string("op_225_cast_fp16")];
            tensor<fp16, [1, 1, 224, 64]> sin_7_to_fp16 = const()[name = string("sin_7_to_fp16"), val = tensor<fp16, [1, 1, 224, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792098880)))];
            tensor<fp16, [1, 16, 224, 64]> var_226_cast_fp16 = mul(x = var_225_cast_fp16, y = sin_7_to_fp16)[name = string("op_226_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_1_cast_fp16 = add(x = var_211_cast_fp16, y = var_226_cast_fp16)[name = string("q_embed_1_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = squeeze_1_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 224, 64]> var_229_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_7_to_fp16)[name = string("op_229_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_241_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_241_cast_fp16")];
            bool var_243_interleave_0 = const()[name = string("op_243_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_243_cast_fp16 = concat(axis = var_49, interleave = var_243_interleave_0, values = (var_241_cast_fp16, x1_3_cast_fp16))[name = string("op_243_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_244_cast_fp16 = mul(x = var_243_cast_fp16, y = sin_7_to_fp16)[name = string("op_244_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_1_cast_fp16 = add(x = var_229_cast_fp16, y = var_244_cast_fp16)[name = string("k_embed_1_cast_fp16")];
            bool var_249_transpose_x_1 = const()[name = string("op_249_transpose_x_1"), val = bool(false)];
            bool var_249_transpose_y_1 = const()[name = string("op_249_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_249_cast_fp16 = matmul(transpose_x = var_249_transpose_x_1, transpose_y = var_249_transpose_y_1, x = q_embed_1_cast_fp16, y = k_embed_1_cast_fp16)[name = string("op_249_cast_fp16")];
            fp16 var_250_to_fp16 = const()[name = string("op_250_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_1_cast_fp16 = mul(x = var_249_cast_fp16, y = var_250_to_fp16)[name = string("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_5_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = key_bias_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_253_cast_fp16 = softmax(axis = var_49, x = input_5_cast_fp16)[name = string("op_253_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_1_cast_fp16 = transpose(perm = value_1_perm_0, x = squeeze_2_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = var_253_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_257_perm_0 = const()[name = string("op_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_259 = const()[name = string("op_259"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_257_cast_fp16 = transpose(perm = var_257_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 224, 1024]> var_260_cast_fp16 = reshape(shape = var_259, x = var_257_cast_fp16)[name = string("op_260_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_0_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_0_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109519360)))];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111616576)))];
            tensor<fp16, [1, 224, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_0_attn_Wo_weight_to_fp16, x = var_260_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_11_cast_fp16 = add(x = input_3_cast_fp16, y = linear_1_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<int32, [1]> input_13_axes_0 = const()[name = string("input_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_0_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_0_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111618688)))];
            tensor<fp16, [1, 224, 1024]> input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_0_mlp_norm_weight_to_fp16, x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_0_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_0_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111620800)))];
            tensor<fp16, [5248]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5248]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122368768)))];
            tensor<fp16, [1, 224, 5248]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_0_mlp_Wi_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [2]> var_272_split_sizes_0 = const()[name = string("op_272_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_272_axis_0 = const()[name = string("op_272_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_272_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_272_cast_fp16_1 = split(axis = var_272_axis_0, split_sizes = var_272_split_sizes_0, x = linear_2_cast_fp16)[name = string("op_272_cast_fp16")];
            string var_274_mode_0 = const()[name = string("op_274_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_274_cast_fp16 = gelu(mode = var_274_mode_0, x = var_272_cast_fp16_0)[name = string("op_274_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_17_cast_fp16 = mul(x = var_274_cast_fp16, y = var_272_cast_fp16_1)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_0_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_0_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122379328)))];
            tensor<fp16, [1, 224, 1024]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_0_mlp_Wo_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_21_cast_fp16 = add(x = input_11_cast_fp16, y = linear_3_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [1]> hidden_states_3_axes_0 = const()[name = string("hidden_states_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_1_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_1_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127753344)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_3_cast_fp16 = layer_norm(axes = hidden_states_3_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_1_attn_norm_weight_to_fp16, x = input_21_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_1_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_1_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127755456)))];
            tensor<fp16, [1, 224, 3072]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_1_attn_Wqkv_weight_to_fp16, x = hidden_states_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [5]> var_293 = const()[name = string("op_293"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_7_cast_fp16 = reshape(shape = var_293, x = linear_4_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> var_295_split_sizes_0 = const()[name = string("op_295_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_295_axis_0 = const()[name = string("op_295_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_295_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_295_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_295_cast_fp16_2 = split(axis = var_295_axis_0, split_sizes = var_295_split_sizes_0, x = qkv_7_cast_fp16)[name = string("op_295_cast_fp16")];
            tensor<int32, [1]> squeeze_3_axes_0 = const()[name = string("squeeze_3_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_3_cast_fp16 = squeeze(axes = squeeze_3_axes_0, x = var_295_cast_fp16_0)[name = string("squeeze_3_cast_fp16")];
            tensor<int32, [1]> squeeze_4_axes_0 = const()[name = string("squeeze_4_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_4_cast_fp16 = squeeze(axes = squeeze_4_axes_0, x = var_295_cast_fp16_1)[name = string("squeeze_4_cast_fp16")];
            tensor<int32, [1]> squeeze_5_axes_0 = const()[name = string("squeeze_5_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_5_cast_fp16 = squeeze(axes = squeeze_5_axes_0, x = var_295_cast_fp16_2)[name = string("squeeze_5_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 1, 224, 64]> cos_11_to_fp16 = const()[name = string("cos_11_to_fp16"), val = tensor<fp16, [1, 1, 224, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792127616)))];
            tensor<fp16, [1, 16, 224, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = squeeze_3_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 16, 224, 64]> var_305_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_11_to_fp16)[name = string("op_305_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_5_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_5_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_317_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_317_cast_fp16")];
            bool var_319_interleave_0 = const()[name = string("op_319_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_319_cast_fp16 = concat(axis = var_49, interleave = var_319_interleave_0, values = (var_317_cast_fp16, x1_5_cast_fp16))[name = string("op_319_cast_fp16")];
            tensor<fp16, [1, 1, 224, 64]> sin_11_to_fp16 = const()[name = string("sin_11_to_fp16"), val = tensor<fp16, [1, 1, 224, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792156352)))];
            tensor<fp16, [1, 16, 224, 64]> var_320_cast_fp16 = mul(x = var_319_cast_fp16, y = sin_11_to_fp16)[name = string("op_320_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_3_cast_fp16 = add(x = var_305_cast_fp16, y = var_320_cast_fp16)[name = string("q_embed_3_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = squeeze_4_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 16, 224, 64]> var_323_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_11_to_fp16)[name = string("op_323_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_5_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_5_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_335_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_335_cast_fp16")];
            bool var_337_interleave_0 = const()[name = string("op_337_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_337_cast_fp16 = concat(axis = var_49, interleave = var_337_interleave_0, values = (var_335_cast_fp16, x1_7_cast_fp16))[name = string("op_337_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_338_cast_fp16 = mul(x = var_337_cast_fp16, y = sin_11_to_fp16)[name = string("op_338_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_3_cast_fp16 = add(x = var_323_cast_fp16, y = var_338_cast_fp16)[name = string("k_embed_3_cast_fp16")];
            bool var_343_transpose_x_1 = const()[name = string("op_343_transpose_x_1"), val = bool(false)];
            bool var_343_transpose_y_1 = const()[name = string("op_343_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_343_cast_fp16 = matmul(transpose_x = var_343_transpose_x_1, transpose_y = var_343_transpose_y_1, x = q_embed_3_cast_fp16, y = k_embed_3_cast_fp16)[name = string("op_343_cast_fp16")];
            fp16 var_344_to_fp16 = const()[name = string("op_344_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_5_cast_fp16 = mul(x = var_343_cast_fp16, y = var_344_to_fp16)[name = string("attn_weights_5_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_23_cast_fp16 = add(x = attn_weights_5_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_347_cast_fp16 = softmax(axis = var_49, x = input_23_cast_fp16)[name = string("op_347_cast_fp16")];
            bool attn_output_7_transpose_x_0 = const()[name = string("attn_output_7_transpose_x_0"), val = bool(false)];
            bool attn_output_7_transpose_y_0 = const()[name = string("attn_output_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = squeeze_5_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = var_347_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<int32, [4]> var_351_perm_0 = const()[name = string("op_351_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_353 = const()[name = string("op_353"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_351_cast_fp16 = transpose(perm = var_351_perm_0, x = attn_output_7_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 224, 1024]> var_354_cast_fp16 = reshape(shape = var_353, x = var_351_cast_fp16)[name = string("op_354_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_1_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_1_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134079872)))];
            tensor<fp16, [1, 224, 1024]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_1_attn_Wo_weight_to_fp16, x = var_354_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_29_cast_fp16 = add(x = input_21_cast_fp16, y = linear_5_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_1_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_1_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136177088)))];
            tensor<fp16, [1, 224, 1024]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_1_mlp_norm_weight_to_fp16, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_1_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_1_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136179200)))];
            tensor<fp16, [1, 224, 5248]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_1_mlp_Wi_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [2]> var_366_split_sizes_0 = const()[name = string("op_366_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_366_axis_0 = const()[name = string("op_366_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_366_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_366_cast_fp16_1 = split(axis = var_366_axis_0, split_sizes = var_366_split_sizes_0, x = linear_6_cast_fp16)[name = string("op_366_cast_fp16")];
            string var_368_mode_0 = const()[name = string("op_368_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_368_cast_fp16 = gelu(mode = var_368_mode_0, x = var_366_cast_fp16_0)[name = string("op_368_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_35_cast_fp16 = mul(x = var_368_cast_fp16, y = var_366_cast_fp16_1)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_1_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_1_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146927168)))];
            tensor<fp16, [1, 224, 1024]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_1_mlp_Wo_weight_to_fp16, x = input_35_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_39_cast_fp16 = add(x = input_29_cast_fp16, y = linear_7_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<int32, [1]> hidden_states_5_axes_0 = const()[name = string("hidden_states_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_2_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_2_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152301184)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_5_cast_fp16 = layer_norm(axes = hidden_states_5_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_2_attn_norm_weight_to_fp16, x = input_39_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_2_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_2_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152303296)))];
            tensor<fp16, [1, 224, 3072]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_2_attn_Wqkv_weight_to_fp16, x = hidden_states_5_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [5]> var_387 = const()[name = string("op_387"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_11_cast_fp16 = reshape(shape = var_387, x = linear_8_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> var_389_split_sizes_0 = const()[name = string("op_389_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_389_axis_0 = const()[name = string("op_389_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_389_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_389_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_389_cast_fp16_2 = split(axis = var_389_axis_0, split_sizes = var_389_split_sizes_0, x = qkv_11_cast_fp16)[name = string("op_389_cast_fp16")];
            tensor<int32, [1]> squeeze_6_axes_0 = const()[name = string("squeeze_6_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_6_cast_fp16 = squeeze(axes = squeeze_6_axes_0, x = var_389_cast_fp16_0)[name = string("squeeze_6_cast_fp16")];
            tensor<int32, [1]> squeeze_7_axes_0 = const()[name = string("squeeze_7_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_7_cast_fp16 = squeeze(axes = squeeze_7_axes_0, x = var_389_cast_fp16_1)[name = string("squeeze_7_cast_fp16")];
            tensor<int32, [1]> squeeze_8_axes_0 = const()[name = string("squeeze_8_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_8_cast_fp16 = squeeze(axes = squeeze_8_axes_0, x = var_389_cast_fp16_2)[name = string("squeeze_8_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = squeeze_6_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 16, 224, 64]> var_399_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_11_to_fp16)[name = string("op_399_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_9_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_9_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_411_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_411_cast_fp16")];
            bool var_413_interleave_0 = const()[name = string("op_413_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_413_cast_fp16 = concat(axis = var_49, interleave = var_413_interleave_0, values = (var_411_cast_fp16, x1_9_cast_fp16))[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_414_cast_fp16 = mul(x = var_413_cast_fp16, y = sin_11_to_fp16)[name = string("op_414_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_5_cast_fp16 = add(x = var_399_cast_fp16, y = var_414_cast_fp16)[name = string("q_embed_5_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = squeeze_7_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 16, 224, 64]> var_417_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_11_to_fp16)[name = string("op_417_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_9_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_9_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_429_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_429_cast_fp16")];
            bool var_431_interleave_0 = const()[name = string("op_431_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_431_cast_fp16 = concat(axis = var_49, interleave = var_431_interleave_0, values = (var_429_cast_fp16, x1_11_cast_fp16))[name = string("op_431_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_432_cast_fp16 = mul(x = var_431_cast_fp16, y = sin_11_to_fp16)[name = string("op_432_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_5_cast_fp16 = add(x = var_417_cast_fp16, y = var_432_cast_fp16)[name = string("k_embed_5_cast_fp16")];
            bool var_437_transpose_x_1 = const()[name = string("op_437_transpose_x_1"), val = bool(false)];
            bool var_437_transpose_y_1 = const()[name = string("op_437_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_437_cast_fp16 = matmul(transpose_x = var_437_transpose_x_1, transpose_y = var_437_transpose_y_1, x = q_embed_5_cast_fp16, y = k_embed_5_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 var_438_to_fp16 = const()[name = string("op_438_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_9_cast_fp16 = mul(x = var_437_cast_fp16, y = var_438_to_fp16)[name = string("attn_weights_9_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_41_cast_fp16 = add(x = attn_weights_9_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_441_cast_fp16 = softmax(axis = var_49, x = input_41_cast_fp16)[name = string("op_441_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = squeeze_8_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = var_441_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_445_perm_0 = const()[name = string("op_445_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_447 = const()[name = string("op_447"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_445_cast_fp16 = transpose(perm = var_445_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 224, 1024]> var_448_cast_fp16 = reshape(shape = var_447, x = var_445_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_2_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_2_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158594816)))];
            tensor<fp16, [1, 224, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_2_attn_Wo_weight_to_fp16, x = var_448_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_47_cast_fp16 = add(x = input_39_cast_fp16, y = linear_9_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<int32, [1]> input_49_axes_0 = const()[name = string("input_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_2_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_2_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160692032)))];
            tensor<fp16, [1, 224, 1024]> input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_2_mlp_norm_weight_to_fp16, x = input_47_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_2_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_2_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160694144)))];
            tensor<fp16, [1, 224, 5248]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_2_mlp_Wi_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [2]> var_460_split_sizes_0 = const()[name = string("op_460_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_460_axis_0 = const()[name = string("op_460_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_460_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_460_cast_fp16_1 = split(axis = var_460_axis_0, split_sizes = var_460_split_sizes_0, x = linear_10_cast_fp16)[name = string("op_460_cast_fp16")];
            string var_462_mode_0 = const()[name = string("op_462_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_462_cast_fp16 = gelu(mode = var_462_mode_0, x = var_460_cast_fp16_0)[name = string("op_462_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_53_cast_fp16 = mul(x = var_462_cast_fp16, y = var_460_cast_fp16_1)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_2_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_2_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171442112)))];
            tensor<fp16, [1, 224, 1024]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_2_mlp_Wo_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_57_cast_fp16 = add(x = input_47_cast_fp16, y = linear_11_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<int32, [1]> hidden_states_7_axes_0 = const()[name = string("hidden_states_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_3_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_3_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176816128)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_7_cast_fp16 = layer_norm(axes = hidden_states_7_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_3_attn_norm_weight_to_fp16, x = input_57_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_3_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_3_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176818240)))];
            tensor<fp16, [1, 224, 3072]> linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_3_attn_Wqkv_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [5]> var_481 = const()[name = string("op_481"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_15_cast_fp16 = reshape(shape = var_481, x = linear_12_cast_fp16)[name = string("qkv_15_cast_fp16")];
            tensor<int32, [3]> var_483_split_sizes_0 = const()[name = string("op_483_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_483_axis_0 = const()[name = string("op_483_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_483_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_483_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_483_cast_fp16_2 = split(axis = var_483_axis_0, split_sizes = var_483_split_sizes_0, x = qkv_15_cast_fp16)[name = string("op_483_cast_fp16")];
            tensor<int32, [1]> squeeze_9_axes_0 = const()[name = string("squeeze_9_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_9_cast_fp16 = squeeze(axes = squeeze_9_axes_0, x = var_483_cast_fp16_0)[name = string("squeeze_9_cast_fp16")];
            tensor<int32, [1]> squeeze_10_axes_0 = const()[name = string("squeeze_10_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_10_cast_fp16 = squeeze(axes = squeeze_10_axes_0, x = var_483_cast_fp16_1)[name = string("squeeze_10_cast_fp16")];
            tensor<int32, [1]> squeeze_11_axes_0 = const()[name = string("squeeze_11_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_11_cast_fp16 = squeeze(axes = squeeze_11_axes_0, x = var_483_cast_fp16_2)[name = string("squeeze_11_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_7_perm_0 = const()[name = string("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = squeeze_9_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 224, 64]> var_493_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_7_to_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_13_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_13_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_505_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_505_cast_fp16")];
            bool var_507_interleave_0 = const()[name = string("op_507_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_507_cast_fp16 = concat(axis = var_49, interleave = var_507_interleave_0, values = (var_505_cast_fp16, x1_13_cast_fp16))[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_508_cast_fp16 = mul(x = var_507_cast_fp16, y = sin_7_to_fp16)[name = string("op_508_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_7_cast_fp16 = add(x = var_493_cast_fp16, y = var_508_cast_fp16)[name = string("q_embed_7_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = squeeze_10_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 16, 224, 64]> var_511_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_7_to_fp16)[name = string("op_511_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_13_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_13_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_523_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_523_cast_fp16")];
            bool var_525_interleave_0 = const()[name = string("op_525_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_525_cast_fp16 = concat(axis = var_49, interleave = var_525_interleave_0, values = (var_523_cast_fp16, x1_15_cast_fp16))[name = string("op_525_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_526_cast_fp16 = mul(x = var_525_cast_fp16, y = sin_7_to_fp16)[name = string("op_526_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_7_cast_fp16 = add(x = var_511_cast_fp16, y = var_526_cast_fp16)[name = string("k_embed_7_cast_fp16")];
            bool var_531_transpose_x_1 = const()[name = string("op_531_transpose_x_1"), val = bool(false)];
            bool var_531_transpose_y_1 = const()[name = string("op_531_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_531_cast_fp16 = matmul(transpose_x = var_531_transpose_x_1, transpose_y = var_531_transpose_y_1, x = q_embed_7_cast_fp16, y = k_embed_7_cast_fp16)[name = string("op_531_cast_fp16")];
            fp16 var_532_to_fp16 = const()[name = string("op_532_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_13_cast_fp16 = mul(x = var_531_cast_fp16, y = var_532_to_fp16)[name = string("attn_weights_13_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_59_cast_fp16 = add(x = attn_weights_13_cast_fp16, y = key_bias_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_535_cast_fp16 = softmax(axis = var_49, x = input_59_cast_fp16)[name = string("op_535_cast_fp16")];
            bool attn_output_19_transpose_x_0 = const()[name = string("attn_output_19_transpose_x_0"), val = bool(false)];
            bool attn_output_19_transpose_y_0 = const()[name = string("attn_output_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = squeeze_11_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = var_535_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_539_perm_0 = const()[name = string("op_539_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_541 = const()[name = string("op_541"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_539_cast_fp16 = transpose(perm = var_539_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 224, 1024]> var_542_cast_fp16 = reshape(shape = var_541, x = var_539_cast_fp16)[name = string("op_542_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_3_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_3_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183109760)))];
            tensor<fp16, [1, 224, 1024]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_3_attn_Wo_weight_to_fp16, x = var_542_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_65_cast_fp16 = add(x = input_57_cast_fp16, y = linear_13_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_3_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_3_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185206976)))];
            tensor<fp16, [1, 224, 1024]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_3_mlp_norm_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_3_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_3_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185209088)))];
            tensor<fp16, [1, 224, 5248]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_3_mlp_Wi_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [2]> var_554_split_sizes_0 = const()[name = string("op_554_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_554_axis_0 = const()[name = string("op_554_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_554_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_554_cast_fp16_1 = split(axis = var_554_axis_0, split_sizes = var_554_split_sizes_0, x = linear_14_cast_fp16)[name = string("op_554_cast_fp16")];
            string var_556_mode_0 = const()[name = string("op_556_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_556_cast_fp16 = gelu(mode = var_556_mode_0, x = var_554_cast_fp16_0)[name = string("op_556_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_71_cast_fp16 = mul(x = var_556_cast_fp16, y = var_554_cast_fp16_1)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_3_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_3_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195957056)))];
            tensor<fp16, [1, 224, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_3_mlp_Wo_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_75_cast_fp16 = add(x = input_65_cast_fp16, y = linear_15_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<int32, [1]> hidden_states_9_axes_0 = const()[name = string("hidden_states_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_4_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_4_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201331072)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_9_cast_fp16 = layer_norm(axes = hidden_states_9_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_4_attn_norm_weight_to_fp16, x = input_75_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_4_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_4_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201333184)))];
            tensor<fp16, [1, 224, 3072]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_4_attn_Wqkv_weight_to_fp16, x = hidden_states_9_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [5]> var_575 = const()[name = string("op_575"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_19_cast_fp16 = reshape(shape = var_575, x = linear_16_cast_fp16)[name = string("qkv_19_cast_fp16")];
            tensor<int32, [3]> var_577_split_sizes_0 = const()[name = string("op_577_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_577_axis_0 = const()[name = string("op_577_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_577_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_577_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_577_cast_fp16_2 = split(axis = var_577_axis_0, split_sizes = var_577_split_sizes_0, x = qkv_19_cast_fp16)[name = string("op_577_cast_fp16")];
            tensor<int32, [1]> squeeze_12_axes_0 = const()[name = string("squeeze_12_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_12_cast_fp16 = squeeze(axes = squeeze_12_axes_0, x = var_577_cast_fp16_0)[name = string("squeeze_12_cast_fp16")];
            tensor<int32, [1]> squeeze_13_axes_0 = const()[name = string("squeeze_13_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_13_cast_fp16 = squeeze(axes = squeeze_13_axes_0, x = var_577_cast_fp16_1)[name = string("squeeze_13_cast_fp16")];
            tensor<int32, [1]> squeeze_14_axes_0 = const()[name = string("squeeze_14_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_14_cast_fp16 = squeeze(axes = squeeze_14_axes_0, x = var_577_cast_fp16_2)[name = string("squeeze_14_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = squeeze_12_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 224, 64]> var_587_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_11_to_fp16)[name = string("op_587_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_17_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_17_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_599_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_55_promoted_to_fp16)[name = string("op_599_cast_fp16")];
            bool var_601_interleave_0 = const()[name = string("op_601_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_601_cast_fp16 = concat(axis = var_49, interleave = var_601_interleave_0, values = (var_599_cast_fp16, x1_17_cast_fp16))[name = string("op_601_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_602_cast_fp16 = mul(x = var_601_cast_fp16, y = sin_11_to_fp16)[name = string("op_602_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_9_cast_fp16 = add(x = var_587_cast_fp16, y = var_602_cast_fp16)[name = string("q_embed_9_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = squeeze_13_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 224, 64]> var_605_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_11_to_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_17_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_17_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_58_promoted_to_fp16 = const()[name = string("const_58_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_617_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_58_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            bool var_619_interleave_0 = const()[name = string("op_619_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_619_cast_fp16 = concat(axis = var_49, interleave = var_619_interleave_0, values = (var_617_cast_fp16, x1_19_cast_fp16))[name = string("op_619_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_620_cast_fp16 = mul(x = var_619_cast_fp16, y = sin_11_to_fp16)[name = string("op_620_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_9_cast_fp16 = add(x = var_605_cast_fp16, y = var_620_cast_fp16)[name = string("k_embed_9_cast_fp16")];
            bool var_625_transpose_x_1 = const()[name = string("op_625_transpose_x_1"), val = bool(false)];
            bool var_625_transpose_y_1 = const()[name = string("op_625_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_625_cast_fp16 = matmul(transpose_x = var_625_transpose_x_1, transpose_y = var_625_transpose_y_1, x = q_embed_9_cast_fp16, y = k_embed_9_cast_fp16)[name = string("op_625_cast_fp16")];
            fp16 var_626_to_fp16 = const()[name = string("op_626_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_17_cast_fp16 = mul(x = var_625_cast_fp16, y = var_626_to_fp16)[name = string("attn_weights_17_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_77_cast_fp16 = add(x = attn_weights_17_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_629_cast_fp16 = softmax(axis = var_49, x = input_77_cast_fp16)[name = string("op_629_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = squeeze_14_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = var_629_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_633_perm_0 = const()[name = string("op_633_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_635 = const()[name = string("op_635"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_633_cast_fp16 = transpose(perm = var_633_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 224, 1024]> var_636_cast_fp16 = reshape(shape = var_635, x = var_633_cast_fp16)[name = string("op_636_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_4_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_4_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207624704)))];
            tensor<fp16, [1, 224, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_4_attn_Wo_weight_to_fp16, x = var_636_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_83_cast_fp16 = add(x = input_75_cast_fp16, y = linear_17_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<int32, [1]> input_85_axes_0 = const()[name = string("input_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_4_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_4_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209721920)))];
            tensor<fp16, [1, 224, 1024]> input_85_cast_fp16 = layer_norm(axes = input_85_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_4_mlp_norm_weight_to_fp16, x = input_83_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_4_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_4_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209724032)))];
            tensor<fp16, [1, 224, 5248]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_4_mlp_Wi_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [2]> var_648_split_sizes_0 = const()[name = string("op_648_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_648_axis_0 = const()[name = string("op_648_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_648_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_648_cast_fp16_1 = split(axis = var_648_axis_0, split_sizes = var_648_split_sizes_0, x = linear_18_cast_fp16)[name = string("op_648_cast_fp16")];
            string var_650_mode_0 = const()[name = string("op_650_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_650_cast_fp16 = gelu(mode = var_650_mode_0, x = var_648_cast_fp16_0)[name = string("op_650_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_89_cast_fp16 = mul(x = var_650_cast_fp16, y = var_648_cast_fp16_1)[name = string("input_89_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_4_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_4_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220472000)))];
            tensor<fp16, [1, 224, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_4_mlp_Wo_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_93_cast_fp16 = add(x = input_83_cast_fp16, y = linear_19_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<int32, [1]> hidden_states_11_axes_0 = const()[name = string("hidden_states_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_5_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_5_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225846016)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_11_cast_fp16 = layer_norm(axes = hidden_states_11_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_5_attn_norm_weight_to_fp16, x = input_93_cast_fp16)[name = string("hidden_states_11_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_5_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_5_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225848128)))];
            tensor<fp16, [1, 224, 3072]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_5_attn_Wqkv_weight_to_fp16, x = hidden_states_11_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [5]> var_669 = const()[name = string("op_669"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_23_cast_fp16 = reshape(shape = var_669, x = linear_20_cast_fp16)[name = string("qkv_23_cast_fp16")];
            tensor<int32, [3]> var_671_split_sizes_0 = const()[name = string("op_671_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_671_axis_0 = const()[name = string("op_671_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_671_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_671_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_671_cast_fp16_2 = split(axis = var_671_axis_0, split_sizes = var_671_split_sizes_0, x = qkv_23_cast_fp16)[name = string("op_671_cast_fp16")];
            tensor<int32, [1]> squeeze_15_axes_0 = const()[name = string("squeeze_15_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_15_cast_fp16 = squeeze(axes = squeeze_15_axes_0, x = var_671_cast_fp16_0)[name = string("squeeze_15_cast_fp16")];
            tensor<int32, [1]> squeeze_16_axes_0 = const()[name = string("squeeze_16_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_16_cast_fp16 = squeeze(axes = squeeze_16_axes_0, x = var_671_cast_fp16_1)[name = string("squeeze_16_cast_fp16")];
            tensor<int32, [1]> squeeze_17_axes_0 = const()[name = string("squeeze_17_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_17_cast_fp16 = squeeze(axes = squeeze_17_axes_0, x = var_671_cast_fp16_2)[name = string("squeeze_17_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = squeeze_15_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 16, 224, 64]> var_681_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_11_to_fp16)[name = string("op_681_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_21_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_21_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_63_promoted_to_fp16 = const()[name = string("const_63_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_693_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_63_promoted_to_fp16)[name = string("op_693_cast_fp16")];
            bool var_695_interleave_0 = const()[name = string("op_695_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_695_cast_fp16 = concat(axis = var_49, interleave = var_695_interleave_0, values = (var_693_cast_fp16, x1_21_cast_fp16))[name = string("op_695_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_696_cast_fp16 = mul(x = var_695_cast_fp16, y = sin_11_to_fp16)[name = string("op_696_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_11_cast_fp16 = add(x = var_681_cast_fp16, y = var_696_cast_fp16)[name = string("q_embed_11_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = squeeze_16_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 224, 64]> var_699_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_11_to_fp16)[name = string("op_699_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_21_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_21_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_66_promoted_to_fp16 = const()[name = string("const_66_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_711_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_66_promoted_to_fp16)[name = string("op_711_cast_fp16")];
            bool var_713_interleave_0 = const()[name = string("op_713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_713_cast_fp16 = concat(axis = var_49, interleave = var_713_interleave_0, values = (var_711_cast_fp16, x1_23_cast_fp16))[name = string("op_713_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_714_cast_fp16 = mul(x = var_713_cast_fp16, y = sin_11_to_fp16)[name = string("op_714_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_11_cast_fp16 = add(x = var_699_cast_fp16, y = var_714_cast_fp16)[name = string("k_embed_11_cast_fp16")];
            bool var_719_transpose_x_1 = const()[name = string("op_719_transpose_x_1"), val = bool(false)];
            bool var_719_transpose_y_1 = const()[name = string("op_719_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_719_cast_fp16 = matmul(transpose_x = var_719_transpose_x_1, transpose_y = var_719_transpose_y_1, x = q_embed_11_cast_fp16, y = k_embed_11_cast_fp16)[name = string("op_719_cast_fp16")];
            fp16 var_720_to_fp16 = const()[name = string("op_720_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_21_cast_fp16 = mul(x = var_719_cast_fp16, y = var_720_to_fp16)[name = string("attn_weights_21_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_95_cast_fp16 = add(x = attn_weights_21_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_723_cast_fp16 = softmax(axis = var_49, x = input_95_cast_fp16)[name = string("op_723_cast_fp16")];
            bool attn_output_31_transpose_x_0 = const()[name = string("attn_output_31_transpose_x_0"), val = bool(false)];
            bool attn_output_31_transpose_y_0 = const()[name = string("attn_output_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = squeeze_17_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = var_723_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [4]> var_727_perm_0 = const()[name = string("op_727_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_729 = const()[name = string("op_729"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_727_cast_fp16 = transpose(perm = var_727_perm_0, x = attn_output_31_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 224, 1024]> var_730_cast_fp16 = reshape(shape = var_729, x = var_727_cast_fp16)[name = string("op_730_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_5_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_5_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232139648)))];
            tensor<fp16, [1, 224, 1024]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_5_attn_Wo_weight_to_fp16, x = var_730_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_101_cast_fp16 = add(x = input_93_cast_fp16, y = linear_21_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<int32, [1]> input_103_axes_0 = const()[name = string("input_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_5_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_5_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234236864)))];
            tensor<fp16, [1, 224, 1024]> input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_5_mlp_norm_weight_to_fp16, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_5_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_5_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234238976)))];
            tensor<fp16, [1, 224, 5248]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_5_mlp_Wi_weight_to_fp16, x = input_103_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [2]> var_742_split_sizes_0 = const()[name = string("op_742_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_742_axis_0 = const()[name = string("op_742_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_742_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_742_cast_fp16_1 = split(axis = var_742_axis_0, split_sizes = var_742_split_sizes_0, x = linear_22_cast_fp16)[name = string("op_742_cast_fp16")];
            string var_744_mode_0 = const()[name = string("op_744_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_744_cast_fp16 = gelu(mode = var_744_mode_0, x = var_742_cast_fp16_0)[name = string("op_744_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_107_cast_fp16 = mul(x = var_744_cast_fp16, y = var_742_cast_fp16_1)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_5_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_5_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244986944)))];
            tensor<fp16, [1, 224, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_5_mlp_Wo_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_111_cast_fp16 = add(x = input_101_cast_fp16, y = linear_23_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<int32, [1]> hidden_states_13_axes_0 = const()[name = string("hidden_states_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_6_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_6_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250360960)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_13_cast_fp16 = layer_norm(axes = hidden_states_13_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_6_attn_norm_weight_to_fp16, x = input_111_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_6_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_6_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250363072)))];
            tensor<fp16, [1, 224, 3072]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_6_attn_Wqkv_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [5]> var_763 = const()[name = string("op_763"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_27_cast_fp16 = reshape(shape = var_763, x = linear_24_cast_fp16)[name = string("qkv_27_cast_fp16")];
            tensor<int32, [3]> var_765_split_sizes_0 = const()[name = string("op_765_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_765_axis_0 = const()[name = string("op_765_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_765_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_765_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_765_cast_fp16_2 = split(axis = var_765_axis_0, split_sizes = var_765_split_sizes_0, x = qkv_27_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<int32, [1]> squeeze_18_axes_0 = const()[name = string("squeeze_18_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_18_cast_fp16 = squeeze(axes = squeeze_18_axes_0, x = var_765_cast_fp16_0)[name = string("squeeze_18_cast_fp16")];
            tensor<int32, [1]> squeeze_19_axes_0 = const()[name = string("squeeze_19_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_19_cast_fp16 = squeeze(axes = squeeze_19_axes_0, x = var_765_cast_fp16_1)[name = string("squeeze_19_cast_fp16")];
            tensor<int32, [1]> squeeze_20_axes_0 = const()[name = string("squeeze_20_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_20_cast_fp16 = squeeze(axes = squeeze_20_axes_0, x = var_765_cast_fp16_2)[name = string("squeeze_20_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = squeeze_18_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 16, 224, 64]> var_775_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_7_to_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_25_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_25_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_71_promoted_to_fp16 = const()[name = string("const_71_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_787_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_71_promoted_to_fp16)[name = string("op_787_cast_fp16")];
            bool var_789_interleave_0 = const()[name = string("op_789_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_789_cast_fp16 = concat(axis = var_49, interleave = var_789_interleave_0, values = (var_787_cast_fp16, x1_25_cast_fp16))[name = string("op_789_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_790_cast_fp16 = mul(x = var_789_cast_fp16, y = sin_7_to_fp16)[name = string("op_790_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_13_cast_fp16 = add(x = var_775_cast_fp16, y = var_790_cast_fp16)[name = string("q_embed_13_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = squeeze_19_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 16, 224, 64]> var_793_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_7_to_fp16)[name = string("op_793_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_25_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_25_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_805_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_74_promoted_to_fp16)[name = string("op_805_cast_fp16")];
            bool var_807_interleave_0 = const()[name = string("op_807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_807_cast_fp16 = concat(axis = var_49, interleave = var_807_interleave_0, values = (var_805_cast_fp16, x1_27_cast_fp16))[name = string("op_807_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_808_cast_fp16 = mul(x = var_807_cast_fp16, y = sin_7_to_fp16)[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_13_cast_fp16 = add(x = var_793_cast_fp16, y = var_808_cast_fp16)[name = string("k_embed_13_cast_fp16")];
            bool var_813_transpose_x_1 = const()[name = string("op_813_transpose_x_1"), val = bool(false)];
            bool var_813_transpose_y_1 = const()[name = string("op_813_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_813_cast_fp16 = matmul(transpose_x = var_813_transpose_x_1, transpose_y = var_813_transpose_y_1, x = q_embed_13_cast_fp16, y = k_embed_13_cast_fp16)[name = string("op_813_cast_fp16")];
            fp16 var_814_to_fp16 = const()[name = string("op_814_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_25_cast_fp16 = mul(x = var_813_cast_fp16, y = var_814_to_fp16)[name = string("attn_weights_25_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_113_cast_fp16 = add(x = attn_weights_25_cast_fp16, y = key_bias_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_817_cast_fp16 = softmax(axis = var_49, x = input_113_cast_fp16)[name = string("op_817_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = squeeze_20_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = var_817_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_821_perm_0 = const()[name = string("op_821_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_823 = const()[name = string("op_823"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_821_cast_fp16 = transpose(perm = var_821_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 224, 1024]> var_824_cast_fp16 = reshape(shape = var_823, x = var_821_cast_fp16)[name = string("op_824_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_6_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_6_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256654592)))];
            tensor<fp16, [1, 224, 1024]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_6_attn_Wo_weight_to_fp16, x = var_824_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_119_cast_fp16 = add(x = input_111_cast_fp16, y = linear_25_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_6_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_6_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258751808)))];
            tensor<fp16, [1, 224, 1024]> input_121_cast_fp16 = layer_norm(axes = input_121_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_6_mlp_norm_weight_to_fp16, x = input_119_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_6_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_6_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258753920)))];
            tensor<fp16, [1, 224, 5248]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_6_mlp_Wi_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [2]> var_836_split_sizes_0 = const()[name = string("op_836_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_836_axis_0 = const()[name = string("op_836_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_836_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_836_cast_fp16_1 = split(axis = var_836_axis_0, split_sizes = var_836_split_sizes_0, x = linear_26_cast_fp16)[name = string("op_836_cast_fp16")];
            string var_838_mode_0 = const()[name = string("op_838_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_838_cast_fp16 = gelu(mode = var_838_mode_0, x = var_836_cast_fp16_0)[name = string("op_838_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_125_cast_fp16 = mul(x = var_838_cast_fp16, y = var_836_cast_fp16_1)[name = string("input_125_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_6_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_6_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269501888)))];
            tensor<fp16, [1, 224, 1024]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_6_mlp_Wo_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_129_cast_fp16 = add(x = input_119_cast_fp16, y = linear_27_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<int32, [1]> hidden_states_15_axes_0 = const()[name = string("hidden_states_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_7_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_7_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274875904)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_15_cast_fp16 = layer_norm(axes = hidden_states_15_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_7_attn_norm_weight_to_fp16, x = input_129_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_7_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_7_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274878016)))];
            tensor<fp16, [1, 224, 3072]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_7_attn_Wqkv_weight_to_fp16, x = hidden_states_15_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [5]> var_857 = const()[name = string("op_857"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_31_cast_fp16 = reshape(shape = var_857, x = linear_28_cast_fp16)[name = string("qkv_31_cast_fp16")];
            tensor<int32, [3]> var_859_split_sizes_0 = const()[name = string("op_859_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_859_axis_0 = const()[name = string("op_859_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_859_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_859_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_859_cast_fp16_2 = split(axis = var_859_axis_0, split_sizes = var_859_split_sizes_0, x = qkv_31_cast_fp16)[name = string("op_859_cast_fp16")];
            tensor<int32, [1]> squeeze_21_axes_0 = const()[name = string("squeeze_21_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_21_cast_fp16 = squeeze(axes = squeeze_21_axes_0, x = var_859_cast_fp16_0)[name = string("squeeze_21_cast_fp16")];
            tensor<int32, [1]> squeeze_22_axes_0 = const()[name = string("squeeze_22_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_22_cast_fp16 = squeeze(axes = squeeze_22_axes_0, x = var_859_cast_fp16_1)[name = string("squeeze_22_cast_fp16")];
            tensor<int32, [1]> squeeze_23_axes_0 = const()[name = string("squeeze_23_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_23_cast_fp16 = squeeze(axes = squeeze_23_axes_0, x = var_859_cast_fp16_2)[name = string("squeeze_23_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = squeeze_21_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 16, 224, 64]> var_869_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_11_to_fp16)[name = string("op_869_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_29_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_29_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_79_promoted_to_fp16 = const()[name = string("const_79_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_881_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_79_promoted_to_fp16)[name = string("op_881_cast_fp16")];
            bool var_883_interleave_0 = const()[name = string("op_883_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_883_cast_fp16 = concat(axis = var_49, interleave = var_883_interleave_0, values = (var_881_cast_fp16, x1_29_cast_fp16))[name = string("op_883_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_884_cast_fp16 = mul(x = var_883_cast_fp16, y = sin_11_to_fp16)[name = string("op_884_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_15_cast_fp16 = add(x = var_869_cast_fp16, y = var_884_cast_fp16)[name = string("q_embed_15_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = squeeze_22_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 16, 224, 64]> var_887_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_11_to_fp16)[name = string("op_887_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_29_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_29_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_899_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_899_cast_fp16")];
            bool var_901_interleave_0 = const()[name = string("op_901_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_901_cast_fp16 = concat(axis = var_49, interleave = var_901_interleave_0, values = (var_899_cast_fp16, x1_31_cast_fp16))[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_902_cast_fp16 = mul(x = var_901_cast_fp16, y = sin_11_to_fp16)[name = string("op_902_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_15_cast_fp16 = add(x = var_887_cast_fp16, y = var_902_cast_fp16)[name = string("k_embed_15_cast_fp16")];
            bool var_907_transpose_x_1 = const()[name = string("op_907_transpose_x_1"), val = bool(false)];
            bool var_907_transpose_y_1 = const()[name = string("op_907_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_907_cast_fp16 = matmul(transpose_x = var_907_transpose_x_1, transpose_y = var_907_transpose_y_1, x = q_embed_15_cast_fp16, y = k_embed_15_cast_fp16)[name = string("op_907_cast_fp16")];
            fp16 var_908_to_fp16 = const()[name = string("op_908_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_29_cast_fp16 = mul(x = var_907_cast_fp16, y = var_908_to_fp16)[name = string("attn_weights_29_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_131_cast_fp16 = add(x = attn_weights_29_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_911_cast_fp16 = softmax(axis = var_49, x = input_131_cast_fp16)[name = string("op_911_cast_fp16")];
            bool attn_output_43_transpose_x_0 = const()[name = string("attn_output_43_transpose_x_0"), val = bool(false)];
            bool attn_output_43_transpose_y_0 = const()[name = string("attn_output_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = squeeze_23_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = var_911_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_915_perm_0 = const()[name = string("op_915_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_917 = const()[name = string("op_917"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_915_cast_fp16 = transpose(perm = var_915_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 224, 1024]> var_918_cast_fp16 = reshape(shape = var_917, x = var_915_cast_fp16)[name = string("op_918_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_7_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_7_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281169536)))];
            tensor<fp16, [1, 224, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_7_attn_Wo_weight_to_fp16, x = var_918_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_137_cast_fp16 = add(x = input_129_cast_fp16, y = linear_29_cast_fp16)[name = string("input_137_cast_fp16")];
            tensor<int32, [1]> input_139_axes_0 = const()[name = string("input_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_7_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_7_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283266752)))];
            tensor<fp16, [1, 224, 1024]> input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_7_mlp_norm_weight_to_fp16, x = input_137_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_7_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_7_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283268864)))];
            tensor<fp16, [1, 224, 5248]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_7_mlp_Wi_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [2]> var_930_split_sizes_0 = const()[name = string("op_930_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_930_axis_0 = const()[name = string("op_930_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_930_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_930_cast_fp16_1 = split(axis = var_930_axis_0, split_sizes = var_930_split_sizes_0, x = linear_30_cast_fp16)[name = string("op_930_cast_fp16")];
            string var_932_mode_0 = const()[name = string("op_932_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_932_cast_fp16 = gelu(mode = var_932_mode_0, x = var_930_cast_fp16_0)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_143_cast_fp16 = mul(x = var_932_cast_fp16, y = var_930_cast_fp16_1)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_7_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_7_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294016832)))];
            tensor<fp16, [1, 224, 1024]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_7_mlp_Wo_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_147_cast_fp16 = add(x = input_137_cast_fp16, y = linear_31_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<int32, [1]> hidden_states_17_axes_0 = const()[name = string("hidden_states_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_8_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_8_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299390848)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_17_cast_fp16 = layer_norm(axes = hidden_states_17_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_8_attn_norm_weight_to_fp16, x = input_147_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_8_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_8_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299392960)))];
            tensor<fp16, [1, 224, 3072]> linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_8_attn_Wqkv_weight_to_fp16, x = hidden_states_17_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<int32, [5]> var_951 = const()[name = string("op_951"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_35_cast_fp16 = reshape(shape = var_951, x = linear_32_cast_fp16)[name = string("qkv_35_cast_fp16")];
            tensor<int32, [3]> var_953_split_sizes_0 = const()[name = string("op_953_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_953_axis_0 = const()[name = string("op_953_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_953_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_953_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_953_cast_fp16_2 = split(axis = var_953_axis_0, split_sizes = var_953_split_sizes_0, x = qkv_35_cast_fp16)[name = string("op_953_cast_fp16")];
            tensor<int32, [1]> squeeze_24_axes_0 = const()[name = string("squeeze_24_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_24_cast_fp16 = squeeze(axes = squeeze_24_axes_0, x = var_953_cast_fp16_0)[name = string("squeeze_24_cast_fp16")];
            tensor<int32, [1]> squeeze_25_axes_0 = const()[name = string("squeeze_25_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_25_cast_fp16 = squeeze(axes = squeeze_25_axes_0, x = var_953_cast_fp16_1)[name = string("squeeze_25_cast_fp16")];
            tensor<int32, [1]> squeeze_26_axes_0 = const()[name = string("squeeze_26_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_26_cast_fp16 = squeeze(axes = squeeze_26_axes_0, x = var_953_cast_fp16_2)[name = string("squeeze_26_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = squeeze_24_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 224, 64]> var_963_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_11_to_fp16)[name = string("op_963_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_33_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_33_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_975_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_975_cast_fp16")];
            bool var_977_interleave_0 = const()[name = string("op_977_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_977_cast_fp16 = concat(axis = var_49, interleave = var_977_interleave_0, values = (var_975_cast_fp16, x1_33_cast_fp16))[name = string("op_977_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_978_cast_fp16 = mul(x = var_977_cast_fp16, y = sin_11_to_fp16)[name = string("op_978_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_17_cast_fp16 = add(x = var_963_cast_fp16, y = var_978_cast_fp16)[name = string("q_embed_17_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = squeeze_25_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 16, 224, 64]> var_981_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_11_to_fp16)[name = string("op_981_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_33_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_33_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_993_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_993_cast_fp16")];
            bool var_995_interleave_0 = const()[name = string("op_995_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_995_cast_fp16 = concat(axis = var_49, interleave = var_995_interleave_0, values = (var_993_cast_fp16, x1_35_cast_fp16))[name = string("op_995_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_996_cast_fp16 = mul(x = var_995_cast_fp16, y = sin_11_to_fp16)[name = string("op_996_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_17_cast_fp16 = add(x = var_981_cast_fp16, y = var_996_cast_fp16)[name = string("k_embed_17_cast_fp16")];
            bool var_1001_transpose_x_1 = const()[name = string("op_1001_transpose_x_1"), val = bool(false)];
            bool var_1001_transpose_y_1 = const()[name = string("op_1001_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1001_cast_fp16 = matmul(transpose_x = var_1001_transpose_x_1, transpose_y = var_1001_transpose_y_1, x = q_embed_17_cast_fp16, y = k_embed_17_cast_fp16)[name = string("op_1001_cast_fp16")];
            fp16 var_1002_to_fp16 = const()[name = string("op_1002_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_33_cast_fp16 = mul(x = var_1001_cast_fp16, y = var_1002_to_fp16)[name = string("attn_weights_33_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_149_cast_fp16 = add(x = attn_weights_33_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1005_cast_fp16 = softmax(axis = var_49, x = input_149_cast_fp16)[name = string("op_1005_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = squeeze_26_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = var_1005_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_1009_perm_0 = const()[name = string("op_1009_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1011 = const()[name = string("op_1011"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1009_cast_fp16 = transpose(perm = var_1009_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 224, 1024]> var_1012_cast_fp16 = reshape(shape = var_1011, x = var_1009_cast_fp16)[name = string("op_1012_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_8_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_8_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305684480)))];
            tensor<fp16, [1, 224, 1024]> linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_8_attn_Wo_weight_to_fp16, x = var_1012_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_155_cast_fp16 = add(x = input_147_cast_fp16, y = linear_33_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<int32, [1]> input_157_axes_0 = const()[name = string("input_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_8_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_8_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307781696)))];
            tensor<fp16, [1, 224, 1024]> input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_8_mlp_norm_weight_to_fp16, x = input_155_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_8_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_8_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307783808)))];
            tensor<fp16, [1, 224, 5248]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_8_mlp_Wi_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<int32, [2]> var_1024_split_sizes_0 = const()[name = string("op_1024_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1024_axis_0 = const()[name = string("op_1024_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1024_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1024_cast_fp16_1 = split(axis = var_1024_axis_0, split_sizes = var_1024_split_sizes_0, x = linear_34_cast_fp16)[name = string("op_1024_cast_fp16")];
            string var_1026_mode_0 = const()[name = string("op_1026_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1026_cast_fp16 = gelu(mode = var_1026_mode_0, x = var_1024_cast_fp16_0)[name = string("op_1026_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_161_cast_fp16 = mul(x = var_1026_cast_fp16, y = var_1024_cast_fp16_1)[name = string("input_161_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_8_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_8_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318531776)))];
            tensor<fp16, [1, 224, 1024]> linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_8_mlp_Wo_weight_to_fp16, x = input_161_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_165_cast_fp16 = add(x = input_155_cast_fp16, y = linear_35_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<int32, [1]> hidden_states_19_axes_0 = const()[name = string("hidden_states_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_9_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_9_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323905792)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_19_cast_fp16 = layer_norm(axes = hidden_states_19_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_9_attn_norm_weight_to_fp16, x = input_165_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_9_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_9_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323907904)))];
            tensor<fp16, [1, 224, 3072]> linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_9_attn_Wqkv_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [5]> var_1045 = const()[name = string("op_1045"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_39_cast_fp16 = reshape(shape = var_1045, x = linear_36_cast_fp16)[name = string("qkv_39_cast_fp16")];
            tensor<int32, [3]> var_1047_split_sizes_0 = const()[name = string("op_1047_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1047_axis_0 = const()[name = string("op_1047_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1047_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1047_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1047_cast_fp16_2 = split(axis = var_1047_axis_0, split_sizes = var_1047_split_sizes_0, x = qkv_39_cast_fp16)[name = string("op_1047_cast_fp16")];
            tensor<int32, [1]> squeeze_27_axes_0 = const()[name = string("squeeze_27_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_27_cast_fp16 = squeeze(axes = squeeze_27_axes_0, x = var_1047_cast_fp16_0)[name = string("squeeze_27_cast_fp16")];
            tensor<int32, [1]> squeeze_28_axes_0 = const()[name = string("squeeze_28_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_28_cast_fp16 = squeeze(axes = squeeze_28_axes_0, x = var_1047_cast_fp16_1)[name = string("squeeze_28_cast_fp16")];
            tensor<int32, [1]> squeeze_29_axes_0 = const()[name = string("squeeze_29_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_29_cast_fp16 = squeeze(axes = squeeze_29_axes_0, x = var_1047_cast_fp16_2)[name = string("squeeze_29_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = squeeze_27_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 224, 64]> var_1057_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_7_to_fp16)[name = string("op_1057_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_37_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_37_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1069_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_95_promoted_to_fp16)[name = string("op_1069_cast_fp16")];
            bool var_1071_interleave_0 = const()[name = string("op_1071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1071_cast_fp16 = concat(axis = var_49, interleave = var_1071_interleave_0, values = (var_1069_cast_fp16, x1_37_cast_fp16))[name = string("op_1071_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1072_cast_fp16 = mul(x = var_1071_cast_fp16, y = sin_7_to_fp16)[name = string("op_1072_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_19_cast_fp16 = add(x = var_1057_cast_fp16, y = var_1072_cast_fp16)[name = string("q_embed_19_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = squeeze_28_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 224, 64]> var_1075_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_7_to_fp16)[name = string("op_1075_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_37_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_37_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_98_promoted_to_fp16 = const()[name = string("const_98_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1087_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_98_promoted_to_fp16)[name = string("op_1087_cast_fp16")];
            bool var_1089_interleave_0 = const()[name = string("op_1089_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1089_cast_fp16 = concat(axis = var_49, interleave = var_1089_interleave_0, values = (var_1087_cast_fp16, x1_39_cast_fp16))[name = string("op_1089_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1090_cast_fp16 = mul(x = var_1089_cast_fp16, y = sin_7_to_fp16)[name = string("op_1090_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_19_cast_fp16 = add(x = var_1075_cast_fp16, y = var_1090_cast_fp16)[name = string("k_embed_19_cast_fp16")];
            bool var_1095_transpose_x_1 = const()[name = string("op_1095_transpose_x_1"), val = bool(false)];
            bool var_1095_transpose_y_1 = const()[name = string("op_1095_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1095_cast_fp16 = matmul(transpose_x = var_1095_transpose_x_1, transpose_y = var_1095_transpose_y_1, x = q_embed_19_cast_fp16, y = k_embed_19_cast_fp16)[name = string("op_1095_cast_fp16")];
            fp16 var_1096_to_fp16 = const()[name = string("op_1096_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_37_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_1096_to_fp16)[name = string("attn_weights_37_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_167_cast_fp16 = add(x = attn_weights_37_cast_fp16, y = key_bias_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1099_cast_fp16 = softmax(axis = var_49, x = input_167_cast_fp16)[name = string("op_1099_cast_fp16")];
            bool attn_output_55_transpose_x_0 = const()[name = string("attn_output_55_transpose_x_0"), val = bool(false)];
            bool attn_output_55_transpose_y_0 = const()[name = string("attn_output_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = squeeze_29_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_55_cast_fp16 = matmul(transpose_x = attn_output_55_transpose_x_0, transpose_y = attn_output_55_transpose_y_0, x = var_1099_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [4]> var_1103_perm_0 = const()[name = string("op_1103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1105 = const()[name = string("op_1105"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1103_cast_fp16 = transpose(perm = var_1103_perm_0, x = attn_output_55_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 224, 1024]> var_1106_cast_fp16 = reshape(shape = var_1105, x = var_1103_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_9_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_9_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330199424)))];
            tensor<fp16, [1, 224, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_9_attn_Wo_weight_to_fp16, x = var_1106_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_173_cast_fp16 = add(x = input_165_cast_fp16, y = linear_37_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<int32, [1]> input_175_axes_0 = const()[name = string("input_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_9_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_9_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332296640)))];
            tensor<fp16, [1, 224, 1024]> input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_9_mlp_norm_weight_to_fp16, x = input_173_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_9_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_9_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332298752)))];
            tensor<fp16, [1, 224, 5248]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_9_mlp_Wi_weight_to_fp16, x = input_175_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1118_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = linear_38_cast_fp16)[name = string("op_1118_cast_fp16")];
            string var_1120_mode_0 = const()[name = string("op_1120_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1120_cast_fp16 = gelu(mode = var_1120_mode_0, x = var_1118_cast_fp16_0)[name = string("op_1120_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_179_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1118_cast_fp16_1)[name = string("input_179_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_9_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_9_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343046720)))];
            tensor<fp16, [1, 224, 1024]> linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_9_mlp_Wo_weight_to_fp16, x = input_179_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_183_cast_fp16 = add(x = input_173_cast_fp16, y = linear_39_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<int32, [1]> hidden_states_21_axes_0 = const()[name = string("hidden_states_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_10_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_10_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348420736)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_21_cast_fp16 = layer_norm(axes = hidden_states_21_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_10_attn_norm_weight_to_fp16, x = input_183_cast_fp16)[name = string("hidden_states_21_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_10_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_10_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348422848)))];
            tensor<fp16, [1, 224, 3072]> linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_10_attn_Wqkv_weight_to_fp16, x = hidden_states_21_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<int32, [5]> var_1139 = const()[name = string("op_1139"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_43_cast_fp16 = reshape(shape = var_1139, x = linear_40_cast_fp16)[name = string("qkv_43_cast_fp16")];
            tensor<int32, [3]> var_1141_split_sizes_0 = const()[name = string("op_1141_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1141_axis_0 = const()[name = string("op_1141_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1141_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1141_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1141_cast_fp16_2 = split(axis = var_1141_axis_0, split_sizes = var_1141_split_sizes_0, x = qkv_43_cast_fp16)[name = string("op_1141_cast_fp16")];
            tensor<int32, [1]> squeeze_30_axes_0 = const()[name = string("squeeze_30_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_30_cast_fp16 = squeeze(axes = squeeze_30_axes_0, x = var_1141_cast_fp16_0)[name = string("squeeze_30_cast_fp16")];
            tensor<int32, [1]> squeeze_31_axes_0 = const()[name = string("squeeze_31_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_31_cast_fp16 = squeeze(axes = squeeze_31_axes_0, x = var_1141_cast_fp16_1)[name = string("squeeze_31_cast_fp16")];
            tensor<int32, [1]> squeeze_32_axes_0 = const()[name = string("squeeze_32_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_32_cast_fp16 = squeeze(axes = squeeze_32_axes_0, x = var_1141_cast_fp16_2)[name = string("squeeze_32_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = squeeze_30_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 16, 224, 64]> var_1151_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_11_to_fp16)[name = string("op_1151_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_41_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_41_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_103_promoted_to_fp16 = const()[name = string("const_103_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1163_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_103_promoted_to_fp16)[name = string("op_1163_cast_fp16")];
            bool var_1165_interleave_0 = const()[name = string("op_1165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1165_cast_fp16 = concat(axis = var_49, interleave = var_1165_interleave_0, values = (var_1163_cast_fp16, x1_41_cast_fp16))[name = string("op_1165_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1166_cast_fp16 = mul(x = var_1165_cast_fp16, y = sin_11_to_fp16)[name = string("op_1166_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_21_cast_fp16 = add(x = var_1151_cast_fp16, y = var_1166_cast_fp16)[name = string("q_embed_21_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = squeeze_31_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 224, 64]> var_1169_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_11_to_fp16)[name = string("op_1169_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_41_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_41_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_106_promoted_to_fp16 = const()[name = string("const_106_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1181_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_106_promoted_to_fp16)[name = string("op_1181_cast_fp16")];
            bool var_1183_interleave_0 = const()[name = string("op_1183_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1183_cast_fp16 = concat(axis = var_49, interleave = var_1183_interleave_0, values = (var_1181_cast_fp16, x1_43_cast_fp16))[name = string("op_1183_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1184_cast_fp16 = mul(x = var_1183_cast_fp16, y = sin_11_to_fp16)[name = string("op_1184_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_21_cast_fp16 = add(x = var_1169_cast_fp16, y = var_1184_cast_fp16)[name = string("k_embed_21_cast_fp16")];
            bool var_1189_transpose_x_1 = const()[name = string("op_1189_transpose_x_1"), val = bool(false)];
            bool var_1189_transpose_y_1 = const()[name = string("op_1189_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1189_cast_fp16 = matmul(transpose_x = var_1189_transpose_x_1, transpose_y = var_1189_transpose_y_1, x = q_embed_21_cast_fp16, y = k_embed_21_cast_fp16)[name = string("op_1189_cast_fp16")];
            fp16 var_1190_to_fp16 = const()[name = string("op_1190_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_41_cast_fp16 = mul(x = var_1189_cast_fp16, y = var_1190_to_fp16)[name = string("attn_weights_41_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_185_cast_fp16 = add(x = attn_weights_41_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1193_cast_fp16 = softmax(axis = var_49, x = input_185_cast_fp16)[name = string("op_1193_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = squeeze_32_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = var_1193_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_1197_perm_0 = const()[name = string("op_1197_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1199 = const()[name = string("op_1199"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1197_cast_fp16 = transpose(perm = var_1197_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 224, 1024]> var_1200_cast_fp16 = reshape(shape = var_1199, x = var_1197_cast_fp16)[name = string("op_1200_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_10_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_10_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354714368)))];
            tensor<fp16, [1, 224, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_10_attn_Wo_weight_to_fp16, x = var_1200_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_191_cast_fp16 = add(x = input_183_cast_fp16, y = linear_41_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<int32, [1]> input_193_axes_0 = const()[name = string("input_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_10_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_10_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356811584)))];
            tensor<fp16, [1, 224, 1024]> input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_10_mlp_norm_weight_to_fp16, x = input_191_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_10_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_10_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356813696)))];
            tensor<fp16, [1, 224, 5248]> linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_10_mlp_Wi_weight_to_fp16, x = input_193_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [2]> var_1212_split_sizes_0 = const()[name = string("op_1212_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1212_axis_0 = const()[name = string("op_1212_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1212_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1212_cast_fp16_1 = split(axis = var_1212_axis_0, split_sizes = var_1212_split_sizes_0, x = linear_42_cast_fp16)[name = string("op_1212_cast_fp16")];
            string var_1214_mode_0 = const()[name = string("op_1214_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1214_cast_fp16 = gelu(mode = var_1214_mode_0, x = var_1212_cast_fp16_0)[name = string("op_1214_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_197_cast_fp16 = mul(x = var_1214_cast_fp16, y = var_1212_cast_fp16_1)[name = string("input_197_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_10_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_10_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367561664)))];
            tensor<fp16, [1, 224, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_10_mlp_Wo_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_201_cast_fp16 = add(x = input_191_cast_fp16, y = linear_43_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<int32, [1]> hidden_states_23_axes_0 = const()[name = string("hidden_states_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_11_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_11_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372935680)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_23_cast_fp16 = layer_norm(axes = hidden_states_23_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_11_attn_norm_weight_to_fp16, x = input_201_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_11_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_11_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372937792)))];
            tensor<fp16, [1, 224, 3072]> linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_11_attn_Wqkv_weight_to_fp16, x = hidden_states_23_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [5]> var_1233 = const()[name = string("op_1233"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_47_cast_fp16 = reshape(shape = var_1233, x = linear_44_cast_fp16)[name = string("qkv_47_cast_fp16")];
            tensor<int32, [3]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1235_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1235_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1235_cast_fp16_2 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = qkv_47_cast_fp16)[name = string("op_1235_cast_fp16")];
            tensor<int32, [1]> squeeze_33_axes_0 = const()[name = string("squeeze_33_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_33_cast_fp16 = squeeze(axes = squeeze_33_axes_0, x = var_1235_cast_fp16_0)[name = string("squeeze_33_cast_fp16")];
            tensor<int32, [1]> squeeze_34_axes_0 = const()[name = string("squeeze_34_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_34_cast_fp16 = squeeze(axes = squeeze_34_axes_0, x = var_1235_cast_fp16_1)[name = string("squeeze_34_cast_fp16")];
            tensor<int32, [1]> squeeze_35_axes_0 = const()[name = string("squeeze_35_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_35_cast_fp16 = squeeze(axes = squeeze_35_axes_0, x = var_1235_cast_fp16_2)[name = string("squeeze_35_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = squeeze_33_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 16, 224, 64]> var_1245_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_11_to_fp16)[name = string("op_1245_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_45_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_45_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_111_promoted_to_fp16 = const()[name = string("const_111_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1257_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_111_promoted_to_fp16)[name = string("op_1257_cast_fp16")];
            bool var_1259_interleave_0 = const()[name = string("op_1259_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1259_cast_fp16 = concat(axis = var_49, interleave = var_1259_interleave_0, values = (var_1257_cast_fp16, x1_45_cast_fp16))[name = string("op_1259_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1260_cast_fp16 = mul(x = var_1259_cast_fp16, y = sin_11_to_fp16)[name = string("op_1260_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_23_cast_fp16 = add(x = var_1245_cast_fp16, y = var_1260_cast_fp16)[name = string("q_embed_23_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = squeeze_34_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 16, 224, 64]> var_1263_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_11_to_fp16)[name = string("op_1263_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_45_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_45_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1275_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_114_promoted_to_fp16)[name = string("op_1275_cast_fp16")];
            bool var_1277_interleave_0 = const()[name = string("op_1277_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1277_cast_fp16 = concat(axis = var_49, interleave = var_1277_interleave_0, values = (var_1275_cast_fp16, x1_47_cast_fp16))[name = string("op_1277_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1278_cast_fp16 = mul(x = var_1277_cast_fp16, y = sin_11_to_fp16)[name = string("op_1278_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_23_cast_fp16 = add(x = var_1263_cast_fp16, y = var_1278_cast_fp16)[name = string("k_embed_23_cast_fp16")];
            bool var_1283_transpose_x_1 = const()[name = string("op_1283_transpose_x_1"), val = bool(false)];
            bool var_1283_transpose_y_1 = const()[name = string("op_1283_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1283_cast_fp16 = matmul(transpose_x = var_1283_transpose_x_1, transpose_y = var_1283_transpose_y_1, x = q_embed_23_cast_fp16, y = k_embed_23_cast_fp16)[name = string("op_1283_cast_fp16")];
            fp16 var_1284_to_fp16 = const()[name = string("op_1284_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_45_cast_fp16 = mul(x = var_1283_cast_fp16, y = var_1284_to_fp16)[name = string("attn_weights_45_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_203_cast_fp16 = add(x = attn_weights_45_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1287_cast_fp16 = softmax(axis = var_49, x = input_203_cast_fp16)[name = string("op_1287_cast_fp16")];
            bool attn_output_67_transpose_x_0 = const()[name = string("attn_output_67_transpose_x_0"), val = bool(false)];
            bool attn_output_67_transpose_y_0 = const()[name = string("attn_output_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = squeeze_35_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_67_cast_fp16 = matmul(transpose_x = attn_output_67_transpose_x_0, transpose_y = attn_output_67_transpose_y_0, x = var_1287_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_1291_perm_0 = const()[name = string("op_1291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1293 = const()[name = string("op_1293"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1291_cast_fp16 = transpose(perm = var_1291_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 224, 1024]> var_1294_cast_fp16 = reshape(shape = var_1293, x = var_1291_cast_fp16)[name = string("op_1294_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_11_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_11_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379229312)))];
            tensor<fp16, [1, 224, 1024]> linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_11_attn_Wo_weight_to_fp16, x = var_1294_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_209_cast_fp16 = add(x = input_201_cast_fp16, y = linear_45_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_11_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_11_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381326528)))];
            tensor<fp16, [1, 224, 1024]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_11_mlp_norm_weight_to_fp16, x = input_209_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_11_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_11_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381328640)))];
            tensor<fp16, [1, 224, 5248]> linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_11_mlp_Wi_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [2]> var_1306_split_sizes_0 = const()[name = string("op_1306_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1306_axis_0 = const()[name = string("op_1306_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1306_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1306_cast_fp16_1 = split(axis = var_1306_axis_0, split_sizes = var_1306_split_sizes_0, x = linear_46_cast_fp16)[name = string("op_1306_cast_fp16")];
            string var_1308_mode_0 = const()[name = string("op_1308_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1308_cast_fp16 = gelu(mode = var_1308_mode_0, x = var_1306_cast_fp16_0)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_215_cast_fp16 = mul(x = var_1308_cast_fp16, y = var_1306_cast_fp16_1)[name = string("input_215_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_11_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_11_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392076608)))];
            tensor<fp16, [1, 224, 1024]> linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_11_mlp_Wo_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_219_cast_fp16 = add(x = input_209_cast_fp16, y = linear_47_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<int32, [1]> hidden_states_25_axes_0 = const()[name = string("hidden_states_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_12_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_12_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397450624)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_25_cast_fp16 = layer_norm(axes = hidden_states_25_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_12_attn_norm_weight_to_fp16, x = input_219_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_12_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_12_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397452736)))];
            tensor<fp16, [1, 224, 3072]> linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_12_attn_Wqkv_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [5]> var_1327 = const()[name = string("op_1327"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_51_cast_fp16 = reshape(shape = var_1327, x = linear_48_cast_fp16)[name = string("qkv_51_cast_fp16")];
            tensor<int32, [3]> var_1329_split_sizes_0 = const()[name = string("op_1329_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1329_axis_0 = const()[name = string("op_1329_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1329_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1329_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1329_cast_fp16_2 = split(axis = var_1329_axis_0, split_sizes = var_1329_split_sizes_0, x = qkv_51_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<int32, [1]> squeeze_36_axes_0 = const()[name = string("squeeze_36_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_36_cast_fp16 = squeeze(axes = squeeze_36_axes_0, x = var_1329_cast_fp16_0)[name = string("squeeze_36_cast_fp16")];
            tensor<int32, [1]> squeeze_37_axes_0 = const()[name = string("squeeze_37_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_37_cast_fp16 = squeeze(axes = squeeze_37_axes_0, x = var_1329_cast_fp16_1)[name = string("squeeze_37_cast_fp16")];
            tensor<int32, [1]> squeeze_38_axes_0 = const()[name = string("squeeze_38_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_38_cast_fp16 = squeeze(axes = squeeze_38_axes_0, x = var_1329_cast_fp16_2)[name = string("squeeze_38_cast_fp16")];
            tensor<int32, [4]> q_49_perm_0 = const()[name = string("q_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_49_perm_0 = const()[name = string("k_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_49_cast_fp16 = transpose(perm = q_49_perm_0, x = squeeze_36_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 16, 224, 64]> var_1339_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_7_to_fp16)[name = string("op_1339_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_49_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_49_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_119_promoted_to_fp16 = const()[name = string("const_119_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1351_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_119_promoted_to_fp16)[name = string("op_1351_cast_fp16")];
            bool var_1353_interleave_0 = const()[name = string("op_1353_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1353_cast_fp16 = concat(axis = var_49, interleave = var_1353_interleave_0, values = (var_1351_cast_fp16, x1_49_cast_fp16))[name = string("op_1353_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1354_cast_fp16 = mul(x = var_1353_cast_fp16, y = sin_7_to_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_25_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1354_cast_fp16)[name = string("q_embed_25_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_49_cast_fp16 = transpose(perm = k_49_perm_0, x = squeeze_37_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 16, 224, 64]> var_1357_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_7_to_fp16)[name = string("op_1357_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_49_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_49_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1369_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_1369_cast_fp16")];
            bool var_1371_interleave_0 = const()[name = string("op_1371_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1371_cast_fp16 = concat(axis = var_49, interleave = var_1371_interleave_0, values = (var_1369_cast_fp16, x1_51_cast_fp16))[name = string("op_1371_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1372_cast_fp16 = mul(x = var_1371_cast_fp16, y = sin_7_to_fp16)[name = string("op_1372_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_25_cast_fp16 = add(x = var_1357_cast_fp16, y = var_1372_cast_fp16)[name = string("k_embed_25_cast_fp16")];
            bool var_1377_transpose_x_1 = const()[name = string("op_1377_transpose_x_1"), val = bool(false)];
            bool var_1377_transpose_y_1 = const()[name = string("op_1377_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1377_cast_fp16 = matmul(transpose_x = var_1377_transpose_x_1, transpose_y = var_1377_transpose_y_1, x = q_embed_25_cast_fp16, y = k_embed_25_cast_fp16)[name = string("op_1377_cast_fp16")];
            fp16 var_1378_to_fp16 = const()[name = string("op_1378_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_49_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1378_to_fp16)[name = string("attn_weights_49_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_221_cast_fp16 = add(x = attn_weights_49_cast_fp16, y = key_bias_cast_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1381_cast_fp16 = softmax(axis = var_49, x = input_221_cast_fp16)[name = string("op_1381_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = squeeze_38_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = var_1381_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_1385_perm_0 = const()[name = string("op_1385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1387 = const()[name = string("op_1387"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1385_cast_fp16 = transpose(perm = var_1385_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 224, 1024]> var_1388_cast_fp16 = reshape(shape = var_1387, x = var_1385_cast_fp16)[name = string("op_1388_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_12_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_12_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403744256)))];
            tensor<fp16, [1, 224, 1024]> linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_12_attn_Wo_weight_to_fp16, x = var_1388_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_227_cast_fp16 = add(x = input_219_cast_fp16, y = linear_49_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<int32, [1]> input_229_axes_0 = const()[name = string("input_229_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_12_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_12_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405841472)))];
            tensor<fp16, [1, 224, 1024]> input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_12_mlp_norm_weight_to_fp16, x = input_227_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_12_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_12_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405843584)))];
            tensor<fp16, [1, 224, 5248]> linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_12_mlp_Wi_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [2]> var_1400_split_sizes_0 = const()[name = string("op_1400_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1400_axis_0 = const()[name = string("op_1400_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1400_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1400_cast_fp16_1 = split(axis = var_1400_axis_0, split_sizes = var_1400_split_sizes_0, x = linear_50_cast_fp16)[name = string("op_1400_cast_fp16")];
            string var_1402_mode_0 = const()[name = string("op_1402_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1402_cast_fp16 = gelu(mode = var_1402_mode_0, x = var_1400_cast_fp16_0)[name = string("op_1402_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_233_cast_fp16 = mul(x = var_1402_cast_fp16, y = var_1400_cast_fp16_1)[name = string("input_233_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_12_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_12_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416591552)))];
            tensor<fp16, [1, 224, 1024]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_12_mlp_Wo_weight_to_fp16, x = input_233_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_237_cast_fp16 = add(x = input_227_cast_fp16, y = linear_51_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<int32, [1]> hidden_states_27_axes_0 = const()[name = string("hidden_states_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_13_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_13_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421965568)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_27_cast_fp16 = layer_norm(axes = hidden_states_27_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_13_attn_norm_weight_to_fp16, x = input_237_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_13_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_13_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421967680)))];
            tensor<fp16, [1, 224, 3072]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_13_attn_Wqkv_weight_to_fp16, x = hidden_states_27_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<int32, [5]> var_1421 = const()[name = string("op_1421"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_55_cast_fp16 = reshape(shape = var_1421, x = linear_52_cast_fp16)[name = string("qkv_55_cast_fp16")];
            tensor<int32, [3]> var_1423_split_sizes_0 = const()[name = string("op_1423_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1423_axis_0 = const()[name = string("op_1423_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1423_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1423_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1423_cast_fp16_2 = split(axis = var_1423_axis_0, split_sizes = var_1423_split_sizes_0, x = qkv_55_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [1]> squeeze_39_axes_0 = const()[name = string("squeeze_39_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_39_cast_fp16 = squeeze(axes = squeeze_39_axes_0, x = var_1423_cast_fp16_0)[name = string("squeeze_39_cast_fp16")];
            tensor<int32, [1]> squeeze_40_axes_0 = const()[name = string("squeeze_40_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_40_cast_fp16 = squeeze(axes = squeeze_40_axes_0, x = var_1423_cast_fp16_1)[name = string("squeeze_40_cast_fp16")];
            tensor<int32, [1]> squeeze_41_axes_0 = const()[name = string("squeeze_41_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_41_cast_fp16 = squeeze(axes = squeeze_41_axes_0, x = var_1423_cast_fp16_2)[name = string("squeeze_41_cast_fp16")];
            tensor<int32, [4]> q_53_perm_0 = const()[name = string("q_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_53_perm_0 = const()[name = string("k_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_53_cast_fp16 = transpose(perm = q_53_perm_0, x = squeeze_39_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 224, 64]> var_1433_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_11_to_fp16)[name = string("op_1433_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_53_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_53_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_127_promoted_to_fp16 = const()[name = string("const_127_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1445_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_127_promoted_to_fp16)[name = string("op_1445_cast_fp16")];
            bool var_1447_interleave_0 = const()[name = string("op_1447_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1447_cast_fp16 = concat(axis = var_49, interleave = var_1447_interleave_0, values = (var_1445_cast_fp16, x1_53_cast_fp16))[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1448_cast_fp16 = mul(x = var_1447_cast_fp16, y = sin_11_to_fp16)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_27_cast_fp16 = add(x = var_1433_cast_fp16, y = var_1448_cast_fp16)[name = string("q_embed_27_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_53_cast_fp16 = transpose(perm = k_53_perm_0, x = squeeze_40_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 16, 224, 64]> var_1451_cast_fp16 = mul(x = k_53_cast_fp16, y = cos_11_to_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_53_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_53_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1463_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_1463_cast_fp16")];
            bool var_1465_interleave_0 = const()[name = string("op_1465_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1465_cast_fp16 = concat(axis = var_49, interleave = var_1465_interleave_0, values = (var_1463_cast_fp16, x1_55_cast_fp16))[name = string("op_1465_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1466_cast_fp16 = mul(x = var_1465_cast_fp16, y = sin_11_to_fp16)[name = string("op_1466_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_27_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1466_cast_fp16)[name = string("k_embed_27_cast_fp16")];
            bool var_1471_transpose_x_1 = const()[name = string("op_1471_transpose_x_1"), val = bool(false)];
            bool var_1471_transpose_y_1 = const()[name = string("op_1471_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1471_cast_fp16 = matmul(transpose_x = var_1471_transpose_x_1, transpose_y = var_1471_transpose_y_1, x = q_embed_27_cast_fp16, y = k_embed_27_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp16 var_1472_to_fp16 = const()[name = string("op_1472_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_53_cast_fp16 = mul(x = var_1471_cast_fp16, y = var_1472_to_fp16)[name = string("attn_weights_53_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_239_cast_fp16 = add(x = attn_weights_53_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1475_cast_fp16 = softmax(axis = var_49, x = input_239_cast_fp16)[name = string("op_1475_cast_fp16")];
            bool attn_output_79_transpose_x_0 = const()[name = string("attn_output_79_transpose_x_0"), val = bool(false)];
            bool attn_output_79_transpose_y_0 = const()[name = string("attn_output_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = squeeze_41_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_79_cast_fp16 = matmul(transpose_x = attn_output_79_transpose_x_0, transpose_y = attn_output_79_transpose_y_0, x = var_1475_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [4]> var_1479_perm_0 = const()[name = string("op_1479_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1481 = const()[name = string("op_1481"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1479_cast_fp16 = transpose(perm = var_1479_perm_0, x = attn_output_79_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 224, 1024]> var_1482_cast_fp16 = reshape(shape = var_1481, x = var_1479_cast_fp16)[name = string("op_1482_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_13_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_13_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428259200)))];
            tensor<fp16, [1, 224, 1024]> linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_13_attn_Wo_weight_to_fp16, x = var_1482_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_245_cast_fp16 = add(x = input_237_cast_fp16, y = linear_53_cast_fp16)[name = string("input_245_cast_fp16")];
            tensor<int32, [1]> input_247_axes_0 = const()[name = string("input_247_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_13_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_13_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430356416)))];
            tensor<fp16, [1, 224, 1024]> input_247_cast_fp16 = layer_norm(axes = input_247_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_13_mlp_norm_weight_to_fp16, x = input_245_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_13_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_13_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430358528)))];
            tensor<fp16, [1, 224, 5248]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_13_mlp_Wi_weight_to_fp16, x = input_247_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<int32, [2]> var_1494_split_sizes_0 = const()[name = string("op_1494_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1494_axis_0 = const()[name = string("op_1494_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1494_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1494_cast_fp16_1 = split(axis = var_1494_axis_0, split_sizes = var_1494_split_sizes_0, x = linear_54_cast_fp16)[name = string("op_1494_cast_fp16")];
            string var_1496_mode_0 = const()[name = string("op_1496_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1496_cast_fp16 = gelu(mode = var_1496_mode_0, x = var_1494_cast_fp16_0)[name = string("op_1496_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_251_cast_fp16 = mul(x = var_1496_cast_fp16, y = var_1494_cast_fp16_1)[name = string("input_251_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_13_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_13_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441106496)))];
            tensor<fp16, [1, 224, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_13_mlp_Wo_weight_to_fp16, x = input_251_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_55_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<int32, [1]> hidden_states_29_axes_0 = const()[name = string("hidden_states_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_14_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_14_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446480512)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_29_cast_fp16 = layer_norm(axes = hidden_states_29_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_14_attn_norm_weight_to_fp16, x = input_255_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_14_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_14_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446482624)))];
            tensor<fp16, [1, 224, 3072]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_14_attn_Wqkv_weight_to_fp16, x = hidden_states_29_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [5]> var_1515 = const()[name = string("op_1515"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_59_cast_fp16 = reshape(shape = var_1515, x = linear_56_cast_fp16)[name = string("qkv_59_cast_fp16")];
            tensor<int32, [3]> var_1517_split_sizes_0 = const()[name = string("op_1517_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1517_axis_0 = const()[name = string("op_1517_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1517_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1517_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1517_cast_fp16_2 = split(axis = var_1517_axis_0, split_sizes = var_1517_split_sizes_0, x = qkv_59_cast_fp16)[name = string("op_1517_cast_fp16")];
            tensor<int32, [1]> squeeze_42_axes_0 = const()[name = string("squeeze_42_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_42_cast_fp16 = squeeze(axes = squeeze_42_axes_0, x = var_1517_cast_fp16_0)[name = string("squeeze_42_cast_fp16")];
            tensor<int32, [1]> squeeze_43_axes_0 = const()[name = string("squeeze_43_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_43_cast_fp16 = squeeze(axes = squeeze_43_axes_0, x = var_1517_cast_fp16_1)[name = string("squeeze_43_cast_fp16")];
            tensor<int32, [1]> squeeze_44_axes_0 = const()[name = string("squeeze_44_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_44_cast_fp16 = squeeze(axes = squeeze_44_axes_0, x = var_1517_cast_fp16_2)[name = string("squeeze_44_cast_fp16")];
            tensor<int32, [4]> q_57_perm_0 = const()[name = string("q_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_57_perm_0 = const()[name = string("k_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_57_cast_fp16 = transpose(perm = q_57_perm_0, x = squeeze_42_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 224, 64]> var_1527_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_11_to_fp16)[name = string("op_1527_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_57_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_57_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1539_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_135_promoted_to_fp16)[name = string("op_1539_cast_fp16")];
            bool var_1541_interleave_0 = const()[name = string("op_1541_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1541_cast_fp16 = concat(axis = var_49, interleave = var_1541_interleave_0, values = (var_1539_cast_fp16, x1_57_cast_fp16))[name = string("op_1541_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1542_cast_fp16 = mul(x = var_1541_cast_fp16, y = sin_11_to_fp16)[name = string("op_1542_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_29_cast_fp16 = add(x = var_1527_cast_fp16, y = var_1542_cast_fp16)[name = string("q_embed_29_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_57_cast_fp16 = transpose(perm = k_57_perm_0, x = squeeze_43_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 224, 64]> var_1545_cast_fp16 = mul(x = k_57_cast_fp16, y = cos_11_to_fp16)[name = string("op_1545_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_57_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_57_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_138_promoted_to_fp16 = const()[name = string("const_138_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1557_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_138_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            bool var_1559_interleave_0 = const()[name = string("op_1559_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1559_cast_fp16 = concat(axis = var_49, interleave = var_1559_interleave_0, values = (var_1557_cast_fp16, x1_59_cast_fp16))[name = string("op_1559_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1560_cast_fp16 = mul(x = var_1559_cast_fp16, y = sin_11_to_fp16)[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_29_cast_fp16 = add(x = var_1545_cast_fp16, y = var_1560_cast_fp16)[name = string("k_embed_29_cast_fp16")];
            bool var_1565_transpose_x_1 = const()[name = string("op_1565_transpose_x_1"), val = bool(false)];
            bool var_1565_transpose_y_1 = const()[name = string("op_1565_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1565_cast_fp16 = matmul(transpose_x = var_1565_transpose_x_1, transpose_y = var_1565_transpose_y_1, x = q_embed_29_cast_fp16, y = k_embed_29_cast_fp16)[name = string("op_1565_cast_fp16")];
            fp16 var_1566_to_fp16 = const()[name = string("op_1566_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_57_cast_fp16 = mul(x = var_1565_cast_fp16, y = var_1566_to_fp16)[name = string("attn_weights_57_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_257_cast_fp16 = add(x = attn_weights_57_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1569_cast_fp16 = softmax(axis = var_49, x = input_257_cast_fp16)[name = string("op_1569_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = squeeze_44_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = var_1569_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_1573_perm_0 = const()[name = string("op_1573_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1575 = const()[name = string("op_1575"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1573_cast_fp16 = transpose(perm = var_1573_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 224, 1024]> var_1576_cast_fp16 = reshape(shape = var_1575, x = var_1573_cast_fp16)[name = string("op_1576_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_14_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_14_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452774144)))];
            tensor<fp16, [1, 224, 1024]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_14_attn_Wo_weight_to_fp16, x = var_1576_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_263_cast_fp16 = add(x = input_255_cast_fp16, y = linear_57_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<int32, [1]> input_265_axes_0 = const()[name = string("input_265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_14_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_14_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454871360)))];
            tensor<fp16, [1, 224, 1024]> input_265_cast_fp16 = layer_norm(axes = input_265_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_14_mlp_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_14_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_14_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454873472)))];
            tensor<fp16, [1, 224, 5248]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_14_mlp_Wi_weight_to_fp16, x = input_265_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [2]> var_1588_split_sizes_0 = const()[name = string("op_1588_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1588_axis_0 = const()[name = string("op_1588_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1588_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1588_cast_fp16_1 = split(axis = var_1588_axis_0, split_sizes = var_1588_split_sizes_0, x = linear_58_cast_fp16)[name = string("op_1588_cast_fp16")];
            string var_1590_mode_0 = const()[name = string("op_1590_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1590_cast_fp16 = gelu(mode = var_1590_mode_0, x = var_1588_cast_fp16_0)[name = string("op_1590_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_269_cast_fp16 = mul(x = var_1590_cast_fp16, y = var_1588_cast_fp16_1)[name = string("input_269_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_14_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_14_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465621440)))];
            tensor<fp16, [1, 224, 1024]> linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_14_mlp_Wo_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_273_cast_fp16 = add(x = input_263_cast_fp16, y = linear_59_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<int32, [1]> hidden_states_31_axes_0 = const()[name = string("hidden_states_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_15_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_15_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470995456)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_31_cast_fp16 = layer_norm(axes = hidden_states_31_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_15_attn_norm_weight_to_fp16, x = input_273_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_15_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_15_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470997568)))];
            tensor<fp16, [1, 224, 3072]> linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_15_attn_Wqkv_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<int32, [5]> var_1609 = const()[name = string("op_1609"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_63_cast_fp16 = reshape(shape = var_1609, x = linear_60_cast_fp16)[name = string("qkv_63_cast_fp16")];
            tensor<int32, [3]> var_1611_split_sizes_0 = const()[name = string("op_1611_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1611_axis_0 = const()[name = string("op_1611_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1611_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1611_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1611_cast_fp16_2 = split(axis = var_1611_axis_0, split_sizes = var_1611_split_sizes_0, x = qkv_63_cast_fp16)[name = string("op_1611_cast_fp16")];
            tensor<int32, [1]> squeeze_45_axes_0 = const()[name = string("squeeze_45_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_45_cast_fp16 = squeeze(axes = squeeze_45_axes_0, x = var_1611_cast_fp16_0)[name = string("squeeze_45_cast_fp16")];
            tensor<int32, [1]> squeeze_46_axes_0 = const()[name = string("squeeze_46_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_46_cast_fp16 = squeeze(axes = squeeze_46_axes_0, x = var_1611_cast_fp16_1)[name = string("squeeze_46_cast_fp16")];
            tensor<int32, [1]> squeeze_47_axes_0 = const()[name = string("squeeze_47_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_47_cast_fp16 = squeeze(axes = squeeze_47_axes_0, x = var_1611_cast_fp16_2)[name = string("squeeze_47_cast_fp16")];
            tensor<int32, [4]> q_61_perm_0 = const()[name = string("q_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_61_perm_0 = const()[name = string("k_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_61_cast_fp16 = transpose(perm = q_61_perm_0, x = squeeze_45_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 16, 224, 64]> var_1621_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_7_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_61_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_61_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_143_promoted_to_fp16 = const()[name = string("const_143_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1633_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_143_promoted_to_fp16)[name = string("op_1633_cast_fp16")];
            bool var_1635_interleave_0 = const()[name = string("op_1635_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1635_cast_fp16 = concat(axis = var_49, interleave = var_1635_interleave_0, values = (var_1633_cast_fp16, x1_61_cast_fp16))[name = string("op_1635_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1636_cast_fp16 = mul(x = var_1635_cast_fp16, y = sin_7_to_fp16)[name = string("op_1636_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_31_cast_fp16 = add(x = var_1621_cast_fp16, y = var_1636_cast_fp16)[name = string("q_embed_31_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_61_cast_fp16 = transpose(perm = k_61_perm_0, x = squeeze_46_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 224, 64]> var_1639_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_7_to_fp16)[name = string("op_1639_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_61_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_61_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_146_promoted_to_fp16 = const()[name = string("const_146_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1651_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_146_promoted_to_fp16)[name = string("op_1651_cast_fp16")];
            bool var_1653_interleave_0 = const()[name = string("op_1653_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1653_cast_fp16 = concat(axis = var_49, interleave = var_1653_interleave_0, values = (var_1651_cast_fp16, x1_63_cast_fp16))[name = string("op_1653_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1654_cast_fp16 = mul(x = var_1653_cast_fp16, y = sin_7_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_31_cast_fp16 = add(x = var_1639_cast_fp16, y = var_1654_cast_fp16)[name = string("k_embed_31_cast_fp16")];
            bool var_1659_transpose_x_1 = const()[name = string("op_1659_transpose_x_1"), val = bool(false)];
            bool var_1659_transpose_y_1 = const()[name = string("op_1659_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1659_cast_fp16 = matmul(transpose_x = var_1659_transpose_x_1, transpose_y = var_1659_transpose_y_1, x = q_embed_31_cast_fp16, y = k_embed_31_cast_fp16)[name = string("op_1659_cast_fp16")];
            fp16 var_1660_to_fp16 = const()[name = string("op_1660_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_61_cast_fp16 = mul(x = var_1659_cast_fp16, y = var_1660_to_fp16)[name = string("attn_weights_61_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_275_cast_fp16 = add(x = attn_weights_61_cast_fp16, y = key_bias_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1663_cast_fp16 = softmax(axis = var_49, x = input_275_cast_fp16)[name = string("op_1663_cast_fp16")];
            bool attn_output_91_transpose_x_0 = const()[name = string("attn_output_91_transpose_x_0"), val = bool(false)];
            bool attn_output_91_transpose_y_0 = const()[name = string("attn_output_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = squeeze_47_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_91_cast_fp16 = matmul(transpose_x = attn_output_91_transpose_x_0, transpose_y = attn_output_91_transpose_y_0, x = var_1663_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_1667_perm_0 = const()[name = string("op_1667_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1669 = const()[name = string("op_1669"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1667_cast_fp16 = transpose(perm = var_1667_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 224, 1024]> var_1670_cast_fp16 = reshape(shape = var_1669, x = var_1667_cast_fp16)[name = string("op_1670_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_15_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_15_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477289088)))];
            tensor<fp16, [1, 224, 1024]> linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_15_attn_Wo_weight_to_fp16, x = var_1670_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_281_cast_fp16 = add(x = input_273_cast_fp16, y = linear_61_cast_fp16)[name = string("input_281_cast_fp16")];
            tensor<int32, [1]> input_283_axes_0 = const()[name = string("input_283_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_15_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_15_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(479386304)))];
            tensor<fp16, [1, 224, 1024]> input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_15_mlp_norm_weight_to_fp16, x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_15_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_15_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(479388416)))];
            tensor<fp16, [1, 224, 5248]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_15_mlp_Wi_weight_to_fp16, x = input_283_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<int32, [2]> var_1682_split_sizes_0 = const()[name = string("op_1682_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1682_axis_0 = const()[name = string("op_1682_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1682_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1682_cast_fp16_1 = split(axis = var_1682_axis_0, split_sizes = var_1682_split_sizes_0, x = linear_62_cast_fp16)[name = string("op_1682_cast_fp16")];
            string var_1684_mode_0 = const()[name = string("op_1684_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1684_cast_fp16 = gelu(mode = var_1684_mode_0, x = var_1682_cast_fp16_0)[name = string("op_1684_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_287_cast_fp16 = mul(x = var_1684_cast_fp16, y = var_1682_cast_fp16_1)[name = string("input_287_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_15_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_15_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490136384)))];
            tensor<fp16, [1, 224, 1024]> linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_15_mlp_Wo_weight_to_fp16, x = input_287_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_291_cast_fp16 = add(x = input_281_cast_fp16, y = linear_63_cast_fp16)[name = string("input_291_cast_fp16")];
            tensor<int32, [1]> hidden_states_33_axes_0 = const()[name = string("hidden_states_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_16_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_16_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495510400)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_33_cast_fp16 = layer_norm(axes = hidden_states_33_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_16_attn_norm_weight_to_fp16, x = input_291_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_16_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_16_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495512512)))];
            tensor<fp16, [1, 224, 3072]> linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_16_attn_Wqkv_weight_to_fp16, x = hidden_states_33_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [5]> var_1703 = const()[name = string("op_1703"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_67_cast_fp16 = reshape(shape = var_1703, x = linear_64_cast_fp16)[name = string("qkv_67_cast_fp16")];
            tensor<int32, [3]> var_1705_split_sizes_0 = const()[name = string("op_1705_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1705_axis_0 = const()[name = string("op_1705_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1705_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1705_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1705_cast_fp16_2 = split(axis = var_1705_axis_0, split_sizes = var_1705_split_sizes_0, x = qkv_67_cast_fp16)[name = string("op_1705_cast_fp16")];
            tensor<int32, [1]> squeeze_48_axes_0 = const()[name = string("squeeze_48_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_48_cast_fp16 = squeeze(axes = squeeze_48_axes_0, x = var_1705_cast_fp16_0)[name = string("squeeze_48_cast_fp16")];
            tensor<int32, [1]> squeeze_49_axes_0 = const()[name = string("squeeze_49_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_49_cast_fp16 = squeeze(axes = squeeze_49_axes_0, x = var_1705_cast_fp16_1)[name = string("squeeze_49_cast_fp16")];
            tensor<int32, [1]> squeeze_50_axes_0 = const()[name = string("squeeze_50_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_50_cast_fp16 = squeeze(axes = squeeze_50_axes_0, x = var_1705_cast_fp16_2)[name = string("squeeze_50_cast_fp16")];
            tensor<int32, [4]> q_65_perm_0 = const()[name = string("q_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_65_perm_0 = const()[name = string("k_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_65_cast_fp16 = transpose(perm = q_65_perm_0, x = squeeze_48_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 16, 224, 64]> var_1715_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_11_to_fp16)[name = string("op_1715_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_65_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_65_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_151_promoted_to_fp16 = const()[name = string("const_151_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1727_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_151_promoted_to_fp16)[name = string("op_1727_cast_fp16")];
            bool var_1729_interleave_0 = const()[name = string("op_1729_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1729_cast_fp16 = concat(axis = var_49, interleave = var_1729_interleave_0, values = (var_1727_cast_fp16, x1_65_cast_fp16))[name = string("op_1729_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1730_cast_fp16 = mul(x = var_1729_cast_fp16, y = sin_11_to_fp16)[name = string("op_1730_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_33_cast_fp16 = add(x = var_1715_cast_fp16, y = var_1730_cast_fp16)[name = string("q_embed_33_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_65_cast_fp16 = transpose(perm = k_65_perm_0, x = squeeze_49_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 16, 224, 64]> var_1733_cast_fp16 = mul(x = k_65_cast_fp16, y = cos_11_to_fp16)[name = string("op_1733_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_65_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_65_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_154_promoted_to_fp16 = const()[name = string("const_154_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1745_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_154_promoted_to_fp16)[name = string("op_1745_cast_fp16")];
            bool var_1747_interleave_0 = const()[name = string("op_1747_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1747_cast_fp16 = concat(axis = var_49, interleave = var_1747_interleave_0, values = (var_1745_cast_fp16, x1_67_cast_fp16))[name = string("op_1747_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1748_cast_fp16 = mul(x = var_1747_cast_fp16, y = sin_11_to_fp16)[name = string("op_1748_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_33_cast_fp16 = add(x = var_1733_cast_fp16, y = var_1748_cast_fp16)[name = string("k_embed_33_cast_fp16")];
            bool var_1753_transpose_x_1 = const()[name = string("op_1753_transpose_x_1"), val = bool(false)];
            bool var_1753_transpose_y_1 = const()[name = string("op_1753_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1753_cast_fp16 = matmul(transpose_x = var_1753_transpose_x_1, transpose_y = var_1753_transpose_y_1, x = q_embed_33_cast_fp16, y = k_embed_33_cast_fp16)[name = string("op_1753_cast_fp16")];
            fp16 var_1754_to_fp16 = const()[name = string("op_1754_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_65_cast_fp16 = mul(x = var_1753_cast_fp16, y = var_1754_to_fp16)[name = string("attn_weights_65_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_293_cast_fp16 = add(x = attn_weights_65_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_293_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1757_cast_fp16 = softmax(axis = var_49, x = input_293_cast_fp16)[name = string("op_1757_cast_fp16")];
            bool attn_output_97_transpose_x_0 = const()[name = string("attn_output_97_transpose_x_0"), val = bool(false)];
            bool attn_output_97_transpose_y_0 = const()[name = string("attn_output_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = squeeze_50_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_0, transpose_y = attn_output_97_transpose_y_0, x = var_1757_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_97_cast_fp16")];
            tensor<int32, [4]> var_1761_perm_0 = const()[name = string("op_1761_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1763 = const()[name = string("op_1763"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1761_cast_fp16 = transpose(perm = var_1761_perm_0, x = attn_output_97_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 224, 1024]> var_1764_cast_fp16 = reshape(shape = var_1763, x = var_1761_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_16_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_16_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501804032)))];
            tensor<fp16, [1, 224, 1024]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_16_attn_Wo_weight_to_fp16, x = var_1764_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_299_cast_fp16 = add(x = input_291_cast_fp16, y = linear_65_cast_fp16)[name = string("input_299_cast_fp16")];
            tensor<int32, [1]> input_301_axes_0 = const()[name = string("input_301_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_16_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_16_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503901248)))];
            tensor<fp16, [1, 224, 1024]> input_301_cast_fp16 = layer_norm(axes = input_301_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_16_mlp_norm_weight_to_fp16, x = input_299_cast_fp16)[name = string("input_301_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_16_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_16_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503903360)))];
            tensor<fp16, [1, 224, 5248]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_16_mlp_Wi_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [2]> var_1776_split_sizes_0 = const()[name = string("op_1776_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1776_axis_0 = const()[name = string("op_1776_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1776_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1776_cast_fp16_1 = split(axis = var_1776_axis_0, split_sizes = var_1776_split_sizes_0, x = linear_66_cast_fp16)[name = string("op_1776_cast_fp16")];
            string var_1778_mode_0 = const()[name = string("op_1778_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1778_cast_fp16 = gelu(mode = var_1778_mode_0, x = var_1776_cast_fp16_0)[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_305_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_1776_cast_fp16_1)[name = string("input_305_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_16_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_16_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514651328)))];
            tensor<fp16, [1, 224, 1024]> linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_16_mlp_Wo_weight_to_fp16, x = input_305_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_309_cast_fp16 = add(x = input_299_cast_fp16, y = linear_67_cast_fp16)[name = string("input_309_cast_fp16")];
            tensor<int32, [1]> hidden_states_35_axes_0 = const()[name = string("hidden_states_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_17_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_17_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520025344)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_35_cast_fp16 = layer_norm(axes = hidden_states_35_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_17_attn_norm_weight_to_fp16, x = input_309_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_17_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_17_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520027456)))];
            tensor<fp16, [1, 224, 3072]> linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_17_attn_Wqkv_weight_to_fp16, x = hidden_states_35_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [5]> var_1797 = const()[name = string("op_1797"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_71_cast_fp16 = reshape(shape = var_1797, x = linear_68_cast_fp16)[name = string("qkv_71_cast_fp16")];
            tensor<int32, [3]> var_1799_split_sizes_0 = const()[name = string("op_1799_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1799_axis_0 = const()[name = string("op_1799_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1799_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1799_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1799_cast_fp16_2 = split(axis = var_1799_axis_0, split_sizes = var_1799_split_sizes_0, x = qkv_71_cast_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> squeeze_51_axes_0 = const()[name = string("squeeze_51_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_51_cast_fp16 = squeeze(axes = squeeze_51_axes_0, x = var_1799_cast_fp16_0)[name = string("squeeze_51_cast_fp16")];
            tensor<int32, [1]> squeeze_52_axes_0 = const()[name = string("squeeze_52_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_52_cast_fp16 = squeeze(axes = squeeze_52_axes_0, x = var_1799_cast_fp16_1)[name = string("squeeze_52_cast_fp16")];
            tensor<int32, [1]> squeeze_53_axes_0 = const()[name = string("squeeze_53_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_53_cast_fp16 = squeeze(axes = squeeze_53_axes_0, x = var_1799_cast_fp16_2)[name = string("squeeze_53_cast_fp16")];
            tensor<int32, [4]> q_69_perm_0 = const()[name = string("q_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_69_perm_0 = const()[name = string("k_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_69_cast_fp16 = transpose(perm = q_69_perm_0, x = squeeze_51_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 16, 224, 64]> var_1809_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_11_to_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_69_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_69_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1821_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1821_cast_fp16")];
            bool var_1823_interleave_0 = const()[name = string("op_1823_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1823_cast_fp16 = concat(axis = var_49, interleave = var_1823_interleave_0, values = (var_1821_cast_fp16, x1_69_cast_fp16))[name = string("op_1823_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1824_cast_fp16 = mul(x = var_1823_cast_fp16, y = sin_11_to_fp16)[name = string("op_1824_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_35_cast_fp16 = add(x = var_1809_cast_fp16, y = var_1824_cast_fp16)[name = string("q_embed_35_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_69_cast_fp16 = transpose(perm = k_69_perm_0, x = squeeze_52_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 16, 224, 64]> var_1827_cast_fp16 = mul(x = k_69_cast_fp16, y = cos_11_to_fp16)[name = string("op_1827_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_69_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_69_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1839_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1839_cast_fp16")];
            bool var_1841_interleave_0 = const()[name = string("op_1841_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1841_cast_fp16 = concat(axis = var_49, interleave = var_1841_interleave_0, values = (var_1839_cast_fp16, x1_71_cast_fp16))[name = string("op_1841_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1842_cast_fp16 = mul(x = var_1841_cast_fp16, y = sin_11_to_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_35_cast_fp16 = add(x = var_1827_cast_fp16, y = var_1842_cast_fp16)[name = string("k_embed_35_cast_fp16")];
            bool var_1847_transpose_x_1 = const()[name = string("op_1847_transpose_x_1"), val = bool(false)];
            bool var_1847_transpose_y_1 = const()[name = string("op_1847_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1847_cast_fp16 = matmul(transpose_x = var_1847_transpose_x_1, transpose_y = var_1847_transpose_y_1, x = q_embed_35_cast_fp16, y = k_embed_35_cast_fp16)[name = string("op_1847_cast_fp16")];
            fp16 var_1848_to_fp16 = const()[name = string("op_1848_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_69_cast_fp16 = mul(x = var_1847_cast_fp16, y = var_1848_to_fp16)[name = string("attn_weights_69_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_311_cast_fp16 = add(x = attn_weights_69_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_311_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1851_cast_fp16 = softmax(axis = var_49, x = input_311_cast_fp16)[name = string("op_1851_cast_fp16")];
            bool attn_output_103_transpose_x_0 = const()[name = string("attn_output_103_transpose_x_0"), val = bool(false)];
            bool attn_output_103_transpose_y_0 = const()[name = string("attn_output_103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = squeeze_53_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_103_cast_fp16 = matmul(transpose_x = attn_output_103_transpose_x_0, transpose_y = attn_output_103_transpose_y_0, x = var_1851_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [4]> var_1855_perm_0 = const()[name = string("op_1855_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1857 = const()[name = string("op_1857"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1855_cast_fp16 = transpose(perm = var_1855_perm_0, x = attn_output_103_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 224, 1024]> var_1858_cast_fp16 = reshape(shape = var_1857, x = var_1855_cast_fp16)[name = string("op_1858_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_17_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_17_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526318976)))];
            tensor<fp16, [1, 224, 1024]> linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_17_attn_Wo_weight_to_fp16, x = var_1858_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_317_cast_fp16 = add(x = input_309_cast_fp16, y = linear_69_cast_fp16)[name = string("input_317_cast_fp16")];
            tensor<int32, [1]> input_319_axes_0 = const()[name = string("input_319_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_17_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_17_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528416192)))];
            tensor<fp16, [1, 224, 1024]> input_319_cast_fp16 = layer_norm(axes = input_319_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_17_mlp_norm_weight_to_fp16, x = input_317_cast_fp16)[name = string("input_319_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_17_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_17_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528418304)))];
            tensor<fp16, [1, 224, 5248]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_17_mlp_Wi_weight_to_fp16, x = input_319_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [2]> var_1870_split_sizes_0 = const()[name = string("op_1870_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1870_axis_0 = const()[name = string("op_1870_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1870_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1870_cast_fp16_1 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_70_cast_fp16)[name = string("op_1870_cast_fp16")];
            string var_1872_mode_0 = const()[name = string("op_1872_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1872_cast_fp16 = gelu(mode = var_1872_mode_0, x = var_1870_cast_fp16_0)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_323_cast_fp16 = mul(x = var_1872_cast_fp16, y = var_1870_cast_fp16_1)[name = string("input_323_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_17_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_17_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539166272)))];
            tensor<fp16, [1, 224, 1024]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_17_mlp_Wo_weight_to_fp16, x = input_323_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_327_cast_fp16 = add(x = input_317_cast_fp16, y = linear_71_cast_fp16)[name = string("input_327_cast_fp16")];
            tensor<int32, [1]> hidden_states_37_axes_0 = const()[name = string("hidden_states_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_18_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_18_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544540288)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_37_cast_fp16 = layer_norm(axes = hidden_states_37_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_18_attn_norm_weight_to_fp16, x = input_327_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_18_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_18_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544542400)))];
            tensor<fp16, [1, 224, 3072]> linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_18_attn_Wqkv_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [5]> var_1891 = const()[name = string("op_1891"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_75_cast_fp16 = reshape(shape = var_1891, x = linear_72_cast_fp16)[name = string("qkv_75_cast_fp16")];
            tensor<int32, [3]> var_1893_split_sizes_0 = const()[name = string("op_1893_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1893_axis_0 = const()[name = string("op_1893_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1893_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1893_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1893_cast_fp16_2 = split(axis = var_1893_axis_0, split_sizes = var_1893_split_sizes_0, x = qkv_75_cast_fp16)[name = string("op_1893_cast_fp16")];
            tensor<int32, [1]> squeeze_54_axes_0 = const()[name = string("squeeze_54_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_54_cast_fp16 = squeeze(axes = squeeze_54_axes_0, x = var_1893_cast_fp16_0)[name = string("squeeze_54_cast_fp16")];
            tensor<int32, [1]> squeeze_55_axes_0 = const()[name = string("squeeze_55_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_55_cast_fp16 = squeeze(axes = squeeze_55_axes_0, x = var_1893_cast_fp16_1)[name = string("squeeze_55_cast_fp16")];
            tensor<int32, [1]> squeeze_56_axes_0 = const()[name = string("squeeze_56_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_56_cast_fp16 = squeeze(axes = squeeze_56_axes_0, x = var_1893_cast_fp16_2)[name = string("squeeze_56_cast_fp16")];
            tensor<int32, [4]> q_73_perm_0 = const()[name = string("q_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_73_perm_0 = const()[name = string("k_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_73_cast_fp16 = transpose(perm = q_73_perm_0, x = squeeze_54_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 224, 64]> var_1903_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_7_to_fp16)[name = string("op_1903_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_73_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_73_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_167_promoted_to_fp16 = const()[name = string("const_167_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1915_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_167_promoted_to_fp16)[name = string("op_1915_cast_fp16")];
            bool var_1917_interleave_0 = const()[name = string("op_1917_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1917_cast_fp16 = concat(axis = var_49, interleave = var_1917_interleave_0, values = (var_1915_cast_fp16, x1_73_cast_fp16))[name = string("op_1917_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1918_cast_fp16 = mul(x = var_1917_cast_fp16, y = sin_7_to_fp16)[name = string("op_1918_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_37_cast_fp16 = add(x = var_1903_cast_fp16, y = var_1918_cast_fp16)[name = string("q_embed_37_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_73_cast_fp16 = transpose(perm = k_73_perm_0, x = squeeze_55_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 16, 224, 64]> var_1921_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_7_to_fp16)[name = string("op_1921_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_73_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_73_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_170_promoted_to_fp16 = const()[name = string("const_170_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_1933_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_170_promoted_to_fp16)[name = string("op_1933_cast_fp16")];
            bool var_1935_interleave_0 = const()[name = string("op_1935_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_1935_cast_fp16 = concat(axis = var_49, interleave = var_1935_interleave_0, values = (var_1933_cast_fp16, x1_75_cast_fp16))[name = string("op_1935_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_1936_cast_fp16 = mul(x = var_1935_cast_fp16, y = sin_7_to_fp16)[name = string("op_1936_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_37_cast_fp16 = add(x = var_1921_cast_fp16, y = var_1936_cast_fp16)[name = string("k_embed_37_cast_fp16")];
            bool var_1941_transpose_x_1 = const()[name = string("op_1941_transpose_x_1"), val = bool(false)];
            bool var_1941_transpose_y_1 = const()[name = string("op_1941_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_1941_cast_fp16 = matmul(transpose_x = var_1941_transpose_x_1, transpose_y = var_1941_transpose_y_1, x = q_embed_37_cast_fp16, y = k_embed_37_cast_fp16)[name = string("op_1941_cast_fp16")];
            fp16 var_1942_to_fp16 = const()[name = string("op_1942_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_73_cast_fp16 = mul(x = var_1941_cast_fp16, y = var_1942_to_fp16)[name = string("attn_weights_73_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_329_cast_fp16 = add(x = attn_weights_73_cast_fp16, y = key_bias_cast_fp16)[name = string("input_329_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_1945_cast_fp16 = softmax(axis = var_49, x = input_329_cast_fp16)[name = string("op_1945_cast_fp16")];
            bool attn_output_109_transpose_x_0 = const()[name = string("attn_output_109_transpose_x_0"), val = bool(false)];
            bool attn_output_109_transpose_y_0 = const()[name = string("attn_output_109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = squeeze_56_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_109_cast_fp16 = matmul(transpose_x = attn_output_109_transpose_x_0, transpose_y = attn_output_109_transpose_y_0, x = var_1945_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_109_cast_fp16")];
            tensor<int32, [4]> var_1949_perm_0 = const()[name = string("op_1949_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1951 = const()[name = string("op_1951"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_1949_cast_fp16 = transpose(perm = var_1949_perm_0, x = attn_output_109_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 224, 1024]> var_1952_cast_fp16 = reshape(shape = var_1951, x = var_1949_cast_fp16)[name = string("op_1952_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_18_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_18_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550833920)))];
            tensor<fp16, [1, 224, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_18_attn_Wo_weight_to_fp16, x = var_1952_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_335_cast_fp16 = add(x = input_327_cast_fp16, y = linear_73_cast_fp16)[name = string("input_335_cast_fp16")];
            tensor<int32, [1]> input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_18_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_18_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552931136)))];
            tensor<fp16, [1, 224, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_18_mlp_norm_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_18_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_18_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552933248)))];
            tensor<fp16, [1, 224, 5248]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_18_mlp_Wi_weight_to_fp16, x = input_337_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [2]> var_1964_split_sizes_0 = const()[name = string("op_1964_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_1964_axis_0 = const()[name = string("op_1964_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_1964_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_1964_cast_fp16_1 = split(axis = var_1964_axis_0, split_sizes = var_1964_split_sizes_0, x = linear_74_cast_fp16)[name = string("op_1964_cast_fp16")];
            string var_1966_mode_0 = const()[name = string("op_1966_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_1966_cast_fp16 = gelu(mode = var_1966_mode_0, x = var_1964_cast_fp16_0)[name = string("op_1966_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_341_cast_fp16 = mul(x = var_1966_cast_fp16, y = var_1964_cast_fp16_1)[name = string("input_341_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_18_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_18_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563681216)))];
            tensor<fp16, [1, 224, 1024]> linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_18_mlp_Wo_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_345_cast_fp16 = add(x = input_335_cast_fp16, y = linear_75_cast_fp16)[name = string("input_345_cast_fp16")];
            tensor<int32, [1]> hidden_states_39_axes_0 = const()[name = string("hidden_states_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_19_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_19_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569055232)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_39_cast_fp16 = layer_norm(axes = hidden_states_39_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_19_attn_norm_weight_to_fp16, x = input_345_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_19_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_19_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569057344)))];
            tensor<fp16, [1, 224, 3072]> linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_19_attn_Wqkv_weight_to_fp16, x = hidden_states_39_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<int32, [5]> var_1985 = const()[name = string("op_1985"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_79_cast_fp16 = reshape(shape = var_1985, x = linear_76_cast_fp16)[name = string("qkv_79_cast_fp16")];
            tensor<int32, [3]> var_1987_split_sizes_0 = const()[name = string("op_1987_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_1987_axis_0 = const()[name = string("op_1987_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_1987_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_1987_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_1987_cast_fp16_2 = split(axis = var_1987_axis_0, split_sizes = var_1987_split_sizes_0, x = qkv_79_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [1]> squeeze_57_axes_0 = const()[name = string("squeeze_57_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_57_cast_fp16 = squeeze(axes = squeeze_57_axes_0, x = var_1987_cast_fp16_0)[name = string("squeeze_57_cast_fp16")];
            tensor<int32, [1]> squeeze_58_axes_0 = const()[name = string("squeeze_58_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_58_cast_fp16 = squeeze(axes = squeeze_58_axes_0, x = var_1987_cast_fp16_1)[name = string("squeeze_58_cast_fp16")];
            tensor<int32, [1]> squeeze_59_axes_0 = const()[name = string("squeeze_59_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_59_cast_fp16 = squeeze(axes = squeeze_59_axes_0, x = var_1987_cast_fp16_2)[name = string("squeeze_59_cast_fp16")];
            tensor<int32, [4]> q_77_perm_0 = const()[name = string("q_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_77_perm_0 = const()[name = string("k_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_77_cast_fp16 = transpose(perm = q_77_perm_0, x = squeeze_57_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 224, 64]> var_1997_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_11_to_fp16)[name = string("op_1997_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_77_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_77_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_175_promoted_to_fp16 = const()[name = string("const_175_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2009_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_175_promoted_to_fp16)[name = string("op_2009_cast_fp16")];
            bool var_2011_interleave_0 = const()[name = string("op_2011_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2011_cast_fp16 = concat(axis = var_49, interleave = var_2011_interleave_0, values = (var_2009_cast_fp16, x1_77_cast_fp16))[name = string("op_2011_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2012_cast_fp16 = mul(x = var_2011_cast_fp16, y = sin_11_to_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_39_cast_fp16 = add(x = var_1997_cast_fp16, y = var_2012_cast_fp16)[name = string("q_embed_39_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_77_cast_fp16 = transpose(perm = k_77_perm_0, x = squeeze_58_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 224, 64]> var_2015_cast_fp16 = mul(x = k_77_cast_fp16, y = cos_11_to_fp16)[name = string("op_2015_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_77_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_77_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_178_promoted_to_fp16 = const()[name = string("const_178_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2027_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_178_promoted_to_fp16)[name = string("op_2027_cast_fp16")];
            bool var_2029_interleave_0 = const()[name = string("op_2029_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2029_cast_fp16 = concat(axis = var_49, interleave = var_2029_interleave_0, values = (var_2027_cast_fp16, x1_79_cast_fp16))[name = string("op_2029_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2030_cast_fp16 = mul(x = var_2029_cast_fp16, y = sin_11_to_fp16)[name = string("op_2030_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_39_cast_fp16 = add(x = var_2015_cast_fp16, y = var_2030_cast_fp16)[name = string("k_embed_39_cast_fp16")];
            bool var_2035_transpose_x_1 = const()[name = string("op_2035_transpose_x_1"), val = bool(false)];
            bool var_2035_transpose_y_1 = const()[name = string("op_2035_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_2035_cast_fp16 = matmul(transpose_x = var_2035_transpose_x_1, transpose_y = var_2035_transpose_y_1, x = q_embed_39_cast_fp16, y = k_embed_39_cast_fp16)[name = string("op_2035_cast_fp16")];
            fp16 var_2036_to_fp16 = const()[name = string("op_2036_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_77_cast_fp16 = mul(x = var_2035_cast_fp16, y = var_2036_to_fp16)[name = string("attn_weights_77_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_347_cast_fp16 = add(x = attn_weights_77_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_347_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_2039_cast_fp16 = softmax(axis = var_49, x = input_347_cast_fp16)[name = string("op_2039_cast_fp16")];
            bool attn_output_115_transpose_x_0 = const()[name = string("attn_output_115_transpose_x_0"), val = bool(false)];
            bool attn_output_115_transpose_y_0 = const()[name = string("attn_output_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = squeeze_59_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_115_cast_fp16 = matmul(transpose_x = attn_output_115_transpose_x_0, transpose_y = attn_output_115_transpose_y_0, x = var_2039_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_115_cast_fp16")];
            tensor<int32, [4]> var_2043_perm_0 = const()[name = string("op_2043_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2045 = const()[name = string("op_2045"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_2043_cast_fp16 = transpose(perm = var_2043_perm_0, x = attn_output_115_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 224, 1024]> var_2046_cast_fp16 = reshape(shape = var_2045, x = var_2043_cast_fp16)[name = string("op_2046_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_19_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_19_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575348864)))];
            tensor<fp16, [1, 224, 1024]> linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_19_attn_Wo_weight_to_fp16, x = var_2046_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_353_cast_fp16 = add(x = input_345_cast_fp16, y = linear_77_cast_fp16)[name = string("input_353_cast_fp16")];
            tensor<int32, [1]> input_355_axes_0 = const()[name = string("input_355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_19_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_19_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577446080)))];
            tensor<fp16, [1, 224, 1024]> input_355_cast_fp16 = layer_norm(axes = input_355_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_19_mlp_norm_weight_to_fp16, x = input_353_cast_fp16)[name = string("input_355_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_19_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_19_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577448192)))];
            tensor<fp16, [1, 224, 5248]> linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_19_mlp_Wi_weight_to_fp16, x = input_355_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [2]> var_2058_split_sizes_0 = const()[name = string("op_2058_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2058_axis_0 = const()[name = string("op_2058_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_2058_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_2058_cast_fp16_1 = split(axis = var_2058_axis_0, split_sizes = var_2058_split_sizes_0, x = linear_78_cast_fp16)[name = string("op_2058_cast_fp16")];
            string var_2060_mode_0 = const()[name = string("op_2060_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_2060_cast_fp16 = gelu(mode = var_2060_mode_0, x = var_2058_cast_fp16_0)[name = string("op_2060_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_359_cast_fp16 = mul(x = var_2060_cast_fp16, y = var_2058_cast_fp16_1)[name = string("input_359_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_19_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_19_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588196160)))];
            tensor<fp16, [1, 224, 1024]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_19_mlp_Wo_weight_to_fp16, x = input_359_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_363_cast_fp16 = add(x = input_353_cast_fp16, y = linear_79_cast_fp16)[name = string("input_363_cast_fp16")];
            tensor<int32, [1]> hidden_states_41_axes_0 = const()[name = string("hidden_states_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_20_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_20_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593570176)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_41_cast_fp16 = layer_norm(axes = hidden_states_41_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_20_attn_norm_weight_to_fp16, x = input_363_cast_fp16)[name = string("hidden_states_41_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_20_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_20_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593572288)))];
            tensor<fp16, [1, 224, 3072]> linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_20_attn_Wqkv_weight_to_fp16, x = hidden_states_41_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [5]> var_2079 = const()[name = string("op_2079"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_83_cast_fp16 = reshape(shape = var_2079, x = linear_80_cast_fp16)[name = string("qkv_83_cast_fp16")];
            tensor<int32, [3]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_2081_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_2081_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_2081_cast_fp16_2 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = qkv_83_cast_fp16)[name = string("op_2081_cast_fp16")];
            tensor<int32, [1]> squeeze_60_axes_0 = const()[name = string("squeeze_60_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_60_cast_fp16 = squeeze(axes = squeeze_60_axes_0, x = var_2081_cast_fp16_0)[name = string("squeeze_60_cast_fp16")];
            tensor<int32, [1]> squeeze_61_axes_0 = const()[name = string("squeeze_61_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_61_cast_fp16 = squeeze(axes = squeeze_61_axes_0, x = var_2081_cast_fp16_1)[name = string("squeeze_61_cast_fp16")];
            tensor<int32, [1]> squeeze_62_axes_0 = const()[name = string("squeeze_62_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_62_cast_fp16 = squeeze(axes = squeeze_62_axes_0, x = var_2081_cast_fp16_2)[name = string("squeeze_62_cast_fp16")];
            tensor<int32, [4]> q_81_perm_0 = const()[name = string("q_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_81_perm_0 = const()[name = string("k_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_81_cast_fp16 = transpose(perm = q_81_perm_0, x = squeeze_60_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 16, 224, 64]> var_2091_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_11_to_fp16)[name = string("op_2091_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_81_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_81_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_183_promoted_to_fp16 = const()[name = string("const_183_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2103_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_183_promoted_to_fp16)[name = string("op_2103_cast_fp16")];
            bool var_2105_interleave_0 = const()[name = string("op_2105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2105_cast_fp16 = concat(axis = var_49, interleave = var_2105_interleave_0, values = (var_2103_cast_fp16, x1_81_cast_fp16))[name = string("op_2105_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2106_cast_fp16 = mul(x = var_2105_cast_fp16, y = sin_11_to_fp16)[name = string("op_2106_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_41_cast_fp16 = add(x = var_2091_cast_fp16, y = var_2106_cast_fp16)[name = string("q_embed_41_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_81_cast_fp16 = transpose(perm = k_81_perm_0, x = squeeze_61_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 224, 64]> var_2109_cast_fp16 = mul(x = k_81_cast_fp16, y = cos_11_to_fp16)[name = string("op_2109_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_81_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_81_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_186_promoted_to_fp16 = const()[name = string("const_186_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2121_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_186_promoted_to_fp16)[name = string("op_2121_cast_fp16")];
            bool var_2123_interleave_0 = const()[name = string("op_2123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2123_cast_fp16 = concat(axis = var_49, interleave = var_2123_interleave_0, values = (var_2121_cast_fp16, x1_83_cast_fp16))[name = string("op_2123_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2124_cast_fp16 = mul(x = var_2123_cast_fp16, y = sin_11_to_fp16)[name = string("op_2124_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_41_cast_fp16 = add(x = var_2109_cast_fp16, y = var_2124_cast_fp16)[name = string("k_embed_41_cast_fp16")];
            bool var_2129_transpose_x_1 = const()[name = string("op_2129_transpose_x_1"), val = bool(false)];
            bool var_2129_transpose_y_1 = const()[name = string("op_2129_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_2129_cast_fp16 = matmul(transpose_x = var_2129_transpose_x_1, transpose_y = var_2129_transpose_y_1, x = q_embed_41_cast_fp16, y = k_embed_41_cast_fp16)[name = string("op_2129_cast_fp16")];
            fp16 var_2130_to_fp16 = const()[name = string("op_2130_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_81_cast_fp16 = mul(x = var_2129_cast_fp16, y = var_2130_to_fp16)[name = string("attn_weights_81_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_365_cast_fp16 = add(x = attn_weights_81_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_365_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_2133_cast_fp16 = softmax(axis = var_49, x = input_365_cast_fp16)[name = string("op_2133_cast_fp16")];
            bool attn_output_121_transpose_x_0 = const()[name = string("attn_output_121_transpose_x_0"), val = bool(false)];
            bool attn_output_121_transpose_y_0 = const()[name = string("attn_output_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = squeeze_62_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_121_cast_fp16 = matmul(transpose_x = attn_output_121_transpose_x_0, transpose_y = attn_output_121_transpose_y_0, x = var_2133_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_121_cast_fp16")];
            tensor<int32, [4]> var_2137_perm_0 = const()[name = string("op_2137_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2139 = const()[name = string("op_2139"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_2137_cast_fp16 = transpose(perm = var_2137_perm_0, x = attn_output_121_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 224, 1024]> var_2140_cast_fp16 = reshape(shape = var_2139, x = var_2137_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_20_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_20_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599863808)))];
            tensor<fp16, [1, 224, 1024]> linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_20_attn_Wo_weight_to_fp16, x = var_2140_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_371_cast_fp16 = add(x = input_363_cast_fp16, y = linear_81_cast_fp16)[name = string("input_371_cast_fp16")];
            tensor<int32, [1]> input_373_axes_0 = const()[name = string("input_373_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_20_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_20_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601961024)))];
            tensor<fp16, [1, 224, 1024]> input_373_cast_fp16 = layer_norm(axes = input_373_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_20_mlp_norm_weight_to_fp16, x = input_371_cast_fp16)[name = string("input_373_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_20_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_20_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601963136)))];
            tensor<fp16, [1, 224, 5248]> linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_20_mlp_Wi_weight_to_fp16, x = input_373_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [2]> var_2152_split_sizes_0 = const()[name = string("op_2152_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2152_axis_0 = const()[name = string("op_2152_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_2152_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_2152_cast_fp16_1 = split(axis = var_2152_axis_0, split_sizes = var_2152_split_sizes_0, x = linear_82_cast_fp16)[name = string("op_2152_cast_fp16")];
            string var_2154_mode_0 = const()[name = string("op_2154_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_2154_cast_fp16 = gelu(mode = var_2154_mode_0, x = var_2152_cast_fp16_0)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_377_cast_fp16 = mul(x = var_2154_cast_fp16, y = var_2152_cast_fp16_1)[name = string("input_377_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_20_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_20_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612711104)))];
            tensor<fp16, [1, 224, 1024]> linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_20_mlp_Wo_weight_to_fp16, x = input_377_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_381_cast_fp16 = add(x = input_371_cast_fp16, y = linear_83_cast_fp16)[name = string("input_381_cast_fp16")];
            tensor<int32, [1]> hidden_states_43_axes_0 = const()[name = string("hidden_states_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_21_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_21_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618085120)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_43_cast_fp16 = layer_norm(axes = hidden_states_43_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_21_attn_norm_weight_to_fp16, x = input_381_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_21_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_21_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618087232)))];
            tensor<fp16, [1, 224, 3072]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_21_attn_Wqkv_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [5]> var_2173 = const()[name = string("op_2173"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_87_cast_fp16 = reshape(shape = var_2173, x = linear_84_cast_fp16)[name = string("qkv_87_cast_fp16")];
            tensor<int32, [3]> var_2175_split_sizes_0 = const()[name = string("op_2175_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2175_axis_0 = const()[name = string("op_2175_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_2175_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_2175_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_2175_cast_fp16_2 = split(axis = var_2175_axis_0, split_sizes = var_2175_split_sizes_0, x = qkv_87_cast_fp16)[name = string("op_2175_cast_fp16")];
            tensor<int32, [1]> squeeze_63_axes_0 = const()[name = string("squeeze_63_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_63_cast_fp16 = squeeze(axes = squeeze_63_axes_0, x = var_2175_cast_fp16_0)[name = string("squeeze_63_cast_fp16")];
            tensor<int32, [1]> squeeze_64_axes_0 = const()[name = string("squeeze_64_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_64_cast_fp16 = squeeze(axes = squeeze_64_axes_0, x = var_2175_cast_fp16_1)[name = string("squeeze_64_cast_fp16")];
            tensor<int32, [1]> squeeze_65_axes_0 = const()[name = string("squeeze_65_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_65_cast_fp16 = squeeze(axes = squeeze_65_axes_0, x = var_2175_cast_fp16_2)[name = string("squeeze_65_cast_fp16")];
            tensor<int32, [4]> q_85_perm_0 = const()[name = string("q_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_85_perm_0 = const()[name = string("k_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_85_cast_fp16 = transpose(perm = q_85_perm_0, x = squeeze_63_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 16, 224, 64]> var_2185_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_7_to_fp16)[name = string("op_2185_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_85_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_85_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_191_promoted_to_fp16 = const()[name = string("const_191_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2197_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_191_promoted_to_fp16)[name = string("op_2197_cast_fp16")];
            bool var_2199_interleave_0 = const()[name = string("op_2199_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2199_cast_fp16 = concat(axis = var_49, interleave = var_2199_interleave_0, values = (var_2197_cast_fp16, x1_85_cast_fp16))[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2200_cast_fp16 = mul(x = var_2199_cast_fp16, y = sin_7_to_fp16)[name = string("op_2200_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_43_cast_fp16 = add(x = var_2185_cast_fp16, y = var_2200_cast_fp16)[name = string("q_embed_43_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_85_cast_fp16 = transpose(perm = k_85_perm_0, x = squeeze_64_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 16, 224, 64]> var_2203_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_7_to_fp16)[name = string("op_2203_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_85_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_85_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_194_promoted_to_fp16 = const()[name = string("const_194_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2215_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_194_promoted_to_fp16)[name = string("op_2215_cast_fp16")];
            bool var_2217_interleave_0 = const()[name = string("op_2217_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2217_cast_fp16 = concat(axis = var_49, interleave = var_2217_interleave_0, values = (var_2215_cast_fp16, x1_87_cast_fp16))[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2218_cast_fp16 = mul(x = var_2217_cast_fp16, y = sin_7_to_fp16)[name = string("op_2218_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_43_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2218_cast_fp16)[name = string("k_embed_43_cast_fp16")];
            bool var_2223_transpose_x_1 = const()[name = string("op_2223_transpose_x_1"), val = bool(false)];
            bool var_2223_transpose_y_1 = const()[name = string("op_2223_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_2223_cast_fp16 = matmul(transpose_x = var_2223_transpose_x_1, transpose_y = var_2223_transpose_y_1, x = q_embed_43_cast_fp16, y = k_embed_43_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_85_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("attn_weights_85_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_383_cast_fp16 = add(x = attn_weights_85_cast_fp16, y = key_bias_cast_fp16)[name = string("input_383_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_2227_cast_fp16 = softmax(axis = var_49, x = input_383_cast_fp16)[name = string("op_2227_cast_fp16")];
            bool attn_output_127_transpose_x_0 = const()[name = string("attn_output_127_transpose_x_0"), val = bool(false)];
            bool attn_output_127_transpose_y_0 = const()[name = string("attn_output_127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = squeeze_65_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_127_cast_fp16 = matmul(transpose_x = attn_output_127_transpose_x_0, transpose_y = attn_output_127_transpose_y_0, x = var_2227_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_127_cast_fp16")];
            tensor<int32, [4]> var_2231_perm_0 = const()[name = string("op_2231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_2231_cast_fp16 = transpose(perm = var_2231_perm_0, x = attn_output_127_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 224, 1024]> var_2234_cast_fp16 = reshape(shape = var_2233, x = var_2231_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_21_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_21_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624378752)))];
            tensor<fp16, [1, 224, 1024]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_21_attn_Wo_weight_to_fp16, x = var_2234_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_389_cast_fp16 = add(x = input_381_cast_fp16, y = linear_85_cast_fp16)[name = string("input_389_cast_fp16")];
            tensor<int32, [1]> input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_21_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_21_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626475968)))];
            tensor<fp16, [1, 224, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_21_mlp_norm_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_21_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_21_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626478080)))];
            tensor<fp16, [1, 224, 5248]> linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_21_mlp_Wi_weight_to_fp16, x = input_391_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [2]> var_2246_split_sizes_0 = const()[name = string("op_2246_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2246_axis_0 = const()[name = string("op_2246_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_2246_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_2246_cast_fp16_1 = split(axis = var_2246_axis_0, split_sizes = var_2246_split_sizes_0, x = linear_86_cast_fp16)[name = string("op_2246_cast_fp16")];
            string var_2248_mode_0 = const()[name = string("op_2248_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_2248_cast_fp16 = gelu(mode = var_2248_mode_0, x = var_2246_cast_fp16_0)[name = string("op_2248_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_395_cast_fp16 = mul(x = var_2248_cast_fp16, y = var_2246_cast_fp16_1)[name = string("input_395_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_21_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_21_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637226048)))];
            tensor<fp16, [1, 224, 1024]> linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_21_mlp_Wo_weight_to_fp16, x = input_395_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_399_cast_fp16 = add(x = input_389_cast_fp16, y = linear_87_cast_fp16)[name = string("input_399_cast_fp16")];
            tensor<int32, [1]> hidden_states_45_axes_0 = const()[name = string("hidden_states_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_22_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_22_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642600064)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_45_cast_fp16 = layer_norm(axes = hidden_states_45_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_22_attn_norm_weight_to_fp16, x = input_399_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_22_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_22_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642602176)))];
            tensor<fp16, [1, 224, 3072]> linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_22_attn_Wqkv_weight_to_fp16, x = hidden_states_45_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<int32, [5]> var_2267 = const()[name = string("op_2267"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_91_cast_fp16 = reshape(shape = var_2267, x = linear_88_cast_fp16)[name = string("qkv_91_cast_fp16")];
            tensor<int32, [3]> var_2269_split_sizes_0 = const()[name = string("op_2269_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2269_axis_0 = const()[name = string("op_2269_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_2269_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_2269_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_2269_cast_fp16_2 = split(axis = var_2269_axis_0, split_sizes = var_2269_split_sizes_0, x = qkv_91_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> squeeze_66_axes_0 = const()[name = string("squeeze_66_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_66_cast_fp16 = squeeze(axes = squeeze_66_axes_0, x = var_2269_cast_fp16_0)[name = string("squeeze_66_cast_fp16")];
            tensor<int32, [1]> squeeze_67_axes_0 = const()[name = string("squeeze_67_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_67_cast_fp16 = squeeze(axes = squeeze_67_axes_0, x = var_2269_cast_fp16_1)[name = string("squeeze_67_cast_fp16")];
            tensor<int32, [1]> squeeze_68_axes_0 = const()[name = string("squeeze_68_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_68_cast_fp16 = squeeze(axes = squeeze_68_axes_0, x = var_2269_cast_fp16_2)[name = string("squeeze_68_cast_fp16")];
            tensor<int32, [4]> q_89_perm_0 = const()[name = string("q_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_89_perm_0 = const()[name = string("k_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_89_cast_fp16 = transpose(perm = q_89_perm_0, x = squeeze_66_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 16, 224, 64]> var_2279_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_11_to_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_89_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_89_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_199_promoted_to_fp16 = const()[name = string("const_199_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2291_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_199_promoted_to_fp16)[name = string("op_2291_cast_fp16")];
            bool var_2293_interleave_0 = const()[name = string("op_2293_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2293_cast_fp16 = concat(axis = var_49, interleave = var_2293_interleave_0, values = (var_2291_cast_fp16, x1_89_cast_fp16))[name = string("op_2293_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2294_cast_fp16 = mul(x = var_2293_cast_fp16, y = sin_11_to_fp16)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_45_cast_fp16 = add(x = var_2279_cast_fp16, y = var_2294_cast_fp16)[name = string("q_embed_45_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_89_cast_fp16 = transpose(perm = k_89_perm_0, x = squeeze_67_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 16, 224, 64]> var_2297_cast_fp16 = mul(x = k_89_cast_fp16, y = cos_11_to_fp16)[name = string("op_2297_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_89_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_89_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_202_promoted_to_fp16 = const()[name = string("const_202_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2309_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_202_promoted_to_fp16)[name = string("op_2309_cast_fp16")];
            bool var_2311_interleave_0 = const()[name = string("op_2311_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2311_cast_fp16 = concat(axis = var_49, interleave = var_2311_interleave_0, values = (var_2309_cast_fp16, x1_91_cast_fp16))[name = string("op_2311_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2312_cast_fp16 = mul(x = var_2311_cast_fp16, y = sin_11_to_fp16)[name = string("op_2312_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_45_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2312_cast_fp16)[name = string("k_embed_45_cast_fp16")];
            bool var_2317_transpose_x_1 = const()[name = string("op_2317_transpose_x_1"), val = bool(false)];
            bool var_2317_transpose_y_1 = const()[name = string("op_2317_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_2317_cast_fp16 = matmul(transpose_x = var_2317_transpose_x_1, transpose_y = var_2317_transpose_y_1, x = q_embed_45_cast_fp16, y = k_embed_45_cast_fp16)[name = string("op_2317_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_89_cast_fp16 = mul(x = var_2317_cast_fp16, y = var_2318_to_fp16)[name = string("attn_weights_89_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_401_cast_fp16 = add(x = attn_weights_89_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_401_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_2321_cast_fp16 = softmax(axis = var_49, x = input_401_cast_fp16)[name = string("op_2321_cast_fp16")];
            bool attn_output_133_transpose_x_0 = const()[name = string("attn_output_133_transpose_x_0"), val = bool(false)];
            bool attn_output_133_transpose_y_0 = const()[name = string("attn_output_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = squeeze_68_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_133_cast_fp16 = matmul(transpose_x = attn_output_133_transpose_x_0, transpose_y = attn_output_133_transpose_y_0, x = var_2321_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_133_cast_fp16")];
            tensor<int32, [4]> var_2325_perm_0 = const()[name = string("op_2325_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2327 = const()[name = string("op_2327"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_2325_cast_fp16 = transpose(perm = var_2325_perm_0, x = attn_output_133_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 224, 1024]> var_2328_cast_fp16 = reshape(shape = var_2327, x = var_2325_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_22_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_22_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648893696)))];
            tensor<fp16, [1, 224, 1024]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_22_attn_Wo_weight_to_fp16, x = var_2328_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_407_cast_fp16 = add(x = input_399_cast_fp16, y = linear_89_cast_fp16)[name = string("input_407_cast_fp16")];
            tensor<int32, [1]> input_409_axes_0 = const()[name = string("input_409_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_22_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_22_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650990912)))];
            tensor<fp16, [1, 224, 1024]> input_409_cast_fp16 = layer_norm(axes = input_409_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_22_mlp_norm_weight_to_fp16, x = input_407_cast_fp16)[name = string("input_409_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_22_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_22_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650993024)))];
            tensor<fp16, [1, 224, 5248]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_22_mlp_Wi_weight_to_fp16, x = input_409_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<int32, [2]> var_2340_split_sizes_0 = const()[name = string("op_2340_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2340_axis_0 = const()[name = string("op_2340_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_2340_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_90_cast_fp16)[name = string("op_2340_cast_fp16")];
            string var_2342_mode_0 = const()[name = string("op_2342_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_2342_cast_fp16 = gelu(mode = var_2342_mode_0, x = var_2340_cast_fp16_0)[name = string("op_2342_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_413_cast_fp16 = mul(x = var_2342_cast_fp16, y = var_2340_cast_fp16_1)[name = string("input_413_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_22_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_22_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661740992)))];
            tensor<fp16, [1, 224, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_22_mlp_Wo_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_417_cast_fp16 = add(x = input_407_cast_fp16, y = linear_91_cast_fp16)[name = string("input_417_cast_fp16")];
            tensor<int32, [1]> hidden_states_47_axes_0 = const()[name = string("hidden_states_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_23_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_23_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667115008)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_47_cast_fp16 = layer_norm(axes = hidden_states_47_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_23_attn_norm_weight_to_fp16, x = input_417_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_23_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_23_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667117120)))];
            tensor<fp16, [1, 224, 3072]> linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_23_attn_Wqkv_weight_to_fp16, x = hidden_states_47_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [5]> var_2361 = const()[name = string("op_2361"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_95_cast_fp16 = reshape(shape = var_2361, x = linear_92_cast_fp16)[name = string("qkv_95_cast_fp16")];
            tensor<int32, [3]> var_2363_split_sizes_0 = const()[name = string("op_2363_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2363_axis_0 = const()[name = string("op_2363_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_2363_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_2363_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_2363_cast_fp16_2 = split(axis = var_2363_axis_0, split_sizes = var_2363_split_sizes_0, x = qkv_95_cast_fp16)[name = string("op_2363_cast_fp16")];
            tensor<int32, [1]> squeeze_69_axes_0 = const()[name = string("squeeze_69_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_69_cast_fp16 = squeeze(axes = squeeze_69_axes_0, x = var_2363_cast_fp16_0)[name = string("squeeze_69_cast_fp16")];
            tensor<int32, [1]> squeeze_70_axes_0 = const()[name = string("squeeze_70_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_70_cast_fp16 = squeeze(axes = squeeze_70_axes_0, x = var_2363_cast_fp16_1)[name = string("squeeze_70_cast_fp16")];
            tensor<int32, [1]> squeeze_71_axes_0 = const()[name = string("squeeze_71_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_71_cast_fp16 = squeeze(axes = squeeze_71_axes_0, x = var_2363_cast_fp16_2)[name = string("squeeze_71_cast_fp16")];
            tensor<int32, [4]> q_93_perm_0 = const()[name = string("q_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_93_perm_0 = const()[name = string("k_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_93_cast_fp16 = transpose(perm = q_93_perm_0, x = squeeze_69_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 224, 64]> var_2373_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_11_to_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_93_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_93_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_207_promoted_to_fp16 = const()[name = string("const_207_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2385_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_207_promoted_to_fp16)[name = string("op_2385_cast_fp16")];
            bool var_2387_interleave_0 = const()[name = string("op_2387_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2387_cast_fp16 = concat(axis = var_49, interleave = var_2387_interleave_0, values = (var_2385_cast_fp16, x1_93_cast_fp16))[name = string("op_2387_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2388_cast_fp16 = mul(x = var_2387_cast_fp16, y = sin_11_to_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_47_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2388_cast_fp16)[name = string("q_embed_47_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_93_cast_fp16 = transpose(perm = k_93_perm_0, x = squeeze_70_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 16, 224, 64]> var_2391_cast_fp16 = mul(x = k_93_cast_fp16, y = cos_11_to_fp16)[name = string("op_2391_cast_fp16")];
            tensor<int32, [4]> x1_95_begin_0 = const()[name = string("x1_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_95_end_0 = const()[name = string("x1_95_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_95_end_mask_0 = const()[name = string("x1_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_95_cast_fp16 = slice_by_index(begin = x1_95_begin_0, end = x1_95_end_0, end_mask = x1_95_end_mask_0, x = k_93_cast_fp16)[name = string("x1_95_cast_fp16")];
            tensor<int32, [4]> x2_95_begin_0 = const()[name = string("x2_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_95_end_0 = const()[name = string("x2_95_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_95_end_mask_0 = const()[name = string("x2_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_95_cast_fp16 = slice_by_index(begin = x2_95_begin_0, end = x2_95_end_0, end_mask = x2_95_end_mask_0, x = k_93_cast_fp16)[name = string("x2_95_cast_fp16")];
            fp16 const_210_promoted_to_fp16 = const()[name = string("const_210_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2403_cast_fp16 = mul(x = x2_95_cast_fp16, y = const_210_promoted_to_fp16)[name = string("op_2403_cast_fp16")];
            bool var_2405_interleave_0 = const()[name = string("op_2405_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2405_cast_fp16 = concat(axis = var_49, interleave = var_2405_interleave_0, values = (var_2403_cast_fp16, x1_95_cast_fp16))[name = string("op_2405_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2406_cast_fp16 = mul(x = var_2405_cast_fp16, y = sin_11_to_fp16)[name = string("op_2406_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_47_cast_fp16 = add(x = var_2391_cast_fp16, y = var_2406_cast_fp16)[name = string("k_embed_47_cast_fp16")];
            bool var_2411_transpose_x_1 = const()[name = string("op_2411_transpose_x_1"), val = bool(false)];
            bool var_2411_transpose_y_1 = const()[name = string("op_2411_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_2411_cast_fp16 = matmul(transpose_x = var_2411_transpose_x_1, transpose_y = var_2411_transpose_y_1, x = q_embed_47_cast_fp16, y = k_embed_47_cast_fp16)[name = string("op_2411_cast_fp16")];
            fp16 var_2412_to_fp16 = const()[name = string("op_2412_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_93_cast_fp16 = mul(x = var_2411_cast_fp16, y = var_2412_to_fp16)[name = string("attn_weights_93_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_419_cast_fp16 = add(x = attn_weights_93_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_419_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_2415_cast_fp16 = softmax(axis = var_49, x = input_419_cast_fp16)[name = string("op_2415_cast_fp16")];
            bool attn_output_139_transpose_x_0 = const()[name = string("attn_output_139_transpose_x_0"), val = bool(false)];
            bool attn_output_139_transpose_y_0 = const()[name = string("attn_output_139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = squeeze_71_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_139_cast_fp16 = matmul(transpose_x = attn_output_139_transpose_x_0, transpose_y = attn_output_139_transpose_y_0, x = var_2415_cast_fp16, y = value_47_cast_fp16)[name = string("attn_output_139_cast_fp16")];
            tensor<int32, [4]> var_2419_perm_0 = const()[name = string("op_2419_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2421 = const()[name = string("op_2421"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_2419_cast_fp16 = transpose(perm = var_2419_perm_0, x = attn_output_139_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 224, 1024]> var_2422_cast_fp16 = reshape(shape = var_2421, x = var_2419_cast_fp16)[name = string("op_2422_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_23_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_23_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673408640)))];
            tensor<fp16, [1, 224, 1024]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_23_attn_Wo_weight_to_fp16, x = var_2422_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_425_cast_fp16 = add(x = input_417_cast_fp16, y = linear_93_cast_fp16)[name = string("input_425_cast_fp16")];
            tensor<int32, [1]> input_427_axes_0 = const()[name = string("input_427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_23_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_23_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675505856)))];
            tensor<fp16, [1, 224, 1024]> input_427_cast_fp16 = layer_norm(axes = input_427_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_23_mlp_norm_weight_to_fp16, x = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_23_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_23_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675507968)))];
            tensor<fp16, [1, 224, 5248]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_23_mlp_Wi_weight_to_fp16, x = input_427_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<int32, [2]> var_2434_split_sizes_0 = const()[name = string("op_2434_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2434_axis_0 = const()[name = string("op_2434_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_2434_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_2434_cast_fp16_1 = split(axis = var_2434_axis_0, split_sizes = var_2434_split_sizes_0, x = linear_94_cast_fp16)[name = string("op_2434_cast_fp16")];
            string var_2436_mode_0 = const()[name = string("op_2436_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_2436_cast_fp16 = gelu(mode = var_2436_mode_0, x = var_2434_cast_fp16_0)[name = string("op_2436_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_431_cast_fp16 = mul(x = var_2436_cast_fp16, y = var_2434_cast_fp16_1)[name = string("input_431_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_23_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_23_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686255936)))];
            tensor<fp16, [1, 224, 1024]> linear_95_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_23_mlp_Wo_weight_to_fp16, x = input_431_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_435_cast_fp16 = add(x = input_425_cast_fp16, y = linear_95_cast_fp16)[name = string("input_435_cast_fp16")];
            tensor<int32, [1]> hidden_states_49_axes_0 = const()[name = string("hidden_states_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_24_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_24_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691629952)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_49_cast_fp16 = layer_norm(axes = hidden_states_49_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_24_attn_norm_weight_to_fp16, x = input_435_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_24_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_24_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691632064)))];
            tensor<fp16, [1, 224, 3072]> linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_24_attn_Wqkv_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<int32, [5]> var_2455 = const()[name = string("op_2455"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_99_cast_fp16 = reshape(shape = var_2455, x = linear_96_cast_fp16)[name = string("qkv_99_cast_fp16")];
            tensor<int32, [3]> var_2457_split_sizes_0 = const()[name = string("op_2457_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2457_axis_0 = const()[name = string("op_2457_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_2457_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_2457_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_2457_cast_fp16_2 = split(axis = var_2457_axis_0, split_sizes = var_2457_split_sizes_0, x = qkv_99_cast_fp16)[name = string("op_2457_cast_fp16")];
            tensor<int32, [1]> squeeze_72_axes_0 = const()[name = string("squeeze_72_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_72_cast_fp16 = squeeze(axes = squeeze_72_axes_0, x = var_2457_cast_fp16_0)[name = string("squeeze_72_cast_fp16")];
            tensor<int32, [1]> squeeze_73_axes_0 = const()[name = string("squeeze_73_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_73_cast_fp16 = squeeze(axes = squeeze_73_axes_0, x = var_2457_cast_fp16_1)[name = string("squeeze_73_cast_fp16")];
            tensor<int32, [1]> squeeze_74_axes_0 = const()[name = string("squeeze_74_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_74_cast_fp16 = squeeze(axes = squeeze_74_axes_0, x = var_2457_cast_fp16_2)[name = string("squeeze_74_cast_fp16")];
            tensor<int32, [4]> q_97_perm_0 = const()[name = string("q_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_97_perm_0 = const()[name = string("k_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_97_cast_fp16 = transpose(perm = q_97_perm_0, x = squeeze_72_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, 224, 64]> var_2467_cast_fp16 = mul(x = q_97_cast_fp16, y = cos_7_to_fp16)[name = string("op_2467_cast_fp16")];
            tensor<int32, [4]> x1_97_begin_0 = const()[name = string("x1_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_97_end_0 = const()[name = string("x1_97_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_97_end_mask_0 = const()[name = string("x1_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_97_cast_fp16 = slice_by_index(begin = x1_97_begin_0, end = x1_97_end_0, end_mask = x1_97_end_mask_0, x = q_97_cast_fp16)[name = string("x1_97_cast_fp16")];
            tensor<int32, [4]> x2_97_begin_0 = const()[name = string("x2_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_97_end_0 = const()[name = string("x2_97_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_97_end_mask_0 = const()[name = string("x2_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_97_cast_fp16 = slice_by_index(begin = x2_97_begin_0, end = x2_97_end_0, end_mask = x2_97_end_mask_0, x = q_97_cast_fp16)[name = string("x2_97_cast_fp16")];
            fp16 const_215_promoted_to_fp16 = const()[name = string("const_215_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2479_cast_fp16 = mul(x = x2_97_cast_fp16, y = const_215_promoted_to_fp16)[name = string("op_2479_cast_fp16")];
            bool var_2481_interleave_0 = const()[name = string("op_2481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2481_cast_fp16 = concat(axis = var_49, interleave = var_2481_interleave_0, values = (var_2479_cast_fp16, x1_97_cast_fp16))[name = string("op_2481_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2482_cast_fp16 = mul(x = var_2481_cast_fp16, y = sin_7_to_fp16)[name = string("op_2482_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_49_cast_fp16 = add(x = var_2467_cast_fp16, y = var_2482_cast_fp16)[name = string("q_embed_49_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_97_cast_fp16 = transpose(perm = k_97_perm_0, x = squeeze_73_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 16, 224, 64]> var_2485_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_7_to_fp16)[name = string("op_2485_cast_fp16")];
            tensor<int32, [4]> x1_99_begin_0 = const()[name = string("x1_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_99_end_0 = const()[name = string("x1_99_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_99_end_mask_0 = const()[name = string("x1_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_99_cast_fp16 = slice_by_index(begin = x1_99_begin_0, end = x1_99_end_0, end_mask = x1_99_end_mask_0, x = k_97_cast_fp16)[name = string("x1_99_cast_fp16")];
            tensor<int32, [4]> x2_99_begin_0 = const()[name = string("x2_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_99_end_0 = const()[name = string("x2_99_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_99_end_mask_0 = const()[name = string("x2_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_99_cast_fp16 = slice_by_index(begin = x2_99_begin_0, end = x2_99_end_0, end_mask = x2_99_end_mask_0, x = k_97_cast_fp16)[name = string("x2_99_cast_fp16")];
            fp16 const_218_promoted_to_fp16 = const()[name = string("const_218_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2497_cast_fp16 = mul(x = x2_99_cast_fp16, y = const_218_promoted_to_fp16)[name = string("op_2497_cast_fp16")];
            bool var_2499_interleave_0 = const()[name = string("op_2499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2499_cast_fp16 = concat(axis = var_49, interleave = var_2499_interleave_0, values = (var_2497_cast_fp16, x1_99_cast_fp16))[name = string("op_2499_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2500_cast_fp16 = mul(x = var_2499_cast_fp16, y = sin_7_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_49_cast_fp16 = add(x = var_2485_cast_fp16, y = var_2500_cast_fp16)[name = string("k_embed_49_cast_fp16")];
            bool var_2505_transpose_x_1 = const()[name = string("op_2505_transpose_x_1"), val = bool(false)];
            bool var_2505_transpose_y_1 = const()[name = string("op_2505_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_2505_cast_fp16 = matmul(transpose_x = var_2505_transpose_x_1, transpose_y = var_2505_transpose_y_1, x = q_embed_49_cast_fp16, y = k_embed_49_cast_fp16)[name = string("op_2505_cast_fp16")];
            fp16 var_2506_to_fp16 = const()[name = string("op_2506_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_97_cast_fp16 = mul(x = var_2505_cast_fp16, y = var_2506_to_fp16)[name = string("attn_weights_97_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_437_cast_fp16 = add(x = attn_weights_97_cast_fp16, y = key_bias_cast_fp16)[name = string("input_437_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_2509_cast_fp16 = softmax(axis = var_49, x = input_437_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool attn_output_145_transpose_x_0 = const()[name = string("attn_output_145_transpose_x_0"), val = bool(false)];
            bool attn_output_145_transpose_y_0 = const()[name = string("attn_output_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = squeeze_74_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_145_cast_fp16 = matmul(transpose_x = attn_output_145_transpose_x_0, transpose_y = attn_output_145_transpose_y_0, x = var_2509_cast_fp16, y = value_49_cast_fp16)[name = string("attn_output_145_cast_fp16")];
            tensor<int32, [4]> var_2513_perm_0 = const()[name = string("op_2513_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2515 = const()[name = string("op_2515"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_2513_cast_fp16 = transpose(perm = var_2513_perm_0, x = attn_output_145_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 224, 1024]> var_2516_cast_fp16 = reshape(shape = var_2515, x = var_2513_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_24_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_24_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(697923584)))];
            tensor<fp16, [1, 224, 1024]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_24_attn_Wo_weight_to_fp16, x = var_2516_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_443_cast_fp16 = add(x = input_435_cast_fp16, y = linear_97_cast_fp16)[name = string("input_443_cast_fp16")];
            tensor<int32, [1]> input_445_axes_0 = const()[name = string("input_445_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_24_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_24_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700020800)))];
            tensor<fp16, [1, 224, 1024]> input_445_cast_fp16 = layer_norm(axes = input_445_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_24_mlp_norm_weight_to_fp16, x = input_443_cast_fp16)[name = string("input_445_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_24_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_24_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700022912)))];
            tensor<fp16, [1, 224, 5248]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_24_mlp_Wi_weight_to_fp16, x = input_445_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [2]> var_2528_split_sizes_0 = const()[name = string("op_2528_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2528_axis_0 = const()[name = string("op_2528_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_2528_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_2528_cast_fp16_1 = split(axis = var_2528_axis_0, split_sizes = var_2528_split_sizes_0, x = linear_98_cast_fp16)[name = string("op_2528_cast_fp16")];
            string var_2530_mode_0 = const()[name = string("op_2530_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_2530_cast_fp16 = gelu(mode = var_2530_mode_0, x = var_2528_cast_fp16_0)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_449_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2528_cast_fp16_1)[name = string("input_449_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_24_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_24_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710770880)))];
            tensor<fp16, [1, 224, 1024]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_24_mlp_Wo_weight_to_fp16, x = input_449_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_453_cast_fp16 = add(x = input_443_cast_fp16, y = linear_99_cast_fp16)[name = string("input_453_cast_fp16")];
            tensor<int32, [1]> hidden_states_51_axes_0 = const()[name = string("hidden_states_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_25_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_25_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(716144896)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_51_cast_fp16 = layer_norm(axes = hidden_states_51_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_25_attn_norm_weight_to_fp16, x = input_453_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_25_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_25_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(716147008)))];
            tensor<fp16, [1, 224, 3072]> linear_100_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_25_attn_Wqkv_weight_to_fp16, x = hidden_states_51_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [5]> var_2549 = const()[name = string("op_2549"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_103_cast_fp16 = reshape(shape = var_2549, x = linear_100_cast_fp16)[name = string("qkv_103_cast_fp16")];
            tensor<int32, [3]> var_2551_split_sizes_0 = const()[name = string("op_2551_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2551_axis_0 = const()[name = string("op_2551_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_2551_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_2551_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_2551_cast_fp16_2 = split(axis = var_2551_axis_0, split_sizes = var_2551_split_sizes_0, x = qkv_103_cast_fp16)[name = string("op_2551_cast_fp16")];
            tensor<int32, [1]> squeeze_75_axes_0 = const()[name = string("squeeze_75_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_75_cast_fp16 = squeeze(axes = squeeze_75_axes_0, x = var_2551_cast_fp16_0)[name = string("squeeze_75_cast_fp16")];
            tensor<int32, [1]> squeeze_76_axes_0 = const()[name = string("squeeze_76_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_76_cast_fp16 = squeeze(axes = squeeze_76_axes_0, x = var_2551_cast_fp16_1)[name = string("squeeze_76_cast_fp16")];
            tensor<int32, [1]> squeeze_77_axes_0 = const()[name = string("squeeze_77_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_77_cast_fp16 = squeeze(axes = squeeze_77_axes_0, x = var_2551_cast_fp16_2)[name = string("squeeze_77_cast_fp16")];
            tensor<int32, [4]> q_101_perm_0 = const()[name = string("q_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_101_perm_0 = const()[name = string("k_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_101_cast_fp16 = transpose(perm = q_101_perm_0, x = squeeze_75_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 16, 224, 64]> var_2561_cast_fp16 = mul(x = q_101_cast_fp16, y = cos_11_to_fp16)[name = string("op_2561_cast_fp16")];
            tensor<int32, [4]> x1_101_begin_0 = const()[name = string("x1_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_101_end_0 = const()[name = string("x1_101_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_101_end_mask_0 = const()[name = string("x1_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_101_cast_fp16 = slice_by_index(begin = x1_101_begin_0, end = x1_101_end_0, end_mask = x1_101_end_mask_0, x = q_101_cast_fp16)[name = string("x1_101_cast_fp16")];
            tensor<int32, [4]> x2_101_begin_0 = const()[name = string("x2_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_101_end_0 = const()[name = string("x2_101_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_101_end_mask_0 = const()[name = string("x2_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_101_cast_fp16 = slice_by_index(begin = x2_101_begin_0, end = x2_101_end_0, end_mask = x2_101_end_mask_0, x = q_101_cast_fp16)[name = string("x2_101_cast_fp16")];
            fp16 const_223_promoted_to_fp16 = const()[name = string("const_223_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2573_cast_fp16 = mul(x = x2_101_cast_fp16, y = const_223_promoted_to_fp16)[name = string("op_2573_cast_fp16")];
            bool var_2575_interleave_0 = const()[name = string("op_2575_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2575_cast_fp16 = concat(axis = var_49, interleave = var_2575_interleave_0, values = (var_2573_cast_fp16, x1_101_cast_fp16))[name = string("op_2575_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2576_cast_fp16 = mul(x = var_2575_cast_fp16, y = sin_11_to_fp16)[name = string("op_2576_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_51_cast_fp16 = add(x = var_2561_cast_fp16, y = var_2576_cast_fp16)[name = string("q_embed_51_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_101_cast_fp16 = transpose(perm = k_101_perm_0, x = squeeze_76_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 16, 224, 64]> var_2579_cast_fp16 = mul(x = k_101_cast_fp16, y = cos_11_to_fp16)[name = string("op_2579_cast_fp16")];
            tensor<int32, [4]> x1_103_begin_0 = const()[name = string("x1_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_103_end_0 = const()[name = string("x1_103_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_103_end_mask_0 = const()[name = string("x1_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_103_cast_fp16 = slice_by_index(begin = x1_103_begin_0, end = x1_103_end_0, end_mask = x1_103_end_mask_0, x = k_101_cast_fp16)[name = string("x1_103_cast_fp16")];
            tensor<int32, [4]> x2_103_begin_0 = const()[name = string("x2_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_103_end_0 = const()[name = string("x2_103_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_103_end_mask_0 = const()[name = string("x2_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_103_cast_fp16 = slice_by_index(begin = x2_103_begin_0, end = x2_103_end_0, end_mask = x2_103_end_mask_0, x = k_101_cast_fp16)[name = string("x2_103_cast_fp16")];
            fp16 const_226_promoted_to_fp16 = const()[name = string("const_226_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2591_cast_fp16 = mul(x = x2_103_cast_fp16, y = const_226_promoted_to_fp16)[name = string("op_2591_cast_fp16")];
            bool var_2593_interleave_0 = const()[name = string("op_2593_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2593_cast_fp16 = concat(axis = var_49, interleave = var_2593_interleave_0, values = (var_2591_cast_fp16, x1_103_cast_fp16))[name = string("op_2593_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2594_cast_fp16 = mul(x = var_2593_cast_fp16, y = sin_11_to_fp16)[name = string("op_2594_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_51_cast_fp16 = add(x = var_2579_cast_fp16, y = var_2594_cast_fp16)[name = string("k_embed_51_cast_fp16")];
            bool var_2599_transpose_x_1 = const()[name = string("op_2599_transpose_x_1"), val = bool(false)];
            bool var_2599_transpose_y_1 = const()[name = string("op_2599_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_2599_cast_fp16 = matmul(transpose_x = var_2599_transpose_x_1, transpose_y = var_2599_transpose_y_1, x = q_embed_51_cast_fp16, y = k_embed_51_cast_fp16)[name = string("op_2599_cast_fp16")];
            fp16 var_2600_to_fp16 = const()[name = string("op_2600_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_101_cast_fp16 = mul(x = var_2599_cast_fp16, y = var_2600_to_fp16)[name = string("attn_weights_101_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_455_cast_fp16 = add(x = attn_weights_101_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_455_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_2603_cast_fp16 = softmax(axis = var_49, x = input_455_cast_fp16)[name = string("op_2603_cast_fp16")];
            bool attn_output_151_transpose_x_0 = const()[name = string("attn_output_151_transpose_x_0"), val = bool(false)];
            bool attn_output_151_transpose_y_0 = const()[name = string("attn_output_151_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = squeeze_77_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_151_cast_fp16 = matmul(transpose_x = attn_output_151_transpose_x_0, transpose_y = attn_output_151_transpose_y_0, x = var_2603_cast_fp16, y = value_51_cast_fp16)[name = string("attn_output_151_cast_fp16")];
            tensor<int32, [4]> var_2607_perm_0 = const()[name = string("op_2607_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2609 = const()[name = string("op_2609"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_2607_cast_fp16 = transpose(perm = var_2607_perm_0, x = attn_output_151_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 224, 1024]> var_2610_cast_fp16 = reshape(shape = var_2609, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_25_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_25_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(722438528)))];
            tensor<fp16, [1, 224, 1024]> linear_101_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_25_attn_Wo_weight_to_fp16, x = var_2610_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_461_cast_fp16 = add(x = input_453_cast_fp16, y = linear_101_cast_fp16)[name = string("input_461_cast_fp16")];
            tensor<int32, [1]> input_463_axes_0 = const()[name = string("input_463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_25_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_25_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724535744)))];
            tensor<fp16, [1, 224, 1024]> input_463_cast_fp16 = layer_norm(axes = input_463_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_25_mlp_norm_weight_to_fp16, x = input_461_cast_fp16)[name = string("input_463_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_25_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_25_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724537856)))];
            tensor<fp16, [1, 224, 5248]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_25_mlp_Wi_weight_to_fp16, x = input_463_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [2]> var_2622_split_sizes_0 = const()[name = string("op_2622_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2622_axis_0 = const()[name = string("op_2622_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_2622_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_2622_cast_fp16_1 = split(axis = var_2622_axis_0, split_sizes = var_2622_split_sizes_0, x = linear_102_cast_fp16)[name = string("op_2622_cast_fp16")];
            string var_2624_mode_0 = const()[name = string("op_2624_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_2624_cast_fp16 = gelu(mode = var_2624_mode_0, x = var_2622_cast_fp16_0)[name = string("op_2624_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_467_cast_fp16 = mul(x = var_2624_cast_fp16, y = var_2622_cast_fp16_1)[name = string("input_467_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_25_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_25_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(735285824)))];
            tensor<fp16, [1, 224, 1024]> linear_103_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_25_mlp_Wo_weight_to_fp16, x = input_467_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_471_cast_fp16 = add(x = input_461_cast_fp16, y = linear_103_cast_fp16)[name = string("input_471_cast_fp16")];
            tensor<int32, [1]> hidden_states_53_axes_0 = const()[name = string("hidden_states_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_26_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_26_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740659840)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_53_cast_fp16 = layer_norm(axes = hidden_states_53_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_26_attn_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_26_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_26_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740661952)))];
            tensor<fp16, [1, 224, 3072]> linear_104_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_26_attn_Wqkv_weight_to_fp16, x = hidden_states_53_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<int32, [5]> var_2643 = const()[name = string("op_2643"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_107_cast_fp16 = reshape(shape = var_2643, x = linear_104_cast_fp16)[name = string("qkv_107_cast_fp16")];
            tensor<int32, [3]> var_2645_split_sizes_0 = const()[name = string("op_2645_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2645_axis_0 = const()[name = string("op_2645_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_2645_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_2645_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_2645_cast_fp16_2 = split(axis = var_2645_axis_0, split_sizes = var_2645_split_sizes_0, x = qkv_107_cast_fp16)[name = string("op_2645_cast_fp16")];
            tensor<int32, [1]> squeeze_78_axes_0 = const()[name = string("squeeze_78_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_78_cast_fp16 = squeeze(axes = squeeze_78_axes_0, x = var_2645_cast_fp16_0)[name = string("squeeze_78_cast_fp16")];
            tensor<int32, [1]> squeeze_79_axes_0 = const()[name = string("squeeze_79_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_79_cast_fp16 = squeeze(axes = squeeze_79_axes_0, x = var_2645_cast_fp16_1)[name = string("squeeze_79_cast_fp16")];
            tensor<int32, [1]> squeeze_80_axes_0 = const()[name = string("squeeze_80_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_80_cast_fp16 = squeeze(axes = squeeze_80_axes_0, x = var_2645_cast_fp16_2)[name = string("squeeze_80_cast_fp16")];
            tensor<int32, [4]> q_105_perm_0 = const()[name = string("q_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_105_perm_0 = const()[name = string("k_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_105_cast_fp16 = transpose(perm = q_105_perm_0, x = squeeze_78_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 16, 224, 64]> var_2655_cast_fp16 = mul(x = q_105_cast_fp16, y = cos_11_to_fp16)[name = string("op_2655_cast_fp16")];
            tensor<int32, [4]> x1_105_begin_0 = const()[name = string("x1_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_105_end_0 = const()[name = string("x1_105_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_105_end_mask_0 = const()[name = string("x1_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_105_cast_fp16 = slice_by_index(begin = x1_105_begin_0, end = x1_105_end_0, end_mask = x1_105_end_mask_0, x = q_105_cast_fp16)[name = string("x1_105_cast_fp16")];
            tensor<int32, [4]> x2_105_begin_0 = const()[name = string("x2_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_105_end_0 = const()[name = string("x2_105_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_105_end_mask_0 = const()[name = string("x2_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_105_cast_fp16 = slice_by_index(begin = x2_105_begin_0, end = x2_105_end_0, end_mask = x2_105_end_mask_0, x = q_105_cast_fp16)[name = string("x2_105_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2667_cast_fp16 = mul(x = x2_105_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_2667_cast_fp16")];
            bool var_2669_interleave_0 = const()[name = string("op_2669_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2669_cast_fp16 = concat(axis = var_49, interleave = var_2669_interleave_0, values = (var_2667_cast_fp16, x1_105_cast_fp16))[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2670_cast_fp16 = mul(x = var_2669_cast_fp16, y = sin_11_to_fp16)[name = string("op_2670_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_53_cast_fp16 = add(x = var_2655_cast_fp16, y = var_2670_cast_fp16)[name = string("q_embed_53_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_105_cast_fp16 = transpose(perm = k_105_perm_0, x = squeeze_79_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 16, 224, 64]> var_2673_cast_fp16 = mul(x = k_105_cast_fp16, y = cos_11_to_fp16)[name = string("op_2673_cast_fp16")];
            tensor<int32, [4]> x1_107_begin_0 = const()[name = string("x1_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_107_end_0 = const()[name = string("x1_107_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_107_end_mask_0 = const()[name = string("x1_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_107_cast_fp16 = slice_by_index(begin = x1_107_begin_0, end = x1_107_end_0, end_mask = x1_107_end_mask_0, x = k_105_cast_fp16)[name = string("x1_107_cast_fp16")];
            tensor<int32, [4]> x2_107_begin_0 = const()[name = string("x2_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_107_end_0 = const()[name = string("x2_107_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_107_end_mask_0 = const()[name = string("x2_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_107_cast_fp16 = slice_by_index(begin = x2_107_begin_0, end = x2_107_end_0, end_mask = x2_107_end_mask_0, x = k_105_cast_fp16)[name = string("x2_107_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2685_cast_fp16 = mul(x = x2_107_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2685_cast_fp16")];
            bool var_2687_interleave_0 = const()[name = string("op_2687_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2687_cast_fp16 = concat(axis = var_49, interleave = var_2687_interleave_0, values = (var_2685_cast_fp16, x1_107_cast_fp16))[name = string("op_2687_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2688_cast_fp16 = mul(x = var_2687_cast_fp16, y = sin_11_to_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_53_cast_fp16 = add(x = var_2673_cast_fp16, y = var_2688_cast_fp16)[name = string("k_embed_53_cast_fp16")];
            bool var_2693_transpose_x_1 = const()[name = string("op_2693_transpose_x_1"), val = bool(false)];
            bool var_2693_transpose_y_1 = const()[name = string("op_2693_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_2693_cast_fp16 = matmul(transpose_x = var_2693_transpose_x_1, transpose_y = var_2693_transpose_y_1, x = q_embed_53_cast_fp16, y = k_embed_53_cast_fp16)[name = string("op_2693_cast_fp16")];
            fp16 var_2694_to_fp16 = const()[name = string("op_2694_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_105_cast_fp16 = mul(x = var_2693_cast_fp16, y = var_2694_to_fp16)[name = string("attn_weights_105_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_473_cast_fp16 = add(x = attn_weights_105_cast_fp16, y = attention_mask_cast_fp16)[name = string("input_473_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_2697_cast_fp16 = softmax(axis = var_49, x = input_473_cast_fp16)[name = string("op_2697_cast_fp16")];
            bool attn_output_157_transpose_x_0 = const()[name = string("attn_output_157_transpose_x_0"), val = bool(false)];
            bool attn_output_157_transpose_y_0 = const()[name = string("attn_output_157_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = squeeze_80_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_157_cast_fp16 = matmul(transpose_x = attn_output_157_transpose_x_0, transpose_y = attn_output_157_transpose_y_0, x = var_2697_cast_fp16, y = value_53_cast_fp16)[name = string("attn_output_157_cast_fp16")];
            tensor<int32, [4]> var_2701_perm_0 = const()[name = string("op_2701_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2703 = const()[name = string("op_2703"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_2701_cast_fp16 = transpose(perm = var_2701_perm_0, x = attn_output_157_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 224, 1024]> var_2704_cast_fp16 = reshape(shape = var_2703, x = var_2701_cast_fp16)[name = string("op_2704_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_26_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_26_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746953472)))];
            tensor<fp16, [1, 224, 1024]> linear_105_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_26_attn_Wo_weight_to_fp16, x = var_2704_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_479_cast_fp16 = add(x = input_471_cast_fp16, y = linear_105_cast_fp16)[name = string("input_479_cast_fp16")];
            tensor<int32, [1]> input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_26_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_26_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749050688)))];
            tensor<fp16, [1, 224, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_26_mlp_norm_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_26_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_26_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749052800)))];
            tensor<fp16, [1, 224, 5248]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_26_mlp_Wi_weight_to_fp16, x = input_481_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [2]> var_2716_split_sizes_0 = const()[name = string("op_2716_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2716_axis_0 = const()[name = string("op_2716_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_2716_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_2716_cast_fp16_1 = split(axis = var_2716_axis_0, split_sizes = var_2716_split_sizes_0, x = linear_106_cast_fp16)[name = string("op_2716_cast_fp16")];
            string var_2718_mode_0 = const()[name = string("op_2718_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_2718_cast_fp16 = gelu(mode = var_2718_mode_0, x = var_2716_cast_fp16_0)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_485_cast_fp16 = mul(x = var_2718_cast_fp16, y = var_2716_cast_fp16_1)[name = string("input_485_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_26_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_26_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(759800768)))];
            tensor<fp16, [1, 224, 1024]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_26_mlp_Wo_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_489_cast_fp16 = add(x = input_479_cast_fp16, y = linear_107_cast_fp16)[name = string("input_489_cast_fp16")];
            tensor<int32, [1]> hidden_states_axes_0 = const()[name = string("hidden_states_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_27_attn_norm_weight_to_fp16 = const()[name = string("model_model_layers_27_attn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765174784)))];
            tensor<fp16, [1, 224, 1024]> hidden_states_cast_fp16 = layer_norm(axes = hidden_states_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_27_attn_norm_weight_to_fp16, x = input_489_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_model_layers_27_attn_Wqkv_weight_to_fp16 = const()[name = string("model_model_layers_27_attn_Wqkv_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765176896)))];
            tensor<fp16, [1, 224, 3072]> linear_108_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_27_attn_Wqkv_weight_to_fp16, x = hidden_states_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<int32, [5]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [5]>([1, 224, 3, -1, 64])];
            tensor<fp16, [1, 224, 3, 16, 64]> qkv_cast_fp16 = reshape(shape = var_2737, x = linear_108_cast_fp16)[name = string("qkv_cast_fp16")];
            tensor<int32, [3]> var_2739_split_sizes_0 = const()[name = string("op_2739_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_2739_axis_0 = const()[name = string("op_2739_axis_0"), val = int32(-3)];
            tensor<fp16, [1, 224, 1, 16, 64]> var_2739_cast_fp16_0, tensor<fp16, [1, 224, 1, 16, 64]> var_2739_cast_fp16_1, tensor<fp16, [1, 224, 1, 16, 64]> var_2739_cast_fp16_2 = split(axis = var_2739_axis_0, split_sizes = var_2739_split_sizes_0, x = qkv_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<int32, [1]> squeeze_81_axes_0 = const()[name = string("squeeze_81_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_81_cast_fp16 = squeeze(axes = squeeze_81_axes_0, x = var_2739_cast_fp16_0)[name = string("squeeze_81_cast_fp16")];
            tensor<int32, [1]> squeeze_82_axes_0 = const()[name = string("squeeze_82_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_82_cast_fp16 = squeeze(axes = squeeze_82_axes_0, x = var_2739_cast_fp16_1)[name = string("squeeze_82_cast_fp16")];
            tensor<int32, [1]> squeeze_83_axes_0 = const()[name = string("squeeze_83_axes_0"), val = tensor<int32, [1]>([-3])];
            tensor<fp16, [1, 224, 16, 64]> squeeze_83_cast_fp16 = squeeze(axes = squeeze_83_axes_0, x = var_2739_cast_fp16_2)[name = string("squeeze_83_cast_fp16")];
            tensor<int32, [4]> q_109_perm_0 = const()[name = string("q_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_109_perm_0 = const()[name = string("k_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 224, 64]> q_109_cast_fp16 = transpose(perm = q_109_perm_0, x = squeeze_81_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 16, 224, 64]> var_2749_cast_fp16 = mul(x = q_109_cast_fp16, y = cos_7_to_fp16)[name = string("op_2749_cast_fp16")];
            tensor<int32, [4]> x1_109_begin_0 = const()[name = string("x1_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_109_end_0 = const()[name = string("x1_109_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_109_end_mask_0 = const()[name = string("x1_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_109_cast_fp16 = slice_by_index(begin = x1_109_begin_0, end = x1_109_end_0, end_mask = x1_109_end_mask_0, x = q_109_cast_fp16)[name = string("x1_109_cast_fp16")];
            tensor<int32, [4]> x2_109_begin_0 = const()[name = string("x2_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_109_end_0 = const()[name = string("x2_109_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_109_end_mask_0 = const()[name = string("x2_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_109_cast_fp16 = slice_by_index(begin = x2_109_begin_0, end = x2_109_end_0, end_mask = x2_109_end_mask_0, x = q_109_cast_fp16)[name = string("x2_109_cast_fp16")];
            fp16 const_239_promoted_to_fp16 = const()[name = string("const_239_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2761_cast_fp16 = mul(x = x2_109_cast_fp16, y = const_239_promoted_to_fp16)[name = string("op_2761_cast_fp16")];
            bool var_2763_interleave_0 = const()[name = string("op_2763_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2763_cast_fp16 = concat(axis = var_49, interleave = var_2763_interleave_0, values = (var_2761_cast_fp16, x1_109_cast_fp16))[name = string("op_2763_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2764_cast_fp16 = mul(x = var_2763_cast_fp16, y = sin_7_to_fp16)[name = string("op_2764_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> q_embed_cast_fp16 = add(x = var_2749_cast_fp16, y = var_2764_cast_fp16)[name = string("q_embed_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_109_cast_fp16 = transpose(perm = k_109_perm_0, x = squeeze_82_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 16, 224, 64]> var_2767_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_7_to_fp16)[name = string("op_2767_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 16, 224, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 224, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_109_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 16, 224, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 224, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_109_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_242_promoted_to_fp16 = const()[name = string("const_242_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 224, 32]> var_2779_cast_fp16 = mul(x = x2_cast_fp16, y = const_242_promoted_to_fp16)[name = string("op_2779_cast_fp16")];
            bool var_2781_interleave_0 = const()[name = string("op_2781_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> var_2781_cast_fp16 = concat(axis = var_49, interleave = var_2781_interleave_0, values = (var_2779_cast_fp16, x1_cast_fp16))[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> var_2782_cast_fp16 = mul(x = var_2781_cast_fp16, y = sin_7_to_fp16)[name = string("op_2782_cast_fp16")];
            tensor<fp16, [1, 16, 224, 64]> k_embed_cast_fp16 = add(x = var_2767_cast_fp16, y = var_2782_cast_fp16)[name = string("k_embed_cast_fp16")];
            bool var_2787_transpose_x_1 = const()[name = string("op_2787_transpose_x_1"), val = bool(false)];
            bool var_2787_transpose_y_1 = const()[name = string("op_2787_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 224, 224]> var_2787_cast_fp16 = matmul(transpose_x = var_2787_transpose_x_1, transpose_y = var_2787_transpose_y_1, x = q_embed_cast_fp16, y = k_embed_cast_fp16)[name = string("op_2787_cast_fp16")];
            fp16 var_2788_to_fp16 = const()[name = string("op_2788_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 224, 224]> attn_weights_109_cast_fp16 = mul(x = var_2787_cast_fp16, y = var_2788_to_fp16)[name = string("attn_weights_109_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> input_491_cast_fp16 = add(x = attn_weights_109_cast_fp16, y = key_bias_cast_fp16)[name = string("input_491_cast_fp16")];
            tensor<fp16, [1, 16, 224, 224]> var_2791_cast_fp16 = softmax(axis = var_49, x = input_491_cast_fp16)[name = string("op_2791_cast_fp16")];
            bool attn_output_163_transpose_x_0 = const()[name = string("attn_output_163_transpose_x_0"), val = bool(false)];
            bool attn_output_163_transpose_y_0 = const()[name = string("attn_output_163_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 224, 64]> value_cast_fp16 = transpose(perm = value_perm_0, x = squeeze_83_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 16, 224, 64]> attn_output_163_cast_fp16 = matmul(transpose_x = attn_output_163_transpose_x_0, transpose_y = attn_output_163_transpose_y_0, x = var_2791_cast_fp16, y = value_cast_fp16)[name = string("attn_output_163_cast_fp16")];
            tensor<int32, [4]> var_2795_perm_0 = const()[name = string("op_2795_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2797 = const()[name = string("op_2797"), val = tensor<int32, [3]>([1, 224, -1])];
            tensor<fp16, [1, 224, 16, 64]> var_2795_cast_fp16 = transpose(perm = var_2795_perm_0, x = attn_output_163_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 224, 1024]> var_2798_cast_fp16 = reshape(shape = var_2797, x = var_2795_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_model_layers_27_attn_Wo_weight_to_fp16 = const()[name = string("model_model_layers_27_attn_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(771468416)))];
            tensor<fp16, [1, 224, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_27_attn_Wo_weight_to_fp16, x = var_2798_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_497_cast_fp16 = add(x = input_489_cast_fp16, y = linear_109_cast_fp16)[name = string("input_497_cast_fp16")];
            tensor<int32, [1]> input_499_axes_0 = const()[name = string("input_499_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_layers_27_mlp_norm_weight_to_fp16 = const()[name = string("model_model_layers_27_mlp_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773565632)))];
            tensor<fp16, [1, 224, 1024]> input_499_cast_fp16 = layer_norm(axes = input_499_axes_0, epsilon = var_52_to_fp16, gamma = model_model_layers_27_mlp_norm_weight_to_fp16, x = input_497_cast_fp16)[name = string("input_499_cast_fp16")];
            tensor<fp16, [5248, 1024]> model_model_layers_27_mlp_Wi_weight_to_fp16 = const()[name = string("model_model_layers_27_mlp_Wi_weight_to_fp16"), val = tensor<fp16, [5248, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773567744)))];
            tensor<fp16, [1, 224, 5248]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_model_layers_27_mlp_Wi_weight_to_fp16, x = input_499_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [2]> var_2810_split_sizes_0 = const()[name = string("op_2810_split_sizes_0"), val = tensor<int32, [2]>([2624, 2624])];
            int32 var_2810_axis_0 = const()[name = string("op_2810_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 224, 2624]> var_2810_cast_fp16_0, tensor<fp16, [1, 224, 2624]> var_2810_cast_fp16_1 = split(axis = var_2810_axis_0, split_sizes = var_2810_split_sizes_0, x = linear_110_cast_fp16)[name = string("op_2810_cast_fp16")];
            string var_2812_mode_0 = const()[name = string("op_2812_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 224, 2624]> var_2812_cast_fp16 = gelu(mode = var_2812_mode_0, x = var_2810_cast_fp16_0)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [1, 224, 2624]> input_503_cast_fp16 = mul(x = var_2812_cast_fp16, y = var_2810_cast_fp16_1)[name = string("input_503_cast_fp16")];
            tensor<fp16, [1024, 2624]> model_model_layers_27_mlp_Wo_weight_to_fp16 = const()[name = string("model_model_layers_27_mlp_Wo_weight_to_fp16"), val = tensor<fp16, [1024, 2624]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(784315712)))];
            tensor<fp16, [1, 224, 1024]> linear_111_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_27_mlp_Wo_weight_to_fp16, x = input_503_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> input_507_cast_fp16 = add(x = input_497_cast_fp16, y = linear_111_cast_fp16)[name = string("input_507_cast_fp16")];
            tensor<int32, [1]> hidden_axes_0 = const()[name = string("hidden_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_model_final_norm_weight_to_fp16 = const()[name = string("model_model_final_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789689728)))];
            tensor<fp16, [1, 224, 1024]> hidden_cast_fp16 = layer_norm(axes = hidden_axes_0, epsilon = var_52_to_fp16, gamma = model_model_final_norm_weight_to_fp16, x = input_507_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_2822_axes_0 = const()[name = string("op_2822_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 224, 1]> var_2822_cast_fp16 = expand_dims(axes = var_2822_axes_0, x = attention_mask_to_fp16)[name = string("op_2822_cast_fp16")];
            tensor<fp16, [1, 224, 1024]> var_2823_cast_fp16 = mul(x = hidden_cast_fp16, y = var_2822_cast_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_2828_axes_0 = const()[name = string("op_2828_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_2828_keep_dims_0 = const()[name = string("op_2828_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 1024]> var_2828_cast_fp16 = reduce_sum(axes = var_2828_axes_0, keep_dims = var_2828_keep_dims_0, x = var_2823_cast_fp16)[name = string("op_2828_cast_fp16")];
            tensor<int32, [1]> var_2833_axes_0 = const()[name = string("op_2833_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_2833_keep_dims_0 = const()[name = string("op_2833_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_2833_cast_fp16 = reduce_sum(axes = var_2833_axes_0, keep_dims = var_2833_keep_dims_0, x = attention_mask_to_fp16)[name = string("op_2833_cast_fp16")];
            tensor<fp16, [1, 1024]> input_509_cast_fp16 = real_div(x = var_2828_cast_fp16, y = var_2833_cast_fp16)[name = string("input_509_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_head_dense_weight_to_fp16 = const()[name = string("model_head_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789691840)))];
            tensor<fp16, [1, 1024]> linear_112_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_head_dense_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_112_cast_fp16")];
            string input_513_mode_0 = const()[name = string("input_513_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024]> input_513_cast_fp16 = gelu(mode = input_513_mode_0, x = linear_112_cast_fp16)[name = string("input_513_cast_fp16")];
            tensor<int32, [1]> input_515_axes_0 = const()[name = string("input_515_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> model_head_norm_weight_to_fp16 = const()[name = string("model_head_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791789056)))];
            fp16 var_2836_to_fp16 = const()[name = string("op_2836_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024]> input_515_cast_fp16 = layer_norm(axes = input_515_axes_0, epsilon = var_2836_to_fp16, gamma = model_head_norm_weight_to_fp16, x = input_513_cast_fp16)[name = string("input_515_cast_fp16")];
            tensor<fp16, [3, 1024]> model_classifier_weight_to_fp16 = const()[name = string("model_classifier_weight_to_fp16"), val = tensor<fp16, [3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791791168)))];
            tensor<fp16, [3]> model_classifier_bias_to_fp16 = const()[name = string("model_classifier_bias_to_fp16"), val = tensor<fp16, [3]>([-0x1.284p-12, -0x1.f2p-17, 0x1.41p-12])];
            tensor<fp16, [1, 3]> linear_113_cast_fp16 = linear(bias = model_classifier_bias_to_fp16, weight = model_classifier_weight_to_fp16, x = input_515_cast_fp16)[name = string("linear_113_cast_fp16")];
            string linear_113_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_113_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 3]> logits = cast(dtype = linear_113_cast_fp16_to_fp32_dtype_0, x = linear_113_cast_fp16)[name = string("cast_59")];
        } -> (logits);
}