program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.2.1"}, {"coremlc-version", "3520.2.1"}, {"coremltools-component-torch", "2.1.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.1"}, {"mldb_token", "mldb-27marq7y6j"}})]
{
    func main<ios18>(tensor<fp16, [1, 197, 64]> specgram) {
            tensor<fp16, [384, 1, 16, 8]> embedding_patch_embed_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1, 16, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24704))))[name = string("embedding_patch_embed_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_0_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25536))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99328))))[name = string("embedding_transformer_0_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_0_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100160))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173952))))[name = string("embedding_transformer_0_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_0_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174784))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248576))))[name = string("embedding_transformer_0_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_0_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249408))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323200))))[name = string("embedding_transformer_0_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_0_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324032))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(619008))))[name = string("embedding_transformer_0_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_0_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622144))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(917120))))[name = string("embedding_transformer_0_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_1_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(917952))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(991744))))[name = string("embedding_transformer_1_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_1_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(992576))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1066368))))[name = string("embedding_transformer_1_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_1_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1067200))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1140992))))[name = string("embedding_transformer_1_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_1_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1141824))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1215616))))[name = string("embedding_transformer_1_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_1_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1216448))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1511424))))[name = string("embedding_transformer_1_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_1_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1514560))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1809536))))[name = string("embedding_transformer_1_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_2_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1810368))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1884160))))[name = string("embedding_transformer_2_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_2_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1884992))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1958784))))[name = string("embedding_transformer_2_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_2_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1959616))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2033408))))[name = string("embedding_transformer_2_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_2_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2034240))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2108032))))[name = string("embedding_transformer_2_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_2_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2108864))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2403840))))[name = string("embedding_transformer_2_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_2_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2406976))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2701952))))[name = string("embedding_transformer_2_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_3_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2702784))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2776576))))[name = string("embedding_transformer_3_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_3_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2777408))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2851200))))[name = string("embedding_transformer_3_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_3_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2852032))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2925824))))[name = string("embedding_transformer_3_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_3_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2926656))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3000448))))[name = string("embedding_transformer_3_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_3_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3001280))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3296256))))[name = string("embedding_transformer_3_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_3_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3299392))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3594368))))[name = string("embedding_transformer_3_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_4_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3595200))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3668992))))[name = string("embedding_transformer_4_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_4_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3669824))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3743616))))[name = string("embedding_transformer_4_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_4_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3744448))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3818240))))[name = string("embedding_transformer_4_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_4_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3819072))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3892864))))[name = string("embedding_transformer_4_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_4_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3893696))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4188672))))[name = string("embedding_transformer_4_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_4_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4191808))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4486784))))[name = string("embedding_transformer_4_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_5_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4487616))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4561408))))[name = string("embedding_transformer_5_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_5_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4562240))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4636032))))[name = string("embedding_transformer_5_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_5_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4636864))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4710656))))[name = string("embedding_transformer_5_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_5_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4711488))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4785280))))[name = string("embedding_transformer_5_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_5_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4786112))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5081088))))[name = string("embedding_transformer_5_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_5_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5084224))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5379200))))[name = string("embedding_transformer_5_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_6_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5380032))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5453824))))[name = string("embedding_transformer_6_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_6_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5454656))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5528448))))[name = string("embedding_transformer_6_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_6_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5529280))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5603072))))[name = string("embedding_transformer_6_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_6_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5603904))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5677696))))[name = string("embedding_transformer_6_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_6_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5678528))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5973504))))[name = string("embedding_transformer_6_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_6_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5976640))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6271616))))[name = string("embedding_transformer_6_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_7_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6272448))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6346240))))[name = string("embedding_transformer_7_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_7_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6347072))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6420864))))[name = string("embedding_transformer_7_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_7_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6421696))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6495488))))[name = string("embedding_transformer_7_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_7_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6496320))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6570112))))[name = string("embedding_transformer_7_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_7_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6570944))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6865920))))[name = string("embedding_transformer_7_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_7_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6869056))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7164032))))[name = string("embedding_transformer_7_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_8_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7164864))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7238656))))[name = string("embedding_transformer_8_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_8_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7239488))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7313280))))[name = string("embedding_transformer_8_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_8_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7314112))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7387904))))[name = string("embedding_transformer_8_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_8_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7388736))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7462528))))[name = string("embedding_transformer_8_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_8_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7463360))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7758336))))[name = string("embedding_transformer_8_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_8_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7761472))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8056448))))[name = string("embedding_transformer_8_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_9_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8057280))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8131072))))[name = string("embedding_transformer_9_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_9_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8131904))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8205696))))[name = string("embedding_transformer_9_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_9_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8206528))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8280320))))[name = string("embedding_transformer_9_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_9_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8281152))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8354944))))[name = string("embedding_transformer_9_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_9_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8355776))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8650752))))[name = string("embedding_transformer_9_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_9_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8653888))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8948864))))[name = string("embedding_transformer_9_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_10_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8949696))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9023488))))[name = string("embedding_transformer_10_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_10_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9024320))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9098112))))[name = string("embedding_transformer_10_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_10_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9098944))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9172736))))[name = string("embedding_transformer_10_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_10_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9173568))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9247360))))[name = string("embedding_transformer_10_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_10_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9248192))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9543168))))[name = string("embedding_transformer_10_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_10_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9546304))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9841280))))[name = string("embedding_transformer_10_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_11_attn_query_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9842112))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9915904))))[name = string("embedding_transformer_11_attn_query_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_11_attn_key_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9916736))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9990528))))[name = string("embedding_transformer_11_attn_key_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_11_attn_value_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9991360))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10065152))))[name = string("embedding_transformer_11_attn_value_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 384, 1, 1]> embedding_transformer_11_attn_out_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10065984))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10139776))))[name = string("embedding_transformer_11_attn_out_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [1536, 384, 1, 1]> embedding_transformer_11_mlp_c_fc_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 384, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10140608))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10435584))))[name = string("embedding_transformer_11_mlp_c_fc_weight_palettized_cast_fp16")];
            tensor<fp16, [384, 1536, 1, 1]> embedding_transformer_11_mlp_c_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [384, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10438720))), lut = tensor<fp16, [24, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10733696))))[name = string("embedding_transformer_11_mlp_c_proj_weight_palettized_cast_fp16")];
            tensor<fp16, [512, 384]> audio_projs_0_weight_palettized_cast_fp16 = constexpr_lut_to_dense(indices = tensor<uint4, [512, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10734528))), lut = tensor<fp16, [32, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10832896))))[name = string("audio_projs_0_weight_palettized_cast_fp16")];
            bool var_2 = const()[name = string("op_2"), val = bool(true)];
            tensor<int32, [1]> x0_1_axes_0 = const()[name = string("x0_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> x0_1_cast_fp16 = expand_dims(axes = x0_1_axes_0, x = specgram)[name = string("x0_1_cast_fp16")];
            int32 var_29 = const()[name = string("op_29"), val = int32(1)];
            string var_65_pad_type_0 = const()[name = string("op_65_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_65_strides_0 = const()[name = string("op_65_strides_0"), val = tensor<int32, [2]>([16, 8])];
            tensor<int32, [4]> var_65_pad_0 = const()[name = string("op_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_65_dilations_0 = const()[name = string("op_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_65_groups_0 = const()[name = string("op_65_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_patch_embed_proj_bias_to_fp16 = const()[name = string("embedding_patch_embed_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10833984)))];
            tensor<fp16, [1, 384, 12, 8]> var_65_cast_fp16 = conv(bias = embedding_patch_embed_proj_bias_to_fp16, dilations = var_65_dilations_0, groups = var_65_groups_0, pad = var_65_pad_0, pad_type = var_65_pad_type_0, strides = var_65_strides_0, weight = embedding_patch_embed_proj_weight_palettized_cast_fp16, x = x0_1_cast_fp16)[name = string("op_65_cast_fp16")];
            tensor<int32, [3]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [3]>([1, 384, 96])];
            tensor<fp16, [1, 384, 96]> var_66_cast_fp16 = reshape(shape = concat_0, x = var_65_cast_fp16)[name = string("op_66_cast_fp16")];
            tensor<fp16, [1, 384, 96]> const_74_to_fp16 = const()[name = string("const_74_to_fp16"), val = tensor<fp16, [1, 384, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10834816)))];
            tensor<fp16, [1, 384, 96]> inputs_4_cast_fp16 = add(x = var_66_cast_fp16, y = const_74_to_fp16)[name = string("inputs_4_cast_fp16")];
            tensor<int32, [1]> inputs1_1_axes_0 = const()[name = string("inputs1_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> inputs1_1_cast_fp16 = expand_dims(axes = inputs1_1_axes_0, x = inputs_4_cast_fp16)[name = string("inputs1_1_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_90_to_fp16 = const()[name = string("op_90_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_90_to_fp16, x = inputs1_1_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [384]> out0_3_mean_0_to_fp16 = const()[name = string("out0_3_mean_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10908608)))];
            tensor<fp16, [384]> out0_3_variance_0_to_fp16 = const()[name = string("out0_3_variance_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10909440)))];
            tensor<fp16, [384]> out0_3_gamma_0_to_fp16 = const()[name = string("out0_3_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10910272)))];
            tensor<fp16, [384]> out0_3_beta_0_to_fp16 = const()[name = string("out0_3_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10911104)))];
            fp16 out0_3_epsilon_0_to_fp16 = const()[name = string("out0_3_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_3_cast_fp16 = batch_norm(beta = out0_3_beta_0_to_fp16, epsilon = out0_3_epsilon_0_to_fp16, gamma = out0_3_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_3_cast_fp16)[name = string("out0_3_cast_fp16")];
            tensor<int32, [1]> var_100_axes_0 = const()[name = string("op_100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_100_cast_fp16 = squeeze(axes = var_100_axes_0, x = out0_3_cast_fp16)[name = string("op_100_cast_fp16")];
            tensor<int32, [1]> hidden_states_5_axes_0 = const()[name = string("hidden_states_5_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_5_cast_fp16 = expand_dims(axes = hidden_states_5_axes_0, x = var_100_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            string var_114_pad_type_0 = const()[name = string("op_114_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_114_strides_0 = const()[name = string("op_114_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_114_pad_0 = const()[name = string("op_114_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_114_dilations_0 = const()[name = string("op_114_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_114_groups_0 = const()[name = string("op_114_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_0_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_0_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10911936)))];
            tensor<fp16, [1, 384, 1, 96]> var_114_cast_fp16 = conv(bias = embedding_transformer_0_attn_query_bias_to_fp16, dilations = var_114_dilations_0, groups = var_114_groups_0, pad = var_114_pad_0, pad_type = var_114_pad_type_0, strides = var_114_strides_0, weight = embedding_transformer_0_attn_query_weight_palettized_cast_fp16, x = hidden_states_5_cast_fp16)[name = string("op_114_cast_fp16")];
            string k_2_pad_type_0 = const()[name = string("k_2_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_2_strides_0 = const()[name = string("k_2_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_2_pad_0 = const()[name = string("k_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_2_dilations_0 = const()[name = string("k_2_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_2_groups_0 = const()[name = string("k_2_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_0_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_0_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10912768)))];
            tensor<fp16, [1, 384, 1, 96]> k_2_cast_fp16 = conv(bias = embedding_transformer_0_attn_key_bias_to_fp16, dilations = k_2_dilations_0, groups = k_2_groups_0, pad = k_2_pad_0, pad_type = k_2_pad_type_0, strides = k_2_strides_0, weight = embedding_transformer_0_attn_key_weight_palettized_cast_fp16, x = hidden_states_5_cast_fp16)[name = string("k_2_cast_fp16")];
            string var_128_pad_type_0 = const()[name = string("op_128_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_128_strides_0 = const()[name = string("op_128_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_128_pad_0 = const()[name = string("op_128_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_128_dilations_0 = const()[name = string("op_128_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_128_groups_0 = const()[name = string("op_128_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_0_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_0_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10913600)))];
            tensor<fp16, [1, 384, 1, 96]> var_128_cast_fp16 = conv(bias = embedding_transformer_0_attn_value_bias_to_fp16, dilations = var_128_dilations_0, groups = var_128_groups_0, pad = var_128_pad_0, pad_type = var_128_pad_type_0, strides = var_128_strides_0, weight = embedding_transformer_0_attn_value_weight_palettized_cast_fp16, x = hidden_states_5_cast_fp16)[name = string("op_128_cast_fp16")];
            tensor<int32, [6]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_129_axis_0 = const()[name = string("op_129_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_129_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_129_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_129_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_129_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_129_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_129_cast_fp16_5 = split(axis = var_129_axis_0, split_sizes = tile_0, x = var_114_cast_fp16)[name = string("op_129_cast_fp16")];
            tensor<int32, [4]> var_136_perm_0 = const()[name = string("op_136_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_137_axis_0 = const()[name = string("op_137_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_136_cast_fp16 = transpose(perm = var_136_perm_0, x = k_2_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 96, 1, 64]> var_137_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_137_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_137_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_137_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_137_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_137_cast_fp16_5 = split(axis = var_137_axis_0, split_sizes = tile_1, x = var_136_cast_fp16)[name = string("op_137_cast_fp16")];
            tensor<int32, [6]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_144_axis_0 = const()[name = string("op_144_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_144_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_144_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_144_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_144_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_144_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_144_cast_fp16_5 = split(axis = var_144_axis_0, split_sizes = tile_2, x = var_128_cast_fp16)[name = string("op_144_cast_fp16")];
            string var_152_equation_0 = const()[name = string("op_152_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_152_cast_fp16 = einsum(equation = var_152_equation_0, values = (var_137_cast_fp16_0, var_129_cast_fp16_0))[name = string("op_152_cast_fp16")];
            fp16 var_153_to_fp16 = const()[name = string("op_153_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_2_cast_fp16 = mul(x = var_152_cast_fp16, y = var_153_to_fp16)[name = string("aw_2_cast_fp16")];
            string var_156_equation_0 = const()[name = string("op_156_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_156_cast_fp16 = einsum(equation = var_156_equation_0, values = (var_137_cast_fp16_1, var_129_cast_fp16_1))[name = string("op_156_cast_fp16")];
            fp16 var_157_to_fp16 = const()[name = string("op_157_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_2_cast_fp16 = mul(x = var_156_cast_fp16, y = var_157_to_fp16)[name = string("aw0_2_cast_fp16")];
            string var_160_equation_0 = const()[name = string("op_160_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_160_cast_fp16 = einsum(equation = var_160_equation_0, values = (var_137_cast_fp16_2, var_129_cast_fp16_2))[name = string("op_160_cast_fp16")];
            fp16 var_161_to_fp16 = const()[name = string("op_161_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_2_cast_fp16 = mul(x = var_160_cast_fp16, y = var_161_to_fp16)[name = string("aw1_2_cast_fp16")];
            string var_164_equation_0 = const()[name = string("op_164_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_164_cast_fp16 = einsum(equation = var_164_equation_0, values = (var_137_cast_fp16_3, var_129_cast_fp16_3))[name = string("op_164_cast_fp16")];
            fp16 var_165_to_fp16 = const()[name = string("op_165_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_2_cast_fp16 = mul(x = var_164_cast_fp16, y = var_165_to_fp16)[name = string("aw2_2_cast_fp16")];
            string var_168_equation_0 = const()[name = string("op_168_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_168_cast_fp16 = einsum(equation = var_168_equation_0, values = (var_137_cast_fp16_4, var_129_cast_fp16_4))[name = string("op_168_cast_fp16")];
            fp16 var_169_to_fp16 = const()[name = string("op_169_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_2_cast_fp16 = mul(x = var_168_cast_fp16, y = var_169_to_fp16)[name = string("aw3_2_cast_fp16")];
            string var_172_equation_0 = const()[name = string("op_172_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_172_cast_fp16 = einsum(equation = var_172_equation_0, values = (var_137_cast_fp16_5, var_129_cast_fp16_5))[name = string("op_172_cast_fp16")];
            fp16 var_173_to_fp16 = const()[name = string("op_173_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_2_cast_fp16 = mul(x = var_172_cast_fp16, y = var_173_to_fp16)[name = string("aw4_2_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_175_cast_fp16 = softmax(axis = var_29, x = aw_2_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_176_cast_fp16 = softmax(axis = var_29, x = aw0_2_cast_fp16)[name = string("op_176_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_177_cast_fp16 = softmax(axis = var_29, x = aw1_2_cast_fp16)[name = string("op_177_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_178_cast_fp16 = softmax(axis = var_29, x = aw2_2_cast_fp16)[name = string("op_178_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_179_cast_fp16 = softmax(axis = var_29, x = aw3_2_cast_fp16)[name = string("op_179_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_180_cast_fp16 = softmax(axis = var_29, x = aw4_2_cast_fp16)[name = string("op_180_cast_fp16")];
            string var_182_equation_0 = const()[name = string("op_182_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_182_cast_fp16 = einsum(equation = var_182_equation_0, values = (var_144_cast_fp16_0, var_175_cast_fp16))[name = string("op_182_cast_fp16")];
            string var_184_equation_0 = const()[name = string("op_184_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_184_cast_fp16 = einsum(equation = var_184_equation_0, values = (var_144_cast_fp16_1, var_176_cast_fp16))[name = string("op_184_cast_fp16")];
            string var_186_equation_0 = const()[name = string("op_186_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_186_cast_fp16 = einsum(equation = var_186_equation_0, values = (var_144_cast_fp16_2, var_177_cast_fp16))[name = string("op_186_cast_fp16")];
            string var_188_equation_0 = const()[name = string("op_188_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_188_cast_fp16 = einsum(equation = var_188_equation_0, values = (var_144_cast_fp16_3, var_178_cast_fp16))[name = string("op_188_cast_fp16")];
            string var_190_equation_0 = const()[name = string("op_190_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_190_cast_fp16 = einsum(equation = var_190_equation_0, values = (var_144_cast_fp16_4, var_179_cast_fp16))[name = string("op_190_cast_fp16")];
            string var_192_equation_0 = const()[name = string("op_192_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_192_cast_fp16 = einsum(equation = var_192_equation_0, values = (var_144_cast_fp16_5, var_180_cast_fp16))[name = string("op_192_cast_fp16")];
            bool input_5_interleave_0 = const()[name = string("input_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_5_cast_fp16 = concat(axis = var_29, interleave = input_5_interleave_0, values = (var_182_cast_fp16, var_184_cast_fp16, var_186_cast_fp16, var_188_cast_fp16, var_190_cast_fp16, var_192_cast_fp16))[name = string("input_5_cast_fp16")];
            string attn_5_pad_type_0 = const()[name = string("attn_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_5_strides_0 = const()[name = string("attn_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_5_pad_0 = const()[name = string("attn_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_5_dilations_0 = const()[name = string("attn_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_5_groups_0 = const()[name = string("attn_5_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_0_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_0_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10914432)))];
            tensor<fp16, [1, 384, 1, 96]> attn_5_cast_fp16 = conv(bias = embedding_transformer_0_attn_out_proj_bias_to_fp16, dilations = attn_5_dilations_0, groups = attn_5_groups_0, pad = attn_5_pad_0, pad_type = attn_5_pad_type_0, strides = attn_5_strides_0, weight = embedding_transformer_0_attn_out_proj_weight_palettized_cast_fp16, x = input_5_cast_fp16)[name = string("attn_5_cast_fp16")];
            tensor<int32, [1]> var_202_axes_0 = const()[name = string("op_202_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_202_cast_fp16 = squeeze(axes = var_202_axes_0, x = attn_5_cast_fp16)[name = string("op_202_cast_fp16")];
            tensor<fp16, [1, 384, 96]> inputs0_2_cast_fp16 = add(x = inputs_4_cast_fp16, y = var_202_cast_fp16)[name = string("inputs0_2_cast_fp16")];
            tensor<int32, [1]> inputs0_3_axes_0 = const()[name = string("inputs0_3_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> inputs0_3_cast_fp16 = expand_dims(axes = inputs0_3_axes_0, x = inputs0_2_cast_fp16)[name = string("inputs0_3_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_215_to_fp16 = const()[name = string("op_215_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_215_to_fp16, x = inputs0_3_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [384]> out0_5_gamma_0_to_fp16 = const()[name = string("out0_5_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10915264)))];
            tensor<fp16, [384]> out0_5_beta_0_to_fp16 = const()[name = string("out0_5_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10916096)))];
            fp16 out0_5_epsilon_0_to_fp16 = const()[name = string("out0_5_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_5_cast_fp16 = batch_norm(beta = out0_5_beta_0_to_fp16, epsilon = out0_5_epsilon_0_to_fp16, gamma = out0_5_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_5_cast_fp16)[name = string("out0_5_cast_fp16")];
            tensor<int32, [1]> var_225_axes_0 = const()[name = string("op_225_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_225_cast_fp16 = squeeze(axes = var_225_axes_0, x = out0_5_cast_fp16)[name = string("op_225_cast_fp16")];
            tensor<int32, [3]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_232_axes_0 = const()[name = string("op_232_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = var_225_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 384, 96, 1]> var_232_cast_fp16 = expand_dims(axes = var_232_axes_0, x = transpose_2_cast_fp16)[name = string("op_232_cast_fp16")];
            string var_237_pad_type_0 = const()[name = string("op_237_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_237_strides_0 = const()[name = string("op_237_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_237_pad_0 = const()[name = string("op_237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_237_dilations_0 = const()[name = string("op_237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_237_groups_0 = const()[name = string("op_237_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_0_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_0_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10916928)))];
            tensor<fp16, [1, 1536, 96, 1]> var_237_cast_fp16 = conv(bias = embedding_transformer_0_mlp_c_fc_bias_to_fp16, dilations = var_237_dilations_0, groups = var_237_groups_0, pad = var_237_pad_0, pad_type = var_237_pad_type_0, strides = var_237_strides_0, weight = embedding_transformer_0_mlp_c_fc_weight_palettized_cast_fp16, x = var_232_cast_fp16)[name = string("op_237_cast_fp16")];
            tensor<int32, [1]> var_238_axes_0 = const()[name = string("op_238_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_238_cast_fp16 = squeeze(axes = var_238_axes_0, x = var_237_cast_fp16)[name = string("op_238_cast_fp16")];
            string var_239_mode_0 = const()[name = string("op_239_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_239_cast_fp16 = gelu(mode = var_239_mode_0, x = var_238_cast_fp16)[name = string("op_239_cast_fp16")];
            tensor<int32, [1]> var_243_axes_0 = const()[name = string("op_243_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_243_cast_fp16 = expand_dims(axes = var_243_axes_0, x = var_239_cast_fp16)[name = string("op_243_cast_fp16")];
            string var_248_pad_type_0 = const()[name = string("op_248_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_248_strides_0 = const()[name = string("op_248_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_248_pad_0 = const()[name = string("op_248_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_248_dilations_0 = const()[name = string("op_248_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_248_groups_0 = const()[name = string("op_248_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_0_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_0_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10920064)))];
            tensor<fp16, [1, 384, 96, 1]> var_248_cast_fp16 = conv(bias = embedding_transformer_0_mlp_c_proj_bias_to_fp16, dilations = var_248_dilations_0, groups = var_248_groups_0, pad = var_248_pad_0, pad_type = var_248_pad_type_0, strides = var_248_strides_0, weight = embedding_transformer_0_mlp_c_proj_weight_palettized_cast_fp16, x = var_243_cast_fp16)[name = string("op_248_cast_fp16")];
            tensor<int32, [1]> var_249_axes_0 = const()[name = string("op_249_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_249_cast_fp16 = squeeze(axes = var_249_axes_0, x = var_248_cast_fp16)[name = string("op_249_cast_fp16")];
            tensor<int32, [3]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [3]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> transpose_17 = transpose(perm = transpose_17_perm_0, x = var_249_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 96, 384]> transpose_16 = transpose(perm = transpose_16_perm_0, x = inputs0_2_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 96, 384]> var_251_cast_fp16 = add(x = transpose_16, y = transpose_17)[name = string("op_251_cast_fp16")];
            tensor<int32, [3]> var_258_perm_0 = const()[name = string("op_258_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_6_axes_0 = const()[name = string("inputs_6_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_258_cast_fp16 = transpose(perm = var_258_perm_0, x = var_251_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 384, 1, 96]> inputs_6_cast_fp16 = expand_dims(axes = inputs_6_axes_0, x = var_258_cast_fp16)[name = string("inputs_6_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_266_to_fp16 = const()[name = string("op_266_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_266_to_fp16, x = inputs_6_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [384]> out0_7_gamma_0_to_fp16 = const()[name = string("out0_7_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10920896)))];
            tensor<fp16, [384]> out0_7_beta_0_to_fp16 = const()[name = string("out0_7_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10921728)))];
            fp16 out0_7_epsilon_0_to_fp16 = const()[name = string("out0_7_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_7_cast_fp16 = batch_norm(beta = out0_7_beta_0_to_fp16, epsilon = out0_7_epsilon_0_to_fp16, gamma = out0_7_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_7_cast_fp16)[name = string("out0_7_cast_fp16")];
            tensor<int32, [1]> var_276_axes_0 = const()[name = string("op_276_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_276_cast_fp16 = squeeze(axes = var_276_axes_0, x = out0_7_cast_fp16)[name = string("op_276_cast_fp16")];
            tensor<int32, [1]> hidden_states_9_axes_0 = const()[name = string("hidden_states_9_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_9_cast_fp16 = expand_dims(axes = hidden_states_9_axes_0, x = var_276_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            string var_290_pad_type_0 = const()[name = string("op_290_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_290_strides_0 = const()[name = string("op_290_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_290_pad_0 = const()[name = string("op_290_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_290_dilations_0 = const()[name = string("op_290_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_290_groups_0 = const()[name = string("op_290_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_1_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_1_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922560)))];
            tensor<fp16, [1, 384, 1, 96]> var_290_cast_fp16 = conv(bias = embedding_transformer_1_attn_query_bias_to_fp16, dilations = var_290_dilations_0, groups = var_290_groups_0, pad = var_290_pad_0, pad_type = var_290_pad_type_0, strides = var_290_strides_0, weight = embedding_transformer_1_attn_query_weight_palettized_cast_fp16, x = hidden_states_9_cast_fp16)[name = string("op_290_cast_fp16")];
            string k_4_pad_type_0 = const()[name = string("k_4_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_4_strides_0 = const()[name = string("k_4_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_4_pad_0 = const()[name = string("k_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_4_dilations_0 = const()[name = string("k_4_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_4_groups_0 = const()[name = string("k_4_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_1_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_1_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10923392)))];
            tensor<fp16, [1, 384, 1, 96]> k_4_cast_fp16 = conv(bias = embedding_transformer_1_attn_key_bias_to_fp16, dilations = k_4_dilations_0, groups = k_4_groups_0, pad = k_4_pad_0, pad_type = k_4_pad_type_0, strides = k_4_strides_0, weight = embedding_transformer_1_attn_key_weight_palettized_cast_fp16, x = hidden_states_9_cast_fp16)[name = string("k_4_cast_fp16")];
            string var_304_pad_type_0 = const()[name = string("op_304_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_304_strides_0 = const()[name = string("op_304_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_304_pad_0 = const()[name = string("op_304_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_304_dilations_0 = const()[name = string("op_304_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_304_groups_0 = const()[name = string("op_304_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_1_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_1_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10924224)))];
            tensor<fp16, [1, 384, 1, 96]> var_304_cast_fp16 = conv(bias = embedding_transformer_1_attn_value_bias_to_fp16, dilations = var_304_dilations_0, groups = var_304_groups_0, pad = var_304_pad_0, pad_type = var_304_pad_type_0, strides = var_304_strides_0, weight = embedding_transformer_1_attn_value_weight_palettized_cast_fp16, x = hidden_states_9_cast_fp16)[name = string("op_304_cast_fp16")];
            tensor<int32, [6]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_305_axis_0 = const()[name = string("op_305_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_305_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_305_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_305_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_305_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_305_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_305_cast_fp16_5 = split(axis = var_305_axis_0, split_sizes = tile_3, x = var_290_cast_fp16)[name = string("op_305_cast_fp16")];
            tensor<int32, [4]> var_312_perm_0 = const()[name = string("op_312_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_313_axis_0 = const()[name = string("op_313_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_312_cast_fp16 = transpose(perm = var_312_perm_0, x = k_4_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 96, 1, 64]> var_313_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_313_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_313_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_313_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_313_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_313_cast_fp16_5 = split(axis = var_313_axis_0, split_sizes = tile_4, x = var_312_cast_fp16)[name = string("op_313_cast_fp16")];
            tensor<int32, [6]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_320_axis_0 = const()[name = string("op_320_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_320_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_320_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_320_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_320_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_320_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_320_cast_fp16_5 = split(axis = var_320_axis_0, split_sizes = tile_5, x = var_304_cast_fp16)[name = string("op_320_cast_fp16")];
            string var_328_equation_0 = const()[name = string("op_328_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_328_cast_fp16 = einsum(equation = var_328_equation_0, values = (var_313_cast_fp16_0, var_305_cast_fp16_0))[name = string("op_328_cast_fp16")];
            fp16 var_329_to_fp16 = const()[name = string("op_329_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_4_cast_fp16 = mul(x = var_328_cast_fp16, y = var_329_to_fp16)[name = string("aw_4_cast_fp16")];
            string var_332_equation_0 = const()[name = string("op_332_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_332_cast_fp16 = einsum(equation = var_332_equation_0, values = (var_313_cast_fp16_1, var_305_cast_fp16_1))[name = string("op_332_cast_fp16")];
            fp16 var_333_to_fp16 = const()[name = string("op_333_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_4_cast_fp16 = mul(x = var_332_cast_fp16, y = var_333_to_fp16)[name = string("aw0_4_cast_fp16")];
            string var_336_equation_0 = const()[name = string("op_336_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_336_cast_fp16 = einsum(equation = var_336_equation_0, values = (var_313_cast_fp16_2, var_305_cast_fp16_2))[name = string("op_336_cast_fp16")];
            fp16 var_337_to_fp16 = const()[name = string("op_337_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_4_cast_fp16 = mul(x = var_336_cast_fp16, y = var_337_to_fp16)[name = string("aw1_4_cast_fp16")];
            string var_340_equation_0 = const()[name = string("op_340_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_340_cast_fp16 = einsum(equation = var_340_equation_0, values = (var_313_cast_fp16_3, var_305_cast_fp16_3))[name = string("op_340_cast_fp16")];
            fp16 var_341_to_fp16 = const()[name = string("op_341_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_4_cast_fp16 = mul(x = var_340_cast_fp16, y = var_341_to_fp16)[name = string("aw2_4_cast_fp16")];
            string var_344_equation_0 = const()[name = string("op_344_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_344_cast_fp16 = einsum(equation = var_344_equation_0, values = (var_313_cast_fp16_4, var_305_cast_fp16_4))[name = string("op_344_cast_fp16")];
            fp16 var_345_to_fp16 = const()[name = string("op_345_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_4_cast_fp16 = mul(x = var_344_cast_fp16, y = var_345_to_fp16)[name = string("aw3_4_cast_fp16")];
            string var_348_equation_0 = const()[name = string("op_348_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_348_cast_fp16 = einsum(equation = var_348_equation_0, values = (var_313_cast_fp16_5, var_305_cast_fp16_5))[name = string("op_348_cast_fp16")];
            fp16 var_349_to_fp16 = const()[name = string("op_349_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_4_cast_fp16 = mul(x = var_348_cast_fp16, y = var_349_to_fp16)[name = string("aw4_4_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_351_cast_fp16 = softmax(axis = var_29, x = aw_4_cast_fp16)[name = string("op_351_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_352_cast_fp16 = softmax(axis = var_29, x = aw0_4_cast_fp16)[name = string("op_352_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_353_cast_fp16 = softmax(axis = var_29, x = aw1_4_cast_fp16)[name = string("op_353_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_354_cast_fp16 = softmax(axis = var_29, x = aw2_4_cast_fp16)[name = string("op_354_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_355_cast_fp16 = softmax(axis = var_29, x = aw3_4_cast_fp16)[name = string("op_355_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_356_cast_fp16 = softmax(axis = var_29, x = aw4_4_cast_fp16)[name = string("op_356_cast_fp16")];
            string var_358_equation_0 = const()[name = string("op_358_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_358_cast_fp16 = einsum(equation = var_358_equation_0, values = (var_320_cast_fp16_0, var_351_cast_fp16))[name = string("op_358_cast_fp16")];
            string var_360_equation_0 = const()[name = string("op_360_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_360_cast_fp16 = einsum(equation = var_360_equation_0, values = (var_320_cast_fp16_1, var_352_cast_fp16))[name = string("op_360_cast_fp16")];
            string var_362_equation_0 = const()[name = string("op_362_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_362_cast_fp16 = einsum(equation = var_362_equation_0, values = (var_320_cast_fp16_2, var_353_cast_fp16))[name = string("op_362_cast_fp16")];
            string var_364_equation_0 = const()[name = string("op_364_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_364_cast_fp16 = einsum(equation = var_364_equation_0, values = (var_320_cast_fp16_3, var_354_cast_fp16))[name = string("op_364_cast_fp16")];
            string var_366_equation_0 = const()[name = string("op_366_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_366_cast_fp16 = einsum(equation = var_366_equation_0, values = (var_320_cast_fp16_4, var_355_cast_fp16))[name = string("op_366_cast_fp16")];
            string var_368_equation_0 = const()[name = string("op_368_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_368_cast_fp16 = einsum(equation = var_368_equation_0, values = (var_320_cast_fp16_5, var_356_cast_fp16))[name = string("op_368_cast_fp16")];
            bool input_11_interleave_0 = const()[name = string("input_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_11_cast_fp16 = concat(axis = var_29, interleave = input_11_interleave_0, values = (var_358_cast_fp16, var_360_cast_fp16, var_362_cast_fp16, var_364_cast_fp16, var_366_cast_fp16, var_368_cast_fp16))[name = string("input_11_cast_fp16")];
            string attn_9_pad_type_0 = const()[name = string("attn_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_9_strides_0 = const()[name = string("attn_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_9_pad_0 = const()[name = string("attn_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_9_dilations_0 = const()[name = string("attn_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_9_groups_0 = const()[name = string("attn_9_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_1_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_1_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10925056)))];
            tensor<fp16, [1, 384, 1, 96]> attn_9_cast_fp16 = conv(bias = embedding_transformer_1_attn_out_proj_bias_to_fp16, dilations = attn_9_dilations_0, groups = attn_9_groups_0, pad = attn_9_pad_0, pad_type = attn_9_pad_type_0, strides = attn_9_strides_0, weight = embedding_transformer_1_attn_out_proj_weight_palettized_cast_fp16, x = input_11_cast_fp16)[name = string("attn_9_cast_fp16")];
            tensor<int32, [1]> var_378_axes_0 = const()[name = string("op_378_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_378_cast_fp16 = squeeze(axes = var_378_axes_0, x = attn_9_cast_fp16)[name = string("op_378_cast_fp16")];
            tensor<int32, [3]> var_379_perm_0 = const()[name = string("op_379_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_379_cast_fp16 = transpose(perm = var_379_perm_0, x = var_378_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 96, 384]> inputs_8_cast_fp16 = add(x = var_251_cast_fp16, y = var_379_cast_fp16)[name = string("inputs_8_cast_fp16")];
            tensor<int32, [3]> var_383_perm_0 = const()[name = string("op_383_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_5_axes_0 = const()[name = string("inputs0_5_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_383_cast_fp16 = transpose(perm = var_383_perm_0, x = inputs_8_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_5_cast_fp16 = expand_dims(axes = inputs0_5_axes_0, x = var_383_cast_fp16)[name = string("inputs0_5_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_391_to_fp16 = const()[name = string("op_391_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_391_to_fp16, x = inputs0_5_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [384]> out0_9_gamma_0_to_fp16 = const()[name = string("out0_9_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10925888)))];
            tensor<fp16, [384]> out0_9_beta_0_to_fp16 = const()[name = string("out0_9_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10926720)))];
            fp16 out0_9_epsilon_0_to_fp16 = const()[name = string("out0_9_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_9_cast_fp16 = batch_norm(beta = out0_9_beta_0_to_fp16, epsilon = out0_9_epsilon_0_to_fp16, gamma = out0_9_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_9_cast_fp16)[name = string("out0_9_cast_fp16")];
            tensor<int32, [1]> var_401_axes_0 = const()[name = string("op_401_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_401_cast_fp16 = squeeze(axes = var_401_axes_0, x = out0_9_cast_fp16)[name = string("op_401_cast_fp16")];
            tensor<int32, [3]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_408_axes_0 = const()[name = string("op_408_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = var_401_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 384, 96, 1]> var_408_cast_fp16 = expand_dims(axes = var_408_axes_0, x = transpose_3_cast_fp16)[name = string("op_408_cast_fp16")];
            string var_413_pad_type_0 = const()[name = string("op_413_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_413_strides_0 = const()[name = string("op_413_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_413_pad_0 = const()[name = string("op_413_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_413_dilations_0 = const()[name = string("op_413_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_413_groups_0 = const()[name = string("op_413_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_1_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_1_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10927552)))];
            tensor<fp16, [1, 1536, 96, 1]> var_413_cast_fp16 = conv(bias = embedding_transformer_1_mlp_c_fc_bias_to_fp16, dilations = var_413_dilations_0, groups = var_413_groups_0, pad = var_413_pad_0, pad_type = var_413_pad_type_0, strides = var_413_strides_0, weight = embedding_transformer_1_mlp_c_fc_weight_palettized_cast_fp16, x = var_408_cast_fp16)[name = string("op_413_cast_fp16")];
            tensor<int32, [1]> var_414_axes_0 = const()[name = string("op_414_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_414_cast_fp16 = squeeze(axes = var_414_axes_0, x = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            string var_415_mode_0 = const()[name = string("op_415_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_415_cast_fp16 = gelu(mode = var_415_mode_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<int32, [1]> var_419_axes_0 = const()[name = string("op_419_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_419_cast_fp16 = expand_dims(axes = var_419_axes_0, x = var_415_cast_fp16)[name = string("op_419_cast_fp16")];
            string var_424_pad_type_0 = const()[name = string("op_424_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_424_strides_0 = const()[name = string("op_424_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_424_pad_0 = const()[name = string("op_424_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_424_dilations_0 = const()[name = string("op_424_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_424_groups_0 = const()[name = string("op_424_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_1_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_1_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10930688)))];
            tensor<fp16, [1, 384, 96, 1]> var_424_cast_fp16 = conv(bias = embedding_transformer_1_mlp_c_proj_bias_to_fp16, dilations = var_424_dilations_0, groups = var_424_groups_0, pad = var_424_pad_0, pad_type = var_424_pad_type_0, strides = var_424_strides_0, weight = embedding_transformer_1_mlp_c_proj_weight_palettized_cast_fp16, x = var_419_cast_fp16)[name = string("op_424_cast_fp16")];
            tensor<int32, [1]> var_425_axes_0 = const()[name = string("op_425_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_425_cast_fp16 = squeeze(axes = var_425_axes_0, x = var_424_cast_fp16)[name = string("op_425_cast_fp16")];
            tensor<int32, [3]> var_426_perm_0 = const()[name = string("op_426_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_426_cast_fp16 = transpose(perm = var_426_perm_0, x = var_425_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 96, 384]> var_427_cast_fp16 = add(x = inputs_8_cast_fp16, y = var_426_cast_fp16)[name = string("op_427_cast_fp16")];
            tensor<int32, [3]> var_434_perm_0 = const()[name = string("op_434_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_10_axes_0 = const()[name = string("inputs_10_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_434_cast_fp16 = transpose(perm = var_434_perm_0, x = var_427_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 384, 1, 96]> inputs_10_cast_fp16 = expand_dims(axes = inputs_10_axes_0, x = var_434_cast_fp16)[name = string("inputs_10_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_442_to_fp16 = const()[name = string("op_442_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_442_to_fp16, x = inputs_10_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [384]> out0_11_gamma_0_to_fp16 = const()[name = string("out0_11_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10931520)))];
            tensor<fp16, [384]> out0_11_beta_0_to_fp16 = const()[name = string("out0_11_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10932352)))];
            fp16 out0_11_epsilon_0_to_fp16 = const()[name = string("out0_11_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_11_cast_fp16 = batch_norm(beta = out0_11_beta_0_to_fp16, epsilon = out0_11_epsilon_0_to_fp16, gamma = out0_11_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_11_cast_fp16)[name = string("out0_11_cast_fp16")];
            tensor<int32, [1]> var_452_axes_0 = const()[name = string("op_452_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_452_cast_fp16 = squeeze(axes = var_452_axes_0, x = out0_11_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<int32, [1]> hidden_states_13_axes_0 = const()[name = string("hidden_states_13_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_13_cast_fp16 = expand_dims(axes = hidden_states_13_axes_0, x = var_452_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            string var_466_pad_type_0 = const()[name = string("op_466_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_466_strides_0 = const()[name = string("op_466_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_466_pad_0 = const()[name = string("op_466_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_466_dilations_0 = const()[name = string("op_466_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_466_groups_0 = const()[name = string("op_466_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_2_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_2_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10933184)))];
            tensor<fp16, [1, 384, 1, 96]> var_466_cast_fp16 = conv(bias = embedding_transformer_2_attn_query_bias_to_fp16, dilations = var_466_dilations_0, groups = var_466_groups_0, pad = var_466_pad_0, pad_type = var_466_pad_type_0, strides = var_466_strides_0, weight = embedding_transformer_2_attn_query_weight_palettized_cast_fp16, x = hidden_states_13_cast_fp16)[name = string("op_466_cast_fp16")];
            string k_6_pad_type_0 = const()[name = string("k_6_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_6_strides_0 = const()[name = string("k_6_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_6_pad_0 = const()[name = string("k_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_6_dilations_0 = const()[name = string("k_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_6_groups_0 = const()[name = string("k_6_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_2_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_2_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10934016)))];
            tensor<fp16, [1, 384, 1, 96]> k_6_cast_fp16 = conv(bias = embedding_transformer_2_attn_key_bias_to_fp16, dilations = k_6_dilations_0, groups = k_6_groups_0, pad = k_6_pad_0, pad_type = k_6_pad_type_0, strides = k_6_strides_0, weight = embedding_transformer_2_attn_key_weight_palettized_cast_fp16, x = hidden_states_13_cast_fp16)[name = string("k_6_cast_fp16")];
            string var_480_pad_type_0 = const()[name = string("op_480_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_480_strides_0 = const()[name = string("op_480_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_480_pad_0 = const()[name = string("op_480_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_480_dilations_0 = const()[name = string("op_480_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_480_groups_0 = const()[name = string("op_480_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_2_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_2_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10934848)))];
            tensor<fp16, [1, 384, 1, 96]> var_480_cast_fp16 = conv(bias = embedding_transformer_2_attn_value_bias_to_fp16, dilations = var_480_dilations_0, groups = var_480_groups_0, pad = var_480_pad_0, pad_type = var_480_pad_type_0, strides = var_480_strides_0, weight = embedding_transformer_2_attn_value_weight_palettized_cast_fp16, x = hidden_states_13_cast_fp16)[name = string("op_480_cast_fp16")];
            tensor<int32, [6]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_481_axis_0 = const()[name = string("op_481_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_481_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_481_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_481_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_481_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_481_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_481_cast_fp16_5 = split(axis = var_481_axis_0, split_sizes = tile_6, x = var_466_cast_fp16)[name = string("op_481_cast_fp16")];
            tensor<int32, [4]> var_488_perm_0 = const()[name = string("op_488_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_488_cast_fp16 = transpose(perm = var_488_perm_0, x = k_6_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 96, 1, 64]> var_489_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_489_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_489_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_489_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_489_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_489_cast_fp16_5 = split(axis = var_489_axis_0, split_sizes = tile_7, x = var_488_cast_fp16)[name = string("op_489_cast_fp16")];
            tensor<int32, [6]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_496_axis_0 = const()[name = string("op_496_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_496_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_496_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_496_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_496_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_496_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_496_cast_fp16_5 = split(axis = var_496_axis_0, split_sizes = tile_8, x = var_480_cast_fp16)[name = string("op_496_cast_fp16")];
            string var_504_equation_0 = const()[name = string("op_504_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_504_cast_fp16 = einsum(equation = var_504_equation_0, values = (var_489_cast_fp16_0, var_481_cast_fp16_0))[name = string("op_504_cast_fp16")];
            fp16 var_505_to_fp16 = const()[name = string("op_505_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_6_cast_fp16 = mul(x = var_504_cast_fp16, y = var_505_to_fp16)[name = string("aw_6_cast_fp16")];
            string var_508_equation_0 = const()[name = string("op_508_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_508_cast_fp16 = einsum(equation = var_508_equation_0, values = (var_489_cast_fp16_1, var_481_cast_fp16_1))[name = string("op_508_cast_fp16")];
            fp16 var_509_to_fp16 = const()[name = string("op_509_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_6_cast_fp16 = mul(x = var_508_cast_fp16, y = var_509_to_fp16)[name = string("aw0_6_cast_fp16")];
            string var_512_equation_0 = const()[name = string("op_512_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_512_cast_fp16 = einsum(equation = var_512_equation_0, values = (var_489_cast_fp16_2, var_481_cast_fp16_2))[name = string("op_512_cast_fp16")];
            fp16 var_513_to_fp16 = const()[name = string("op_513_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_6_cast_fp16 = mul(x = var_512_cast_fp16, y = var_513_to_fp16)[name = string("aw1_6_cast_fp16")];
            string var_516_equation_0 = const()[name = string("op_516_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_516_cast_fp16 = einsum(equation = var_516_equation_0, values = (var_489_cast_fp16_3, var_481_cast_fp16_3))[name = string("op_516_cast_fp16")];
            fp16 var_517_to_fp16 = const()[name = string("op_517_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_6_cast_fp16 = mul(x = var_516_cast_fp16, y = var_517_to_fp16)[name = string("aw2_6_cast_fp16")];
            string var_520_equation_0 = const()[name = string("op_520_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_520_cast_fp16 = einsum(equation = var_520_equation_0, values = (var_489_cast_fp16_4, var_481_cast_fp16_4))[name = string("op_520_cast_fp16")];
            fp16 var_521_to_fp16 = const()[name = string("op_521_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_6_cast_fp16 = mul(x = var_520_cast_fp16, y = var_521_to_fp16)[name = string("aw3_6_cast_fp16")];
            string var_524_equation_0 = const()[name = string("op_524_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_524_cast_fp16 = einsum(equation = var_524_equation_0, values = (var_489_cast_fp16_5, var_481_cast_fp16_5))[name = string("op_524_cast_fp16")];
            fp16 var_525_to_fp16 = const()[name = string("op_525_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_6_cast_fp16 = mul(x = var_524_cast_fp16, y = var_525_to_fp16)[name = string("aw4_6_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_527_cast_fp16 = softmax(axis = var_29, x = aw_6_cast_fp16)[name = string("op_527_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_528_cast_fp16 = softmax(axis = var_29, x = aw0_6_cast_fp16)[name = string("op_528_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_529_cast_fp16 = softmax(axis = var_29, x = aw1_6_cast_fp16)[name = string("op_529_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_530_cast_fp16 = softmax(axis = var_29, x = aw2_6_cast_fp16)[name = string("op_530_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_531_cast_fp16 = softmax(axis = var_29, x = aw3_6_cast_fp16)[name = string("op_531_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_532_cast_fp16 = softmax(axis = var_29, x = aw4_6_cast_fp16)[name = string("op_532_cast_fp16")];
            string var_534_equation_0 = const()[name = string("op_534_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_534_cast_fp16 = einsum(equation = var_534_equation_0, values = (var_496_cast_fp16_0, var_527_cast_fp16))[name = string("op_534_cast_fp16")];
            string var_536_equation_0 = const()[name = string("op_536_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_536_cast_fp16 = einsum(equation = var_536_equation_0, values = (var_496_cast_fp16_1, var_528_cast_fp16))[name = string("op_536_cast_fp16")];
            string var_538_equation_0 = const()[name = string("op_538_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_538_cast_fp16 = einsum(equation = var_538_equation_0, values = (var_496_cast_fp16_2, var_529_cast_fp16))[name = string("op_538_cast_fp16")];
            string var_540_equation_0 = const()[name = string("op_540_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_540_cast_fp16 = einsum(equation = var_540_equation_0, values = (var_496_cast_fp16_3, var_530_cast_fp16))[name = string("op_540_cast_fp16")];
            string var_542_equation_0 = const()[name = string("op_542_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_542_cast_fp16 = einsum(equation = var_542_equation_0, values = (var_496_cast_fp16_4, var_531_cast_fp16))[name = string("op_542_cast_fp16")];
            string var_544_equation_0 = const()[name = string("op_544_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_544_cast_fp16 = einsum(equation = var_544_equation_0, values = (var_496_cast_fp16_5, var_532_cast_fp16))[name = string("op_544_cast_fp16")];
            bool input_17_interleave_0 = const()[name = string("input_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_17_cast_fp16 = concat(axis = var_29, interleave = input_17_interleave_0, values = (var_534_cast_fp16, var_536_cast_fp16, var_538_cast_fp16, var_540_cast_fp16, var_542_cast_fp16, var_544_cast_fp16))[name = string("input_17_cast_fp16")];
            string attn_13_pad_type_0 = const()[name = string("attn_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_13_strides_0 = const()[name = string("attn_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_13_pad_0 = const()[name = string("attn_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_13_dilations_0 = const()[name = string("attn_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_13_groups_0 = const()[name = string("attn_13_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_2_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_2_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10935680)))];
            tensor<fp16, [1, 384, 1, 96]> attn_13_cast_fp16 = conv(bias = embedding_transformer_2_attn_out_proj_bias_to_fp16, dilations = attn_13_dilations_0, groups = attn_13_groups_0, pad = attn_13_pad_0, pad_type = attn_13_pad_type_0, strides = attn_13_strides_0, weight = embedding_transformer_2_attn_out_proj_weight_palettized_cast_fp16, x = input_17_cast_fp16)[name = string("attn_13_cast_fp16")];
            tensor<int32, [1]> var_554_axes_0 = const()[name = string("op_554_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_554_cast_fp16 = squeeze(axes = var_554_axes_0, x = attn_13_cast_fp16)[name = string("op_554_cast_fp16")];
            tensor<int32, [3]> var_555_perm_0 = const()[name = string("op_555_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_555_cast_fp16 = transpose(perm = var_555_perm_0, x = var_554_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 96, 384]> inputs_12_cast_fp16 = add(x = var_427_cast_fp16, y = var_555_cast_fp16)[name = string("inputs_12_cast_fp16")];
            tensor<int32, [3]> var_559_perm_0 = const()[name = string("op_559_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_7_axes_0 = const()[name = string("inputs0_7_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_559_cast_fp16 = transpose(perm = var_559_perm_0, x = inputs_12_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_7_cast_fp16 = expand_dims(axes = inputs0_7_axes_0, x = var_559_cast_fp16)[name = string("inputs0_7_cast_fp16")];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_567_to_fp16 = const()[name = string("op_567_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_567_to_fp16, x = inputs0_7_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [384]> out0_13_gamma_0_to_fp16 = const()[name = string("out0_13_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10936512)))];
            tensor<fp16, [384]> out0_13_beta_0_to_fp16 = const()[name = string("out0_13_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10937344)))];
            fp16 out0_13_epsilon_0_to_fp16 = const()[name = string("out0_13_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_13_cast_fp16 = batch_norm(beta = out0_13_beta_0_to_fp16, epsilon = out0_13_epsilon_0_to_fp16, gamma = out0_13_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_13_cast_fp16)[name = string("out0_13_cast_fp16")];
            tensor<int32, [1]> var_577_axes_0 = const()[name = string("op_577_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_577_cast_fp16 = squeeze(axes = var_577_axes_0, x = out0_13_cast_fp16)[name = string("op_577_cast_fp16")];
            tensor<int32, [3]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_584_axes_0 = const()[name = string("op_584_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = var_577_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 384, 96, 1]> var_584_cast_fp16 = expand_dims(axes = var_584_axes_0, x = transpose_4_cast_fp16)[name = string("op_584_cast_fp16")];
            string var_589_pad_type_0 = const()[name = string("op_589_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_589_strides_0 = const()[name = string("op_589_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_589_pad_0 = const()[name = string("op_589_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_589_dilations_0 = const()[name = string("op_589_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_589_groups_0 = const()[name = string("op_589_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_2_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_2_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10938176)))];
            tensor<fp16, [1, 1536, 96, 1]> var_589_cast_fp16 = conv(bias = embedding_transformer_2_mlp_c_fc_bias_to_fp16, dilations = var_589_dilations_0, groups = var_589_groups_0, pad = var_589_pad_0, pad_type = var_589_pad_type_0, strides = var_589_strides_0, weight = embedding_transformer_2_mlp_c_fc_weight_palettized_cast_fp16, x = var_584_cast_fp16)[name = string("op_589_cast_fp16")];
            tensor<int32, [1]> var_590_axes_0 = const()[name = string("op_590_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_590_cast_fp16 = squeeze(axes = var_590_axes_0, x = var_589_cast_fp16)[name = string("op_590_cast_fp16")];
            string var_591_mode_0 = const()[name = string("op_591_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_591_cast_fp16 = gelu(mode = var_591_mode_0, x = var_590_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_595_axes_0 = const()[name = string("op_595_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_595_cast_fp16 = expand_dims(axes = var_595_axes_0, x = var_591_cast_fp16)[name = string("op_595_cast_fp16")];
            string var_600_pad_type_0 = const()[name = string("op_600_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_600_strides_0 = const()[name = string("op_600_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_600_pad_0 = const()[name = string("op_600_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_600_dilations_0 = const()[name = string("op_600_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_600_groups_0 = const()[name = string("op_600_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_2_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_2_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10941312)))];
            tensor<fp16, [1, 384, 96, 1]> var_600_cast_fp16 = conv(bias = embedding_transformer_2_mlp_c_proj_bias_to_fp16, dilations = var_600_dilations_0, groups = var_600_groups_0, pad = var_600_pad_0, pad_type = var_600_pad_type_0, strides = var_600_strides_0, weight = embedding_transformer_2_mlp_c_proj_weight_palettized_cast_fp16, x = var_595_cast_fp16)[name = string("op_600_cast_fp16")];
            tensor<int32, [1]> var_601_axes_0 = const()[name = string("op_601_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_601_cast_fp16 = squeeze(axes = var_601_axes_0, x = var_600_cast_fp16)[name = string("op_601_cast_fp16")];
            tensor<int32, [3]> var_602_perm_0 = const()[name = string("op_602_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_602_cast_fp16 = transpose(perm = var_602_perm_0, x = var_601_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 96, 384]> var_603_cast_fp16 = add(x = inputs_12_cast_fp16, y = var_602_cast_fp16)[name = string("op_603_cast_fp16")];
            tensor<int32, [3]> var_610_perm_0 = const()[name = string("op_610_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_14_axes_0 = const()[name = string("inputs_14_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_610_cast_fp16 = transpose(perm = var_610_perm_0, x = var_603_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 384, 1, 96]> inputs_14_cast_fp16 = expand_dims(axes = inputs_14_axes_0, x = var_610_cast_fp16)[name = string("inputs_14_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_618_to_fp16 = const()[name = string("op_618_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_618_to_fp16, x = inputs_14_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [384]> out0_15_gamma_0_to_fp16 = const()[name = string("out0_15_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10942144)))];
            tensor<fp16, [384]> out0_15_beta_0_to_fp16 = const()[name = string("out0_15_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10942976)))];
            fp16 out0_15_epsilon_0_to_fp16 = const()[name = string("out0_15_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_15_cast_fp16 = batch_norm(beta = out0_15_beta_0_to_fp16, epsilon = out0_15_epsilon_0_to_fp16, gamma = out0_15_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_15_cast_fp16)[name = string("out0_15_cast_fp16")];
            tensor<int32, [1]> var_628_axes_0 = const()[name = string("op_628_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_628_cast_fp16 = squeeze(axes = var_628_axes_0, x = out0_15_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [1]> hidden_states_17_axes_0 = const()[name = string("hidden_states_17_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_17_cast_fp16 = expand_dims(axes = hidden_states_17_axes_0, x = var_628_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            string var_642_pad_type_0 = const()[name = string("op_642_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_642_strides_0 = const()[name = string("op_642_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_642_pad_0 = const()[name = string("op_642_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_642_dilations_0 = const()[name = string("op_642_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_642_groups_0 = const()[name = string("op_642_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_3_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_3_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10943808)))];
            tensor<fp16, [1, 384, 1, 96]> var_642_cast_fp16 = conv(bias = embedding_transformer_3_attn_query_bias_to_fp16, dilations = var_642_dilations_0, groups = var_642_groups_0, pad = var_642_pad_0, pad_type = var_642_pad_type_0, strides = var_642_strides_0, weight = embedding_transformer_3_attn_query_weight_palettized_cast_fp16, x = hidden_states_17_cast_fp16)[name = string("op_642_cast_fp16")];
            string k_8_pad_type_0 = const()[name = string("k_8_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_8_strides_0 = const()[name = string("k_8_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_8_pad_0 = const()[name = string("k_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_8_dilations_0 = const()[name = string("k_8_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_8_groups_0 = const()[name = string("k_8_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_3_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_3_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10944640)))];
            tensor<fp16, [1, 384, 1, 96]> k_8_cast_fp16 = conv(bias = embedding_transformer_3_attn_key_bias_to_fp16, dilations = k_8_dilations_0, groups = k_8_groups_0, pad = k_8_pad_0, pad_type = k_8_pad_type_0, strides = k_8_strides_0, weight = embedding_transformer_3_attn_key_weight_palettized_cast_fp16, x = hidden_states_17_cast_fp16)[name = string("k_8_cast_fp16")];
            string var_656_pad_type_0 = const()[name = string("op_656_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_656_strides_0 = const()[name = string("op_656_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_656_pad_0 = const()[name = string("op_656_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_656_dilations_0 = const()[name = string("op_656_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_656_groups_0 = const()[name = string("op_656_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_3_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_3_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10945472)))];
            tensor<fp16, [1, 384, 1, 96]> var_656_cast_fp16 = conv(bias = embedding_transformer_3_attn_value_bias_to_fp16, dilations = var_656_dilations_0, groups = var_656_groups_0, pad = var_656_pad_0, pad_type = var_656_pad_type_0, strides = var_656_strides_0, weight = embedding_transformer_3_attn_value_weight_palettized_cast_fp16, x = hidden_states_17_cast_fp16)[name = string("op_656_cast_fp16")];
            tensor<int32, [6]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_657_axis_0 = const()[name = string("op_657_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_657_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_657_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_657_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_657_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_657_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_657_cast_fp16_5 = split(axis = var_657_axis_0, split_sizes = tile_9, x = var_642_cast_fp16)[name = string("op_657_cast_fp16")];
            tensor<int32, [4]> var_664_perm_0 = const()[name = string("op_664_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_665_axis_0 = const()[name = string("op_665_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_664_cast_fp16 = transpose(perm = var_664_perm_0, x = k_8_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 96, 1, 64]> var_665_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_665_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_665_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_665_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_665_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_665_cast_fp16_5 = split(axis = var_665_axis_0, split_sizes = tile_10, x = var_664_cast_fp16)[name = string("op_665_cast_fp16")];
            tensor<int32, [6]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_672_axis_0 = const()[name = string("op_672_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_672_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_672_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_672_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_672_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_672_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_672_cast_fp16_5 = split(axis = var_672_axis_0, split_sizes = tile_11, x = var_656_cast_fp16)[name = string("op_672_cast_fp16")];
            string var_680_equation_0 = const()[name = string("op_680_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_680_cast_fp16 = einsum(equation = var_680_equation_0, values = (var_665_cast_fp16_0, var_657_cast_fp16_0))[name = string("op_680_cast_fp16")];
            fp16 var_681_to_fp16 = const()[name = string("op_681_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_8_cast_fp16 = mul(x = var_680_cast_fp16, y = var_681_to_fp16)[name = string("aw_8_cast_fp16")];
            string var_684_equation_0 = const()[name = string("op_684_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_684_cast_fp16 = einsum(equation = var_684_equation_0, values = (var_665_cast_fp16_1, var_657_cast_fp16_1))[name = string("op_684_cast_fp16")];
            fp16 var_685_to_fp16 = const()[name = string("op_685_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_8_cast_fp16 = mul(x = var_684_cast_fp16, y = var_685_to_fp16)[name = string("aw0_8_cast_fp16")];
            string var_688_equation_0 = const()[name = string("op_688_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_688_cast_fp16 = einsum(equation = var_688_equation_0, values = (var_665_cast_fp16_2, var_657_cast_fp16_2))[name = string("op_688_cast_fp16")];
            fp16 var_689_to_fp16 = const()[name = string("op_689_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_8_cast_fp16 = mul(x = var_688_cast_fp16, y = var_689_to_fp16)[name = string("aw1_8_cast_fp16")];
            string var_692_equation_0 = const()[name = string("op_692_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_692_cast_fp16 = einsum(equation = var_692_equation_0, values = (var_665_cast_fp16_3, var_657_cast_fp16_3))[name = string("op_692_cast_fp16")];
            fp16 var_693_to_fp16 = const()[name = string("op_693_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_8_cast_fp16 = mul(x = var_692_cast_fp16, y = var_693_to_fp16)[name = string("aw2_8_cast_fp16")];
            string var_696_equation_0 = const()[name = string("op_696_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_696_cast_fp16 = einsum(equation = var_696_equation_0, values = (var_665_cast_fp16_4, var_657_cast_fp16_4))[name = string("op_696_cast_fp16")];
            fp16 var_697_to_fp16 = const()[name = string("op_697_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_8_cast_fp16 = mul(x = var_696_cast_fp16, y = var_697_to_fp16)[name = string("aw3_8_cast_fp16")];
            string var_700_equation_0 = const()[name = string("op_700_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_700_cast_fp16 = einsum(equation = var_700_equation_0, values = (var_665_cast_fp16_5, var_657_cast_fp16_5))[name = string("op_700_cast_fp16")];
            fp16 var_701_to_fp16 = const()[name = string("op_701_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_8_cast_fp16 = mul(x = var_700_cast_fp16, y = var_701_to_fp16)[name = string("aw4_8_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_703_cast_fp16 = softmax(axis = var_29, x = aw_8_cast_fp16)[name = string("op_703_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_704_cast_fp16 = softmax(axis = var_29, x = aw0_8_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_705_cast_fp16 = softmax(axis = var_29, x = aw1_8_cast_fp16)[name = string("op_705_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_706_cast_fp16 = softmax(axis = var_29, x = aw2_8_cast_fp16)[name = string("op_706_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_707_cast_fp16 = softmax(axis = var_29, x = aw3_8_cast_fp16)[name = string("op_707_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_708_cast_fp16 = softmax(axis = var_29, x = aw4_8_cast_fp16)[name = string("op_708_cast_fp16")];
            string var_710_equation_0 = const()[name = string("op_710_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_710_cast_fp16 = einsum(equation = var_710_equation_0, values = (var_672_cast_fp16_0, var_703_cast_fp16))[name = string("op_710_cast_fp16")];
            string var_712_equation_0 = const()[name = string("op_712_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_712_cast_fp16 = einsum(equation = var_712_equation_0, values = (var_672_cast_fp16_1, var_704_cast_fp16))[name = string("op_712_cast_fp16")];
            string var_714_equation_0 = const()[name = string("op_714_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_714_cast_fp16 = einsum(equation = var_714_equation_0, values = (var_672_cast_fp16_2, var_705_cast_fp16))[name = string("op_714_cast_fp16")];
            string var_716_equation_0 = const()[name = string("op_716_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_716_cast_fp16 = einsum(equation = var_716_equation_0, values = (var_672_cast_fp16_3, var_706_cast_fp16))[name = string("op_716_cast_fp16")];
            string var_718_equation_0 = const()[name = string("op_718_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_718_cast_fp16 = einsum(equation = var_718_equation_0, values = (var_672_cast_fp16_4, var_707_cast_fp16))[name = string("op_718_cast_fp16")];
            string var_720_equation_0 = const()[name = string("op_720_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_720_cast_fp16 = einsum(equation = var_720_equation_0, values = (var_672_cast_fp16_5, var_708_cast_fp16))[name = string("op_720_cast_fp16")];
            bool input_23_interleave_0 = const()[name = string("input_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_23_cast_fp16 = concat(axis = var_29, interleave = input_23_interleave_0, values = (var_710_cast_fp16, var_712_cast_fp16, var_714_cast_fp16, var_716_cast_fp16, var_718_cast_fp16, var_720_cast_fp16))[name = string("input_23_cast_fp16")];
            string attn_17_pad_type_0 = const()[name = string("attn_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_17_strides_0 = const()[name = string("attn_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_17_pad_0 = const()[name = string("attn_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_17_dilations_0 = const()[name = string("attn_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_17_groups_0 = const()[name = string("attn_17_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_3_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_3_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10946304)))];
            tensor<fp16, [1, 384, 1, 96]> attn_17_cast_fp16 = conv(bias = embedding_transformer_3_attn_out_proj_bias_to_fp16, dilations = attn_17_dilations_0, groups = attn_17_groups_0, pad = attn_17_pad_0, pad_type = attn_17_pad_type_0, strides = attn_17_strides_0, weight = embedding_transformer_3_attn_out_proj_weight_palettized_cast_fp16, x = input_23_cast_fp16)[name = string("attn_17_cast_fp16")];
            tensor<int32, [1]> var_730_axes_0 = const()[name = string("op_730_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_730_cast_fp16 = squeeze(axes = var_730_axes_0, x = attn_17_cast_fp16)[name = string("op_730_cast_fp16")];
            tensor<int32, [3]> var_731_perm_0 = const()[name = string("op_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_731_cast_fp16 = transpose(perm = var_731_perm_0, x = var_730_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 96, 384]> inputs_16_cast_fp16 = add(x = var_603_cast_fp16, y = var_731_cast_fp16)[name = string("inputs_16_cast_fp16")];
            tensor<int32, [3]> var_735_perm_0 = const()[name = string("op_735_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_9_axes_0 = const()[name = string("inputs0_9_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_735_cast_fp16 = transpose(perm = var_735_perm_0, x = inputs_16_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_9_cast_fp16 = expand_dims(axes = inputs0_9_axes_0, x = var_735_cast_fp16)[name = string("inputs0_9_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_743_to_fp16 = const()[name = string("op_743_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_743_to_fp16, x = inputs0_9_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [384]> out0_17_gamma_0_to_fp16 = const()[name = string("out0_17_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10947136)))];
            tensor<fp16, [384]> out0_17_beta_0_to_fp16 = const()[name = string("out0_17_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10947968)))];
            fp16 out0_17_epsilon_0_to_fp16 = const()[name = string("out0_17_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_17_cast_fp16 = batch_norm(beta = out0_17_beta_0_to_fp16, epsilon = out0_17_epsilon_0_to_fp16, gamma = out0_17_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_17_cast_fp16)[name = string("out0_17_cast_fp16")];
            tensor<int32, [1]> var_753_axes_0 = const()[name = string("op_753_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_753_cast_fp16 = squeeze(axes = var_753_axes_0, x = out0_17_cast_fp16)[name = string("op_753_cast_fp16")];
            tensor<int32, [3]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_760_axes_0 = const()[name = string("op_760_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = var_753_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 384, 96, 1]> var_760_cast_fp16 = expand_dims(axes = var_760_axes_0, x = transpose_5_cast_fp16)[name = string("op_760_cast_fp16")];
            string var_765_pad_type_0 = const()[name = string("op_765_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_765_strides_0 = const()[name = string("op_765_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_765_pad_0 = const()[name = string("op_765_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_765_dilations_0 = const()[name = string("op_765_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_765_groups_0 = const()[name = string("op_765_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_3_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_3_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10948800)))];
            tensor<fp16, [1, 1536, 96, 1]> var_765_cast_fp16 = conv(bias = embedding_transformer_3_mlp_c_fc_bias_to_fp16, dilations = var_765_dilations_0, groups = var_765_groups_0, pad = var_765_pad_0, pad_type = var_765_pad_type_0, strides = var_765_strides_0, weight = embedding_transformer_3_mlp_c_fc_weight_palettized_cast_fp16, x = var_760_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<int32, [1]> var_766_axes_0 = const()[name = string("op_766_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_766_cast_fp16 = squeeze(axes = var_766_axes_0, x = var_765_cast_fp16)[name = string("op_766_cast_fp16")];
            string var_767_mode_0 = const()[name = string("op_767_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_767_cast_fp16 = gelu(mode = var_767_mode_0, x = var_766_cast_fp16)[name = string("op_767_cast_fp16")];
            tensor<int32, [1]> var_771_axes_0 = const()[name = string("op_771_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_771_cast_fp16 = expand_dims(axes = var_771_axes_0, x = var_767_cast_fp16)[name = string("op_771_cast_fp16")];
            string var_776_pad_type_0 = const()[name = string("op_776_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_776_strides_0 = const()[name = string("op_776_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_776_pad_0 = const()[name = string("op_776_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_776_dilations_0 = const()[name = string("op_776_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_776_groups_0 = const()[name = string("op_776_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_3_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_3_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10951936)))];
            tensor<fp16, [1, 384, 96, 1]> var_776_cast_fp16 = conv(bias = embedding_transformer_3_mlp_c_proj_bias_to_fp16, dilations = var_776_dilations_0, groups = var_776_groups_0, pad = var_776_pad_0, pad_type = var_776_pad_type_0, strides = var_776_strides_0, weight = embedding_transformer_3_mlp_c_proj_weight_palettized_cast_fp16, x = var_771_cast_fp16)[name = string("op_776_cast_fp16")];
            tensor<int32, [1]> var_777_axes_0 = const()[name = string("op_777_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_777_cast_fp16 = squeeze(axes = var_777_axes_0, x = var_776_cast_fp16)[name = string("op_777_cast_fp16")];
            tensor<int32, [3]> var_778_perm_0 = const()[name = string("op_778_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_778_cast_fp16 = transpose(perm = var_778_perm_0, x = var_777_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 96, 384]> var_779_cast_fp16 = add(x = inputs_16_cast_fp16, y = var_778_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [3]> var_786_perm_0 = const()[name = string("op_786_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_18_axes_0 = const()[name = string("inputs_18_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_786_cast_fp16 = transpose(perm = var_786_perm_0, x = var_779_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 384, 1, 96]> inputs_18_cast_fp16 = expand_dims(axes = inputs_18_axes_0, x = var_786_cast_fp16)[name = string("inputs_18_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_794_to_fp16 = const()[name = string("op_794_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_794_to_fp16, x = inputs_18_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [384]> out0_19_gamma_0_to_fp16 = const()[name = string("out0_19_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10952768)))];
            tensor<fp16, [384]> out0_19_beta_0_to_fp16 = const()[name = string("out0_19_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10953600)))];
            fp16 out0_19_epsilon_0_to_fp16 = const()[name = string("out0_19_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_19_cast_fp16 = batch_norm(beta = out0_19_beta_0_to_fp16, epsilon = out0_19_epsilon_0_to_fp16, gamma = out0_19_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_19_cast_fp16)[name = string("out0_19_cast_fp16")];
            tensor<int32, [1]> var_804_axes_0 = const()[name = string("op_804_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_804_cast_fp16 = squeeze(axes = var_804_axes_0, x = out0_19_cast_fp16)[name = string("op_804_cast_fp16")];
            tensor<int32, [1]> hidden_states_21_axes_0 = const()[name = string("hidden_states_21_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_21_cast_fp16 = expand_dims(axes = hidden_states_21_axes_0, x = var_804_cast_fp16)[name = string("hidden_states_21_cast_fp16")];
            string var_818_pad_type_0 = const()[name = string("op_818_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_818_strides_0 = const()[name = string("op_818_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_818_pad_0 = const()[name = string("op_818_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_818_dilations_0 = const()[name = string("op_818_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_818_groups_0 = const()[name = string("op_818_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_4_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_4_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10954432)))];
            tensor<fp16, [1, 384, 1, 96]> var_818_cast_fp16 = conv(bias = embedding_transformer_4_attn_query_bias_to_fp16, dilations = var_818_dilations_0, groups = var_818_groups_0, pad = var_818_pad_0, pad_type = var_818_pad_type_0, strides = var_818_strides_0, weight = embedding_transformer_4_attn_query_weight_palettized_cast_fp16, x = hidden_states_21_cast_fp16)[name = string("op_818_cast_fp16")];
            string k_10_pad_type_0 = const()[name = string("k_10_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_10_strides_0 = const()[name = string("k_10_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_10_pad_0 = const()[name = string("k_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_10_dilations_0 = const()[name = string("k_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_10_groups_0 = const()[name = string("k_10_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_4_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_4_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10955264)))];
            tensor<fp16, [1, 384, 1, 96]> k_10_cast_fp16 = conv(bias = embedding_transformer_4_attn_key_bias_to_fp16, dilations = k_10_dilations_0, groups = k_10_groups_0, pad = k_10_pad_0, pad_type = k_10_pad_type_0, strides = k_10_strides_0, weight = embedding_transformer_4_attn_key_weight_palettized_cast_fp16, x = hidden_states_21_cast_fp16)[name = string("k_10_cast_fp16")];
            string var_832_pad_type_0 = const()[name = string("op_832_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_832_strides_0 = const()[name = string("op_832_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_832_pad_0 = const()[name = string("op_832_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_832_dilations_0 = const()[name = string("op_832_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_832_groups_0 = const()[name = string("op_832_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_4_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_4_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10956096)))];
            tensor<fp16, [1, 384, 1, 96]> var_832_cast_fp16 = conv(bias = embedding_transformer_4_attn_value_bias_to_fp16, dilations = var_832_dilations_0, groups = var_832_groups_0, pad = var_832_pad_0, pad_type = var_832_pad_type_0, strides = var_832_strides_0, weight = embedding_transformer_4_attn_value_weight_palettized_cast_fp16, x = hidden_states_21_cast_fp16)[name = string("op_832_cast_fp16")];
            tensor<int32, [6]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_833_axis_0 = const()[name = string("op_833_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_833_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_833_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_833_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_833_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_833_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_833_cast_fp16_5 = split(axis = var_833_axis_0, split_sizes = tile_12, x = var_818_cast_fp16)[name = string("op_833_cast_fp16")];
            tensor<int32, [4]> var_840_perm_0 = const()[name = string("op_840_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_841_axis_0 = const()[name = string("op_841_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_840_cast_fp16 = transpose(perm = var_840_perm_0, x = k_10_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 96, 1, 64]> var_841_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_841_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_841_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_841_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_841_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_841_cast_fp16_5 = split(axis = var_841_axis_0, split_sizes = tile_13, x = var_840_cast_fp16)[name = string("op_841_cast_fp16")];
            tensor<int32, [6]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_848_axis_0 = const()[name = string("op_848_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_848_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_848_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_848_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_848_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_848_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_848_cast_fp16_5 = split(axis = var_848_axis_0, split_sizes = tile_14, x = var_832_cast_fp16)[name = string("op_848_cast_fp16")];
            string var_856_equation_0 = const()[name = string("op_856_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_856_cast_fp16 = einsum(equation = var_856_equation_0, values = (var_841_cast_fp16_0, var_833_cast_fp16_0))[name = string("op_856_cast_fp16")];
            fp16 var_857_to_fp16 = const()[name = string("op_857_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_10_cast_fp16 = mul(x = var_856_cast_fp16, y = var_857_to_fp16)[name = string("aw_10_cast_fp16")];
            string var_860_equation_0 = const()[name = string("op_860_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_860_cast_fp16 = einsum(equation = var_860_equation_0, values = (var_841_cast_fp16_1, var_833_cast_fp16_1))[name = string("op_860_cast_fp16")];
            fp16 var_861_to_fp16 = const()[name = string("op_861_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_10_cast_fp16 = mul(x = var_860_cast_fp16, y = var_861_to_fp16)[name = string("aw0_10_cast_fp16")];
            string var_864_equation_0 = const()[name = string("op_864_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_864_cast_fp16 = einsum(equation = var_864_equation_0, values = (var_841_cast_fp16_2, var_833_cast_fp16_2))[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_10_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("aw1_10_cast_fp16")];
            string var_868_equation_0 = const()[name = string("op_868_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_868_cast_fp16 = einsum(equation = var_868_equation_0, values = (var_841_cast_fp16_3, var_833_cast_fp16_3))[name = string("op_868_cast_fp16")];
            fp16 var_869_to_fp16 = const()[name = string("op_869_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_10_cast_fp16 = mul(x = var_868_cast_fp16, y = var_869_to_fp16)[name = string("aw2_10_cast_fp16")];
            string var_872_equation_0 = const()[name = string("op_872_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_872_cast_fp16 = einsum(equation = var_872_equation_0, values = (var_841_cast_fp16_4, var_833_cast_fp16_4))[name = string("op_872_cast_fp16")];
            fp16 var_873_to_fp16 = const()[name = string("op_873_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_10_cast_fp16 = mul(x = var_872_cast_fp16, y = var_873_to_fp16)[name = string("aw3_10_cast_fp16")];
            string var_876_equation_0 = const()[name = string("op_876_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_876_cast_fp16 = einsum(equation = var_876_equation_0, values = (var_841_cast_fp16_5, var_833_cast_fp16_5))[name = string("op_876_cast_fp16")];
            fp16 var_877_to_fp16 = const()[name = string("op_877_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_10_cast_fp16 = mul(x = var_876_cast_fp16, y = var_877_to_fp16)[name = string("aw4_10_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_879_cast_fp16 = softmax(axis = var_29, x = aw_10_cast_fp16)[name = string("op_879_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_880_cast_fp16 = softmax(axis = var_29, x = aw0_10_cast_fp16)[name = string("op_880_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_881_cast_fp16 = softmax(axis = var_29, x = aw1_10_cast_fp16)[name = string("op_881_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_882_cast_fp16 = softmax(axis = var_29, x = aw2_10_cast_fp16)[name = string("op_882_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_883_cast_fp16 = softmax(axis = var_29, x = aw3_10_cast_fp16)[name = string("op_883_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_884_cast_fp16 = softmax(axis = var_29, x = aw4_10_cast_fp16)[name = string("op_884_cast_fp16")];
            string var_886_equation_0 = const()[name = string("op_886_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_886_cast_fp16 = einsum(equation = var_886_equation_0, values = (var_848_cast_fp16_0, var_879_cast_fp16))[name = string("op_886_cast_fp16")];
            string var_888_equation_0 = const()[name = string("op_888_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_888_cast_fp16 = einsum(equation = var_888_equation_0, values = (var_848_cast_fp16_1, var_880_cast_fp16))[name = string("op_888_cast_fp16")];
            string var_890_equation_0 = const()[name = string("op_890_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_890_cast_fp16 = einsum(equation = var_890_equation_0, values = (var_848_cast_fp16_2, var_881_cast_fp16))[name = string("op_890_cast_fp16")];
            string var_892_equation_0 = const()[name = string("op_892_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_892_cast_fp16 = einsum(equation = var_892_equation_0, values = (var_848_cast_fp16_3, var_882_cast_fp16))[name = string("op_892_cast_fp16")];
            string var_894_equation_0 = const()[name = string("op_894_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_894_cast_fp16 = einsum(equation = var_894_equation_0, values = (var_848_cast_fp16_4, var_883_cast_fp16))[name = string("op_894_cast_fp16")];
            string var_896_equation_0 = const()[name = string("op_896_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_896_cast_fp16 = einsum(equation = var_896_equation_0, values = (var_848_cast_fp16_5, var_884_cast_fp16))[name = string("op_896_cast_fp16")];
            bool input_29_interleave_0 = const()[name = string("input_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_29_cast_fp16 = concat(axis = var_29, interleave = input_29_interleave_0, values = (var_886_cast_fp16, var_888_cast_fp16, var_890_cast_fp16, var_892_cast_fp16, var_894_cast_fp16, var_896_cast_fp16))[name = string("input_29_cast_fp16")];
            string attn_21_pad_type_0 = const()[name = string("attn_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_21_strides_0 = const()[name = string("attn_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_21_pad_0 = const()[name = string("attn_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_21_dilations_0 = const()[name = string("attn_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_21_groups_0 = const()[name = string("attn_21_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_4_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_4_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10956928)))];
            tensor<fp16, [1, 384, 1, 96]> attn_21_cast_fp16 = conv(bias = embedding_transformer_4_attn_out_proj_bias_to_fp16, dilations = attn_21_dilations_0, groups = attn_21_groups_0, pad = attn_21_pad_0, pad_type = attn_21_pad_type_0, strides = attn_21_strides_0, weight = embedding_transformer_4_attn_out_proj_weight_palettized_cast_fp16, x = input_29_cast_fp16)[name = string("attn_21_cast_fp16")];
            tensor<int32, [1]> var_906_axes_0 = const()[name = string("op_906_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_906_cast_fp16 = squeeze(axes = var_906_axes_0, x = attn_21_cast_fp16)[name = string("op_906_cast_fp16")];
            tensor<int32, [3]> var_907_perm_0 = const()[name = string("op_907_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_907_cast_fp16 = transpose(perm = var_907_perm_0, x = var_906_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 96, 384]> inputs_20_cast_fp16 = add(x = var_779_cast_fp16, y = var_907_cast_fp16)[name = string("inputs_20_cast_fp16")];
            tensor<int32, [3]> var_911_perm_0 = const()[name = string("op_911_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_11_axes_0 = const()[name = string("inputs0_11_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_911_cast_fp16 = transpose(perm = var_911_perm_0, x = inputs_20_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_11_cast_fp16 = expand_dims(axes = inputs0_11_axes_0, x = var_911_cast_fp16)[name = string("inputs0_11_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_919_to_fp16 = const()[name = string("op_919_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_919_to_fp16, x = inputs0_11_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [384]> out0_21_gamma_0_to_fp16 = const()[name = string("out0_21_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10957760)))];
            tensor<fp16, [384]> out0_21_beta_0_to_fp16 = const()[name = string("out0_21_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10958592)))];
            fp16 out0_21_epsilon_0_to_fp16 = const()[name = string("out0_21_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_21_cast_fp16 = batch_norm(beta = out0_21_beta_0_to_fp16, epsilon = out0_21_epsilon_0_to_fp16, gamma = out0_21_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_21_cast_fp16)[name = string("out0_21_cast_fp16")];
            tensor<int32, [1]> var_929_axes_0 = const()[name = string("op_929_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_929_cast_fp16 = squeeze(axes = var_929_axes_0, x = out0_21_cast_fp16)[name = string("op_929_cast_fp16")];
            tensor<int32, [3]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_936_axes_0 = const()[name = string("op_936_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = var_929_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 384, 96, 1]> var_936_cast_fp16 = expand_dims(axes = var_936_axes_0, x = transpose_6_cast_fp16)[name = string("op_936_cast_fp16")];
            string var_941_pad_type_0 = const()[name = string("op_941_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_941_strides_0 = const()[name = string("op_941_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_941_pad_0 = const()[name = string("op_941_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_941_dilations_0 = const()[name = string("op_941_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_941_groups_0 = const()[name = string("op_941_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_4_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_4_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10959424)))];
            tensor<fp16, [1, 1536, 96, 1]> var_941_cast_fp16 = conv(bias = embedding_transformer_4_mlp_c_fc_bias_to_fp16, dilations = var_941_dilations_0, groups = var_941_groups_0, pad = var_941_pad_0, pad_type = var_941_pad_type_0, strides = var_941_strides_0, weight = embedding_transformer_4_mlp_c_fc_weight_palettized_cast_fp16, x = var_936_cast_fp16)[name = string("op_941_cast_fp16")];
            tensor<int32, [1]> var_942_axes_0 = const()[name = string("op_942_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_942_cast_fp16 = squeeze(axes = var_942_axes_0, x = var_941_cast_fp16)[name = string("op_942_cast_fp16")];
            string var_943_mode_0 = const()[name = string("op_943_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_943_cast_fp16 = gelu(mode = var_943_mode_0, x = var_942_cast_fp16)[name = string("op_943_cast_fp16")];
            tensor<int32, [1]> var_947_axes_0 = const()[name = string("op_947_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_947_cast_fp16 = expand_dims(axes = var_947_axes_0, x = var_943_cast_fp16)[name = string("op_947_cast_fp16")];
            string var_952_pad_type_0 = const()[name = string("op_952_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_952_strides_0 = const()[name = string("op_952_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_952_pad_0 = const()[name = string("op_952_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_952_dilations_0 = const()[name = string("op_952_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_952_groups_0 = const()[name = string("op_952_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_4_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_4_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10962560)))];
            tensor<fp16, [1, 384, 96, 1]> var_952_cast_fp16 = conv(bias = embedding_transformer_4_mlp_c_proj_bias_to_fp16, dilations = var_952_dilations_0, groups = var_952_groups_0, pad = var_952_pad_0, pad_type = var_952_pad_type_0, strides = var_952_strides_0, weight = embedding_transformer_4_mlp_c_proj_weight_palettized_cast_fp16, x = var_947_cast_fp16)[name = string("op_952_cast_fp16")];
            tensor<int32, [1]> var_953_axes_0 = const()[name = string("op_953_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_953_cast_fp16 = squeeze(axes = var_953_axes_0, x = var_952_cast_fp16)[name = string("op_953_cast_fp16")];
            tensor<int32, [3]> var_954_perm_0 = const()[name = string("op_954_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_954_cast_fp16 = transpose(perm = var_954_perm_0, x = var_953_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 96, 384]> var_955_cast_fp16 = add(x = inputs_20_cast_fp16, y = var_954_cast_fp16)[name = string("op_955_cast_fp16")];
            tensor<int32, [3]> var_962_perm_0 = const()[name = string("op_962_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_22_axes_0 = const()[name = string("inputs_22_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_962_cast_fp16 = transpose(perm = var_962_perm_0, x = var_955_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 384, 1, 96]> inputs_22_cast_fp16 = expand_dims(axes = inputs_22_axes_0, x = var_962_cast_fp16)[name = string("inputs_22_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_970_to_fp16 = const()[name = string("op_970_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_970_to_fp16, x = inputs_22_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [384]> out0_23_gamma_0_to_fp16 = const()[name = string("out0_23_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10963392)))];
            tensor<fp16, [384]> out0_23_beta_0_to_fp16 = const()[name = string("out0_23_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10964224)))];
            fp16 out0_23_epsilon_0_to_fp16 = const()[name = string("out0_23_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_23_cast_fp16 = batch_norm(beta = out0_23_beta_0_to_fp16, epsilon = out0_23_epsilon_0_to_fp16, gamma = out0_23_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_23_cast_fp16)[name = string("out0_23_cast_fp16")];
            tensor<int32, [1]> var_980_axes_0 = const()[name = string("op_980_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_980_cast_fp16 = squeeze(axes = var_980_axes_0, x = out0_23_cast_fp16)[name = string("op_980_cast_fp16")];
            tensor<int32, [1]> hidden_states_25_axes_0 = const()[name = string("hidden_states_25_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_25_cast_fp16 = expand_dims(axes = hidden_states_25_axes_0, x = var_980_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            string var_994_pad_type_0 = const()[name = string("op_994_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_994_strides_0 = const()[name = string("op_994_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_994_pad_0 = const()[name = string("op_994_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_994_dilations_0 = const()[name = string("op_994_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_994_groups_0 = const()[name = string("op_994_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_5_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_5_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10965056)))];
            tensor<fp16, [1, 384, 1, 96]> var_994_cast_fp16 = conv(bias = embedding_transformer_5_attn_query_bias_to_fp16, dilations = var_994_dilations_0, groups = var_994_groups_0, pad = var_994_pad_0, pad_type = var_994_pad_type_0, strides = var_994_strides_0, weight = embedding_transformer_5_attn_query_weight_palettized_cast_fp16, x = hidden_states_25_cast_fp16)[name = string("op_994_cast_fp16")];
            string k_12_pad_type_0 = const()[name = string("k_12_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_12_strides_0 = const()[name = string("k_12_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_12_pad_0 = const()[name = string("k_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_12_dilations_0 = const()[name = string("k_12_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_12_groups_0 = const()[name = string("k_12_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_5_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_5_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10965888)))];
            tensor<fp16, [1, 384, 1, 96]> k_12_cast_fp16 = conv(bias = embedding_transformer_5_attn_key_bias_to_fp16, dilations = k_12_dilations_0, groups = k_12_groups_0, pad = k_12_pad_0, pad_type = k_12_pad_type_0, strides = k_12_strides_0, weight = embedding_transformer_5_attn_key_weight_palettized_cast_fp16, x = hidden_states_25_cast_fp16)[name = string("k_12_cast_fp16")];
            string var_1008_pad_type_0 = const()[name = string("op_1008_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1008_strides_0 = const()[name = string("op_1008_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1008_pad_0 = const()[name = string("op_1008_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1008_dilations_0 = const()[name = string("op_1008_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1008_groups_0 = const()[name = string("op_1008_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_5_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_5_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10966720)))];
            tensor<fp16, [1, 384, 1, 96]> var_1008_cast_fp16 = conv(bias = embedding_transformer_5_attn_value_bias_to_fp16, dilations = var_1008_dilations_0, groups = var_1008_groups_0, pad = var_1008_pad_0, pad_type = var_1008_pad_type_0, strides = var_1008_strides_0, weight = embedding_transformer_5_attn_value_weight_palettized_cast_fp16, x = hidden_states_25_cast_fp16)[name = string("op_1008_cast_fp16")];
            tensor<int32, [6]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1009_axis_0 = const()[name = string("op_1009_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1009_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1009_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1009_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1009_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1009_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1009_cast_fp16_5 = split(axis = var_1009_axis_0, split_sizes = tile_15, x = var_994_cast_fp16)[name = string("op_1009_cast_fp16")];
            tensor<int32, [4]> var_1016_perm_0 = const()[name = string("op_1016_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1017_axis_0 = const()[name = string("op_1017_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_1016_cast_fp16 = transpose(perm = var_1016_perm_0, x = k_12_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 96, 1, 64]> var_1017_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_1017_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_1017_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_1017_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_1017_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_1017_cast_fp16_5 = split(axis = var_1017_axis_0, split_sizes = tile_16, x = var_1016_cast_fp16)[name = string("op_1017_cast_fp16")];
            tensor<int32, [6]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1024_axis_0 = const()[name = string("op_1024_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1024_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1024_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1024_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1024_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1024_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1024_cast_fp16_5 = split(axis = var_1024_axis_0, split_sizes = tile_17, x = var_1008_cast_fp16)[name = string("op_1024_cast_fp16")];
            string var_1032_equation_0 = const()[name = string("op_1032_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1032_cast_fp16 = einsum(equation = var_1032_equation_0, values = (var_1017_cast_fp16_0, var_1009_cast_fp16_0))[name = string("op_1032_cast_fp16")];
            fp16 var_1033_to_fp16 = const()[name = string("op_1033_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_12_cast_fp16 = mul(x = var_1032_cast_fp16, y = var_1033_to_fp16)[name = string("aw_12_cast_fp16")];
            string var_1036_equation_0 = const()[name = string("op_1036_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1036_cast_fp16 = einsum(equation = var_1036_equation_0, values = (var_1017_cast_fp16_1, var_1009_cast_fp16_1))[name = string("op_1036_cast_fp16")];
            fp16 var_1037_to_fp16 = const()[name = string("op_1037_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_12_cast_fp16 = mul(x = var_1036_cast_fp16, y = var_1037_to_fp16)[name = string("aw0_12_cast_fp16")];
            string var_1040_equation_0 = const()[name = string("op_1040_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1040_cast_fp16 = einsum(equation = var_1040_equation_0, values = (var_1017_cast_fp16_2, var_1009_cast_fp16_2))[name = string("op_1040_cast_fp16")];
            fp16 var_1041_to_fp16 = const()[name = string("op_1041_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_12_cast_fp16 = mul(x = var_1040_cast_fp16, y = var_1041_to_fp16)[name = string("aw1_12_cast_fp16")];
            string var_1044_equation_0 = const()[name = string("op_1044_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1044_cast_fp16 = einsum(equation = var_1044_equation_0, values = (var_1017_cast_fp16_3, var_1009_cast_fp16_3))[name = string("op_1044_cast_fp16")];
            fp16 var_1045_to_fp16 = const()[name = string("op_1045_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_12_cast_fp16 = mul(x = var_1044_cast_fp16, y = var_1045_to_fp16)[name = string("aw2_12_cast_fp16")];
            string var_1048_equation_0 = const()[name = string("op_1048_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1048_cast_fp16 = einsum(equation = var_1048_equation_0, values = (var_1017_cast_fp16_4, var_1009_cast_fp16_4))[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_12_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("aw3_12_cast_fp16")];
            string var_1052_equation_0 = const()[name = string("op_1052_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1052_cast_fp16 = einsum(equation = var_1052_equation_0, values = (var_1017_cast_fp16_5, var_1009_cast_fp16_5))[name = string("op_1052_cast_fp16")];
            fp16 var_1053_to_fp16 = const()[name = string("op_1053_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_12_cast_fp16 = mul(x = var_1052_cast_fp16, y = var_1053_to_fp16)[name = string("aw4_12_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1055_cast_fp16 = softmax(axis = var_29, x = aw_12_cast_fp16)[name = string("op_1055_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1056_cast_fp16 = softmax(axis = var_29, x = aw0_12_cast_fp16)[name = string("op_1056_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1057_cast_fp16 = softmax(axis = var_29, x = aw1_12_cast_fp16)[name = string("op_1057_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1058_cast_fp16 = softmax(axis = var_29, x = aw2_12_cast_fp16)[name = string("op_1058_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1059_cast_fp16 = softmax(axis = var_29, x = aw3_12_cast_fp16)[name = string("op_1059_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1060_cast_fp16 = softmax(axis = var_29, x = aw4_12_cast_fp16)[name = string("op_1060_cast_fp16")];
            string var_1062_equation_0 = const()[name = string("op_1062_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1062_cast_fp16 = einsum(equation = var_1062_equation_0, values = (var_1024_cast_fp16_0, var_1055_cast_fp16))[name = string("op_1062_cast_fp16")];
            string var_1064_equation_0 = const()[name = string("op_1064_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1064_cast_fp16 = einsum(equation = var_1064_equation_0, values = (var_1024_cast_fp16_1, var_1056_cast_fp16))[name = string("op_1064_cast_fp16")];
            string var_1066_equation_0 = const()[name = string("op_1066_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1066_cast_fp16 = einsum(equation = var_1066_equation_0, values = (var_1024_cast_fp16_2, var_1057_cast_fp16))[name = string("op_1066_cast_fp16")];
            string var_1068_equation_0 = const()[name = string("op_1068_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1068_cast_fp16 = einsum(equation = var_1068_equation_0, values = (var_1024_cast_fp16_3, var_1058_cast_fp16))[name = string("op_1068_cast_fp16")];
            string var_1070_equation_0 = const()[name = string("op_1070_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1070_cast_fp16 = einsum(equation = var_1070_equation_0, values = (var_1024_cast_fp16_4, var_1059_cast_fp16))[name = string("op_1070_cast_fp16")];
            string var_1072_equation_0 = const()[name = string("op_1072_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1072_cast_fp16 = einsum(equation = var_1072_equation_0, values = (var_1024_cast_fp16_5, var_1060_cast_fp16))[name = string("op_1072_cast_fp16")];
            bool input_35_interleave_0 = const()[name = string("input_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_35_cast_fp16 = concat(axis = var_29, interleave = input_35_interleave_0, values = (var_1062_cast_fp16, var_1064_cast_fp16, var_1066_cast_fp16, var_1068_cast_fp16, var_1070_cast_fp16, var_1072_cast_fp16))[name = string("input_35_cast_fp16")];
            string attn_25_pad_type_0 = const()[name = string("attn_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_25_strides_0 = const()[name = string("attn_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_25_pad_0 = const()[name = string("attn_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_25_dilations_0 = const()[name = string("attn_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_25_groups_0 = const()[name = string("attn_25_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_5_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_5_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10967552)))];
            tensor<fp16, [1, 384, 1, 96]> attn_25_cast_fp16 = conv(bias = embedding_transformer_5_attn_out_proj_bias_to_fp16, dilations = attn_25_dilations_0, groups = attn_25_groups_0, pad = attn_25_pad_0, pad_type = attn_25_pad_type_0, strides = attn_25_strides_0, weight = embedding_transformer_5_attn_out_proj_weight_palettized_cast_fp16, x = input_35_cast_fp16)[name = string("attn_25_cast_fp16")];
            tensor<int32, [1]> var_1082_axes_0 = const()[name = string("op_1082_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1082_cast_fp16 = squeeze(axes = var_1082_axes_0, x = attn_25_cast_fp16)[name = string("op_1082_cast_fp16")];
            tensor<int32, [3]> var_1083_perm_0 = const()[name = string("op_1083_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_1083_cast_fp16 = transpose(perm = var_1083_perm_0, x = var_1082_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 96, 384]> inputs_24_cast_fp16 = add(x = var_955_cast_fp16, y = var_1083_cast_fp16)[name = string("inputs_24_cast_fp16")];
            tensor<int32, [3]> var_1087_perm_0 = const()[name = string("op_1087_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_13_axes_0 = const()[name = string("inputs0_13_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1087_cast_fp16 = transpose(perm = var_1087_perm_0, x = inputs_24_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_13_cast_fp16 = expand_dims(axes = inputs0_13_axes_0, x = var_1087_cast_fp16)[name = string("inputs0_13_cast_fp16")];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1095_to_fp16 = const()[name = string("op_1095_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_1095_to_fp16, x = inputs0_13_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [384]> out0_25_gamma_0_to_fp16 = const()[name = string("out0_25_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10968384)))];
            tensor<fp16, [384]> out0_25_beta_0_to_fp16 = const()[name = string("out0_25_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10969216)))];
            fp16 out0_25_epsilon_0_to_fp16 = const()[name = string("out0_25_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_25_cast_fp16 = batch_norm(beta = out0_25_beta_0_to_fp16, epsilon = out0_25_epsilon_0_to_fp16, gamma = out0_25_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_25_cast_fp16)[name = string("out0_25_cast_fp16")];
            tensor<int32, [1]> var_1105_axes_0 = const()[name = string("op_1105_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1105_cast_fp16 = squeeze(axes = var_1105_axes_0, x = out0_25_cast_fp16)[name = string("op_1105_cast_fp16")];
            tensor<int32, [3]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_1112_axes_0 = const()[name = string("op_1112_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = var_1105_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 384, 96, 1]> var_1112_cast_fp16 = expand_dims(axes = var_1112_axes_0, x = transpose_7_cast_fp16)[name = string("op_1112_cast_fp16")];
            string var_1117_pad_type_0 = const()[name = string("op_1117_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1117_strides_0 = const()[name = string("op_1117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1117_pad_0 = const()[name = string("op_1117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1117_dilations_0 = const()[name = string("op_1117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1117_groups_0 = const()[name = string("op_1117_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_5_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_5_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10970048)))];
            tensor<fp16, [1, 1536, 96, 1]> var_1117_cast_fp16 = conv(bias = embedding_transformer_5_mlp_c_fc_bias_to_fp16, dilations = var_1117_dilations_0, groups = var_1117_groups_0, pad = var_1117_pad_0, pad_type = var_1117_pad_type_0, strides = var_1117_strides_0, weight = embedding_transformer_5_mlp_c_fc_weight_palettized_cast_fp16, x = var_1112_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<int32, [1]> var_1118_axes_0 = const()[name = string("op_1118_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_1118_cast_fp16 = squeeze(axes = var_1118_axes_0, x = var_1117_cast_fp16)[name = string("op_1118_cast_fp16")];
            string var_1119_mode_0 = const()[name = string("op_1119_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_1119_cast_fp16 = gelu(mode = var_1119_mode_0, x = var_1118_cast_fp16)[name = string("op_1119_cast_fp16")];
            tensor<int32, [1]> var_1123_axes_0 = const()[name = string("op_1123_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_1123_cast_fp16 = expand_dims(axes = var_1123_axes_0, x = var_1119_cast_fp16)[name = string("op_1123_cast_fp16")];
            string var_1128_pad_type_0 = const()[name = string("op_1128_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1128_strides_0 = const()[name = string("op_1128_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1128_pad_0 = const()[name = string("op_1128_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1128_dilations_0 = const()[name = string("op_1128_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1128_groups_0 = const()[name = string("op_1128_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_5_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_5_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10973184)))];
            tensor<fp16, [1, 384, 96, 1]> var_1128_cast_fp16 = conv(bias = embedding_transformer_5_mlp_c_proj_bias_to_fp16, dilations = var_1128_dilations_0, groups = var_1128_groups_0, pad = var_1128_pad_0, pad_type = var_1128_pad_type_0, strides = var_1128_strides_0, weight = embedding_transformer_5_mlp_c_proj_weight_palettized_cast_fp16, x = var_1123_cast_fp16)[name = string("op_1128_cast_fp16")];
            tensor<int32, [1]> var_1129_axes_0 = const()[name = string("op_1129_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_1129_cast_fp16 = squeeze(axes = var_1129_axes_0, x = var_1128_cast_fp16)[name = string("op_1129_cast_fp16")];
            tensor<int32, [3]> var_1130_perm_0 = const()[name = string("op_1130_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_1130_cast_fp16 = transpose(perm = var_1130_perm_0, x = var_1129_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 96, 384]> var_1131_cast_fp16 = add(x = inputs_24_cast_fp16, y = var_1130_cast_fp16)[name = string("op_1131_cast_fp16")];
            tensor<int32, [3]> var_1138_perm_0 = const()[name = string("op_1138_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_26_axes_0 = const()[name = string("inputs_26_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1138_cast_fp16 = transpose(perm = var_1138_perm_0, x = var_1131_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 384, 1, 96]> inputs_26_cast_fp16 = expand_dims(axes = inputs_26_axes_0, x = var_1138_cast_fp16)[name = string("inputs_26_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1146_to_fp16 = const()[name = string("op_1146_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1146_to_fp16, x = inputs_26_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [384]> out0_27_gamma_0_to_fp16 = const()[name = string("out0_27_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10974016)))];
            tensor<fp16, [384]> out0_27_beta_0_to_fp16 = const()[name = string("out0_27_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10974848)))];
            fp16 out0_27_epsilon_0_to_fp16 = const()[name = string("out0_27_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_27_cast_fp16 = batch_norm(beta = out0_27_beta_0_to_fp16, epsilon = out0_27_epsilon_0_to_fp16, gamma = out0_27_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_27_cast_fp16)[name = string("out0_27_cast_fp16")];
            tensor<int32, [1]> var_1156_axes_0 = const()[name = string("op_1156_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1156_cast_fp16 = squeeze(axes = var_1156_axes_0, x = out0_27_cast_fp16)[name = string("op_1156_cast_fp16")];
            tensor<int32, [1]> hidden_states_29_axes_0 = const()[name = string("hidden_states_29_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_29_cast_fp16 = expand_dims(axes = hidden_states_29_axes_0, x = var_1156_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            string var_1170_pad_type_0 = const()[name = string("op_1170_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1170_strides_0 = const()[name = string("op_1170_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1170_pad_0 = const()[name = string("op_1170_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1170_dilations_0 = const()[name = string("op_1170_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1170_groups_0 = const()[name = string("op_1170_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_6_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_6_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10975680)))];
            tensor<fp16, [1, 384, 1, 96]> var_1170_cast_fp16 = conv(bias = embedding_transformer_6_attn_query_bias_to_fp16, dilations = var_1170_dilations_0, groups = var_1170_groups_0, pad = var_1170_pad_0, pad_type = var_1170_pad_type_0, strides = var_1170_strides_0, weight = embedding_transformer_6_attn_query_weight_palettized_cast_fp16, x = hidden_states_29_cast_fp16)[name = string("op_1170_cast_fp16")];
            string k_14_pad_type_0 = const()[name = string("k_14_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_14_strides_0 = const()[name = string("k_14_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_14_pad_0 = const()[name = string("k_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_14_dilations_0 = const()[name = string("k_14_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_14_groups_0 = const()[name = string("k_14_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_6_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_6_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10976512)))];
            tensor<fp16, [1, 384, 1, 96]> k_14_cast_fp16 = conv(bias = embedding_transformer_6_attn_key_bias_to_fp16, dilations = k_14_dilations_0, groups = k_14_groups_0, pad = k_14_pad_0, pad_type = k_14_pad_type_0, strides = k_14_strides_0, weight = embedding_transformer_6_attn_key_weight_palettized_cast_fp16, x = hidden_states_29_cast_fp16)[name = string("k_14_cast_fp16")];
            string var_1184_pad_type_0 = const()[name = string("op_1184_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1184_strides_0 = const()[name = string("op_1184_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1184_pad_0 = const()[name = string("op_1184_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1184_dilations_0 = const()[name = string("op_1184_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1184_groups_0 = const()[name = string("op_1184_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_6_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_6_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10977344)))];
            tensor<fp16, [1, 384, 1, 96]> var_1184_cast_fp16 = conv(bias = embedding_transformer_6_attn_value_bias_to_fp16, dilations = var_1184_dilations_0, groups = var_1184_groups_0, pad = var_1184_pad_0, pad_type = var_1184_pad_type_0, strides = var_1184_strides_0, weight = embedding_transformer_6_attn_value_weight_palettized_cast_fp16, x = hidden_states_29_cast_fp16)[name = string("op_1184_cast_fp16")];
            tensor<int32, [6]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1185_axis_0 = const()[name = string("op_1185_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1185_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1185_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1185_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1185_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1185_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1185_cast_fp16_5 = split(axis = var_1185_axis_0, split_sizes = tile_18, x = var_1170_cast_fp16)[name = string("op_1185_cast_fp16")];
            tensor<int32, [4]> var_1192_perm_0 = const()[name = string("op_1192_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1193_axis_0 = const()[name = string("op_1193_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_1192_cast_fp16 = transpose(perm = var_1192_perm_0, x = k_14_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 96, 1, 64]> var_1193_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_1193_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_1193_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_1193_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_1193_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_1193_cast_fp16_5 = split(axis = var_1193_axis_0, split_sizes = tile_19, x = var_1192_cast_fp16)[name = string("op_1193_cast_fp16")];
            tensor<int32, [6]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1200_axis_0 = const()[name = string("op_1200_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1200_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1200_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1200_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1200_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1200_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1200_cast_fp16_5 = split(axis = var_1200_axis_0, split_sizes = tile_20, x = var_1184_cast_fp16)[name = string("op_1200_cast_fp16")];
            string var_1208_equation_0 = const()[name = string("op_1208_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1208_cast_fp16 = einsum(equation = var_1208_equation_0, values = (var_1193_cast_fp16_0, var_1185_cast_fp16_0))[name = string("op_1208_cast_fp16")];
            fp16 var_1209_to_fp16 = const()[name = string("op_1209_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_14_cast_fp16 = mul(x = var_1208_cast_fp16, y = var_1209_to_fp16)[name = string("aw_14_cast_fp16")];
            string var_1212_equation_0 = const()[name = string("op_1212_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1212_cast_fp16 = einsum(equation = var_1212_equation_0, values = (var_1193_cast_fp16_1, var_1185_cast_fp16_1))[name = string("op_1212_cast_fp16")];
            fp16 var_1213_to_fp16 = const()[name = string("op_1213_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_14_cast_fp16 = mul(x = var_1212_cast_fp16, y = var_1213_to_fp16)[name = string("aw0_14_cast_fp16")];
            string var_1216_equation_0 = const()[name = string("op_1216_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1216_cast_fp16 = einsum(equation = var_1216_equation_0, values = (var_1193_cast_fp16_2, var_1185_cast_fp16_2))[name = string("op_1216_cast_fp16")];
            fp16 var_1217_to_fp16 = const()[name = string("op_1217_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_14_cast_fp16 = mul(x = var_1216_cast_fp16, y = var_1217_to_fp16)[name = string("aw1_14_cast_fp16")];
            string var_1220_equation_0 = const()[name = string("op_1220_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1220_cast_fp16 = einsum(equation = var_1220_equation_0, values = (var_1193_cast_fp16_3, var_1185_cast_fp16_3))[name = string("op_1220_cast_fp16")];
            fp16 var_1221_to_fp16 = const()[name = string("op_1221_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_14_cast_fp16 = mul(x = var_1220_cast_fp16, y = var_1221_to_fp16)[name = string("aw2_14_cast_fp16")];
            string var_1224_equation_0 = const()[name = string("op_1224_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1224_cast_fp16 = einsum(equation = var_1224_equation_0, values = (var_1193_cast_fp16_4, var_1185_cast_fp16_4))[name = string("op_1224_cast_fp16")];
            fp16 var_1225_to_fp16 = const()[name = string("op_1225_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_14_cast_fp16 = mul(x = var_1224_cast_fp16, y = var_1225_to_fp16)[name = string("aw3_14_cast_fp16")];
            string var_1228_equation_0 = const()[name = string("op_1228_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1228_cast_fp16 = einsum(equation = var_1228_equation_0, values = (var_1193_cast_fp16_5, var_1185_cast_fp16_5))[name = string("op_1228_cast_fp16")];
            fp16 var_1229_to_fp16 = const()[name = string("op_1229_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_14_cast_fp16 = mul(x = var_1228_cast_fp16, y = var_1229_to_fp16)[name = string("aw4_14_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1231_cast_fp16 = softmax(axis = var_29, x = aw_14_cast_fp16)[name = string("op_1231_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1232_cast_fp16 = softmax(axis = var_29, x = aw0_14_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1233_cast_fp16 = softmax(axis = var_29, x = aw1_14_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1234_cast_fp16 = softmax(axis = var_29, x = aw2_14_cast_fp16)[name = string("op_1234_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1235_cast_fp16 = softmax(axis = var_29, x = aw3_14_cast_fp16)[name = string("op_1235_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1236_cast_fp16 = softmax(axis = var_29, x = aw4_14_cast_fp16)[name = string("op_1236_cast_fp16")];
            string var_1238_equation_0 = const()[name = string("op_1238_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1238_cast_fp16 = einsum(equation = var_1238_equation_0, values = (var_1200_cast_fp16_0, var_1231_cast_fp16))[name = string("op_1238_cast_fp16")];
            string var_1240_equation_0 = const()[name = string("op_1240_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1240_cast_fp16 = einsum(equation = var_1240_equation_0, values = (var_1200_cast_fp16_1, var_1232_cast_fp16))[name = string("op_1240_cast_fp16")];
            string var_1242_equation_0 = const()[name = string("op_1242_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1242_cast_fp16 = einsum(equation = var_1242_equation_0, values = (var_1200_cast_fp16_2, var_1233_cast_fp16))[name = string("op_1242_cast_fp16")];
            string var_1244_equation_0 = const()[name = string("op_1244_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1244_cast_fp16 = einsum(equation = var_1244_equation_0, values = (var_1200_cast_fp16_3, var_1234_cast_fp16))[name = string("op_1244_cast_fp16")];
            string var_1246_equation_0 = const()[name = string("op_1246_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1246_cast_fp16 = einsum(equation = var_1246_equation_0, values = (var_1200_cast_fp16_4, var_1235_cast_fp16))[name = string("op_1246_cast_fp16")];
            string var_1248_equation_0 = const()[name = string("op_1248_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1248_cast_fp16 = einsum(equation = var_1248_equation_0, values = (var_1200_cast_fp16_5, var_1236_cast_fp16))[name = string("op_1248_cast_fp16")];
            bool input_41_interleave_0 = const()[name = string("input_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_41_cast_fp16 = concat(axis = var_29, interleave = input_41_interleave_0, values = (var_1238_cast_fp16, var_1240_cast_fp16, var_1242_cast_fp16, var_1244_cast_fp16, var_1246_cast_fp16, var_1248_cast_fp16))[name = string("input_41_cast_fp16")];
            string attn_29_pad_type_0 = const()[name = string("attn_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_29_strides_0 = const()[name = string("attn_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_29_pad_0 = const()[name = string("attn_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_29_dilations_0 = const()[name = string("attn_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_29_groups_0 = const()[name = string("attn_29_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_6_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_6_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10978176)))];
            tensor<fp16, [1, 384, 1, 96]> attn_29_cast_fp16 = conv(bias = embedding_transformer_6_attn_out_proj_bias_to_fp16, dilations = attn_29_dilations_0, groups = attn_29_groups_0, pad = attn_29_pad_0, pad_type = attn_29_pad_type_0, strides = attn_29_strides_0, weight = embedding_transformer_6_attn_out_proj_weight_palettized_cast_fp16, x = input_41_cast_fp16)[name = string("attn_29_cast_fp16")];
            tensor<int32, [1]> var_1258_axes_0 = const()[name = string("op_1258_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1258_cast_fp16 = squeeze(axes = var_1258_axes_0, x = attn_29_cast_fp16)[name = string("op_1258_cast_fp16")];
            tensor<int32, [3]> var_1259_perm_0 = const()[name = string("op_1259_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_1259_cast_fp16 = transpose(perm = var_1259_perm_0, x = var_1258_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 96, 384]> inputs_28_cast_fp16 = add(x = var_1131_cast_fp16, y = var_1259_cast_fp16)[name = string("inputs_28_cast_fp16")];
            tensor<int32, [3]> var_1263_perm_0 = const()[name = string("op_1263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_15_axes_0 = const()[name = string("inputs0_15_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1263_cast_fp16 = transpose(perm = var_1263_perm_0, x = inputs_28_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_15_cast_fp16 = expand_dims(axes = inputs0_15_axes_0, x = var_1263_cast_fp16)[name = string("inputs0_15_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1271_to_fp16 = const()[name = string("op_1271_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1271_to_fp16, x = inputs0_15_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [384]> out0_29_gamma_0_to_fp16 = const()[name = string("out0_29_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10979008)))];
            tensor<fp16, [384]> out0_29_beta_0_to_fp16 = const()[name = string("out0_29_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10979840)))];
            fp16 out0_29_epsilon_0_to_fp16 = const()[name = string("out0_29_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_29_cast_fp16 = batch_norm(beta = out0_29_beta_0_to_fp16, epsilon = out0_29_epsilon_0_to_fp16, gamma = out0_29_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_29_cast_fp16)[name = string("out0_29_cast_fp16")];
            tensor<int32, [1]> var_1281_axes_0 = const()[name = string("op_1281_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1281_cast_fp16 = squeeze(axes = var_1281_axes_0, x = out0_29_cast_fp16)[name = string("op_1281_cast_fp16")];
            tensor<int32, [3]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = var_1281_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 384, 96, 1]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = transpose_8_cast_fp16)[name = string("op_1288_cast_fp16")];
            string var_1293_pad_type_0 = const()[name = string("op_1293_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1293_strides_0 = const()[name = string("op_1293_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1293_pad_0 = const()[name = string("op_1293_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1293_dilations_0 = const()[name = string("op_1293_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1293_groups_0 = const()[name = string("op_1293_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_6_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_6_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10980672)))];
            tensor<fp16, [1, 1536, 96, 1]> var_1293_cast_fp16 = conv(bias = embedding_transformer_6_mlp_c_fc_bias_to_fp16, dilations = var_1293_dilations_0, groups = var_1293_groups_0, pad = var_1293_pad_0, pad_type = var_1293_pad_type_0, strides = var_1293_strides_0, weight = embedding_transformer_6_mlp_c_fc_weight_palettized_cast_fp16, x = var_1288_cast_fp16)[name = string("op_1293_cast_fp16")];
            tensor<int32, [1]> var_1294_axes_0 = const()[name = string("op_1294_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_1294_cast_fp16 = squeeze(axes = var_1294_axes_0, x = var_1293_cast_fp16)[name = string("op_1294_cast_fp16")];
            string var_1295_mode_0 = const()[name = string("op_1295_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_1295_cast_fp16 = gelu(mode = var_1295_mode_0, x = var_1294_cast_fp16)[name = string("op_1295_cast_fp16")];
            tensor<int32, [1]> var_1299_axes_0 = const()[name = string("op_1299_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_1299_cast_fp16 = expand_dims(axes = var_1299_axes_0, x = var_1295_cast_fp16)[name = string("op_1299_cast_fp16")];
            string var_1304_pad_type_0 = const()[name = string("op_1304_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1304_strides_0 = const()[name = string("op_1304_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1304_pad_0 = const()[name = string("op_1304_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1304_dilations_0 = const()[name = string("op_1304_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1304_groups_0 = const()[name = string("op_1304_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_6_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_6_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10983808)))];
            tensor<fp16, [1, 384, 96, 1]> var_1304_cast_fp16 = conv(bias = embedding_transformer_6_mlp_c_proj_bias_to_fp16, dilations = var_1304_dilations_0, groups = var_1304_groups_0, pad = var_1304_pad_0, pad_type = var_1304_pad_type_0, strides = var_1304_strides_0, weight = embedding_transformer_6_mlp_c_proj_weight_palettized_cast_fp16, x = var_1299_cast_fp16)[name = string("op_1304_cast_fp16")];
            tensor<int32, [1]> var_1305_axes_0 = const()[name = string("op_1305_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_1305_cast_fp16 = squeeze(axes = var_1305_axes_0, x = var_1304_cast_fp16)[name = string("op_1305_cast_fp16")];
            tensor<int32, [3]> var_1306_perm_0 = const()[name = string("op_1306_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_1306_cast_fp16 = transpose(perm = var_1306_perm_0, x = var_1305_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 96, 384]> var_1307_cast_fp16 = add(x = inputs_28_cast_fp16, y = var_1306_cast_fp16)[name = string("op_1307_cast_fp16")];
            tensor<int32, [3]> var_1314_perm_0 = const()[name = string("op_1314_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_30_axes_0 = const()[name = string("inputs_30_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1314_cast_fp16 = transpose(perm = var_1314_perm_0, x = var_1307_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 384, 1, 96]> inputs_30_cast_fp16 = expand_dims(axes = inputs_30_axes_0, x = var_1314_cast_fp16)[name = string("inputs_30_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1322_to_fp16 = const()[name = string("op_1322_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1322_to_fp16, x = inputs_30_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [384]> out0_31_gamma_0_to_fp16 = const()[name = string("out0_31_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10984640)))];
            tensor<fp16, [384]> out0_31_beta_0_to_fp16 = const()[name = string("out0_31_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10985472)))];
            fp16 out0_31_epsilon_0_to_fp16 = const()[name = string("out0_31_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_31_cast_fp16 = batch_norm(beta = out0_31_beta_0_to_fp16, epsilon = out0_31_epsilon_0_to_fp16, gamma = out0_31_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_31_cast_fp16)[name = string("out0_31_cast_fp16")];
            tensor<int32, [1]> var_1332_axes_0 = const()[name = string("op_1332_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1332_cast_fp16 = squeeze(axes = var_1332_axes_0, x = out0_31_cast_fp16)[name = string("op_1332_cast_fp16")];
            tensor<int32, [1]> hidden_states_33_axes_0 = const()[name = string("hidden_states_33_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_33_cast_fp16 = expand_dims(axes = hidden_states_33_axes_0, x = var_1332_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            string var_1346_pad_type_0 = const()[name = string("op_1346_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1346_strides_0 = const()[name = string("op_1346_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1346_pad_0 = const()[name = string("op_1346_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1346_dilations_0 = const()[name = string("op_1346_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1346_groups_0 = const()[name = string("op_1346_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_7_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_7_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10986304)))];
            tensor<fp16, [1, 384, 1, 96]> var_1346_cast_fp16 = conv(bias = embedding_transformer_7_attn_query_bias_to_fp16, dilations = var_1346_dilations_0, groups = var_1346_groups_0, pad = var_1346_pad_0, pad_type = var_1346_pad_type_0, strides = var_1346_strides_0, weight = embedding_transformer_7_attn_query_weight_palettized_cast_fp16, x = hidden_states_33_cast_fp16)[name = string("op_1346_cast_fp16")];
            string k_16_pad_type_0 = const()[name = string("k_16_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_16_strides_0 = const()[name = string("k_16_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_16_pad_0 = const()[name = string("k_16_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_16_dilations_0 = const()[name = string("k_16_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_16_groups_0 = const()[name = string("k_16_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_7_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_7_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10987136)))];
            tensor<fp16, [1, 384, 1, 96]> k_16_cast_fp16 = conv(bias = embedding_transformer_7_attn_key_bias_to_fp16, dilations = k_16_dilations_0, groups = k_16_groups_0, pad = k_16_pad_0, pad_type = k_16_pad_type_0, strides = k_16_strides_0, weight = embedding_transformer_7_attn_key_weight_palettized_cast_fp16, x = hidden_states_33_cast_fp16)[name = string("k_16_cast_fp16")];
            string var_1360_pad_type_0 = const()[name = string("op_1360_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1360_strides_0 = const()[name = string("op_1360_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1360_pad_0 = const()[name = string("op_1360_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1360_dilations_0 = const()[name = string("op_1360_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1360_groups_0 = const()[name = string("op_1360_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_7_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_7_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10987968)))];
            tensor<fp16, [1, 384, 1, 96]> var_1360_cast_fp16 = conv(bias = embedding_transformer_7_attn_value_bias_to_fp16, dilations = var_1360_dilations_0, groups = var_1360_groups_0, pad = var_1360_pad_0, pad_type = var_1360_pad_type_0, strides = var_1360_strides_0, weight = embedding_transformer_7_attn_value_weight_palettized_cast_fp16, x = hidden_states_33_cast_fp16)[name = string("op_1360_cast_fp16")];
            tensor<int32, [6]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1361_axis_0 = const()[name = string("op_1361_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1361_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1361_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1361_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1361_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1361_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1361_cast_fp16_5 = split(axis = var_1361_axis_0, split_sizes = tile_21, x = var_1346_cast_fp16)[name = string("op_1361_cast_fp16")];
            tensor<int32, [4]> var_1368_perm_0 = const()[name = string("op_1368_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1369_axis_0 = const()[name = string("op_1369_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_1368_cast_fp16 = transpose(perm = var_1368_perm_0, x = k_16_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 96, 1, 64]> var_1369_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_1369_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_1369_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_1369_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_1369_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_1369_cast_fp16_5 = split(axis = var_1369_axis_0, split_sizes = tile_22, x = var_1368_cast_fp16)[name = string("op_1369_cast_fp16")];
            tensor<int32, [6]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1376_axis_0 = const()[name = string("op_1376_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1376_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1376_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1376_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1376_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1376_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1376_cast_fp16_5 = split(axis = var_1376_axis_0, split_sizes = tile_23, x = var_1360_cast_fp16)[name = string("op_1376_cast_fp16")];
            string var_1384_equation_0 = const()[name = string("op_1384_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1384_cast_fp16 = einsum(equation = var_1384_equation_0, values = (var_1369_cast_fp16_0, var_1361_cast_fp16_0))[name = string("op_1384_cast_fp16")];
            fp16 var_1385_to_fp16 = const()[name = string("op_1385_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_16_cast_fp16 = mul(x = var_1384_cast_fp16, y = var_1385_to_fp16)[name = string("aw_16_cast_fp16")];
            string var_1388_equation_0 = const()[name = string("op_1388_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1388_cast_fp16 = einsum(equation = var_1388_equation_0, values = (var_1369_cast_fp16_1, var_1361_cast_fp16_1))[name = string("op_1388_cast_fp16")];
            fp16 var_1389_to_fp16 = const()[name = string("op_1389_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_16_cast_fp16 = mul(x = var_1388_cast_fp16, y = var_1389_to_fp16)[name = string("aw0_16_cast_fp16")];
            string var_1392_equation_0 = const()[name = string("op_1392_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1392_cast_fp16 = einsum(equation = var_1392_equation_0, values = (var_1369_cast_fp16_2, var_1361_cast_fp16_2))[name = string("op_1392_cast_fp16")];
            fp16 var_1393_to_fp16 = const()[name = string("op_1393_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_16_cast_fp16 = mul(x = var_1392_cast_fp16, y = var_1393_to_fp16)[name = string("aw1_16_cast_fp16")];
            string var_1396_equation_0 = const()[name = string("op_1396_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1396_cast_fp16 = einsum(equation = var_1396_equation_0, values = (var_1369_cast_fp16_3, var_1361_cast_fp16_3))[name = string("op_1396_cast_fp16")];
            fp16 var_1397_to_fp16 = const()[name = string("op_1397_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_16_cast_fp16 = mul(x = var_1396_cast_fp16, y = var_1397_to_fp16)[name = string("aw2_16_cast_fp16")];
            string var_1400_equation_0 = const()[name = string("op_1400_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1400_cast_fp16 = einsum(equation = var_1400_equation_0, values = (var_1369_cast_fp16_4, var_1361_cast_fp16_4))[name = string("op_1400_cast_fp16")];
            fp16 var_1401_to_fp16 = const()[name = string("op_1401_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_16_cast_fp16 = mul(x = var_1400_cast_fp16, y = var_1401_to_fp16)[name = string("aw3_16_cast_fp16")];
            string var_1404_equation_0 = const()[name = string("op_1404_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1404_cast_fp16 = einsum(equation = var_1404_equation_0, values = (var_1369_cast_fp16_5, var_1361_cast_fp16_5))[name = string("op_1404_cast_fp16")];
            fp16 var_1405_to_fp16 = const()[name = string("op_1405_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_16_cast_fp16 = mul(x = var_1404_cast_fp16, y = var_1405_to_fp16)[name = string("aw4_16_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1407_cast_fp16 = softmax(axis = var_29, x = aw_16_cast_fp16)[name = string("op_1407_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1408_cast_fp16 = softmax(axis = var_29, x = aw0_16_cast_fp16)[name = string("op_1408_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1409_cast_fp16 = softmax(axis = var_29, x = aw1_16_cast_fp16)[name = string("op_1409_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1410_cast_fp16 = softmax(axis = var_29, x = aw2_16_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1411_cast_fp16 = softmax(axis = var_29, x = aw3_16_cast_fp16)[name = string("op_1411_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1412_cast_fp16 = softmax(axis = var_29, x = aw4_16_cast_fp16)[name = string("op_1412_cast_fp16")];
            string var_1414_equation_0 = const()[name = string("op_1414_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1414_cast_fp16 = einsum(equation = var_1414_equation_0, values = (var_1376_cast_fp16_0, var_1407_cast_fp16))[name = string("op_1414_cast_fp16")];
            string var_1416_equation_0 = const()[name = string("op_1416_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1416_cast_fp16 = einsum(equation = var_1416_equation_0, values = (var_1376_cast_fp16_1, var_1408_cast_fp16))[name = string("op_1416_cast_fp16")];
            string var_1418_equation_0 = const()[name = string("op_1418_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1418_cast_fp16 = einsum(equation = var_1418_equation_0, values = (var_1376_cast_fp16_2, var_1409_cast_fp16))[name = string("op_1418_cast_fp16")];
            string var_1420_equation_0 = const()[name = string("op_1420_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1420_cast_fp16 = einsum(equation = var_1420_equation_0, values = (var_1376_cast_fp16_3, var_1410_cast_fp16))[name = string("op_1420_cast_fp16")];
            string var_1422_equation_0 = const()[name = string("op_1422_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1422_cast_fp16 = einsum(equation = var_1422_equation_0, values = (var_1376_cast_fp16_4, var_1411_cast_fp16))[name = string("op_1422_cast_fp16")];
            string var_1424_equation_0 = const()[name = string("op_1424_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1424_cast_fp16 = einsum(equation = var_1424_equation_0, values = (var_1376_cast_fp16_5, var_1412_cast_fp16))[name = string("op_1424_cast_fp16")];
            bool input_47_interleave_0 = const()[name = string("input_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_47_cast_fp16 = concat(axis = var_29, interleave = input_47_interleave_0, values = (var_1414_cast_fp16, var_1416_cast_fp16, var_1418_cast_fp16, var_1420_cast_fp16, var_1422_cast_fp16, var_1424_cast_fp16))[name = string("input_47_cast_fp16")];
            string attn_33_pad_type_0 = const()[name = string("attn_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_33_strides_0 = const()[name = string("attn_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_33_pad_0 = const()[name = string("attn_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_33_dilations_0 = const()[name = string("attn_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_33_groups_0 = const()[name = string("attn_33_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_7_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_7_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10988800)))];
            tensor<fp16, [1, 384, 1, 96]> attn_33_cast_fp16 = conv(bias = embedding_transformer_7_attn_out_proj_bias_to_fp16, dilations = attn_33_dilations_0, groups = attn_33_groups_0, pad = attn_33_pad_0, pad_type = attn_33_pad_type_0, strides = attn_33_strides_0, weight = embedding_transformer_7_attn_out_proj_weight_palettized_cast_fp16, x = input_47_cast_fp16)[name = string("attn_33_cast_fp16")];
            tensor<int32, [1]> var_1434_axes_0 = const()[name = string("op_1434_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1434_cast_fp16 = squeeze(axes = var_1434_axes_0, x = attn_33_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<int32, [3]> var_1435_perm_0 = const()[name = string("op_1435_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_1435_cast_fp16 = transpose(perm = var_1435_perm_0, x = var_1434_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 96, 384]> inputs_32_cast_fp16 = add(x = var_1307_cast_fp16, y = var_1435_cast_fp16)[name = string("inputs_32_cast_fp16")];
            tensor<int32, [3]> var_1439_perm_0 = const()[name = string("op_1439_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_17_axes_0 = const()[name = string("inputs0_17_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1439_cast_fp16 = transpose(perm = var_1439_perm_0, x = inputs_32_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_17_cast_fp16 = expand_dims(axes = inputs0_17_axes_0, x = var_1439_cast_fp16)[name = string("inputs0_17_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1447_to_fp16 = const()[name = string("op_1447_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1447_to_fp16, x = inputs0_17_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [384]> out0_33_gamma_0_to_fp16 = const()[name = string("out0_33_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10989632)))];
            tensor<fp16, [384]> out0_33_beta_0_to_fp16 = const()[name = string("out0_33_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10990464)))];
            fp16 out0_33_epsilon_0_to_fp16 = const()[name = string("out0_33_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_33_cast_fp16 = batch_norm(beta = out0_33_beta_0_to_fp16, epsilon = out0_33_epsilon_0_to_fp16, gamma = out0_33_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_33_cast_fp16)[name = string("out0_33_cast_fp16")];
            tensor<int32, [1]> var_1457_axes_0 = const()[name = string("op_1457_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1457_cast_fp16 = squeeze(axes = var_1457_axes_0, x = out0_33_cast_fp16)[name = string("op_1457_cast_fp16")];
            tensor<int32, [3]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_1464_axes_0 = const()[name = string("op_1464_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = var_1457_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 384, 96, 1]> var_1464_cast_fp16 = expand_dims(axes = var_1464_axes_0, x = transpose_9_cast_fp16)[name = string("op_1464_cast_fp16")];
            string var_1469_pad_type_0 = const()[name = string("op_1469_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1469_strides_0 = const()[name = string("op_1469_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1469_pad_0 = const()[name = string("op_1469_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1469_dilations_0 = const()[name = string("op_1469_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1469_groups_0 = const()[name = string("op_1469_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_7_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_7_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10991296)))];
            tensor<fp16, [1, 1536, 96, 1]> var_1469_cast_fp16 = conv(bias = embedding_transformer_7_mlp_c_fc_bias_to_fp16, dilations = var_1469_dilations_0, groups = var_1469_groups_0, pad = var_1469_pad_0, pad_type = var_1469_pad_type_0, strides = var_1469_strides_0, weight = embedding_transformer_7_mlp_c_fc_weight_palettized_cast_fp16, x = var_1464_cast_fp16)[name = string("op_1469_cast_fp16")];
            tensor<int32, [1]> var_1470_axes_0 = const()[name = string("op_1470_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_1470_cast_fp16 = squeeze(axes = var_1470_axes_0, x = var_1469_cast_fp16)[name = string("op_1470_cast_fp16")];
            string var_1471_mode_0 = const()[name = string("op_1471_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_1471_cast_fp16 = gelu(mode = var_1471_mode_0, x = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            tensor<int32, [1]> var_1475_axes_0 = const()[name = string("op_1475_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_1475_cast_fp16 = expand_dims(axes = var_1475_axes_0, x = var_1471_cast_fp16)[name = string("op_1475_cast_fp16")];
            string var_1480_pad_type_0 = const()[name = string("op_1480_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1480_strides_0 = const()[name = string("op_1480_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1480_pad_0 = const()[name = string("op_1480_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1480_dilations_0 = const()[name = string("op_1480_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1480_groups_0 = const()[name = string("op_1480_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_7_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_7_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10994432)))];
            tensor<fp16, [1, 384, 96, 1]> var_1480_cast_fp16 = conv(bias = embedding_transformer_7_mlp_c_proj_bias_to_fp16, dilations = var_1480_dilations_0, groups = var_1480_groups_0, pad = var_1480_pad_0, pad_type = var_1480_pad_type_0, strides = var_1480_strides_0, weight = embedding_transformer_7_mlp_c_proj_weight_palettized_cast_fp16, x = var_1475_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<int32, [1]> var_1481_axes_0 = const()[name = string("op_1481_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_1481_cast_fp16 = squeeze(axes = var_1481_axes_0, x = var_1480_cast_fp16)[name = string("op_1481_cast_fp16")];
            tensor<int32, [3]> var_1482_perm_0 = const()[name = string("op_1482_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_1482_cast_fp16 = transpose(perm = var_1482_perm_0, x = var_1481_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 96, 384]> var_1483_cast_fp16 = add(x = inputs_32_cast_fp16, y = var_1482_cast_fp16)[name = string("op_1483_cast_fp16")];
            tensor<int32, [3]> var_1490_perm_0 = const()[name = string("op_1490_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_34_axes_0 = const()[name = string("inputs_34_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1490_cast_fp16 = transpose(perm = var_1490_perm_0, x = var_1483_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 384, 1, 96]> inputs_34_cast_fp16 = expand_dims(axes = inputs_34_axes_0, x = var_1490_cast_fp16)[name = string("inputs_34_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1498_to_fp16 = const()[name = string("op_1498_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1498_to_fp16, x = inputs_34_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [384]> out0_35_gamma_0_to_fp16 = const()[name = string("out0_35_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10995264)))];
            tensor<fp16, [384]> out0_35_beta_0_to_fp16 = const()[name = string("out0_35_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10996096)))];
            fp16 out0_35_epsilon_0_to_fp16 = const()[name = string("out0_35_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_35_cast_fp16 = batch_norm(beta = out0_35_beta_0_to_fp16, epsilon = out0_35_epsilon_0_to_fp16, gamma = out0_35_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_35_cast_fp16)[name = string("out0_35_cast_fp16")];
            tensor<int32, [1]> var_1508_axes_0 = const()[name = string("op_1508_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1508_cast_fp16 = squeeze(axes = var_1508_axes_0, x = out0_35_cast_fp16)[name = string("op_1508_cast_fp16")];
            tensor<int32, [1]> hidden_states_37_axes_0 = const()[name = string("hidden_states_37_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_37_cast_fp16 = expand_dims(axes = hidden_states_37_axes_0, x = var_1508_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            string var_1522_pad_type_0 = const()[name = string("op_1522_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1522_strides_0 = const()[name = string("op_1522_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1522_pad_0 = const()[name = string("op_1522_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1522_dilations_0 = const()[name = string("op_1522_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1522_groups_0 = const()[name = string("op_1522_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_8_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_8_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10996928)))];
            tensor<fp16, [1, 384, 1, 96]> var_1522_cast_fp16 = conv(bias = embedding_transformer_8_attn_query_bias_to_fp16, dilations = var_1522_dilations_0, groups = var_1522_groups_0, pad = var_1522_pad_0, pad_type = var_1522_pad_type_0, strides = var_1522_strides_0, weight = embedding_transformer_8_attn_query_weight_palettized_cast_fp16, x = hidden_states_37_cast_fp16)[name = string("op_1522_cast_fp16")];
            string k_18_pad_type_0 = const()[name = string("k_18_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_18_strides_0 = const()[name = string("k_18_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_18_pad_0 = const()[name = string("k_18_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_18_dilations_0 = const()[name = string("k_18_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_18_groups_0 = const()[name = string("k_18_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_8_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_8_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10997760)))];
            tensor<fp16, [1, 384, 1, 96]> k_18_cast_fp16 = conv(bias = embedding_transformer_8_attn_key_bias_to_fp16, dilations = k_18_dilations_0, groups = k_18_groups_0, pad = k_18_pad_0, pad_type = k_18_pad_type_0, strides = k_18_strides_0, weight = embedding_transformer_8_attn_key_weight_palettized_cast_fp16, x = hidden_states_37_cast_fp16)[name = string("k_18_cast_fp16")];
            string var_1536_pad_type_0 = const()[name = string("op_1536_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1536_strides_0 = const()[name = string("op_1536_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1536_pad_0 = const()[name = string("op_1536_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1536_dilations_0 = const()[name = string("op_1536_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1536_groups_0 = const()[name = string("op_1536_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_8_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_8_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10998592)))];
            tensor<fp16, [1, 384, 1, 96]> var_1536_cast_fp16 = conv(bias = embedding_transformer_8_attn_value_bias_to_fp16, dilations = var_1536_dilations_0, groups = var_1536_groups_0, pad = var_1536_pad_0, pad_type = var_1536_pad_type_0, strides = var_1536_strides_0, weight = embedding_transformer_8_attn_value_weight_palettized_cast_fp16, x = hidden_states_37_cast_fp16)[name = string("op_1536_cast_fp16")];
            tensor<int32, [6]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1537_axis_0 = const()[name = string("op_1537_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1537_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1537_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1537_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1537_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1537_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1537_cast_fp16_5 = split(axis = var_1537_axis_0, split_sizes = tile_24, x = var_1522_cast_fp16)[name = string("op_1537_cast_fp16")];
            tensor<int32, [4]> var_1544_perm_0 = const()[name = string("op_1544_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1545_axis_0 = const()[name = string("op_1545_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_1544_cast_fp16 = transpose(perm = var_1544_perm_0, x = k_18_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 96, 1, 64]> var_1545_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_1545_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_1545_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_1545_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_1545_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_1545_cast_fp16_5 = split(axis = var_1545_axis_0, split_sizes = tile_25, x = var_1544_cast_fp16)[name = string("op_1545_cast_fp16")];
            tensor<int32, [6]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1552_axis_0 = const()[name = string("op_1552_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1552_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1552_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1552_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1552_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1552_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1552_cast_fp16_5 = split(axis = var_1552_axis_0, split_sizes = tile_26, x = var_1536_cast_fp16)[name = string("op_1552_cast_fp16")];
            string var_1560_equation_0 = const()[name = string("op_1560_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1560_cast_fp16 = einsum(equation = var_1560_equation_0, values = (var_1545_cast_fp16_0, var_1537_cast_fp16_0))[name = string("op_1560_cast_fp16")];
            fp16 var_1561_to_fp16 = const()[name = string("op_1561_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_18_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_1561_to_fp16)[name = string("aw_18_cast_fp16")];
            string var_1564_equation_0 = const()[name = string("op_1564_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1564_cast_fp16 = einsum(equation = var_1564_equation_0, values = (var_1545_cast_fp16_1, var_1537_cast_fp16_1))[name = string("op_1564_cast_fp16")];
            fp16 var_1565_to_fp16 = const()[name = string("op_1565_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_18_cast_fp16 = mul(x = var_1564_cast_fp16, y = var_1565_to_fp16)[name = string("aw0_18_cast_fp16")];
            string var_1568_equation_0 = const()[name = string("op_1568_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1568_cast_fp16 = einsum(equation = var_1568_equation_0, values = (var_1545_cast_fp16_2, var_1537_cast_fp16_2))[name = string("op_1568_cast_fp16")];
            fp16 var_1569_to_fp16 = const()[name = string("op_1569_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_18_cast_fp16 = mul(x = var_1568_cast_fp16, y = var_1569_to_fp16)[name = string("aw1_18_cast_fp16")];
            string var_1572_equation_0 = const()[name = string("op_1572_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1572_cast_fp16 = einsum(equation = var_1572_equation_0, values = (var_1545_cast_fp16_3, var_1537_cast_fp16_3))[name = string("op_1572_cast_fp16")];
            fp16 var_1573_to_fp16 = const()[name = string("op_1573_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_18_cast_fp16 = mul(x = var_1572_cast_fp16, y = var_1573_to_fp16)[name = string("aw2_18_cast_fp16")];
            string var_1576_equation_0 = const()[name = string("op_1576_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1576_cast_fp16 = einsum(equation = var_1576_equation_0, values = (var_1545_cast_fp16_4, var_1537_cast_fp16_4))[name = string("op_1576_cast_fp16")];
            fp16 var_1577_to_fp16 = const()[name = string("op_1577_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_18_cast_fp16 = mul(x = var_1576_cast_fp16, y = var_1577_to_fp16)[name = string("aw3_18_cast_fp16")];
            string var_1580_equation_0 = const()[name = string("op_1580_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1580_cast_fp16 = einsum(equation = var_1580_equation_0, values = (var_1545_cast_fp16_5, var_1537_cast_fp16_5))[name = string("op_1580_cast_fp16")];
            fp16 var_1581_to_fp16 = const()[name = string("op_1581_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_18_cast_fp16 = mul(x = var_1580_cast_fp16, y = var_1581_to_fp16)[name = string("aw4_18_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1583_cast_fp16 = softmax(axis = var_29, x = aw_18_cast_fp16)[name = string("op_1583_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1584_cast_fp16 = softmax(axis = var_29, x = aw0_18_cast_fp16)[name = string("op_1584_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1585_cast_fp16 = softmax(axis = var_29, x = aw1_18_cast_fp16)[name = string("op_1585_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1586_cast_fp16 = softmax(axis = var_29, x = aw2_18_cast_fp16)[name = string("op_1586_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1587_cast_fp16 = softmax(axis = var_29, x = aw3_18_cast_fp16)[name = string("op_1587_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1588_cast_fp16 = softmax(axis = var_29, x = aw4_18_cast_fp16)[name = string("op_1588_cast_fp16")];
            string var_1590_equation_0 = const()[name = string("op_1590_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1590_cast_fp16 = einsum(equation = var_1590_equation_0, values = (var_1552_cast_fp16_0, var_1583_cast_fp16))[name = string("op_1590_cast_fp16")];
            string var_1592_equation_0 = const()[name = string("op_1592_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1592_cast_fp16 = einsum(equation = var_1592_equation_0, values = (var_1552_cast_fp16_1, var_1584_cast_fp16))[name = string("op_1592_cast_fp16")];
            string var_1594_equation_0 = const()[name = string("op_1594_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1594_cast_fp16 = einsum(equation = var_1594_equation_0, values = (var_1552_cast_fp16_2, var_1585_cast_fp16))[name = string("op_1594_cast_fp16")];
            string var_1596_equation_0 = const()[name = string("op_1596_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1596_cast_fp16 = einsum(equation = var_1596_equation_0, values = (var_1552_cast_fp16_3, var_1586_cast_fp16))[name = string("op_1596_cast_fp16")];
            string var_1598_equation_0 = const()[name = string("op_1598_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1598_cast_fp16 = einsum(equation = var_1598_equation_0, values = (var_1552_cast_fp16_4, var_1587_cast_fp16))[name = string("op_1598_cast_fp16")];
            string var_1600_equation_0 = const()[name = string("op_1600_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1600_cast_fp16 = einsum(equation = var_1600_equation_0, values = (var_1552_cast_fp16_5, var_1588_cast_fp16))[name = string("op_1600_cast_fp16")];
            bool input_53_interleave_0 = const()[name = string("input_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_53_cast_fp16 = concat(axis = var_29, interleave = input_53_interleave_0, values = (var_1590_cast_fp16, var_1592_cast_fp16, var_1594_cast_fp16, var_1596_cast_fp16, var_1598_cast_fp16, var_1600_cast_fp16))[name = string("input_53_cast_fp16")];
            string attn_37_pad_type_0 = const()[name = string("attn_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_37_strides_0 = const()[name = string("attn_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_37_pad_0 = const()[name = string("attn_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_37_dilations_0 = const()[name = string("attn_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_37_groups_0 = const()[name = string("attn_37_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_8_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_8_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10999424)))];
            tensor<fp16, [1, 384, 1, 96]> attn_37_cast_fp16 = conv(bias = embedding_transformer_8_attn_out_proj_bias_to_fp16, dilations = attn_37_dilations_0, groups = attn_37_groups_0, pad = attn_37_pad_0, pad_type = attn_37_pad_type_0, strides = attn_37_strides_0, weight = embedding_transformer_8_attn_out_proj_weight_palettized_cast_fp16, x = input_53_cast_fp16)[name = string("attn_37_cast_fp16")];
            tensor<int32, [1]> var_1610_axes_0 = const()[name = string("op_1610_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1610_cast_fp16 = squeeze(axes = var_1610_axes_0, x = attn_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [3]> var_1611_perm_0 = const()[name = string("op_1611_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_1611_cast_fp16 = transpose(perm = var_1611_perm_0, x = var_1610_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 96, 384]> inputs_36_cast_fp16 = add(x = var_1483_cast_fp16, y = var_1611_cast_fp16)[name = string("inputs_36_cast_fp16")];
            tensor<int32, [3]> var_1615_perm_0 = const()[name = string("op_1615_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_19_axes_0 = const()[name = string("inputs0_19_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1615_cast_fp16 = transpose(perm = var_1615_perm_0, x = inputs_36_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_19_cast_fp16 = expand_dims(axes = inputs0_19_axes_0, x = var_1615_cast_fp16)[name = string("inputs0_19_cast_fp16")];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1623_to_fp16 = const()[name = string("op_1623_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1623_to_fp16, x = inputs0_19_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [384]> out0_37_gamma_0_to_fp16 = const()[name = string("out0_37_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11000256)))];
            tensor<fp16, [384]> out0_37_beta_0_to_fp16 = const()[name = string("out0_37_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11001088)))];
            fp16 out0_37_epsilon_0_to_fp16 = const()[name = string("out0_37_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_37_cast_fp16 = batch_norm(beta = out0_37_beta_0_to_fp16, epsilon = out0_37_epsilon_0_to_fp16, gamma = out0_37_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_37_cast_fp16)[name = string("out0_37_cast_fp16")];
            tensor<int32, [1]> var_1633_axes_0 = const()[name = string("op_1633_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1633_cast_fp16 = squeeze(axes = var_1633_axes_0, x = out0_37_cast_fp16)[name = string("op_1633_cast_fp16")];
            tensor<int32, [3]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_1640_axes_0 = const()[name = string("op_1640_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = var_1633_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 384, 96, 1]> var_1640_cast_fp16 = expand_dims(axes = var_1640_axes_0, x = transpose_10_cast_fp16)[name = string("op_1640_cast_fp16")];
            string var_1645_pad_type_0 = const()[name = string("op_1645_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1645_strides_0 = const()[name = string("op_1645_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1645_pad_0 = const()[name = string("op_1645_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1645_dilations_0 = const()[name = string("op_1645_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1645_groups_0 = const()[name = string("op_1645_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_8_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_8_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11001920)))];
            tensor<fp16, [1, 1536, 96, 1]> var_1645_cast_fp16 = conv(bias = embedding_transformer_8_mlp_c_fc_bias_to_fp16, dilations = var_1645_dilations_0, groups = var_1645_groups_0, pad = var_1645_pad_0, pad_type = var_1645_pad_type_0, strides = var_1645_strides_0, weight = embedding_transformer_8_mlp_c_fc_weight_palettized_cast_fp16, x = var_1640_cast_fp16)[name = string("op_1645_cast_fp16")];
            tensor<int32, [1]> var_1646_axes_0 = const()[name = string("op_1646_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_1646_cast_fp16 = squeeze(axes = var_1646_axes_0, x = var_1645_cast_fp16)[name = string("op_1646_cast_fp16")];
            string var_1647_mode_0 = const()[name = string("op_1647_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_1647_cast_fp16 = gelu(mode = var_1647_mode_0, x = var_1646_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<int32, [1]> var_1651_axes_0 = const()[name = string("op_1651_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_1651_cast_fp16 = expand_dims(axes = var_1651_axes_0, x = var_1647_cast_fp16)[name = string("op_1651_cast_fp16")];
            string var_1656_pad_type_0 = const()[name = string("op_1656_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1656_strides_0 = const()[name = string("op_1656_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1656_pad_0 = const()[name = string("op_1656_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1656_dilations_0 = const()[name = string("op_1656_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1656_groups_0 = const()[name = string("op_1656_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_8_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_8_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11005056)))];
            tensor<fp16, [1, 384, 96, 1]> var_1656_cast_fp16 = conv(bias = embedding_transformer_8_mlp_c_proj_bias_to_fp16, dilations = var_1656_dilations_0, groups = var_1656_groups_0, pad = var_1656_pad_0, pad_type = var_1656_pad_type_0, strides = var_1656_strides_0, weight = embedding_transformer_8_mlp_c_proj_weight_palettized_cast_fp16, x = var_1651_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<int32, [1]> var_1657_axes_0 = const()[name = string("op_1657_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_1657_cast_fp16 = squeeze(axes = var_1657_axes_0, x = var_1656_cast_fp16)[name = string("op_1657_cast_fp16")];
            tensor<int32, [3]> var_1658_perm_0 = const()[name = string("op_1658_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_1658_cast_fp16 = transpose(perm = var_1658_perm_0, x = var_1657_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 96, 384]> var_1659_cast_fp16 = add(x = inputs_36_cast_fp16, y = var_1658_cast_fp16)[name = string("op_1659_cast_fp16")];
            tensor<int32, [3]> var_1666_perm_0 = const()[name = string("op_1666_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_38_axes_0 = const()[name = string("inputs_38_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1666_cast_fp16 = transpose(perm = var_1666_perm_0, x = var_1659_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 384, 1, 96]> inputs_38_cast_fp16 = expand_dims(axes = inputs_38_axes_0, x = var_1666_cast_fp16)[name = string("inputs_38_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1674_to_fp16 = const()[name = string("op_1674_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1674_to_fp16, x = inputs_38_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [384]> out0_39_gamma_0_to_fp16 = const()[name = string("out0_39_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11005888)))];
            tensor<fp16, [384]> out0_39_beta_0_to_fp16 = const()[name = string("out0_39_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11006720)))];
            fp16 out0_39_epsilon_0_to_fp16 = const()[name = string("out0_39_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_39_cast_fp16 = batch_norm(beta = out0_39_beta_0_to_fp16, epsilon = out0_39_epsilon_0_to_fp16, gamma = out0_39_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_39_cast_fp16)[name = string("out0_39_cast_fp16")];
            tensor<int32, [1]> var_1684_axes_0 = const()[name = string("op_1684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1684_cast_fp16 = squeeze(axes = var_1684_axes_0, x = out0_39_cast_fp16)[name = string("op_1684_cast_fp16")];
            tensor<int32, [1]> hidden_states_41_axes_0 = const()[name = string("hidden_states_41_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_41_cast_fp16 = expand_dims(axes = hidden_states_41_axes_0, x = var_1684_cast_fp16)[name = string("hidden_states_41_cast_fp16")];
            string var_1698_pad_type_0 = const()[name = string("op_1698_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1698_strides_0 = const()[name = string("op_1698_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1698_pad_0 = const()[name = string("op_1698_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1698_dilations_0 = const()[name = string("op_1698_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1698_groups_0 = const()[name = string("op_1698_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_9_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_9_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11007552)))];
            tensor<fp16, [1, 384, 1, 96]> var_1698_cast_fp16 = conv(bias = embedding_transformer_9_attn_query_bias_to_fp16, dilations = var_1698_dilations_0, groups = var_1698_groups_0, pad = var_1698_pad_0, pad_type = var_1698_pad_type_0, strides = var_1698_strides_0, weight = embedding_transformer_9_attn_query_weight_palettized_cast_fp16, x = hidden_states_41_cast_fp16)[name = string("op_1698_cast_fp16")];
            string k_20_pad_type_0 = const()[name = string("k_20_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_20_strides_0 = const()[name = string("k_20_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_20_pad_0 = const()[name = string("k_20_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_20_dilations_0 = const()[name = string("k_20_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_20_groups_0 = const()[name = string("k_20_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_9_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_9_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11008384)))];
            tensor<fp16, [1, 384, 1, 96]> k_20_cast_fp16 = conv(bias = embedding_transformer_9_attn_key_bias_to_fp16, dilations = k_20_dilations_0, groups = k_20_groups_0, pad = k_20_pad_0, pad_type = k_20_pad_type_0, strides = k_20_strides_0, weight = embedding_transformer_9_attn_key_weight_palettized_cast_fp16, x = hidden_states_41_cast_fp16)[name = string("k_20_cast_fp16")];
            string var_1712_pad_type_0 = const()[name = string("op_1712_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1712_strides_0 = const()[name = string("op_1712_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1712_pad_0 = const()[name = string("op_1712_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1712_dilations_0 = const()[name = string("op_1712_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1712_groups_0 = const()[name = string("op_1712_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_9_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_9_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11009216)))];
            tensor<fp16, [1, 384, 1, 96]> var_1712_cast_fp16 = conv(bias = embedding_transformer_9_attn_value_bias_to_fp16, dilations = var_1712_dilations_0, groups = var_1712_groups_0, pad = var_1712_pad_0, pad_type = var_1712_pad_type_0, strides = var_1712_strides_0, weight = embedding_transformer_9_attn_value_weight_palettized_cast_fp16, x = hidden_states_41_cast_fp16)[name = string("op_1712_cast_fp16")];
            tensor<int32, [6]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1713_axis_0 = const()[name = string("op_1713_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1713_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1713_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1713_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1713_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1713_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1713_cast_fp16_5 = split(axis = var_1713_axis_0, split_sizes = tile_27, x = var_1698_cast_fp16)[name = string("op_1713_cast_fp16")];
            tensor<int32, [4]> var_1720_perm_0 = const()[name = string("op_1720_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_28 = const()[name = string("tile_28"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1721_axis_0 = const()[name = string("op_1721_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_1720_cast_fp16 = transpose(perm = var_1720_perm_0, x = k_20_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 96, 1, 64]> var_1721_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_1721_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_1721_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_1721_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_1721_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_1721_cast_fp16_5 = split(axis = var_1721_axis_0, split_sizes = tile_28, x = var_1720_cast_fp16)[name = string("op_1721_cast_fp16")];
            tensor<int32, [6]> tile_29 = const()[name = string("tile_29"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1728_axis_0 = const()[name = string("op_1728_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1728_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1728_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1728_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1728_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1728_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1728_cast_fp16_5 = split(axis = var_1728_axis_0, split_sizes = tile_29, x = var_1712_cast_fp16)[name = string("op_1728_cast_fp16")];
            string var_1736_equation_0 = const()[name = string("op_1736_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1736_cast_fp16 = einsum(equation = var_1736_equation_0, values = (var_1721_cast_fp16_0, var_1713_cast_fp16_0))[name = string("op_1736_cast_fp16")];
            fp16 var_1737_to_fp16 = const()[name = string("op_1737_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_20_cast_fp16 = mul(x = var_1736_cast_fp16, y = var_1737_to_fp16)[name = string("aw_20_cast_fp16")];
            string var_1740_equation_0 = const()[name = string("op_1740_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1740_cast_fp16 = einsum(equation = var_1740_equation_0, values = (var_1721_cast_fp16_1, var_1713_cast_fp16_1))[name = string("op_1740_cast_fp16")];
            fp16 var_1741_to_fp16 = const()[name = string("op_1741_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_20_cast_fp16 = mul(x = var_1740_cast_fp16, y = var_1741_to_fp16)[name = string("aw0_20_cast_fp16")];
            string var_1744_equation_0 = const()[name = string("op_1744_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1744_cast_fp16 = einsum(equation = var_1744_equation_0, values = (var_1721_cast_fp16_2, var_1713_cast_fp16_2))[name = string("op_1744_cast_fp16")];
            fp16 var_1745_to_fp16 = const()[name = string("op_1745_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_20_cast_fp16 = mul(x = var_1744_cast_fp16, y = var_1745_to_fp16)[name = string("aw1_20_cast_fp16")];
            string var_1748_equation_0 = const()[name = string("op_1748_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1748_cast_fp16 = einsum(equation = var_1748_equation_0, values = (var_1721_cast_fp16_3, var_1713_cast_fp16_3))[name = string("op_1748_cast_fp16")];
            fp16 var_1749_to_fp16 = const()[name = string("op_1749_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_20_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_1749_to_fp16)[name = string("aw2_20_cast_fp16")];
            string var_1752_equation_0 = const()[name = string("op_1752_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1752_cast_fp16 = einsum(equation = var_1752_equation_0, values = (var_1721_cast_fp16_4, var_1713_cast_fp16_4))[name = string("op_1752_cast_fp16")];
            fp16 var_1753_to_fp16 = const()[name = string("op_1753_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_20_cast_fp16 = mul(x = var_1752_cast_fp16, y = var_1753_to_fp16)[name = string("aw3_20_cast_fp16")];
            string var_1756_equation_0 = const()[name = string("op_1756_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1756_cast_fp16 = einsum(equation = var_1756_equation_0, values = (var_1721_cast_fp16_5, var_1713_cast_fp16_5))[name = string("op_1756_cast_fp16")];
            fp16 var_1757_to_fp16 = const()[name = string("op_1757_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_20_cast_fp16 = mul(x = var_1756_cast_fp16, y = var_1757_to_fp16)[name = string("aw4_20_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1759_cast_fp16 = softmax(axis = var_29, x = aw_20_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1760_cast_fp16 = softmax(axis = var_29, x = aw0_20_cast_fp16)[name = string("op_1760_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1761_cast_fp16 = softmax(axis = var_29, x = aw1_20_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1762_cast_fp16 = softmax(axis = var_29, x = aw2_20_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1763_cast_fp16 = softmax(axis = var_29, x = aw3_20_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1764_cast_fp16 = softmax(axis = var_29, x = aw4_20_cast_fp16)[name = string("op_1764_cast_fp16")];
            string var_1766_equation_0 = const()[name = string("op_1766_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1766_cast_fp16 = einsum(equation = var_1766_equation_0, values = (var_1728_cast_fp16_0, var_1759_cast_fp16))[name = string("op_1766_cast_fp16")];
            string var_1768_equation_0 = const()[name = string("op_1768_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1768_cast_fp16 = einsum(equation = var_1768_equation_0, values = (var_1728_cast_fp16_1, var_1760_cast_fp16))[name = string("op_1768_cast_fp16")];
            string var_1770_equation_0 = const()[name = string("op_1770_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1770_cast_fp16 = einsum(equation = var_1770_equation_0, values = (var_1728_cast_fp16_2, var_1761_cast_fp16))[name = string("op_1770_cast_fp16")];
            string var_1772_equation_0 = const()[name = string("op_1772_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1772_cast_fp16 = einsum(equation = var_1772_equation_0, values = (var_1728_cast_fp16_3, var_1762_cast_fp16))[name = string("op_1772_cast_fp16")];
            string var_1774_equation_0 = const()[name = string("op_1774_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1774_cast_fp16 = einsum(equation = var_1774_equation_0, values = (var_1728_cast_fp16_4, var_1763_cast_fp16))[name = string("op_1774_cast_fp16")];
            string var_1776_equation_0 = const()[name = string("op_1776_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1776_cast_fp16 = einsum(equation = var_1776_equation_0, values = (var_1728_cast_fp16_5, var_1764_cast_fp16))[name = string("op_1776_cast_fp16")];
            bool input_59_interleave_0 = const()[name = string("input_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_59_cast_fp16 = concat(axis = var_29, interleave = input_59_interleave_0, values = (var_1766_cast_fp16, var_1768_cast_fp16, var_1770_cast_fp16, var_1772_cast_fp16, var_1774_cast_fp16, var_1776_cast_fp16))[name = string("input_59_cast_fp16")];
            string attn_41_pad_type_0 = const()[name = string("attn_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_41_strides_0 = const()[name = string("attn_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_41_pad_0 = const()[name = string("attn_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_41_dilations_0 = const()[name = string("attn_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_41_groups_0 = const()[name = string("attn_41_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_9_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_9_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11010048)))];
            tensor<fp16, [1, 384, 1, 96]> attn_41_cast_fp16 = conv(bias = embedding_transformer_9_attn_out_proj_bias_to_fp16, dilations = attn_41_dilations_0, groups = attn_41_groups_0, pad = attn_41_pad_0, pad_type = attn_41_pad_type_0, strides = attn_41_strides_0, weight = embedding_transformer_9_attn_out_proj_weight_palettized_cast_fp16, x = input_59_cast_fp16)[name = string("attn_41_cast_fp16")];
            tensor<int32, [1]> var_1786_axes_0 = const()[name = string("op_1786_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1786_cast_fp16 = squeeze(axes = var_1786_axes_0, x = attn_41_cast_fp16)[name = string("op_1786_cast_fp16")];
            tensor<int32, [3]> var_1787_perm_0 = const()[name = string("op_1787_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_1787_cast_fp16 = transpose(perm = var_1787_perm_0, x = var_1786_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 96, 384]> inputs_40_cast_fp16 = add(x = var_1659_cast_fp16, y = var_1787_cast_fp16)[name = string("inputs_40_cast_fp16")];
            tensor<int32, [3]> var_1791_perm_0 = const()[name = string("op_1791_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_21_axes_0 = const()[name = string("inputs0_21_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1791_cast_fp16 = transpose(perm = var_1791_perm_0, x = inputs_40_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_21_cast_fp16 = expand_dims(axes = inputs0_21_axes_0, x = var_1791_cast_fp16)[name = string("inputs0_21_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1799_to_fp16 = const()[name = string("op_1799_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1799_to_fp16, x = inputs0_21_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [384]> out0_41_gamma_0_to_fp16 = const()[name = string("out0_41_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11010880)))];
            tensor<fp16, [384]> out0_41_beta_0_to_fp16 = const()[name = string("out0_41_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11011712)))];
            fp16 out0_41_epsilon_0_to_fp16 = const()[name = string("out0_41_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_41_cast_fp16 = batch_norm(beta = out0_41_beta_0_to_fp16, epsilon = out0_41_epsilon_0_to_fp16, gamma = out0_41_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_41_cast_fp16)[name = string("out0_41_cast_fp16")];
            tensor<int32, [1]> var_1809_axes_0 = const()[name = string("op_1809_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1809_cast_fp16 = squeeze(axes = var_1809_axes_0, x = out0_41_cast_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [3]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_1816_axes_0 = const()[name = string("op_1816_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = var_1809_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 384, 96, 1]> var_1816_cast_fp16 = expand_dims(axes = var_1816_axes_0, x = transpose_11_cast_fp16)[name = string("op_1816_cast_fp16")];
            string var_1821_pad_type_0 = const()[name = string("op_1821_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1821_strides_0 = const()[name = string("op_1821_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1821_pad_0 = const()[name = string("op_1821_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1821_dilations_0 = const()[name = string("op_1821_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1821_groups_0 = const()[name = string("op_1821_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_9_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_9_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11012544)))];
            tensor<fp16, [1, 1536, 96, 1]> var_1821_cast_fp16 = conv(bias = embedding_transformer_9_mlp_c_fc_bias_to_fp16, dilations = var_1821_dilations_0, groups = var_1821_groups_0, pad = var_1821_pad_0, pad_type = var_1821_pad_type_0, strides = var_1821_strides_0, weight = embedding_transformer_9_mlp_c_fc_weight_palettized_cast_fp16, x = var_1816_cast_fp16)[name = string("op_1821_cast_fp16")];
            tensor<int32, [1]> var_1822_axes_0 = const()[name = string("op_1822_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_1822_cast_fp16 = squeeze(axes = var_1822_axes_0, x = var_1821_cast_fp16)[name = string("op_1822_cast_fp16")];
            string var_1823_mode_0 = const()[name = string("op_1823_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_1823_cast_fp16 = gelu(mode = var_1823_mode_0, x = var_1822_cast_fp16)[name = string("op_1823_cast_fp16")];
            tensor<int32, [1]> var_1827_axes_0 = const()[name = string("op_1827_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_1827_cast_fp16 = expand_dims(axes = var_1827_axes_0, x = var_1823_cast_fp16)[name = string("op_1827_cast_fp16")];
            string var_1832_pad_type_0 = const()[name = string("op_1832_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1832_strides_0 = const()[name = string("op_1832_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1832_pad_0 = const()[name = string("op_1832_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1832_dilations_0 = const()[name = string("op_1832_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1832_groups_0 = const()[name = string("op_1832_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_9_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_9_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11015680)))];
            tensor<fp16, [1, 384, 96, 1]> var_1832_cast_fp16 = conv(bias = embedding_transformer_9_mlp_c_proj_bias_to_fp16, dilations = var_1832_dilations_0, groups = var_1832_groups_0, pad = var_1832_pad_0, pad_type = var_1832_pad_type_0, strides = var_1832_strides_0, weight = embedding_transformer_9_mlp_c_proj_weight_palettized_cast_fp16, x = var_1827_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [1]> var_1833_axes_0 = const()[name = string("op_1833_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_1833_cast_fp16 = squeeze(axes = var_1833_axes_0, x = var_1832_cast_fp16)[name = string("op_1833_cast_fp16")];
            tensor<int32, [3]> var_1834_perm_0 = const()[name = string("op_1834_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_1834_cast_fp16 = transpose(perm = var_1834_perm_0, x = var_1833_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 96, 384]> var_1835_cast_fp16 = add(x = inputs_40_cast_fp16, y = var_1834_cast_fp16)[name = string("op_1835_cast_fp16")];
            tensor<int32, [3]> var_1842_perm_0 = const()[name = string("op_1842_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_42_axes_0 = const()[name = string("inputs_42_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1842_cast_fp16 = transpose(perm = var_1842_perm_0, x = var_1835_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 384, 1, 96]> inputs_42_cast_fp16 = expand_dims(axes = inputs_42_axes_0, x = var_1842_cast_fp16)[name = string("inputs_42_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1850_to_fp16 = const()[name = string("op_1850_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1850_to_fp16, x = inputs_42_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [384]> out0_43_gamma_0_to_fp16 = const()[name = string("out0_43_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11016512)))];
            tensor<fp16, [384]> out0_43_beta_0_to_fp16 = const()[name = string("out0_43_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11017344)))];
            fp16 out0_43_epsilon_0_to_fp16 = const()[name = string("out0_43_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_43_cast_fp16 = batch_norm(beta = out0_43_beta_0_to_fp16, epsilon = out0_43_epsilon_0_to_fp16, gamma = out0_43_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_43_cast_fp16)[name = string("out0_43_cast_fp16")];
            tensor<int32, [1]> var_1860_axes_0 = const()[name = string("op_1860_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1860_cast_fp16 = squeeze(axes = var_1860_axes_0, x = out0_43_cast_fp16)[name = string("op_1860_cast_fp16")];
            tensor<int32, [1]> hidden_states_45_axes_0 = const()[name = string("hidden_states_45_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_45_cast_fp16 = expand_dims(axes = hidden_states_45_axes_0, x = var_1860_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            string var_1874_pad_type_0 = const()[name = string("op_1874_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1874_strides_0 = const()[name = string("op_1874_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1874_pad_0 = const()[name = string("op_1874_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1874_dilations_0 = const()[name = string("op_1874_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1874_groups_0 = const()[name = string("op_1874_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_10_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_10_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11018176)))];
            tensor<fp16, [1, 384, 1, 96]> var_1874_cast_fp16 = conv(bias = embedding_transformer_10_attn_query_bias_to_fp16, dilations = var_1874_dilations_0, groups = var_1874_groups_0, pad = var_1874_pad_0, pad_type = var_1874_pad_type_0, strides = var_1874_strides_0, weight = embedding_transformer_10_attn_query_weight_palettized_cast_fp16, x = hidden_states_45_cast_fp16)[name = string("op_1874_cast_fp16")];
            string k_22_pad_type_0 = const()[name = string("k_22_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_22_strides_0 = const()[name = string("k_22_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_22_pad_0 = const()[name = string("k_22_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_22_dilations_0 = const()[name = string("k_22_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_22_groups_0 = const()[name = string("k_22_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_10_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_10_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11019008)))];
            tensor<fp16, [1, 384, 1, 96]> k_22_cast_fp16 = conv(bias = embedding_transformer_10_attn_key_bias_to_fp16, dilations = k_22_dilations_0, groups = k_22_groups_0, pad = k_22_pad_0, pad_type = k_22_pad_type_0, strides = k_22_strides_0, weight = embedding_transformer_10_attn_key_weight_palettized_cast_fp16, x = hidden_states_45_cast_fp16)[name = string("k_22_cast_fp16")];
            string var_1888_pad_type_0 = const()[name = string("op_1888_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1888_strides_0 = const()[name = string("op_1888_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1888_pad_0 = const()[name = string("op_1888_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1888_dilations_0 = const()[name = string("op_1888_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1888_groups_0 = const()[name = string("op_1888_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_10_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_10_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11019840)))];
            tensor<fp16, [1, 384, 1, 96]> var_1888_cast_fp16 = conv(bias = embedding_transformer_10_attn_value_bias_to_fp16, dilations = var_1888_dilations_0, groups = var_1888_groups_0, pad = var_1888_pad_0, pad_type = var_1888_pad_type_0, strides = var_1888_strides_0, weight = embedding_transformer_10_attn_value_weight_palettized_cast_fp16, x = hidden_states_45_cast_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [6]> tile_30 = const()[name = string("tile_30"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1889_axis_0 = const()[name = string("op_1889_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1889_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1889_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1889_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1889_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1889_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1889_cast_fp16_5 = split(axis = var_1889_axis_0, split_sizes = tile_30, x = var_1874_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [4]> var_1896_perm_0 = const()[name = string("op_1896_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_31 = const()[name = string("tile_31"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1897_axis_0 = const()[name = string("op_1897_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_1896_cast_fp16 = transpose(perm = var_1896_perm_0, x = k_22_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 96, 1, 64]> var_1897_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_1897_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_1897_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_1897_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_1897_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_1897_cast_fp16_5 = split(axis = var_1897_axis_0, split_sizes = tile_31, x = var_1896_cast_fp16)[name = string("op_1897_cast_fp16")];
            tensor<int32, [6]> tile_32 = const()[name = string("tile_32"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_1904_axis_0 = const()[name = string("op_1904_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_1904_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_1904_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_1904_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_1904_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_1904_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_1904_cast_fp16_5 = split(axis = var_1904_axis_0, split_sizes = tile_32, x = var_1888_cast_fp16)[name = string("op_1904_cast_fp16")];
            string var_1912_equation_0 = const()[name = string("op_1912_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1912_cast_fp16 = einsum(equation = var_1912_equation_0, values = (var_1897_cast_fp16_0, var_1889_cast_fp16_0))[name = string("op_1912_cast_fp16")];
            fp16 var_1913_to_fp16 = const()[name = string("op_1913_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_22_cast_fp16 = mul(x = var_1912_cast_fp16, y = var_1913_to_fp16)[name = string("aw_22_cast_fp16")];
            string var_1916_equation_0 = const()[name = string("op_1916_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1916_cast_fp16 = einsum(equation = var_1916_equation_0, values = (var_1897_cast_fp16_1, var_1889_cast_fp16_1))[name = string("op_1916_cast_fp16")];
            fp16 var_1917_to_fp16 = const()[name = string("op_1917_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_22_cast_fp16 = mul(x = var_1916_cast_fp16, y = var_1917_to_fp16)[name = string("aw0_22_cast_fp16")];
            string var_1920_equation_0 = const()[name = string("op_1920_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1920_cast_fp16 = einsum(equation = var_1920_equation_0, values = (var_1897_cast_fp16_2, var_1889_cast_fp16_2))[name = string("op_1920_cast_fp16")];
            fp16 var_1921_to_fp16 = const()[name = string("op_1921_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_22_cast_fp16 = mul(x = var_1920_cast_fp16, y = var_1921_to_fp16)[name = string("aw1_22_cast_fp16")];
            string var_1924_equation_0 = const()[name = string("op_1924_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1924_cast_fp16 = einsum(equation = var_1924_equation_0, values = (var_1897_cast_fp16_3, var_1889_cast_fp16_3))[name = string("op_1924_cast_fp16")];
            fp16 var_1925_to_fp16 = const()[name = string("op_1925_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_22_cast_fp16 = mul(x = var_1924_cast_fp16, y = var_1925_to_fp16)[name = string("aw2_22_cast_fp16")];
            string var_1928_equation_0 = const()[name = string("op_1928_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1928_cast_fp16 = einsum(equation = var_1928_equation_0, values = (var_1897_cast_fp16_4, var_1889_cast_fp16_4))[name = string("op_1928_cast_fp16")];
            fp16 var_1929_to_fp16 = const()[name = string("op_1929_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_22_cast_fp16 = mul(x = var_1928_cast_fp16, y = var_1929_to_fp16)[name = string("aw3_22_cast_fp16")];
            string var_1932_equation_0 = const()[name = string("op_1932_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_1932_cast_fp16 = einsum(equation = var_1932_equation_0, values = (var_1897_cast_fp16_5, var_1889_cast_fp16_5))[name = string("op_1932_cast_fp16")];
            fp16 var_1933_to_fp16 = const()[name = string("op_1933_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_22_cast_fp16 = mul(x = var_1932_cast_fp16, y = var_1933_to_fp16)[name = string("aw4_22_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1935_cast_fp16 = softmax(axis = var_29, x = aw_22_cast_fp16)[name = string("op_1935_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1936_cast_fp16 = softmax(axis = var_29, x = aw0_22_cast_fp16)[name = string("op_1936_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1937_cast_fp16 = softmax(axis = var_29, x = aw1_22_cast_fp16)[name = string("op_1937_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1938_cast_fp16 = softmax(axis = var_29, x = aw2_22_cast_fp16)[name = string("op_1938_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1939_cast_fp16 = softmax(axis = var_29, x = aw3_22_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_1940_cast_fp16 = softmax(axis = var_29, x = aw4_22_cast_fp16)[name = string("op_1940_cast_fp16")];
            string var_1942_equation_0 = const()[name = string("op_1942_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1942_cast_fp16 = einsum(equation = var_1942_equation_0, values = (var_1904_cast_fp16_0, var_1935_cast_fp16))[name = string("op_1942_cast_fp16")];
            string var_1944_equation_0 = const()[name = string("op_1944_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1944_cast_fp16 = einsum(equation = var_1944_equation_0, values = (var_1904_cast_fp16_1, var_1936_cast_fp16))[name = string("op_1944_cast_fp16")];
            string var_1946_equation_0 = const()[name = string("op_1946_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1946_cast_fp16 = einsum(equation = var_1946_equation_0, values = (var_1904_cast_fp16_2, var_1937_cast_fp16))[name = string("op_1946_cast_fp16")];
            string var_1948_equation_0 = const()[name = string("op_1948_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1948_cast_fp16 = einsum(equation = var_1948_equation_0, values = (var_1904_cast_fp16_3, var_1938_cast_fp16))[name = string("op_1948_cast_fp16")];
            string var_1950_equation_0 = const()[name = string("op_1950_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1950_cast_fp16 = einsum(equation = var_1950_equation_0, values = (var_1904_cast_fp16_4, var_1939_cast_fp16))[name = string("op_1950_cast_fp16")];
            string var_1952_equation_0 = const()[name = string("op_1952_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_1952_cast_fp16 = einsum(equation = var_1952_equation_0, values = (var_1904_cast_fp16_5, var_1940_cast_fp16))[name = string("op_1952_cast_fp16")];
            bool input_65_interleave_0 = const()[name = string("input_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_65_cast_fp16 = concat(axis = var_29, interleave = input_65_interleave_0, values = (var_1942_cast_fp16, var_1944_cast_fp16, var_1946_cast_fp16, var_1948_cast_fp16, var_1950_cast_fp16, var_1952_cast_fp16))[name = string("input_65_cast_fp16")];
            string attn_45_pad_type_0 = const()[name = string("attn_45_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_45_strides_0 = const()[name = string("attn_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_45_pad_0 = const()[name = string("attn_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_45_dilations_0 = const()[name = string("attn_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_45_groups_0 = const()[name = string("attn_45_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_10_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_10_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11020672)))];
            tensor<fp16, [1, 384, 1, 96]> attn_45_cast_fp16 = conv(bias = embedding_transformer_10_attn_out_proj_bias_to_fp16, dilations = attn_45_dilations_0, groups = attn_45_groups_0, pad = attn_45_pad_0, pad_type = attn_45_pad_type_0, strides = attn_45_strides_0, weight = embedding_transformer_10_attn_out_proj_weight_palettized_cast_fp16, x = input_65_cast_fp16)[name = string("attn_45_cast_fp16")];
            tensor<int32, [1]> var_1962_axes_0 = const()[name = string("op_1962_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1962_cast_fp16 = squeeze(axes = var_1962_axes_0, x = attn_45_cast_fp16)[name = string("op_1962_cast_fp16")];
            tensor<int32, [3]> var_1963_perm_0 = const()[name = string("op_1963_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_1963_cast_fp16 = transpose(perm = var_1963_perm_0, x = var_1962_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 96, 384]> inputs_44_cast_fp16 = add(x = var_1835_cast_fp16, y = var_1963_cast_fp16)[name = string("inputs_44_cast_fp16")];
            tensor<int32, [3]> var_1967_perm_0 = const()[name = string("op_1967_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_23_axes_0 = const()[name = string("inputs0_23_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1967_cast_fp16 = transpose(perm = var_1967_perm_0, x = inputs_44_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_23_cast_fp16 = expand_dims(axes = inputs0_23_axes_0, x = var_1967_cast_fp16)[name = string("inputs0_23_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_1975_to_fp16 = const()[name = string("op_1975_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1975_to_fp16, x = inputs0_23_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [384]> out0_45_gamma_0_to_fp16 = const()[name = string("out0_45_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11021504)))];
            tensor<fp16, [384]> out0_45_beta_0_to_fp16 = const()[name = string("out0_45_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11022336)))];
            fp16 out0_45_epsilon_0_to_fp16 = const()[name = string("out0_45_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_45_cast_fp16 = batch_norm(beta = out0_45_beta_0_to_fp16, epsilon = out0_45_epsilon_0_to_fp16, gamma = out0_45_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_45_cast_fp16)[name = string("out0_45_cast_fp16")];
            tensor<int32, [1]> var_1985_axes_0 = const()[name = string("op_1985_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_1985_cast_fp16 = squeeze(axes = var_1985_axes_0, x = out0_45_cast_fp16)[name = string("op_1985_cast_fp16")];
            tensor<int32, [3]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_1992_axes_0 = const()[name = string("op_1992_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = var_1985_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 384, 96, 1]> var_1992_cast_fp16 = expand_dims(axes = var_1992_axes_0, x = transpose_12_cast_fp16)[name = string("op_1992_cast_fp16")];
            string var_1997_pad_type_0 = const()[name = string("op_1997_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1997_strides_0 = const()[name = string("op_1997_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1997_pad_0 = const()[name = string("op_1997_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1997_dilations_0 = const()[name = string("op_1997_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1997_groups_0 = const()[name = string("op_1997_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_10_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_10_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11023168)))];
            tensor<fp16, [1, 1536, 96, 1]> var_1997_cast_fp16 = conv(bias = embedding_transformer_10_mlp_c_fc_bias_to_fp16, dilations = var_1997_dilations_0, groups = var_1997_groups_0, pad = var_1997_pad_0, pad_type = var_1997_pad_type_0, strides = var_1997_strides_0, weight = embedding_transformer_10_mlp_c_fc_weight_palettized_cast_fp16, x = var_1992_cast_fp16)[name = string("op_1997_cast_fp16")];
            tensor<int32, [1]> var_1998_axes_0 = const()[name = string("op_1998_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_1998_cast_fp16 = squeeze(axes = var_1998_axes_0, x = var_1997_cast_fp16)[name = string("op_1998_cast_fp16")];
            string var_1999_mode_0 = const()[name = string("op_1999_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_1999_cast_fp16 = gelu(mode = var_1999_mode_0, x = var_1998_cast_fp16)[name = string("op_1999_cast_fp16")];
            tensor<int32, [1]> var_2003_axes_0 = const()[name = string("op_2003_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_2003_cast_fp16 = expand_dims(axes = var_2003_axes_0, x = var_1999_cast_fp16)[name = string("op_2003_cast_fp16")];
            string var_2008_pad_type_0 = const()[name = string("op_2008_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2008_strides_0 = const()[name = string("op_2008_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2008_pad_0 = const()[name = string("op_2008_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2008_dilations_0 = const()[name = string("op_2008_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2008_groups_0 = const()[name = string("op_2008_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_10_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_10_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11026304)))];
            tensor<fp16, [1, 384, 96, 1]> var_2008_cast_fp16 = conv(bias = embedding_transformer_10_mlp_c_proj_bias_to_fp16, dilations = var_2008_dilations_0, groups = var_2008_groups_0, pad = var_2008_pad_0, pad_type = var_2008_pad_type_0, strides = var_2008_strides_0, weight = embedding_transformer_10_mlp_c_proj_weight_palettized_cast_fp16, x = var_2003_cast_fp16)[name = string("op_2008_cast_fp16")];
            tensor<int32, [1]> var_2009_axes_0 = const()[name = string("op_2009_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_2009_cast_fp16 = squeeze(axes = var_2009_axes_0, x = var_2008_cast_fp16)[name = string("op_2009_cast_fp16")];
            tensor<int32, [3]> var_2010_perm_0 = const()[name = string("op_2010_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_2010_cast_fp16 = transpose(perm = var_2010_perm_0, x = var_2009_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 96, 384]> var_2011_cast_fp16 = add(x = inputs_44_cast_fp16, y = var_2010_cast_fp16)[name = string("op_2011_cast_fp16")];
            tensor<int32, [3]> var_2018_perm_0 = const()[name = string("op_2018_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_46_axes_0 = const()[name = string("inputs_46_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_2018_cast_fp16 = transpose(perm = var_2018_perm_0, x = var_2011_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 384, 1, 96]> inputs_46_cast_fp16 = expand_dims(axes = inputs_46_axes_0, x = var_2018_cast_fp16)[name = string("inputs_46_cast_fp16")];
            tensor<int32, [1]> out_2_axes_0 = const()[name = string("out_2_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2026_to_fp16 = const()[name = string("op_2026_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_2_cast_fp16 = layer_norm(axes = out_2_axes_0, epsilon = var_2026_to_fp16, x = inputs_46_cast_fp16)[name = string("out_2_cast_fp16")];
            tensor<fp16, [384]> out0_2_gamma_0_to_fp16 = const()[name = string("out0_2_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11027136)))];
            tensor<fp16, [384]> out0_2_beta_0_to_fp16 = const()[name = string("out0_2_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11027968)))];
            fp16 out0_2_epsilon_0_to_fp16 = const()[name = string("out0_2_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_2_cast_fp16 = batch_norm(beta = out0_2_beta_0_to_fp16, epsilon = out0_2_epsilon_0_to_fp16, gamma = out0_2_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_2_cast_fp16)[name = string("out0_2_cast_fp16")];
            tensor<int32, [1]> var_2036_axes_0 = const()[name = string("op_2036_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_2036_cast_fp16 = squeeze(axes = var_2036_axes_0, x = out0_2_cast_fp16)[name = string("op_2036_cast_fp16")];
            tensor<int32, [1]> hidden_states_1_axes_0 = const()[name = string("hidden_states_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 1, 96]> hidden_states_1_cast_fp16 = expand_dims(axes = hidden_states_1_axes_0, x = var_2036_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            string var_2050_pad_type_0 = const()[name = string("op_2050_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2050_strides_0 = const()[name = string("op_2050_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2050_pad_0 = const()[name = string("op_2050_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2050_dilations_0 = const()[name = string("op_2050_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2050_groups_0 = const()[name = string("op_2050_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_11_attn_query_bias_to_fp16 = const()[name = string("embedding_transformer_11_attn_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11028800)))];
            tensor<fp16, [1, 384, 1, 96]> var_2050_cast_fp16 = conv(bias = embedding_transformer_11_attn_query_bias_to_fp16, dilations = var_2050_dilations_0, groups = var_2050_groups_0, pad = var_2050_pad_0, pad_type = var_2050_pad_type_0, strides = var_2050_strides_0, weight = embedding_transformer_11_attn_query_weight_palettized_cast_fp16, x = hidden_states_1_cast_fp16)[name = string("op_2050_cast_fp16")];
            string k_1_pad_type_0 = const()[name = string("k_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_1_strides_0 = const()[name = string("k_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_1_pad_0 = const()[name = string("k_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_1_dilations_0 = const()[name = string("k_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_1_groups_0 = const()[name = string("k_1_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_11_attn_key_bias_to_fp16 = const()[name = string("embedding_transformer_11_attn_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11029632)))];
            tensor<fp16, [1, 384, 1, 96]> k_1_cast_fp16 = conv(bias = embedding_transformer_11_attn_key_bias_to_fp16, dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = embedding_transformer_11_attn_key_weight_palettized_cast_fp16, x = hidden_states_1_cast_fp16)[name = string("k_1_cast_fp16")];
            string var_2064_pad_type_0 = const()[name = string("op_2064_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2064_strides_0 = const()[name = string("op_2064_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2064_pad_0 = const()[name = string("op_2064_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2064_dilations_0 = const()[name = string("op_2064_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2064_groups_0 = const()[name = string("op_2064_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_11_attn_value_bias_to_fp16 = const()[name = string("embedding_transformer_11_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11030464)))];
            tensor<fp16, [1, 384, 1, 96]> var_2064_cast_fp16 = conv(bias = embedding_transformer_11_attn_value_bias_to_fp16, dilations = var_2064_dilations_0, groups = var_2064_groups_0, pad = var_2064_pad_0, pad_type = var_2064_pad_type_0, strides = var_2064_strides_0, weight = embedding_transformer_11_attn_value_weight_palettized_cast_fp16, x = hidden_states_1_cast_fp16)[name = string("op_2064_cast_fp16")];
            tensor<int32, [6]> tile_33 = const()[name = string("tile_33"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_2065_axis_0 = const()[name = string("op_2065_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_2065_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_2065_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_2065_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_2065_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_2065_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_2065_cast_fp16_5 = split(axis = var_2065_axis_0, split_sizes = tile_33, x = var_2050_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<int32, [4]> var_2072_perm_0 = const()[name = string("op_2072_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [6]> tile_34 = const()[name = string("tile_34"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_2073_axis_0 = const()[name = string("op_2073_axis_0"), val = int32(3)];
            tensor<fp16, [1, 96, 1, 384]> var_2072_cast_fp16 = transpose(perm = var_2072_perm_0, x = k_1_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 96, 1, 64]> var_2073_cast_fp16_0, tensor<fp16, [1, 96, 1, 64]> var_2073_cast_fp16_1, tensor<fp16, [1, 96, 1, 64]> var_2073_cast_fp16_2, tensor<fp16, [1, 96, 1, 64]> var_2073_cast_fp16_3, tensor<fp16, [1, 96, 1, 64]> var_2073_cast_fp16_4, tensor<fp16, [1, 96, 1, 64]> var_2073_cast_fp16_5 = split(axis = var_2073_axis_0, split_sizes = tile_34, x = var_2072_cast_fp16)[name = string("op_2073_cast_fp16")];
            tensor<int32, [6]> tile_35 = const()[name = string("tile_35"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
            int32 var_2080_axis_0 = const()[name = string("op_2080_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 96]> var_2080_cast_fp16_0, tensor<fp16, [1, 64, 1, 96]> var_2080_cast_fp16_1, tensor<fp16, [1, 64, 1, 96]> var_2080_cast_fp16_2, tensor<fp16, [1, 64, 1, 96]> var_2080_cast_fp16_3, tensor<fp16, [1, 64, 1, 96]> var_2080_cast_fp16_4, tensor<fp16, [1, 64, 1, 96]> var_2080_cast_fp16_5 = split(axis = var_2080_axis_0, split_sizes = tile_35, x = var_2064_cast_fp16)[name = string("op_2080_cast_fp16")];
            string var_2088_equation_0 = const()[name = string("op_2088_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_2088_cast_fp16 = einsum(equation = var_2088_equation_0, values = (var_2073_cast_fp16_0, var_2065_cast_fp16_0))[name = string("op_2088_cast_fp16")];
            fp16 var_2089_to_fp16 = const()[name = string("op_2089_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw_1_cast_fp16 = mul(x = var_2088_cast_fp16, y = var_2089_to_fp16)[name = string("aw_1_cast_fp16")];
            string var_2092_equation_0 = const()[name = string("op_2092_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_2092_cast_fp16 = einsum(equation = var_2092_equation_0, values = (var_2073_cast_fp16_1, var_2065_cast_fp16_1))[name = string("op_2092_cast_fp16")];
            fp16 var_2093_to_fp16 = const()[name = string("op_2093_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw0_1_cast_fp16 = mul(x = var_2092_cast_fp16, y = var_2093_to_fp16)[name = string("aw0_1_cast_fp16")];
            string var_2096_equation_0 = const()[name = string("op_2096_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_2096_cast_fp16 = einsum(equation = var_2096_equation_0, values = (var_2073_cast_fp16_2, var_2065_cast_fp16_2))[name = string("op_2096_cast_fp16")];
            fp16 var_2097_to_fp16 = const()[name = string("op_2097_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw1_1_cast_fp16 = mul(x = var_2096_cast_fp16, y = var_2097_to_fp16)[name = string("aw1_1_cast_fp16")];
            string var_2100_equation_0 = const()[name = string("op_2100_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_2100_cast_fp16 = einsum(equation = var_2100_equation_0, values = (var_2073_cast_fp16_3, var_2065_cast_fp16_3))[name = string("op_2100_cast_fp16")];
            fp16 var_2101_to_fp16 = const()[name = string("op_2101_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw2_1_cast_fp16 = mul(x = var_2100_cast_fp16, y = var_2101_to_fp16)[name = string("aw2_1_cast_fp16")];
            string var_2104_equation_0 = const()[name = string("op_2104_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_2104_cast_fp16 = einsum(equation = var_2104_equation_0, values = (var_2073_cast_fp16_4, var_2065_cast_fp16_4))[name = string("op_2104_cast_fp16")];
            fp16 var_2105_to_fp16 = const()[name = string("op_2105_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw3_1_cast_fp16 = mul(x = var_2104_cast_fp16, y = var_2105_to_fp16)[name = string("aw3_1_cast_fp16")];
            string var_2108_equation_0 = const()[name = string("op_2108_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 96, 1, 96]> var_2108_cast_fp16 = einsum(equation = var_2108_equation_0, values = (var_2073_cast_fp16_5, var_2065_cast_fp16_5))[name = string("op_2108_cast_fp16")];
            fp16 var_2109_to_fp16 = const()[name = string("op_2109_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 96, 1, 96]> aw4_1_cast_fp16 = mul(x = var_2108_cast_fp16, y = var_2109_to_fp16)[name = string("aw4_1_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_2111_cast_fp16 = softmax(axis = var_29, x = aw_1_cast_fp16)[name = string("op_2111_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_2112_cast_fp16 = softmax(axis = var_29, x = aw0_1_cast_fp16)[name = string("op_2112_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_2113_cast_fp16 = softmax(axis = var_29, x = aw1_1_cast_fp16)[name = string("op_2113_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_2114_cast_fp16 = softmax(axis = var_29, x = aw2_1_cast_fp16)[name = string("op_2114_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_2115_cast_fp16 = softmax(axis = var_29, x = aw3_1_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<fp16, [1, 96, 1, 96]> var_2116_cast_fp16 = softmax(axis = var_29, x = aw4_1_cast_fp16)[name = string("op_2116_cast_fp16")];
            string var_2118_equation_0 = const()[name = string("op_2118_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_2118_cast_fp16 = einsum(equation = var_2118_equation_0, values = (var_2080_cast_fp16_0, var_2111_cast_fp16))[name = string("op_2118_cast_fp16")];
            string var_2120_equation_0 = const()[name = string("op_2120_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_2120_cast_fp16 = einsum(equation = var_2120_equation_0, values = (var_2080_cast_fp16_1, var_2112_cast_fp16))[name = string("op_2120_cast_fp16")];
            string var_2122_equation_0 = const()[name = string("op_2122_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_2122_cast_fp16 = einsum(equation = var_2122_equation_0, values = (var_2080_cast_fp16_2, var_2113_cast_fp16))[name = string("op_2122_cast_fp16")];
            string var_2124_equation_0 = const()[name = string("op_2124_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_2124_cast_fp16 = einsum(equation = var_2124_equation_0, values = (var_2080_cast_fp16_3, var_2114_cast_fp16))[name = string("op_2124_cast_fp16")];
            string var_2126_equation_0 = const()[name = string("op_2126_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_2126_cast_fp16 = einsum(equation = var_2126_equation_0, values = (var_2080_cast_fp16_4, var_2115_cast_fp16))[name = string("op_2126_cast_fp16")];
            string var_2128_equation_0 = const()[name = string("op_2128_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 96]> var_2128_cast_fp16 = einsum(equation = var_2128_equation_0, values = (var_2080_cast_fp16_5, var_2116_cast_fp16))[name = string("op_2128_cast_fp16")];
            bool input_4_interleave_0 = const()[name = string("input_4_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 384, 1, 96]> input_4_cast_fp16 = concat(axis = var_29, interleave = input_4_interleave_0, values = (var_2118_cast_fp16, var_2120_cast_fp16, var_2122_cast_fp16, var_2124_cast_fp16, var_2126_cast_fp16, var_2128_cast_fp16))[name = string("input_4_cast_fp16")];
            string attn_1_pad_type_0 = const()[name = string("attn_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> attn_1_strides_0 = const()[name = string("attn_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> attn_1_pad_0 = const()[name = string("attn_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> attn_1_dilations_0 = const()[name = string("attn_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 attn_1_groups_0 = const()[name = string("attn_1_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_11_attn_out_proj_bias_to_fp16 = const()[name = string("embedding_transformer_11_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11031296)))];
            tensor<fp16, [1, 384, 1, 96]> attn_1_cast_fp16 = conv(bias = embedding_transformer_11_attn_out_proj_bias_to_fp16, dilations = attn_1_dilations_0, groups = attn_1_groups_0, pad = attn_1_pad_0, pad_type = attn_1_pad_type_0, strides = attn_1_strides_0, weight = embedding_transformer_11_attn_out_proj_weight_palettized_cast_fp16, x = input_4_cast_fp16)[name = string("attn_1_cast_fp16")];
            tensor<int32, [1]> var_2138_axes_0 = const()[name = string("op_2138_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_2138_cast_fp16 = squeeze(axes = var_2138_axes_0, x = attn_1_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [3]> var_2139_perm_0 = const()[name = string("op_2139_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 96, 384]> var_2139_cast_fp16 = transpose(perm = var_2139_perm_0, x = var_2138_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 96, 384]> inputs_2_cast_fp16 = add(x = var_2011_cast_fp16, y = var_2139_cast_fp16)[name = string("inputs_2_cast_fp16")];
            tensor<int32, [3]> var_2143_perm_0 = const()[name = string("op_2143_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs0_1_axes_0 = const()[name = string("inputs0_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_2143_cast_fp16 = transpose(perm = var_2143_perm_0, x = inputs_2_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 384, 1, 96]> inputs0_1_cast_fp16 = expand_dims(axes = inputs0_1_axes_0, x = var_2143_cast_fp16)[name = string("inputs0_1_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2151_to_fp16 = const()[name = string("op_2151_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_2151_to_fp16, x = inputs0_1_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [384]> out0_47_gamma_0_to_fp16 = const()[name = string("out0_47_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11032128)))];
            tensor<fp16, [384]> out0_47_beta_0_to_fp16 = const()[name = string("out0_47_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11032960)))];
            fp16 out0_47_epsilon_0_to_fp16 = const()[name = string("out0_47_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_47_cast_fp16 = batch_norm(beta = out0_47_beta_0_to_fp16, epsilon = out0_47_epsilon_0_to_fp16, gamma = out0_47_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_47_cast_fp16)[name = string("out0_47_cast_fp16")];
            tensor<int32, [1]> var_2161_axes_0 = const()[name = string("op_2161_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_2161_cast_fp16 = squeeze(axes = var_2161_axes_0, x = out0_47_cast_fp16)[name = string("op_2161_cast_fp16")];
            tensor<int32, [3]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_2168_axes_0 = const()[name = string("op_2168_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = var_2161_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 384, 96, 1]> var_2168_cast_fp16 = expand_dims(axes = var_2168_axes_0, x = transpose_13_cast_fp16)[name = string("op_2168_cast_fp16")];
            string var_2173_pad_type_0 = const()[name = string("op_2173_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2173_strides_0 = const()[name = string("op_2173_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2173_pad_0 = const()[name = string("op_2173_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2173_dilations_0 = const()[name = string("op_2173_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2173_groups_0 = const()[name = string("op_2173_groups_0"), val = int32(1)];
            tensor<fp16, [1536]> embedding_transformer_11_mlp_c_fc_bias_to_fp16 = const()[name = string("embedding_transformer_11_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11033792)))];
            tensor<fp16, [1, 1536, 96, 1]> var_2173_cast_fp16 = conv(bias = embedding_transformer_11_mlp_c_fc_bias_to_fp16, dilations = var_2173_dilations_0, groups = var_2173_groups_0, pad = var_2173_pad_0, pad_type = var_2173_pad_type_0, strides = var_2173_strides_0, weight = embedding_transformer_11_mlp_c_fc_weight_palettized_cast_fp16, x = var_2168_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [1]> var_2174_axes_0 = const()[name = string("op_2174_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1536, 96, 1]> var_2174_cast_fp16 = squeeze(axes = var_2174_axes_0, x = var_2173_cast_fp16)[name = string("op_2174_cast_fp16")];
            string var_2175_mode_0 = const()[name = string("op_2175_mode_0"), val = string("EXACT")];
            tensor<fp16, [1536, 96, 1]> var_2175_cast_fp16 = gelu(mode = var_2175_mode_0, x = var_2174_cast_fp16)[name = string("op_2175_cast_fp16")];
            tensor<int32, [1]> var_2179_axes_0 = const()[name = string("op_2179_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1536, 96, 1]> var_2179_cast_fp16 = expand_dims(axes = var_2179_axes_0, x = var_2175_cast_fp16)[name = string("op_2179_cast_fp16")];
            string var_2184_pad_type_0 = const()[name = string("op_2184_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2184_strides_0 = const()[name = string("op_2184_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2184_pad_0 = const()[name = string("op_2184_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2184_dilations_0 = const()[name = string("op_2184_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2184_groups_0 = const()[name = string("op_2184_groups_0"), val = int32(1)];
            tensor<fp16, [384]> embedding_transformer_11_mlp_c_proj_bias_to_fp16 = const()[name = string("embedding_transformer_11_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11036928)))];
            tensor<fp16, [1, 384, 96, 1]> var_2184_cast_fp16 = conv(bias = embedding_transformer_11_mlp_c_proj_bias_to_fp16, dilations = var_2184_dilations_0, groups = var_2184_groups_0, pad = var_2184_pad_0, pad_type = var_2184_pad_type_0, strides = var_2184_strides_0, weight = embedding_transformer_11_mlp_c_proj_weight_palettized_cast_fp16, x = var_2179_cast_fp16)[name = string("op_2184_cast_fp16")];
            tensor<int32, [1]> var_2185_axes_0 = const()[name = string("op_2185_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [384, 96, 1]> var_2185_cast_fp16 = squeeze(axes = var_2185_axes_0, x = var_2184_cast_fp16)[name = string("op_2185_cast_fp16")];
            tensor<int32, [3]> var_2186_perm_0 = const()[name = string("op_2186_perm_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<fp16, [1, 96, 384]> var_2186_cast_fp16 = transpose(perm = var_2186_perm_0, x = var_2185_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 96, 384]> var_2187_cast_fp16 = add(x = inputs_2_cast_fp16, y = var_2186_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [3]> var_2190_perm_0 = const()[name = string("op_2190_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> inputs_1_axes_0 = const()[name = string("inputs_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_2190_cast_fp16 = transpose(perm = var_2190_perm_0, x = var_2187_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 384, 1, 96]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_2190_cast_fp16)[name = string("inputs_1_cast_fp16")];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 var_2198_to_fp16 = const()[name = string("op_2198_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_2198_to_fp16, x = inputs_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [384]> out0_1_gamma_0_to_fp16 = const()[name = string("out0_1_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11037760)))];
            tensor<fp16, [384]> out0_1_beta_0_to_fp16 = const()[name = string("out0_1_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11038592)))];
            fp16 out0_1_epsilon_0_to_fp16 = const()[name = string("out0_1_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 384, 1, 96]> out0_1_cast_fp16 = batch_norm(beta = out0_1_beta_0_to_fp16, epsilon = out0_1_epsilon_0_to_fp16, gamma = out0_1_gamma_0_to_fp16, mean = out0_3_mean_0_to_fp16, variance = out0_3_variance_0_to_fp16, x = out_1_cast_fp16)[name = string("out0_1_cast_fp16")];
            tensor<int32, [1]> var_2208_axes_0 = const()[name = string("op_2208_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 384, 96]> var_2208_cast_fp16 = squeeze(axes = var_2208_axes_0, x = out0_1_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<int32, [3]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_6_axes_0 = const()[name = string("input_6_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_6_keep_dims_0 = const()[name = string("input_6_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 96, 384]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = var_2208_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 384]> input_6_cast_fp16 = reduce_mean(axes = input_6_axes_0, keep_dims = input_6_keep_dims_0, x = x_1_cast_fp16)[name = string("input_6_cast_fp16")];
            tensor<fp16, [512]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039424)))];
            tensor<fp16, [1, 512]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = audio_projs_0_weight_palettized_cast_fp16, x = input_6_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [1]> var_2216 = const()[name = string("op_2216"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> var_2217_cast_fp16 = reduce_l2_norm(axes = var_2216, keep_dims = var_2, x = linear_0_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 512]> embedding = real_div(x = linear_0_cast_fp16, y = var_2217_cast_fp16)[name = string("op_2218_cast_fp16")];
        } -> (embedding);
}