program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios18>(tensor<fp32, [1, 300, 560]> speech) {
            int32 var_28 = const()[name = string("op_28"), val = int32(1)];
            int32 var_29 = const()[name = string("op_29"), val = int32(-1)];
            bool speech_interleave_0 = const()[name = string("speech_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 560]> textnorm_query_to_fp16 = const()[name = string("textnorm_query_to_fp16"), val = tensor<fp16, [1, 1, 560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string speech_to_fp16_dtype_0 = const()[name = string("speech_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 300, 560]> speech_to_fp16 = cast(dtype = speech_to_fp16_dtype_0, x = speech)[name = string("cast_363")];
            tensor<fp16, [1, 301, 560]> speech_cast_fp16 = concat(axis = var_28, interleave = speech_interleave_0, values = (textnorm_query_to_fp16, speech_to_fp16))[name = string("speech_cast_fp16")];
            bool xs_pad_interleave_0 = const()[name = string("xs_pad_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 560]> language_query_to_fp16 = const()[name = string("language_query_to_fp16"), val = tensor<fp16, [1, 1, 560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1280)))];
            tensor<fp16, [1, 2, 560]> var_53_to_fp16 = const()[name = string("op_53_to_fp16"), val = tensor<fp16, [1, 2, 560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2496)))];
            tensor<fp16, [1, 304, 560]> xs_pad_cast_fp16 = concat(axis = var_28, interleave = xs_pad_interleave_0, values = (language_query_to_fp16, var_53_to_fp16, speech_cast_fp16))[name = string("xs_pad_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.6ap+4)];
            tensor<fp16, [1, 304, 560]> x_1_cast_fp16 = mul(x = xs_pad_cast_fp16, y = var_11_to_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 304, 560]> const_19_to_fp16 = const()[name = string("const_19_to_fp16"), val = tensor<fp16, [1, 304, 560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4800)))];
            tensor<fp16, [1, 304, 560]> input_7_cast_fp16 = add(x = x_1_cast_fp16, y = const_19_to_fp16)[name = string("input_7_cast_fp16")];
            tensor<int32, [1]> output_1_axes_0 = const()[name = string("output_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [560]> const_20_to_fp16 = const()[name = string("const_20_to_fp16"), val = tensor<fp16, [560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345344)))];
            tensor<fp16, [560]> const_21_to_fp16 = const()[name = string("const_21_to_fp16"), val = tensor<fp16, [560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346560)))];
            fp16 var_22_to_fp16 = const()[name = string("op_22_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 304, 560]> output_1_cast_fp16 = layer_norm(axes = output_1_axes_0, beta = const_21_to_fp16, epsilon = var_22_to_fp16, gamma = const_20_to_fp16, x = input_7_cast_fp16)[name = string("output_1_cast_fp16")];
            tensor<fp16, [1536, 560]> m_encoder_encoders0_0_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders0_0_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347776)))];
            tensor<fp16, [1536]> m_encoder_encoders0_0_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders0_0_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2068160)))];
            tensor<fp16, [1, 304, 1536]> linear_0_cast_fp16 = linear(bias = m_encoder_encoders0_0_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders0_0_self_attn_linear_q_k_v_weight_to_fp16, x = output_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_277_axis_0 = const()[name = string("op_277_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_277_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_277_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_277_cast_fp16_2 = split(axis = var_277_axis_0, split_sizes = tile_0, x = linear_0_cast_fp16)[name = string("op_277_cast_fp16")];
            tensor<int32, [4]> var_281 = const()[name = string("op_281"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_282_cast_fp16 = reshape(shape = var_281, x = var_277_cast_fp16_0)[name = string("op_282_cast_fp16")];
            tensor<int32, [4]> var_284 = const()[name = string("op_284"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_285_cast_fp16 = reshape(shape = var_284, x = var_277_cast_fp16_1)[name = string("op_285_cast_fp16")];
            tensor<int32, [4]> var_287 = const()[name = string("op_287"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_288_cast_fp16 = reshape(shape = var_287, x = var_277_cast_fp16_2)[name = string("op_288_cast_fp16")];
            tensor<int32, [4]> value_1_perm_0 = const()[name = string("value_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_11_perm_0 = const()[name = string("input_11_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_13_mode_0 = const()[name = string("input_13_mode_0"), val = string("constant")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_11_cast_fp16 = transpose(perm = input_11_perm_0, x = var_277_cast_fp16_2)[name = string("transpose_768")];
            tensor<fp16, [1, 512, 314]> input_13_cast_fp16 = pad(constant_val = const_27_to_fp16, mode = input_13_mode_0, pad = input_13_pad_0, x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
            string x_5_pad_type_0 = const()[name = string("x_5_pad_type_0"), val = string("valid")];
            int32 x_5_groups_0 = const()[name = string("x_5_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_5_strides_0 = const()[name = string("x_5_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_5_pad_0 = const()[name = string("x_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_5_dilations_0 = const()[name = string("x_5_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders0_0_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders0_0_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2071296)))];
            tensor<fp16, [1, 512, 304]> x_5_cast_fp16 = conv(dilations = x_5_dilations_0, groups = x_5_groups_0, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = x_5_strides_0, weight = m_encoder_encoders0_0_self_attn_fsmn_block_weight_to_fp16, x = input_13_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<int32, [3]> x_7_perm_0 = const()[name = string("x_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_7_cast_fp16 = transpose(perm = x_7_perm_0, x = x_5_cast_fp16)[name = string("transpose_767")];
            tensor<fp16, [1, 304, 512]> input_15_cast_fp16 = add(x = x_7_cast_fp16, y = var_277_cast_fp16_2)[name = string("input_15_cast_fp16")];
            fp16 var_307_to_fp16 = const()[name = string("op_307_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_3_cast_fp16 = mul(x = var_282_cast_fp16, y = var_307_to_fp16)[name = string("q_h_3_cast_fp16")];
            bool scores_1_transpose_x_0 = const()[name = string("scores_1_transpose_x_0"), val = bool(false)];
            bool scores_1_transpose_y_0 = const()[name = string("scores_1_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_210_perm_0 = const()[name = string("transpose_210_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_211_perm_0 = const()[name = string("transpose_211_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_211 = transpose(perm = transpose_211_perm_0, x = var_285_cast_fp16)[name = string("transpose_765")];
            tensor<fp16, [1, 4, 304, 128]> transpose_210 = transpose(perm = transpose_210_perm_0, x = q_h_3_cast_fp16)[name = string("transpose_766")];
            tensor<fp16, [1, 4, 304, 304]> scores_1_cast_fp16 = matmul(transpose_x = scores_1_transpose_x_0, transpose_y = scores_1_transpose_y_0, x = transpose_210, y = transpose_211)[name = string("scores_1_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_315_cast_fp16 = softmax(axis = var_29, x = scores_1_cast_fp16)[name = string("op_315_cast_fp16")];
            bool x_11_transpose_x_0 = const()[name = string("x_11_transpose_x_0"), val = bool(false)];
            bool x_11_transpose_y_0 = const()[name = string("x_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_1_cast_fp16 = transpose(perm = value_1_perm_0, x = var_288_cast_fp16)[name = string("transpose_769")];
            tensor<fp16, [1, 4, 304, 128]> x_11_cast_fp16 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = var_315_cast_fp16, y = value_1_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> var_319_perm_0 = const()[name = string("op_319_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_321 = const()[name = string("op_321"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_319_cast_fp16 = transpose(perm = var_319_perm_0, x = x_11_cast_fp16)[name = string("transpose_764")];
            tensor<fp16, [1, 304, 512]> input_19_cast_fp16 = reshape(shape = var_321, x = var_319_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders0_0_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders0_0_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2082624)))];
            tensor<fp16, [512]> m_encoder_encoders0_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders0_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2606976)))];
            tensor<fp16, [1, 304, 512]> linear_1_cast_fp16 = linear(bias = m_encoder_encoders0_0_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders0_0_self_attn_linear_out_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_21_cast_fp16 = add(x = linear_1_cast_fp16, y = input_15_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [1]> output_3_axes_0 = const()[name = string("output_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_29_to_fp16 = const()[name = string("const_29_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2608064)))];
            tensor<fp16, [512]> const_30_to_fp16 = const()[name = string("const_30_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2609152)))];
            tensor<fp16, [1, 304, 512]> output_3_cast_fp16 = layer_norm(axes = output_3_axes_0, beta = const_30_to_fp16, epsilon = var_22_to_fp16, gamma = const_29_to_fp16, x = input_21_cast_fp16)[name = string("output_3_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders0_0_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders0_0_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2610240)))];
            tensor<fp16, [2048]> m_encoder_encoders0_0_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders0_0_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4707456)))];
            tensor<fp16, [1, 304, 2048]> linear_2_cast_fp16 = linear(bias = m_encoder_encoders0_0_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders0_0_feed_forward_w_1_weight_to_fp16, x = output_3_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_31_cast_fp16 = relu(x = linear_2_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders0_0_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders0_0_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4711616)))];
            tensor<fp16, [512]> m_encoder_encoders0_0_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders0_0_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6808832)))];
            tensor<fp16, [1, 304, 512]> linear_3_cast_fp16 = linear(bias = m_encoder_encoders0_0_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders0_0_feed_forward_w_2_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_37_cast_fp16 = add(x = input_21_cast_fp16, y = linear_3_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<int32, [1]> output_5_axes_0 = const()[name = string("output_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_31_to_fp16 = const()[name = string("const_31_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6809920)))];
            tensor<fp16, [512]> const_32_to_fp16 = const()[name = string("const_32_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6811008)))];
            tensor<fp16, [1, 304, 512]> output_5_cast_fp16 = layer_norm(axes = output_5_axes_0, beta = const_32_to_fp16, epsilon = var_22_to_fp16, gamma = const_31_to_fp16, x = input_37_cast_fp16)[name = string("output_5_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_0_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_0_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6812096)))];
            tensor<fp16, [1536]> m_encoder_encoders_0_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_0_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8385024)))];
            tensor<fp16, [1, 304, 1536]> linear_4_cast_fp16 = linear(bias = m_encoder_encoders_0_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_0_self_attn_linear_q_k_v_weight_to_fp16, x = output_5_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_378_axis_0 = const()[name = string("op_378_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_378_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_378_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_378_cast_fp16_2 = split(axis = var_378_axis_0, split_sizes = tile_1, x = linear_4_cast_fp16)[name = string("op_378_cast_fp16")];
            tensor<int32, [4]> var_382 = const()[name = string("op_382"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_383_cast_fp16 = reshape(shape = var_382, x = var_378_cast_fp16_0)[name = string("op_383_cast_fp16")];
            tensor<int32, [4]> var_385 = const()[name = string("op_385"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_386_cast_fp16 = reshape(shape = var_385, x = var_378_cast_fp16_1)[name = string("op_386_cast_fp16")];
            tensor<int32, [4]> var_388 = const()[name = string("op_388"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_389_cast_fp16 = reshape(shape = var_388, x = var_378_cast_fp16_2)[name = string("op_389_cast_fp16")];
            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_41_perm_0 = const()[name = string("input_41_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_43_pad_0 = const()[name = string("input_43_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_43_mode_0 = const()[name = string("input_43_mode_0"), val = string("constant")];
            fp16 const_38_to_fp16 = const()[name = string("const_38_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_41_cast_fp16 = transpose(perm = input_41_perm_0, x = var_378_cast_fp16_2)[name = string("transpose_762")];
            tensor<fp16, [1, 512, 314]> input_43_cast_fp16 = pad(constant_val = const_38_to_fp16, mode = input_43_mode_0, pad = input_43_pad_0, x = input_41_cast_fp16)[name = string("input_43_cast_fp16")];
            string x_15_pad_type_0 = const()[name = string("x_15_pad_type_0"), val = string("valid")];
            int32 x_15_groups_0 = const()[name = string("x_15_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_15_strides_0 = const()[name = string("x_15_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_15_pad_0 = const()[name = string("x_15_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_15_dilations_0 = const()[name = string("x_15_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_0_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_0_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8388160)))];
            tensor<fp16, [1, 512, 304]> x_15_cast_fp16 = conv(dilations = x_15_dilations_0, groups = x_15_groups_0, pad = x_15_pad_0, pad_type = x_15_pad_type_0, strides = x_15_strides_0, weight = m_encoder_encoders_0_self_attn_fsmn_block_weight_to_fp16, x = input_43_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [3]> x_17_perm_0 = const()[name = string("x_17_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_17_cast_fp16 = transpose(perm = x_17_perm_0, x = x_15_cast_fp16)[name = string("transpose_761")];
            tensor<fp16, [1, 304, 512]> input_45_cast_fp16 = add(x = x_17_cast_fp16, y = var_378_cast_fp16_2)[name = string("input_45_cast_fp16")];
            fp16 var_408_to_fp16 = const()[name = string("op_408_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_7_cast_fp16 = mul(x = var_383_cast_fp16, y = var_408_to_fp16)[name = string("q_h_7_cast_fp16")];
            bool scores_5_transpose_x_0 = const()[name = string("scores_5_transpose_x_0"), val = bool(false)];
            bool scores_5_transpose_y_0 = const()[name = string("scores_5_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_212_perm_0 = const()[name = string("transpose_212_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_213_perm_0 = const()[name = string("transpose_213_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_213 = transpose(perm = transpose_213_perm_0, x = var_386_cast_fp16)[name = string("transpose_759")];
            tensor<fp16, [1, 4, 304, 128]> transpose_212 = transpose(perm = transpose_212_perm_0, x = q_h_7_cast_fp16)[name = string("transpose_760")];
            tensor<fp16, [1, 4, 304, 304]> scores_5_cast_fp16 = matmul(transpose_x = scores_5_transpose_x_0, transpose_y = scores_5_transpose_y_0, x = transpose_212, y = transpose_213)[name = string("scores_5_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_416_cast_fp16 = softmax(axis = var_29, x = scores_5_cast_fp16)[name = string("op_416_cast_fp16")];
            bool x_21_transpose_x_0 = const()[name = string("x_21_transpose_x_0"), val = bool(false)];
            bool x_21_transpose_y_0 = const()[name = string("x_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = var_389_cast_fp16)[name = string("transpose_763")];
            tensor<fp16, [1, 4, 304, 128]> x_21_cast_fp16 = matmul(transpose_x = x_21_transpose_x_0, transpose_y = x_21_transpose_y_0, x = var_416_cast_fp16, y = value_3_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> var_420_perm_0 = const()[name = string("op_420_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_422 = const()[name = string("op_422"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_420_cast_fp16 = transpose(perm = var_420_perm_0, x = x_21_cast_fp16)[name = string("transpose_758")];
            tensor<fp16, [1, 304, 512]> input_49_cast_fp16 = reshape(shape = var_422, x = var_420_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_0_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_0_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8399488)))];
            tensor<fp16, [512]> m_encoder_encoders_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8923840)))];
            tensor<fp16, [1, 304, 512]> linear_5_cast_fp16 = linear(bias = m_encoder_encoders_0_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_0_self_attn_linear_out_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_51_cast_fp16 = add(x = linear_5_cast_fp16, y = input_45_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_53_cast_fp16 = add(x = input_37_cast_fp16, y = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<int32, [1]> output_7_axes_0 = const()[name = string("output_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_40_to_fp16 = const()[name = string("const_40_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8924928)))];
            tensor<fp16, [512]> const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8926016)))];
            tensor<fp16, [1, 304, 512]> output_7_cast_fp16 = layer_norm(axes = output_7_axes_0, beta = const_41_to_fp16, epsilon = var_22_to_fp16, gamma = const_40_to_fp16, x = input_53_cast_fp16)[name = string("output_7_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_0_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_0_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8927104)))];
            tensor<fp16, [2048]> m_encoder_encoders_0_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_0_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11024320)))];
            tensor<fp16, [1, 304, 2048]> linear_6_cast_fp16 = linear(bias = m_encoder_encoders_0_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_0_feed_forward_w_1_weight_to_fp16, x = output_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_61_cast_fp16 = relu(x = linear_6_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_0_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_0_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11028480)))];
            tensor<fp16, [512]> m_encoder_encoders_0_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_0_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13125696)))];
            tensor<fp16, [1, 304, 512]> linear_7_cast_fp16 = linear(bias = m_encoder_encoders_0_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_0_feed_forward_w_2_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_67_cast_fp16 = add(x = input_53_cast_fp16, y = linear_7_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<int32, [1]> output_9_axes_0 = const()[name = string("output_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_42_to_fp16 = const()[name = string("const_42_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13126784)))];
            tensor<fp16, [512]> const_43_to_fp16 = const()[name = string("const_43_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13127872)))];
            tensor<fp16, [1, 304, 512]> output_9_cast_fp16 = layer_norm(axes = output_9_axes_0, beta = const_43_to_fp16, epsilon = var_22_to_fp16, gamma = const_42_to_fp16, x = input_67_cast_fp16)[name = string("output_9_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_1_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_1_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13128960)))];
            tensor<fp16, [1536]> m_encoder_encoders_1_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_1_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14701888)))];
            tensor<fp16, [1, 304, 1536]> linear_8_cast_fp16 = linear(bias = m_encoder_encoders_1_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_1_self_attn_linear_q_k_v_weight_to_fp16, x = output_9_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_480_axis_0 = const()[name = string("op_480_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_480_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_480_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_480_cast_fp16_2 = split(axis = var_480_axis_0, split_sizes = tile_2, x = linear_8_cast_fp16)[name = string("op_480_cast_fp16")];
            tensor<int32, [4]> var_484 = const()[name = string("op_484"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_485_cast_fp16 = reshape(shape = var_484, x = var_480_cast_fp16_0)[name = string("op_485_cast_fp16")];
            tensor<int32, [4]> var_487 = const()[name = string("op_487"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_488_cast_fp16 = reshape(shape = var_487, x = var_480_cast_fp16_1)[name = string("op_488_cast_fp16")];
            tensor<int32, [4]> var_490 = const()[name = string("op_490"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_491_cast_fp16 = reshape(shape = var_490, x = var_480_cast_fp16_2)[name = string("op_491_cast_fp16")];
            tensor<int32, [4]> value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_71_perm_0 = const()[name = string("input_71_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_73_pad_0 = const()[name = string("input_73_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_73_mode_0 = const()[name = string("input_73_mode_0"), val = string("constant")];
            fp16 const_49_to_fp16 = const()[name = string("const_49_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_71_cast_fp16 = transpose(perm = input_71_perm_0, x = var_480_cast_fp16_2)[name = string("transpose_756")];
            tensor<fp16, [1, 512, 314]> input_73_cast_fp16 = pad(constant_val = const_49_to_fp16, mode = input_73_mode_0, pad = input_73_pad_0, x = input_71_cast_fp16)[name = string("input_73_cast_fp16")];
            string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")];
            int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_1_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_1_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14705024)))];
            tensor<fp16, [1, 512, 304]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = m_encoder_encoders_1_self_attn_fsmn_block_weight_to_fp16, x = input_73_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<int32, [3]> x_27_perm_0 = const()[name = string("x_27_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_27_cast_fp16 = transpose(perm = x_27_perm_0, x = x_25_cast_fp16)[name = string("transpose_755")];
            tensor<fp16, [1, 304, 512]> input_75_cast_fp16 = add(x = x_27_cast_fp16, y = var_480_cast_fp16_2)[name = string("input_75_cast_fp16")];
            fp16 var_510_to_fp16 = const()[name = string("op_510_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_11_cast_fp16 = mul(x = var_485_cast_fp16, y = var_510_to_fp16)[name = string("q_h_11_cast_fp16")];
            bool scores_9_transpose_x_0 = const()[name = string("scores_9_transpose_x_0"), val = bool(false)];
            bool scores_9_transpose_y_0 = const()[name = string("scores_9_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_214_perm_0 = const()[name = string("transpose_214_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_215_perm_0 = const()[name = string("transpose_215_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_215 = transpose(perm = transpose_215_perm_0, x = var_488_cast_fp16)[name = string("transpose_753")];
            tensor<fp16, [1, 4, 304, 128]> transpose_214 = transpose(perm = transpose_214_perm_0, x = q_h_11_cast_fp16)[name = string("transpose_754")];
            tensor<fp16, [1, 4, 304, 304]> scores_9_cast_fp16 = matmul(transpose_x = scores_9_transpose_x_0, transpose_y = scores_9_transpose_y_0, x = transpose_214, y = transpose_215)[name = string("scores_9_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_518_cast_fp16 = softmax(axis = var_29, x = scores_9_cast_fp16)[name = string("op_518_cast_fp16")];
            bool x_31_transpose_x_0 = const()[name = string("x_31_transpose_x_0"), val = bool(false)];
            bool x_31_transpose_y_0 = const()[name = string("x_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = var_491_cast_fp16)[name = string("transpose_757")];
            tensor<fp16, [1, 4, 304, 128]> x_31_cast_fp16 = matmul(transpose_x = x_31_transpose_x_0, transpose_y = x_31_transpose_y_0, x = var_518_cast_fp16, y = value_5_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> var_522_perm_0 = const()[name = string("op_522_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_524 = const()[name = string("op_524"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_522_cast_fp16 = transpose(perm = var_522_perm_0, x = x_31_cast_fp16)[name = string("transpose_752")];
            tensor<fp16, [1, 304, 512]> input_79_cast_fp16 = reshape(shape = var_524, x = var_522_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_1_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_1_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14716352)))];
            tensor<fp16, [512]> m_encoder_encoders_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15240704)))];
            tensor<fp16, [1, 304, 512]> linear_9_cast_fp16 = linear(bias = m_encoder_encoders_1_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_1_self_attn_linear_out_weight_to_fp16, x = input_79_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_81_cast_fp16 = add(x = linear_9_cast_fp16, y = input_75_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_83_cast_fp16 = add(x = input_67_cast_fp16, y = input_81_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<int32, [1]> output_11_axes_0 = const()[name = string("output_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_51_to_fp16 = const()[name = string("const_51_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15241792)))];
            tensor<fp16, [512]> const_52_to_fp16 = const()[name = string("const_52_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15242880)))];
            tensor<fp16, [1, 304, 512]> output_11_cast_fp16 = layer_norm(axes = output_11_axes_0, beta = const_52_to_fp16, epsilon = var_22_to_fp16, gamma = const_51_to_fp16, x = input_83_cast_fp16)[name = string("output_11_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_1_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_1_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15243968)))];
            tensor<fp16, [2048]> m_encoder_encoders_1_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_1_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17341184)))];
            tensor<fp16, [1, 304, 2048]> linear_10_cast_fp16 = linear(bias = m_encoder_encoders_1_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_1_feed_forward_w_1_weight_to_fp16, x = output_11_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_91_cast_fp16 = relu(x = linear_10_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_1_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_1_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17345344)))];
            tensor<fp16, [512]> m_encoder_encoders_1_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_1_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19442560)))];
            tensor<fp16, [1, 304, 512]> linear_11_cast_fp16 = linear(bias = m_encoder_encoders_1_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_1_feed_forward_w_2_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_97_cast_fp16 = add(x = input_83_cast_fp16, y = linear_11_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<int32, [1]> output_13_axes_0 = const()[name = string("output_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_53_to_fp16 = const()[name = string("const_53_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19443648)))];
            tensor<fp16, [512]> const_54_to_fp16 = const()[name = string("const_54_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19444736)))];
            tensor<fp16, [1, 304, 512]> output_13_cast_fp16 = layer_norm(axes = output_13_axes_0, beta = const_54_to_fp16, epsilon = var_22_to_fp16, gamma = const_53_to_fp16, x = input_97_cast_fp16)[name = string("output_13_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_2_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_2_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19445824)))];
            tensor<fp16, [1536]> m_encoder_encoders_2_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_2_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21018752)))];
            tensor<fp16, [1, 304, 1536]> linear_12_cast_fp16 = linear(bias = m_encoder_encoders_2_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_2_self_attn_linear_q_k_v_weight_to_fp16, x = output_13_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_582_axis_0 = const()[name = string("op_582_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_582_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_582_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_582_cast_fp16_2 = split(axis = var_582_axis_0, split_sizes = tile_3, x = linear_12_cast_fp16)[name = string("op_582_cast_fp16")];
            tensor<int32, [4]> var_586 = const()[name = string("op_586"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_587_cast_fp16 = reshape(shape = var_586, x = var_582_cast_fp16_0)[name = string("op_587_cast_fp16")];
            tensor<int32, [4]> var_589 = const()[name = string("op_589"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_590_cast_fp16 = reshape(shape = var_589, x = var_582_cast_fp16_1)[name = string("op_590_cast_fp16")];
            tensor<int32, [4]> var_592 = const()[name = string("op_592"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_593_cast_fp16 = reshape(shape = var_592, x = var_582_cast_fp16_2)[name = string("op_593_cast_fp16")];
            tensor<int32, [4]> value_7_perm_0 = const()[name = string("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_103_mode_0 = const()[name = string("input_103_mode_0"), val = string("constant")];
            fp16 const_60_to_fp16 = const()[name = string("const_60_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = var_582_cast_fp16_2)[name = string("transpose_750")];
            tensor<fp16, [1, 512, 314]> input_103_cast_fp16 = pad(constant_val = const_60_to_fp16, mode = input_103_mode_0, pad = input_103_pad_0, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
            string x_35_pad_type_0 = const()[name = string("x_35_pad_type_0"), val = string("valid")];
            int32 x_35_groups_0 = const()[name = string("x_35_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_35_strides_0 = const()[name = string("x_35_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_35_dilations_0 = const()[name = string("x_35_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_2_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_2_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21021888)))];
            tensor<fp16, [1, 512, 304]> x_35_cast_fp16 = conv(dilations = x_35_dilations_0, groups = x_35_groups_0, pad = x_35_pad_0, pad_type = x_35_pad_type_0, strides = x_35_strides_0, weight = m_encoder_encoders_2_self_attn_fsmn_block_weight_to_fp16, x = input_103_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> x_37_perm_0 = const()[name = string("x_37_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_37_cast_fp16 = transpose(perm = x_37_perm_0, x = x_35_cast_fp16)[name = string("transpose_749")];
            tensor<fp16, [1, 304, 512]> input_105_cast_fp16 = add(x = x_37_cast_fp16, y = var_582_cast_fp16_2)[name = string("input_105_cast_fp16")];
            fp16 var_612_to_fp16 = const()[name = string("op_612_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_15_cast_fp16 = mul(x = var_587_cast_fp16, y = var_612_to_fp16)[name = string("q_h_15_cast_fp16")];
            bool scores_13_transpose_x_0 = const()[name = string("scores_13_transpose_x_0"), val = bool(false)];
            bool scores_13_transpose_y_0 = const()[name = string("scores_13_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_216_perm_0 = const()[name = string("transpose_216_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_217_perm_0 = const()[name = string("transpose_217_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_217 = transpose(perm = transpose_217_perm_0, x = var_590_cast_fp16)[name = string("transpose_747")];
            tensor<fp16, [1, 4, 304, 128]> transpose_216 = transpose(perm = transpose_216_perm_0, x = q_h_15_cast_fp16)[name = string("transpose_748")];
            tensor<fp16, [1, 4, 304, 304]> scores_13_cast_fp16 = matmul(transpose_x = scores_13_transpose_x_0, transpose_y = scores_13_transpose_y_0, x = transpose_216, y = transpose_217)[name = string("scores_13_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_620_cast_fp16 = softmax(axis = var_29, x = scores_13_cast_fp16)[name = string("op_620_cast_fp16")];
            bool x_41_transpose_x_0 = const()[name = string("x_41_transpose_x_0"), val = bool(false)];
            bool x_41_transpose_y_0 = const()[name = string("x_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = var_593_cast_fp16)[name = string("transpose_751")];
            tensor<fp16, [1, 4, 304, 128]> x_41_cast_fp16 = matmul(transpose_x = x_41_transpose_x_0, transpose_y = x_41_transpose_y_0, x = var_620_cast_fp16, y = value_7_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> var_624_perm_0 = const()[name = string("op_624_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_626 = const()[name = string("op_626"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_624_cast_fp16 = transpose(perm = var_624_perm_0, x = x_41_cast_fp16)[name = string("transpose_746")];
            tensor<fp16, [1, 304, 512]> input_109_cast_fp16 = reshape(shape = var_626, x = var_624_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_2_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_2_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21033216)))];
            tensor<fp16, [512]> m_encoder_encoders_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21557568)))];
            tensor<fp16, [1, 304, 512]> linear_13_cast_fp16 = linear(bias = m_encoder_encoders_2_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_2_self_attn_linear_out_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_111_cast_fp16 = add(x = linear_13_cast_fp16, y = input_105_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_113_cast_fp16 = add(x = input_97_cast_fp16, y = input_111_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<int32, [1]> output_15_axes_0 = const()[name = string("output_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_62_to_fp16 = const()[name = string("const_62_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21558656)))];
            tensor<fp16, [512]> const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21559744)))];
            tensor<fp16, [1, 304, 512]> output_15_cast_fp16 = layer_norm(axes = output_15_axes_0, beta = const_63_to_fp16, epsilon = var_22_to_fp16, gamma = const_62_to_fp16, x = input_113_cast_fp16)[name = string("output_15_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_2_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_2_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21560832)))];
            tensor<fp16, [2048]> m_encoder_encoders_2_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_2_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23658048)))];
            tensor<fp16, [1, 304, 2048]> linear_14_cast_fp16 = linear(bias = m_encoder_encoders_2_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_2_feed_forward_w_1_weight_to_fp16, x = output_15_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_121_cast_fp16 = relu(x = linear_14_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_2_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_2_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23662208)))];
            tensor<fp16, [512]> m_encoder_encoders_2_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_2_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25759424)))];
            tensor<fp16, [1, 304, 512]> linear_15_cast_fp16 = linear(bias = m_encoder_encoders_2_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_2_feed_forward_w_2_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_127_cast_fp16 = add(x = input_113_cast_fp16, y = linear_15_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<int32, [1]> output_17_axes_0 = const()[name = string("output_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25760512)))];
            tensor<fp16, [512]> const_65_to_fp16 = const()[name = string("const_65_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25761600)))];
            tensor<fp16, [1, 304, 512]> output_17_cast_fp16 = layer_norm(axes = output_17_axes_0, beta = const_65_to_fp16, epsilon = var_22_to_fp16, gamma = const_64_to_fp16, x = input_127_cast_fp16)[name = string("output_17_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_3_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_3_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25762688)))];
            tensor<fp16, [1536]> m_encoder_encoders_3_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_3_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27335616)))];
            tensor<fp16, [1, 304, 1536]> linear_16_cast_fp16 = linear(bias = m_encoder_encoders_3_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_3_self_attn_linear_q_k_v_weight_to_fp16, x = output_17_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [3]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_684_axis_0 = const()[name = string("op_684_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_684_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_684_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_684_cast_fp16_2 = split(axis = var_684_axis_0, split_sizes = tile_4, x = linear_16_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [4]> var_688 = const()[name = string("op_688"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_689_cast_fp16 = reshape(shape = var_688, x = var_684_cast_fp16_0)[name = string("op_689_cast_fp16")];
            tensor<int32, [4]> var_691 = const()[name = string("op_691"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_692_cast_fp16 = reshape(shape = var_691, x = var_684_cast_fp16_1)[name = string("op_692_cast_fp16")];
            tensor<int32, [4]> var_694 = const()[name = string("op_694"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_695_cast_fp16 = reshape(shape = var_694, x = var_684_cast_fp16_2)[name = string("op_695_cast_fp16")];
            tensor<int32, [4]> value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_131_perm_0 = const()[name = string("input_131_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_133_pad_0 = const()[name = string("input_133_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_133_mode_0 = const()[name = string("input_133_mode_0"), val = string("constant")];
            fp16 const_71_to_fp16 = const()[name = string("const_71_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_131_cast_fp16 = transpose(perm = input_131_perm_0, x = var_684_cast_fp16_2)[name = string("transpose_744")];
            tensor<fp16, [1, 512, 314]> input_133_cast_fp16 = pad(constant_val = const_71_to_fp16, mode = input_133_mode_0, pad = input_133_pad_0, x = input_131_cast_fp16)[name = string("input_133_cast_fp16")];
            string x_45_pad_type_0 = const()[name = string("x_45_pad_type_0"), val = string("valid")];
            int32 x_45_groups_0 = const()[name = string("x_45_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_45_strides_0 = const()[name = string("x_45_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_45_pad_0 = const()[name = string("x_45_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_45_dilations_0 = const()[name = string("x_45_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_3_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_3_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27338752)))];
            tensor<fp16, [1, 512, 304]> x_45_cast_fp16 = conv(dilations = x_45_dilations_0, groups = x_45_groups_0, pad = x_45_pad_0, pad_type = x_45_pad_type_0, strides = x_45_strides_0, weight = m_encoder_encoders_3_self_attn_fsmn_block_weight_to_fp16, x = input_133_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<int32, [3]> x_47_perm_0 = const()[name = string("x_47_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_47_cast_fp16 = transpose(perm = x_47_perm_0, x = x_45_cast_fp16)[name = string("transpose_743")];
            tensor<fp16, [1, 304, 512]> input_135_cast_fp16 = add(x = x_47_cast_fp16, y = var_684_cast_fp16_2)[name = string("input_135_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_19_cast_fp16 = mul(x = var_689_cast_fp16, y = var_714_to_fp16)[name = string("q_h_19_cast_fp16")];
            bool scores_17_transpose_x_0 = const()[name = string("scores_17_transpose_x_0"), val = bool(false)];
            bool scores_17_transpose_y_0 = const()[name = string("scores_17_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_218_perm_0 = const()[name = string("transpose_218_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_219_perm_0 = const()[name = string("transpose_219_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_219 = transpose(perm = transpose_219_perm_0, x = var_692_cast_fp16)[name = string("transpose_741")];
            tensor<fp16, [1, 4, 304, 128]> transpose_218 = transpose(perm = transpose_218_perm_0, x = q_h_19_cast_fp16)[name = string("transpose_742")];
            tensor<fp16, [1, 4, 304, 304]> scores_17_cast_fp16 = matmul(transpose_x = scores_17_transpose_x_0, transpose_y = scores_17_transpose_y_0, x = transpose_218, y = transpose_219)[name = string("scores_17_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_722_cast_fp16 = softmax(axis = var_29, x = scores_17_cast_fp16)[name = string("op_722_cast_fp16")];
            bool x_51_transpose_x_0 = const()[name = string("x_51_transpose_x_0"), val = bool(false)];
            bool x_51_transpose_y_0 = const()[name = string("x_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = var_695_cast_fp16)[name = string("transpose_745")];
            tensor<fp16, [1, 4, 304, 128]> x_51_cast_fp16 = matmul(transpose_x = x_51_transpose_x_0, transpose_y = x_51_transpose_y_0, x = var_722_cast_fp16, y = value_9_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> var_726_perm_0 = const()[name = string("op_726_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_728 = const()[name = string("op_728"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_726_cast_fp16 = transpose(perm = var_726_perm_0, x = x_51_cast_fp16)[name = string("transpose_740")];
            tensor<fp16, [1, 304, 512]> input_139_cast_fp16 = reshape(shape = var_728, x = var_726_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_3_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_3_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27350080)))];
            tensor<fp16, [512]> m_encoder_encoders_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27874432)))];
            tensor<fp16, [1, 304, 512]> linear_17_cast_fp16 = linear(bias = m_encoder_encoders_3_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_3_self_attn_linear_out_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_141_cast_fp16 = add(x = linear_17_cast_fp16, y = input_135_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_143_cast_fp16 = add(x = input_127_cast_fp16, y = input_141_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<int32, [1]> output_19_axes_0 = const()[name = string("output_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_73_to_fp16 = const()[name = string("const_73_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27875520)))];
            tensor<fp16, [512]> const_74_to_fp16 = const()[name = string("const_74_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27876608)))];
            tensor<fp16, [1, 304, 512]> output_19_cast_fp16 = layer_norm(axes = output_19_axes_0, beta = const_74_to_fp16, epsilon = var_22_to_fp16, gamma = const_73_to_fp16, x = input_143_cast_fp16)[name = string("output_19_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_3_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_3_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27877696)))];
            tensor<fp16, [2048]> m_encoder_encoders_3_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_3_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29974912)))];
            tensor<fp16, [1, 304, 2048]> linear_18_cast_fp16 = linear(bias = m_encoder_encoders_3_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_3_feed_forward_w_1_weight_to_fp16, x = output_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_151_cast_fp16 = relu(x = linear_18_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_3_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_3_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29979072)))];
            tensor<fp16, [512]> m_encoder_encoders_3_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_3_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32076288)))];
            tensor<fp16, [1, 304, 512]> linear_19_cast_fp16 = linear(bias = m_encoder_encoders_3_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_3_feed_forward_w_2_weight_to_fp16, x = input_151_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_157_cast_fp16 = add(x = input_143_cast_fp16, y = linear_19_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<int32, [1]> output_21_axes_0 = const()[name = string("output_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_75_to_fp16 = const()[name = string("const_75_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32077376)))];
            tensor<fp16, [512]> const_76_to_fp16 = const()[name = string("const_76_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32078464)))];
            tensor<fp16, [1, 304, 512]> output_21_cast_fp16 = layer_norm(axes = output_21_axes_0, beta = const_76_to_fp16, epsilon = var_22_to_fp16, gamma = const_75_to_fp16, x = input_157_cast_fp16)[name = string("output_21_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_4_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_4_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32079552)))];
            tensor<fp16, [1536]> m_encoder_encoders_4_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_4_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33652480)))];
            tensor<fp16, [1, 304, 1536]> linear_20_cast_fp16 = linear(bias = m_encoder_encoders_4_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_4_self_attn_linear_q_k_v_weight_to_fp16, x = output_21_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_786_axis_0 = const()[name = string("op_786_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_786_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_786_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_786_cast_fp16_2 = split(axis = var_786_axis_0, split_sizes = tile_5, x = linear_20_cast_fp16)[name = string("op_786_cast_fp16")];
            tensor<int32, [4]> var_790 = const()[name = string("op_790"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_791_cast_fp16 = reshape(shape = var_790, x = var_786_cast_fp16_0)[name = string("op_791_cast_fp16")];
            tensor<int32, [4]> var_793 = const()[name = string("op_793"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_794_cast_fp16 = reshape(shape = var_793, x = var_786_cast_fp16_1)[name = string("op_794_cast_fp16")];
            tensor<int32, [4]> var_796 = const()[name = string("op_796"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_797_cast_fp16 = reshape(shape = var_796, x = var_786_cast_fp16_2)[name = string("op_797_cast_fp16")];
            tensor<int32, [4]> value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_163_pad_0 = const()[name = string("input_163_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_163_mode_0 = const()[name = string("input_163_mode_0"), val = string("constant")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = var_786_cast_fp16_2)[name = string("transpose_738")];
            tensor<fp16, [1, 512, 314]> input_163_cast_fp16 = pad(constant_val = const_82_to_fp16, mode = input_163_mode_0, pad = input_163_pad_0, x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
            string x_55_pad_type_0 = const()[name = string("x_55_pad_type_0"), val = string("valid")];
            int32 x_55_groups_0 = const()[name = string("x_55_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_55_strides_0 = const()[name = string("x_55_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_55_pad_0 = const()[name = string("x_55_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_55_dilations_0 = const()[name = string("x_55_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_4_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_4_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33655616)))];
            tensor<fp16, [1, 512, 304]> x_55_cast_fp16 = conv(dilations = x_55_dilations_0, groups = x_55_groups_0, pad = x_55_pad_0, pad_type = x_55_pad_type_0, strides = x_55_strides_0, weight = m_encoder_encoders_4_self_attn_fsmn_block_weight_to_fp16, x = input_163_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<int32, [3]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = x_55_cast_fp16)[name = string("transpose_737")];
            tensor<fp16, [1, 304, 512]> input_165_cast_fp16 = add(x = x_57_cast_fp16, y = var_786_cast_fp16_2)[name = string("input_165_cast_fp16")];
            fp16 var_816_to_fp16 = const()[name = string("op_816_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_23_cast_fp16 = mul(x = var_791_cast_fp16, y = var_816_to_fp16)[name = string("q_h_23_cast_fp16")];
            bool scores_21_transpose_x_0 = const()[name = string("scores_21_transpose_x_0"), val = bool(false)];
            bool scores_21_transpose_y_0 = const()[name = string("scores_21_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_220_perm_0 = const()[name = string("transpose_220_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_221_perm_0 = const()[name = string("transpose_221_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_221 = transpose(perm = transpose_221_perm_0, x = var_794_cast_fp16)[name = string("transpose_735")];
            tensor<fp16, [1, 4, 304, 128]> transpose_220 = transpose(perm = transpose_220_perm_0, x = q_h_23_cast_fp16)[name = string("transpose_736")];
            tensor<fp16, [1, 4, 304, 304]> scores_21_cast_fp16 = matmul(transpose_x = scores_21_transpose_x_0, transpose_y = scores_21_transpose_y_0, x = transpose_220, y = transpose_221)[name = string("scores_21_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_824_cast_fp16 = softmax(axis = var_29, x = scores_21_cast_fp16)[name = string("op_824_cast_fp16")];
            bool x_61_transpose_x_0 = const()[name = string("x_61_transpose_x_0"), val = bool(false)];
            bool x_61_transpose_y_0 = const()[name = string("x_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = var_797_cast_fp16)[name = string("transpose_739")];
            tensor<fp16, [1, 4, 304, 128]> x_61_cast_fp16 = matmul(transpose_x = x_61_transpose_x_0, transpose_y = x_61_transpose_y_0, x = var_824_cast_fp16, y = value_11_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> var_828_perm_0 = const()[name = string("op_828_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_830 = const()[name = string("op_830"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_828_cast_fp16 = transpose(perm = var_828_perm_0, x = x_61_cast_fp16)[name = string("transpose_734")];
            tensor<fp16, [1, 304, 512]> input_169_cast_fp16 = reshape(shape = var_830, x = var_828_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_4_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_4_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33666944)))];
            tensor<fp16, [512]> m_encoder_encoders_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34191296)))];
            tensor<fp16, [1, 304, 512]> linear_21_cast_fp16 = linear(bias = m_encoder_encoders_4_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_4_self_attn_linear_out_weight_to_fp16, x = input_169_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_171_cast_fp16 = add(x = linear_21_cast_fp16, y = input_165_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_173_cast_fp16 = add(x = input_157_cast_fp16, y = input_171_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<int32, [1]> output_23_axes_0 = const()[name = string("output_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_84_to_fp16 = const()[name = string("const_84_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34192384)))];
            tensor<fp16, [512]> const_85_to_fp16 = const()[name = string("const_85_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34193472)))];
            tensor<fp16, [1, 304, 512]> output_23_cast_fp16 = layer_norm(axes = output_23_axes_0, beta = const_85_to_fp16, epsilon = var_22_to_fp16, gamma = const_84_to_fp16, x = input_173_cast_fp16)[name = string("output_23_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_4_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_4_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34194560)))];
            tensor<fp16, [2048]> m_encoder_encoders_4_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_4_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36291776)))];
            tensor<fp16, [1, 304, 2048]> linear_22_cast_fp16 = linear(bias = m_encoder_encoders_4_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_4_feed_forward_w_1_weight_to_fp16, x = output_23_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_181_cast_fp16 = relu(x = linear_22_cast_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_4_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_4_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36295936)))];
            tensor<fp16, [512]> m_encoder_encoders_4_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_4_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38393152)))];
            tensor<fp16, [1, 304, 512]> linear_23_cast_fp16 = linear(bias = m_encoder_encoders_4_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_4_feed_forward_w_2_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_187_cast_fp16 = add(x = input_173_cast_fp16, y = linear_23_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<int32, [1]> output_25_axes_0 = const()[name = string("output_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_86_to_fp16 = const()[name = string("const_86_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38394240)))];
            tensor<fp16, [512]> const_87_to_fp16 = const()[name = string("const_87_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38395328)))];
            tensor<fp16, [1, 304, 512]> output_25_cast_fp16 = layer_norm(axes = output_25_axes_0, beta = const_87_to_fp16, epsilon = var_22_to_fp16, gamma = const_86_to_fp16, x = input_187_cast_fp16)[name = string("output_25_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_5_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_5_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38396416)))];
            tensor<fp16, [1536]> m_encoder_encoders_5_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_5_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39969344)))];
            tensor<fp16, [1, 304, 1536]> linear_24_cast_fp16 = linear(bias = m_encoder_encoders_5_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_5_self_attn_linear_q_k_v_weight_to_fp16, x = output_25_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_888_axis_0 = const()[name = string("op_888_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_888_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_888_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_888_cast_fp16_2 = split(axis = var_888_axis_0, split_sizes = tile_6, x = linear_24_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [4]> var_892 = const()[name = string("op_892"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_893_cast_fp16 = reshape(shape = var_892, x = var_888_cast_fp16_0)[name = string("op_893_cast_fp16")];
            tensor<int32, [4]> var_895 = const()[name = string("op_895"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_896_cast_fp16 = reshape(shape = var_895, x = var_888_cast_fp16_1)[name = string("op_896_cast_fp16")];
            tensor<int32, [4]> var_898 = const()[name = string("op_898"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_899_cast_fp16 = reshape(shape = var_898, x = var_888_cast_fp16_2)[name = string("op_899_cast_fp16")];
            tensor<int32, [4]> value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_191_perm_0 = const()[name = string("input_191_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_193_pad_0 = const()[name = string("input_193_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_193_mode_0 = const()[name = string("input_193_mode_0"), val = string("constant")];
            fp16 const_93_to_fp16 = const()[name = string("const_93_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_191_cast_fp16 = transpose(perm = input_191_perm_0, x = var_888_cast_fp16_2)[name = string("transpose_732")];
            tensor<fp16, [1, 512, 314]> input_193_cast_fp16 = pad(constant_val = const_93_to_fp16, mode = input_193_mode_0, pad = input_193_pad_0, x = input_191_cast_fp16)[name = string("input_193_cast_fp16")];
            string x_65_pad_type_0 = const()[name = string("x_65_pad_type_0"), val = string("valid")];
            int32 x_65_groups_0 = const()[name = string("x_65_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_65_strides_0 = const()[name = string("x_65_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_65_pad_0 = const()[name = string("x_65_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_65_dilations_0 = const()[name = string("x_65_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_5_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_5_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39972480)))];
            tensor<fp16, [1, 512, 304]> x_65_cast_fp16 = conv(dilations = x_65_dilations_0, groups = x_65_groups_0, pad = x_65_pad_0, pad_type = x_65_pad_type_0, strides = x_65_strides_0, weight = m_encoder_encoders_5_self_attn_fsmn_block_weight_to_fp16, x = input_193_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<int32, [3]> x_67_perm_0 = const()[name = string("x_67_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_67_cast_fp16 = transpose(perm = x_67_perm_0, x = x_65_cast_fp16)[name = string("transpose_731")];
            tensor<fp16, [1, 304, 512]> input_195_cast_fp16 = add(x = x_67_cast_fp16, y = var_888_cast_fp16_2)[name = string("input_195_cast_fp16")];
            fp16 var_918_to_fp16 = const()[name = string("op_918_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_27_cast_fp16 = mul(x = var_893_cast_fp16, y = var_918_to_fp16)[name = string("q_h_27_cast_fp16")];
            bool scores_25_transpose_x_0 = const()[name = string("scores_25_transpose_x_0"), val = bool(false)];
            bool scores_25_transpose_y_0 = const()[name = string("scores_25_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_222_perm_0 = const()[name = string("transpose_222_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_223_perm_0 = const()[name = string("transpose_223_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_223 = transpose(perm = transpose_223_perm_0, x = var_896_cast_fp16)[name = string("transpose_729")];
            tensor<fp16, [1, 4, 304, 128]> transpose_222 = transpose(perm = transpose_222_perm_0, x = q_h_27_cast_fp16)[name = string("transpose_730")];
            tensor<fp16, [1, 4, 304, 304]> scores_25_cast_fp16 = matmul(transpose_x = scores_25_transpose_x_0, transpose_y = scores_25_transpose_y_0, x = transpose_222, y = transpose_223)[name = string("scores_25_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_926_cast_fp16 = softmax(axis = var_29, x = scores_25_cast_fp16)[name = string("op_926_cast_fp16")];
            bool x_71_transpose_x_0 = const()[name = string("x_71_transpose_x_0"), val = bool(false)];
            bool x_71_transpose_y_0 = const()[name = string("x_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = var_899_cast_fp16)[name = string("transpose_733")];
            tensor<fp16, [1, 4, 304, 128]> x_71_cast_fp16 = matmul(transpose_x = x_71_transpose_x_0, transpose_y = x_71_transpose_y_0, x = var_926_cast_fp16, y = value_13_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> var_930_perm_0 = const()[name = string("op_930_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_932 = const()[name = string("op_932"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_930_cast_fp16 = transpose(perm = var_930_perm_0, x = x_71_cast_fp16)[name = string("transpose_728")];
            tensor<fp16, [1, 304, 512]> input_199_cast_fp16 = reshape(shape = var_932, x = var_930_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_5_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_5_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39983808)))];
            tensor<fp16, [512]> m_encoder_encoders_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40508160)))];
            tensor<fp16, [1, 304, 512]> linear_25_cast_fp16 = linear(bias = m_encoder_encoders_5_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_5_self_attn_linear_out_weight_to_fp16, x = input_199_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_201_cast_fp16 = add(x = linear_25_cast_fp16, y = input_195_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_203_cast_fp16 = add(x = input_187_cast_fp16, y = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<int32, [1]> output_27_axes_0 = const()[name = string("output_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_95_to_fp16 = const()[name = string("const_95_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40509248)))];
            tensor<fp16, [512]> const_96_to_fp16 = const()[name = string("const_96_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40510336)))];
            tensor<fp16, [1, 304, 512]> output_27_cast_fp16 = layer_norm(axes = output_27_axes_0, beta = const_96_to_fp16, epsilon = var_22_to_fp16, gamma = const_95_to_fp16, x = input_203_cast_fp16)[name = string("output_27_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_5_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_5_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40511424)))];
            tensor<fp16, [2048]> m_encoder_encoders_5_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_5_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42608640)))];
            tensor<fp16, [1, 304, 2048]> linear_26_cast_fp16 = linear(bias = m_encoder_encoders_5_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_5_feed_forward_w_1_weight_to_fp16, x = output_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_211_cast_fp16 = relu(x = linear_26_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_5_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_5_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42612800)))];
            tensor<fp16, [512]> m_encoder_encoders_5_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_5_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44710016)))];
            tensor<fp16, [1, 304, 512]> linear_27_cast_fp16 = linear(bias = m_encoder_encoders_5_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_5_feed_forward_w_2_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_217_cast_fp16 = add(x = input_203_cast_fp16, y = linear_27_cast_fp16)[name = string("input_217_cast_fp16")];
            tensor<int32, [1]> output_29_axes_0 = const()[name = string("output_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_97_to_fp16 = const()[name = string("const_97_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44711104)))];
            tensor<fp16, [512]> const_98_to_fp16 = const()[name = string("const_98_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44712192)))];
            tensor<fp16, [1, 304, 512]> output_29_cast_fp16 = layer_norm(axes = output_29_axes_0, beta = const_98_to_fp16, epsilon = var_22_to_fp16, gamma = const_97_to_fp16, x = input_217_cast_fp16)[name = string("output_29_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_6_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_6_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44713280)))];
            tensor<fp16, [1536]> m_encoder_encoders_6_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_6_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46286208)))];
            tensor<fp16, [1, 304, 1536]> linear_28_cast_fp16 = linear(bias = m_encoder_encoders_6_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_6_self_attn_linear_q_k_v_weight_to_fp16, x = output_29_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_990_axis_0 = const()[name = string("op_990_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_990_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_990_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_990_cast_fp16_2 = split(axis = var_990_axis_0, split_sizes = tile_7, x = linear_28_cast_fp16)[name = string("op_990_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_995_cast_fp16 = reshape(shape = var_994, x = var_990_cast_fp16_0)[name = string("op_995_cast_fp16")];
            tensor<int32, [4]> var_997 = const()[name = string("op_997"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_998_cast_fp16 = reshape(shape = var_997, x = var_990_cast_fp16_1)[name = string("op_998_cast_fp16")];
            tensor<int32, [4]> var_1000 = const()[name = string("op_1000"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1001_cast_fp16 = reshape(shape = var_1000, x = var_990_cast_fp16_2)[name = string("op_1001_cast_fp16")];
            tensor<int32, [4]> value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_221_perm_0 = const()[name = string("input_221_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_223_pad_0 = const()[name = string("input_223_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_223_mode_0 = const()[name = string("input_223_mode_0"), val = string("constant")];
            fp16 const_104_to_fp16 = const()[name = string("const_104_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_221_cast_fp16 = transpose(perm = input_221_perm_0, x = var_990_cast_fp16_2)[name = string("transpose_726")];
            tensor<fp16, [1, 512, 314]> input_223_cast_fp16 = pad(constant_val = const_104_to_fp16, mode = input_223_mode_0, pad = input_223_pad_0, x = input_221_cast_fp16)[name = string("input_223_cast_fp16")];
            string x_75_pad_type_0 = const()[name = string("x_75_pad_type_0"), val = string("valid")];
            int32 x_75_groups_0 = const()[name = string("x_75_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_75_strides_0 = const()[name = string("x_75_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_75_pad_0 = const()[name = string("x_75_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_75_dilations_0 = const()[name = string("x_75_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_6_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_6_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46289344)))];
            tensor<fp16, [1, 512, 304]> x_75_cast_fp16 = conv(dilations = x_75_dilations_0, groups = x_75_groups_0, pad = x_75_pad_0, pad_type = x_75_pad_type_0, strides = x_75_strides_0, weight = m_encoder_encoders_6_self_attn_fsmn_block_weight_to_fp16, x = input_223_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<int32, [3]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = x_75_cast_fp16)[name = string("transpose_725")];
            tensor<fp16, [1, 304, 512]> input_225_cast_fp16 = add(x = x_77_cast_fp16, y = var_990_cast_fp16_2)[name = string("input_225_cast_fp16")];
            fp16 var_1020_to_fp16 = const()[name = string("op_1020_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_31_cast_fp16 = mul(x = var_995_cast_fp16, y = var_1020_to_fp16)[name = string("q_h_31_cast_fp16")];
            bool scores_29_transpose_x_0 = const()[name = string("scores_29_transpose_x_0"), val = bool(false)];
            bool scores_29_transpose_y_0 = const()[name = string("scores_29_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_224_perm_0 = const()[name = string("transpose_224_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_225_perm_0 = const()[name = string("transpose_225_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_225 = transpose(perm = transpose_225_perm_0, x = var_998_cast_fp16)[name = string("transpose_723")];
            tensor<fp16, [1, 4, 304, 128]> transpose_224 = transpose(perm = transpose_224_perm_0, x = q_h_31_cast_fp16)[name = string("transpose_724")];
            tensor<fp16, [1, 4, 304, 304]> scores_29_cast_fp16 = matmul(transpose_x = scores_29_transpose_x_0, transpose_y = scores_29_transpose_y_0, x = transpose_224, y = transpose_225)[name = string("scores_29_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1028_cast_fp16 = softmax(axis = var_29, x = scores_29_cast_fp16)[name = string("op_1028_cast_fp16")];
            bool x_81_transpose_x_0 = const()[name = string("x_81_transpose_x_0"), val = bool(false)];
            bool x_81_transpose_y_0 = const()[name = string("x_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = var_1001_cast_fp16)[name = string("transpose_727")];
            tensor<fp16, [1, 4, 304, 128]> x_81_cast_fp16 = matmul(transpose_x = x_81_transpose_x_0, transpose_y = x_81_transpose_y_0, x = var_1028_cast_fp16, y = value_15_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> var_1032_perm_0 = const()[name = string("op_1032_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1034 = const()[name = string("op_1034"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1032_cast_fp16 = transpose(perm = var_1032_perm_0, x = x_81_cast_fp16)[name = string("transpose_722")];
            tensor<fp16, [1, 304, 512]> input_229_cast_fp16 = reshape(shape = var_1034, x = var_1032_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_6_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_6_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46300672)))];
            tensor<fp16, [512]> m_encoder_encoders_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_6_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46825024)))];
            tensor<fp16, [1, 304, 512]> linear_29_cast_fp16 = linear(bias = m_encoder_encoders_6_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_6_self_attn_linear_out_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_231_cast_fp16 = add(x = linear_29_cast_fp16, y = input_225_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_233_cast_fp16 = add(x = input_217_cast_fp16, y = input_231_cast_fp16)[name = string("input_233_cast_fp16")];
            tensor<int32, [1]> output_31_axes_0 = const()[name = string("output_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_106_to_fp16 = const()[name = string("const_106_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46826112)))];
            tensor<fp16, [512]> const_107_to_fp16 = const()[name = string("const_107_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46827200)))];
            tensor<fp16, [1, 304, 512]> output_31_cast_fp16 = layer_norm(axes = output_31_axes_0, beta = const_107_to_fp16, epsilon = var_22_to_fp16, gamma = const_106_to_fp16, x = input_233_cast_fp16)[name = string("output_31_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_6_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_6_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46828288)))];
            tensor<fp16, [2048]> m_encoder_encoders_6_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_6_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48925504)))];
            tensor<fp16, [1, 304, 2048]> linear_30_cast_fp16 = linear(bias = m_encoder_encoders_6_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_6_feed_forward_w_1_weight_to_fp16, x = output_31_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_241_cast_fp16 = relu(x = linear_30_cast_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_6_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_6_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48929664)))];
            tensor<fp16, [512]> m_encoder_encoders_6_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_6_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51026880)))];
            tensor<fp16, [1, 304, 512]> linear_31_cast_fp16 = linear(bias = m_encoder_encoders_6_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_6_feed_forward_w_2_weight_to_fp16, x = input_241_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_247_cast_fp16 = add(x = input_233_cast_fp16, y = linear_31_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<int32, [1]> output_33_axes_0 = const()[name = string("output_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_108_to_fp16 = const()[name = string("const_108_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51027968)))];
            tensor<fp16, [512]> const_109_to_fp16 = const()[name = string("const_109_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51029056)))];
            tensor<fp16, [1, 304, 512]> output_33_cast_fp16 = layer_norm(axes = output_33_axes_0, beta = const_109_to_fp16, epsilon = var_22_to_fp16, gamma = const_108_to_fp16, x = input_247_cast_fp16)[name = string("output_33_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_7_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_7_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51030144)))];
            tensor<fp16, [1536]> m_encoder_encoders_7_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_7_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52603072)))];
            tensor<fp16, [1, 304, 1536]> linear_32_cast_fp16 = linear(bias = m_encoder_encoders_7_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_7_self_attn_linear_q_k_v_weight_to_fp16, x = output_33_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<int32, [3]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_1092_axis_0 = const()[name = string("op_1092_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_1092_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_1092_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_1092_cast_fp16_2 = split(axis = var_1092_axis_0, split_sizes = tile_8, x = linear_32_cast_fp16)[name = string("op_1092_cast_fp16")];
            tensor<int32, [4]> var_1096 = const()[name = string("op_1096"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1097_cast_fp16 = reshape(shape = var_1096, x = var_1092_cast_fp16_0)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> var_1099 = const()[name = string("op_1099"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1100_cast_fp16 = reshape(shape = var_1099, x = var_1092_cast_fp16_1)[name = string("op_1100_cast_fp16")];
            tensor<int32, [4]> var_1102 = const()[name = string("op_1102"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1103_cast_fp16 = reshape(shape = var_1102, x = var_1092_cast_fp16_2)[name = string("op_1103_cast_fp16")];
            tensor<int32, [4]> value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_251_perm_0 = const()[name = string("input_251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_253_pad_0 = const()[name = string("input_253_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_253_mode_0 = const()[name = string("input_253_mode_0"), val = string("constant")];
            fp16 const_115_to_fp16 = const()[name = string("const_115_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_251_cast_fp16 = transpose(perm = input_251_perm_0, x = var_1092_cast_fp16_2)[name = string("transpose_720")];
            tensor<fp16, [1, 512, 314]> input_253_cast_fp16 = pad(constant_val = const_115_to_fp16, mode = input_253_mode_0, pad = input_253_pad_0, x = input_251_cast_fp16)[name = string("input_253_cast_fp16")];
            string x_85_pad_type_0 = const()[name = string("x_85_pad_type_0"), val = string("valid")];
            int32 x_85_groups_0 = const()[name = string("x_85_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_85_strides_0 = const()[name = string("x_85_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_85_pad_0 = const()[name = string("x_85_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_85_dilations_0 = const()[name = string("x_85_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_7_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_7_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52606208)))];
            tensor<fp16, [1, 512, 304]> x_85_cast_fp16 = conv(dilations = x_85_dilations_0, groups = x_85_groups_0, pad = x_85_pad_0, pad_type = x_85_pad_type_0, strides = x_85_strides_0, weight = m_encoder_encoders_7_self_attn_fsmn_block_weight_to_fp16, x = input_253_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<int32, [3]> x_87_perm_0 = const()[name = string("x_87_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = x_85_cast_fp16)[name = string("transpose_719")];
            tensor<fp16, [1, 304, 512]> input_255_cast_fp16 = add(x = x_87_cast_fp16, y = var_1092_cast_fp16_2)[name = string("input_255_cast_fp16")];
            fp16 var_1122_to_fp16 = const()[name = string("op_1122_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_35_cast_fp16 = mul(x = var_1097_cast_fp16, y = var_1122_to_fp16)[name = string("q_h_35_cast_fp16")];
            bool scores_33_transpose_x_0 = const()[name = string("scores_33_transpose_x_0"), val = bool(false)];
            bool scores_33_transpose_y_0 = const()[name = string("scores_33_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_226_perm_0 = const()[name = string("transpose_226_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_227_perm_0 = const()[name = string("transpose_227_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_227 = transpose(perm = transpose_227_perm_0, x = var_1100_cast_fp16)[name = string("transpose_717")];
            tensor<fp16, [1, 4, 304, 128]> transpose_226 = transpose(perm = transpose_226_perm_0, x = q_h_35_cast_fp16)[name = string("transpose_718")];
            tensor<fp16, [1, 4, 304, 304]> scores_33_cast_fp16 = matmul(transpose_x = scores_33_transpose_x_0, transpose_y = scores_33_transpose_y_0, x = transpose_226, y = transpose_227)[name = string("scores_33_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1130_cast_fp16 = softmax(axis = var_29, x = scores_33_cast_fp16)[name = string("op_1130_cast_fp16")];
            bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)];
            bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = var_1103_cast_fp16)[name = string("transpose_721")];
            tensor<fp16, [1, 4, 304, 128]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = var_1130_cast_fp16, y = value_17_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> var_1134_perm_0 = const()[name = string("op_1134_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1136 = const()[name = string("op_1136"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1134_cast_fp16 = transpose(perm = var_1134_perm_0, x = x_91_cast_fp16)[name = string("transpose_716")];
            tensor<fp16, [1, 304, 512]> input_259_cast_fp16 = reshape(shape = var_1136, x = var_1134_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_7_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_7_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52617536)))];
            tensor<fp16, [512]> m_encoder_encoders_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_7_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53141888)))];
            tensor<fp16, [1, 304, 512]> linear_33_cast_fp16 = linear(bias = m_encoder_encoders_7_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_7_self_attn_linear_out_weight_to_fp16, x = input_259_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_261_cast_fp16 = add(x = linear_33_cast_fp16, y = input_255_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_263_cast_fp16 = add(x = input_247_cast_fp16, y = input_261_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<int32, [1]> output_35_axes_0 = const()[name = string("output_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53142976)))];
            tensor<fp16, [512]> const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53144064)))];
            tensor<fp16, [1, 304, 512]> output_35_cast_fp16 = layer_norm(axes = output_35_axes_0, beta = const_118_to_fp16, epsilon = var_22_to_fp16, gamma = const_117_to_fp16, x = input_263_cast_fp16)[name = string("output_35_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_7_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_7_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53145152)))];
            tensor<fp16, [2048]> m_encoder_encoders_7_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_7_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55242368)))];
            tensor<fp16, [1, 304, 2048]> linear_34_cast_fp16 = linear(bias = m_encoder_encoders_7_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_7_feed_forward_w_1_weight_to_fp16, x = output_35_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_271_cast_fp16 = relu(x = linear_34_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_7_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_7_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55246528)))];
            tensor<fp16, [512]> m_encoder_encoders_7_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_7_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57343744)))];
            tensor<fp16, [1, 304, 512]> linear_35_cast_fp16 = linear(bias = m_encoder_encoders_7_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_7_feed_forward_w_2_weight_to_fp16, x = input_271_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_277_cast_fp16 = add(x = input_263_cast_fp16, y = linear_35_cast_fp16)[name = string("input_277_cast_fp16")];
            tensor<int32, [1]> output_37_axes_0 = const()[name = string("output_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_119_to_fp16 = const()[name = string("const_119_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57344832)))];
            tensor<fp16, [512]> const_120_to_fp16 = const()[name = string("const_120_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57345920)))];
            tensor<fp16, [1, 304, 512]> output_37_cast_fp16 = layer_norm(axes = output_37_axes_0, beta = const_120_to_fp16, epsilon = var_22_to_fp16, gamma = const_119_to_fp16, x = input_277_cast_fp16)[name = string("output_37_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_8_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_8_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57347008)))];
            tensor<fp16, [1536]> m_encoder_encoders_8_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_8_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58919936)))];
            tensor<fp16, [1, 304, 1536]> linear_36_cast_fp16 = linear(bias = m_encoder_encoders_8_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_8_self_attn_linear_q_k_v_weight_to_fp16, x = output_37_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [3]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_1194_axis_0 = const()[name = string("op_1194_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_1194_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_1194_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_1194_cast_fp16_2 = split(axis = var_1194_axis_0, split_sizes = tile_9, x = linear_36_cast_fp16)[name = string("op_1194_cast_fp16")];
            tensor<int32, [4]> var_1198 = const()[name = string("op_1198"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1199_cast_fp16 = reshape(shape = var_1198, x = var_1194_cast_fp16_0)[name = string("op_1199_cast_fp16")];
            tensor<int32, [4]> var_1201 = const()[name = string("op_1201"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1202_cast_fp16 = reshape(shape = var_1201, x = var_1194_cast_fp16_1)[name = string("op_1202_cast_fp16")];
            tensor<int32, [4]> var_1204 = const()[name = string("op_1204"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1205_cast_fp16 = reshape(shape = var_1204, x = var_1194_cast_fp16_2)[name = string("op_1205_cast_fp16")];
            tensor<int32, [4]> value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_281_perm_0 = const()[name = string("input_281_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_283_pad_0 = const()[name = string("input_283_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_283_mode_0 = const()[name = string("input_283_mode_0"), val = string("constant")];
            fp16 const_126_to_fp16 = const()[name = string("const_126_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_281_cast_fp16 = transpose(perm = input_281_perm_0, x = var_1194_cast_fp16_2)[name = string("transpose_714")];
            tensor<fp16, [1, 512, 314]> input_283_cast_fp16 = pad(constant_val = const_126_to_fp16, mode = input_283_mode_0, pad = input_283_pad_0, x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
            string x_95_pad_type_0 = const()[name = string("x_95_pad_type_0"), val = string("valid")];
            int32 x_95_groups_0 = const()[name = string("x_95_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_95_strides_0 = const()[name = string("x_95_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_95_pad_0 = const()[name = string("x_95_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_95_dilations_0 = const()[name = string("x_95_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_8_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_8_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58923072)))];
            tensor<fp16, [1, 512, 304]> x_95_cast_fp16 = conv(dilations = x_95_dilations_0, groups = x_95_groups_0, pad = x_95_pad_0, pad_type = x_95_pad_type_0, strides = x_95_strides_0, weight = m_encoder_encoders_8_self_attn_fsmn_block_weight_to_fp16, x = input_283_cast_fp16)[name = string("x_95_cast_fp16")];
            tensor<int32, [3]> x_97_perm_0 = const()[name = string("x_97_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = x_95_cast_fp16)[name = string("transpose_713")];
            tensor<fp16, [1, 304, 512]> input_285_cast_fp16 = add(x = x_97_cast_fp16, y = var_1194_cast_fp16_2)[name = string("input_285_cast_fp16")];
            fp16 var_1224_to_fp16 = const()[name = string("op_1224_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_39_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1224_to_fp16)[name = string("q_h_39_cast_fp16")];
            bool scores_37_transpose_x_0 = const()[name = string("scores_37_transpose_x_0"), val = bool(false)];
            bool scores_37_transpose_y_0 = const()[name = string("scores_37_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_228_perm_0 = const()[name = string("transpose_228_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_229_perm_0 = const()[name = string("transpose_229_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_229 = transpose(perm = transpose_229_perm_0, x = var_1202_cast_fp16)[name = string("transpose_711")];
            tensor<fp16, [1, 4, 304, 128]> transpose_228 = transpose(perm = transpose_228_perm_0, x = q_h_39_cast_fp16)[name = string("transpose_712")];
            tensor<fp16, [1, 4, 304, 304]> scores_37_cast_fp16 = matmul(transpose_x = scores_37_transpose_x_0, transpose_y = scores_37_transpose_y_0, x = transpose_228, y = transpose_229)[name = string("scores_37_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1232_cast_fp16 = softmax(axis = var_29, x = scores_37_cast_fp16)[name = string("op_1232_cast_fp16")];
            bool x_101_transpose_x_0 = const()[name = string("x_101_transpose_x_0"), val = bool(false)];
            bool x_101_transpose_y_0 = const()[name = string("x_101_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = var_1205_cast_fp16)[name = string("transpose_715")];
            tensor<fp16, [1, 4, 304, 128]> x_101_cast_fp16 = matmul(transpose_x = x_101_transpose_x_0, transpose_y = x_101_transpose_y_0, x = var_1232_cast_fp16, y = value_19_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> var_1236_perm_0 = const()[name = string("op_1236_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1238 = const()[name = string("op_1238"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1236_cast_fp16 = transpose(perm = var_1236_perm_0, x = x_101_cast_fp16)[name = string("transpose_710")];
            tensor<fp16, [1, 304, 512]> input_289_cast_fp16 = reshape(shape = var_1238, x = var_1236_cast_fp16)[name = string("input_289_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_8_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_8_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58934400)))];
            tensor<fp16, [512]> m_encoder_encoders_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_8_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59458752)))];
            tensor<fp16, [1, 304, 512]> linear_37_cast_fp16 = linear(bias = m_encoder_encoders_8_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_8_self_attn_linear_out_weight_to_fp16, x = input_289_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_291_cast_fp16 = add(x = linear_37_cast_fp16, y = input_285_cast_fp16)[name = string("input_291_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_293_cast_fp16 = add(x = input_277_cast_fp16, y = input_291_cast_fp16)[name = string("input_293_cast_fp16")];
            tensor<int32, [1]> output_39_axes_0 = const()[name = string("output_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_128_to_fp16 = const()[name = string("const_128_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59459840)))];
            tensor<fp16, [512]> const_129_to_fp16 = const()[name = string("const_129_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59460928)))];
            tensor<fp16, [1, 304, 512]> output_39_cast_fp16 = layer_norm(axes = output_39_axes_0, beta = const_129_to_fp16, epsilon = var_22_to_fp16, gamma = const_128_to_fp16, x = input_293_cast_fp16)[name = string("output_39_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_8_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_8_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59462016)))];
            tensor<fp16, [2048]> m_encoder_encoders_8_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_8_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61559232)))];
            tensor<fp16, [1, 304, 2048]> linear_38_cast_fp16 = linear(bias = m_encoder_encoders_8_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_8_feed_forward_w_1_weight_to_fp16, x = output_39_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_301_cast_fp16 = relu(x = linear_38_cast_fp16)[name = string("input_301_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_8_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_8_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61563392)))];
            tensor<fp16, [512]> m_encoder_encoders_8_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_8_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63660608)))];
            tensor<fp16, [1, 304, 512]> linear_39_cast_fp16 = linear(bias = m_encoder_encoders_8_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_8_feed_forward_w_2_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_307_cast_fp16 = add(x = input_293_cast_fp16, y = linear_39_cast_fp16)[name = string("input_307_cast_fp16")];
            tensor<int32, [1]> output_41_axes_0 = const()[name = string("output_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_130_to_fp16 = const()[name = string("const_130_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63661696)))];
            tensor<fp16, [512]> const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63662784)))];
            tensor<fp16, [1, 304, 512]> output_41_cast_fp16 = layer_norm(axes = output_41_axes_0, beta = const_131_to_fp16, epsilon = var_22_to_fp16, gamma = const_130_to_fp16, x = input_307_cast_fp16)[name = string("output_41_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_9_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_9_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63663872)))];
            tensor<fp16, [1536]> m_encoder_encoders_9_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_9_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65236800)))];
            tensor<fp16, [1, 304, 1536]> linear_40_cast_fp16 = linear(bias = m_encoder_encoders_9_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_9_self_attn_linear_q_k_v_weight_to_fp16, x = output_41_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<int32, [3]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_1296_axis_0 = const()[name = string("op_1296_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_1296_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_1296_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_1296_cast_fp16_2 = split(axis = var_1296_axis_0, split_sizes = tile_10, x = linear_40_cast_fp16)[name = string("op_1296_cast_fp16")];
            tensor<int32, [4]> var_1300 = const()[name = string("op_1300"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1301_cast_fp16 = reshape(shape = var_1300, x = var_1296_cast_fp16_0)[name = string("op_1301_cast_fp16")];
            tensor<int32, [4]> var_1303 = const()[name = string("op_1303"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1304_cast_fp16 = reshape(shape = var_1303, x = var_1296_cast_fp16_1)[name = string("op_1304_cast_fp16")];
            tensor<int32, [4]> var_1306 = const()[name = string("op_1306"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1307_cast_fp16 = reshape(shape = var_1306, x = var_1296_cast_fp16_2)[name = string("op_1307_cast_fp16")];
            tensor<int32, [4]> value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_311_perm_0 = const()[name = string("input_311_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_313_pad_0 = const()[name = string("input_313_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_313_mode_0 = const()[name = string("input_313_mode_0"), val = string("constant")];
            fp16 const_137_to_fp16 = const()[name = string("const_137_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_311_cast_fp16 = transpose(perm = input_311_perm_0, x = var_1296_cast_fp16_2)[name = string("transpose_708")];
            tensor<fp16, [1, 512, 314]> input_313_cast_fp16 = pad(constant_val = const_137_to_fp16, mode = input_313_mode_0, pad = input_313_pad_0, x = input_311_cast_fp16)[name = string("input_313_cast_fp16")];
            string x_105_pad_type_0 = const()[name = string("x_105_pad_type_0"), val = string("valid")];
            int32 x_105_groups_0 = const()[name = string("x_105_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_105_strides_0 = const()[name = string("x_105_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_105_pad_0 = const()[name = string("x_105_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_105_dilations_0 = const()[name = string("x_105_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_9_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_9_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65239936)))];
            tensor<fp16, [1, 512, 304]> x_105_cast_fp16 = conv(dilations = x_105_dilations_0, groups = x_105_groups_0, pad = x_105_pad_0, pad_type = x_105_pad_type_0, strides = x_105_strides_0, weight = m_encoder_encoders_9_self_attn_fsmn_block_weight_to_fp16, x = input_313_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [3]> x_107_perm_0 = const()[name = string("x_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_107_cast_fp16 = transpose(perm = x_107_perm_0, x = x_105_cast_fp16)[name = string("transpose_707")];
            tensor<fp16, [1, 304, 512]> input_315_cast_fp16 = add(x = x_107_cast_fp16, y = var_1296_cast_fp16_2)[name = string("input_315_cast_fp16")];
            fp16 var_1326_to_fp16 = const()[name = string("op_1326_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_43_cast_fp16 = mul(x = var_1301_cast_fp16, y = var_1326_to_fp16)[name = string("q_h_43_cast_fp16")];
            bool scores_41_transpose_x_0 = const()[name = string("scores_41_transpose_x_0"), val = bool(false)];
            bool scores_41_transpose_y_0 = const()[name = string("scores_41_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_230_perm_0 = const()[name = string("transpose_230_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_231_perm_0 = const()[name = string("transpose_231_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_231 = transpose(perm = transpose_231_perm_0, x = var_1304_cast_fp16)[name = string("transpose_705")];
            tensor<fp16, [1, 4, 304, 128]> transpose_230 = transpose(perm = transpose_230_perm_0, x = q_h_43_cast_fp16)[name = string("transpose_706")];
            tensor<fp16, [1, 4, 304, 304]> scores_41_cast_fp16 = matmul(transpose_x = scores_41_transpose_x_0, transpose_y = scores_41_transpose_y_0, x = transpose_230, y = transpose_231)[name = string("scores_41_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1334_cast_fp16 = softmax(axis = var_29, x = scores_41_cast_fp16)[name = string("op_1334_cast_fp16")];
            bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)];
            bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = var_1307_cast_fp16)[name = string("transpose_709")];
            tensor<fp16, [1, 4, 304, 128]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = var_1334_cast_fp16, y = value_21_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> var_1338_perm_0 = const()[name = string("op_1338_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1340 = const()[name = string("op_1340"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1338_cast_fp16 = transpose(perm = var_1338_perm_0, x = x_111_cast_fp16)[name = string("transpose_704")];
            tensor<fp16, [1, 304, 512]> input_319_cast_fp16 = reshape(shape = var_1340, x = var_1338_cast_fp16)[name = string("input_319_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_9_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_9_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65251264)))];
            tensor<fp16, [512]> m_encoder_encoders_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_9_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65775616)))];
            tensor<fp16, [1, 304, 512]> linear_41_cast_fp16 = linear(bias = m_encoder_encoders_9_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_9_self_attn_linear_out_weight_to_fp16, x = input_319_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_321_cast_fp16 = add(x = linear_41_cast_fp16, y = input_315_cast_fp16)[name = string("input_321_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
            tensor<int32, [1]> output_43_axes_0 = const()[name = string("output_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_139_to_fp16 = const()[name = string("const_139_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65776704)))];
            tensor<fp16, [512]> const_140_to_fp16 = const()[name = string("const_140_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65777792)))];
            tensor<fp16, [1, 304, 512]> output_43_cast_fp16 = layer_norm(axes = output_43_axes_0, beta = const_140_to_fp16, epsilon = var_22_to_fp16, gamma = const_139_to_fp16, x = input_323_cast_fp16)[name = string("output_43_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_9_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_9_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65778880)))];
            tensor<fp16, [2048]> m_encoder_encoders_9_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_9_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67876096)))];
            tensor<fp16, [1, 304, 2048]> linear_42_cast_fp16 = linear(bias = m_encoder_encoders_9_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_9_feed_forward_w_1_weight_to_fp16, x = output_43_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_331_cast_fp16 = relu(x = linear_42_cast_fp16)[name = string("input_331_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_9_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_9_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67880256)))];
            tensor<fp16, [512]> m_encoder_encoders_9_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_9_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69977472)))];
            tensor<fp16, [1, 304, 512]> linear_43_cast_fp16 = linear(bias = m_encoder_encoders_9_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_9_feed_forward_w_2_weight_to_fp16, x = input_331_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_337_cast_fp16 = add(x = input_323_cast_fp16, y = linear_43_cast_fp16)[name = string("input_337_cast_fp16")];
            tensor<int32, [1]> output_45_axes_0 = const()[name = string("output_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_141_to_fp16 = const()[name = string("const_141_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69978560)))];
            tensor<fp16, [512]> const_142_to_fp16 = const()[name = string("const_142_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69979648)))];
            tensor<fp16, [1, 304, 512]> output_45_cast_fp16 = layer_norm(axes = output_45_axes_0, beta = const_142_to_fp16, epsilon = var_22_to_fp16, gamma = const_141_to_fp16, x = input_337_cast_fp16)[name = string("output_45_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_10_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_10_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69980736)))];
            tensor<fp16, [1536]> m_encoder_encoders_10_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_10_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71553664)))];
            tensor<fp16, [1, 304, 1536]> linear_44_cast_fp16 = linear(bias = m_encoder_encoders_10_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_10_self_attn_linear_q_k_v_weight_to_fp16, x = output_45_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [3]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_1398_axis_0 = const()[name = string("op_1398_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_1398_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_1398_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_1398_cast_fp16_2 = split(axis = var_1398_axis_0, split_sizes = tile_11, x = linear_44_cast_fp16)[name = string("op_1398_cast_fp16")];
            tensor<int32, [4]> var_1402 = const()[name = string("op_1402"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1403_cast_fp16 = reshape(shape = var_1402, x = var_1398_cast_fp16_0)[name = string("op_1403_cast_fp16")];
            tensor<int32, [4]> var_1405 = const()[name = string("op_1405"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1406_cast_fp16 = reshape(shape = var_1405, x = var_1398_cast_fp16_1)[name = string("op_1406_cast_fp16")];
            tensor<int32, [4]> var_1408 = const()[name = string("op_1408"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1409_cast_fp16 = reshape(shape = var_1408, x = var_1398_cast_fp16_2)[name = string("op_1409_cast_fp16")];
            tensor<int32, [4]> value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_341_perm_0 = const()[name = string("input_341_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_343_pad_0 = const()[name = string("input_343_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_343_mode_0 = const()[name = string("input_343_mode_0"), val = string("constant")];
            fp16 const_148_to_fp16 = const()[name = string("const_148_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_341_cast_fp16 = transpose(perm = input_341_perm_0, x = var_1398_cast_fp16_2)[name = string("transpose_702")];
            tensor<fp16, [1, 512, 314]> input_343_cast_fp16 = pad(constant_val = const_148_to_fp16, mode = input_343_mode_0, pad = input_343_pad_0, x = input_341_cast_fp16)[name = string("input_343_cast_fp16")];
            string x_115_pad_type_0 = const()[name = string("x_115_pad_type_0"), val = string("valid")];
            int32 x_115_groups_0 = const()[name = string("x_115_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_115_strides_0 = const()[name = string("x_115_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_115_pad_0 = const()[name = string("x_115_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_115_dilations_0 = const()[name = string("x_115_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_10_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_10_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71556800)))];
            tensor<fp16, [1, 512, 304]> x_115_cast_fp16 = conv(dilations = x_115_dilations_0, groups = x_115_groups_0, pad = x_115_pad_0, pad_type = x_115_pad_type_0, strides = x_115_strides_0, weight = m_encoder_encoders_10_self_attn_fsmn_block_weight_to_fp16, x = input_343_cast_fp16)[name = string("x_115_cast_fp16")];
            tensor<int32, [3]> x_117_perm_0 = const()[name = string("x_117_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_117_cast_fp16 = transpose(perm = x_117_perm_0, x = x_115_cast_fp16)[name = string("transpose_701")];
            tensor<fp16, [1, 304, 512]> input_345_cast_fp16 = add(x = x_117_cast_fp16, y = var_1398_cast_fp16_2)[name = string("input_345_cast_fp16")];
            fp16 var_1428_to_fp16 = const()[name = string("op_1428_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_47_cast_fp16 = mul(x = var_1403_cast_fp16, y = var_1428_to_fp16)[name = string("q_h_47_cast_fp16")];
            bool scores_45_transpose_x_0 = const()[name = string("scores_45_transpose_x_0"), val = bool(false)];
            bool scores_45_transpose_y_0 = const()[name = string("scores_45_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_232_perm_0 = const()[name = string("transpose_232_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_233_perm_0 = const()[name = string("transpose_233_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_233 = transpose(perm = transpose_233_perm_0, x = var_1406_cast_fp16)[name = string("transpose_699")];
            tensor<fp16, [1, 4, 304, 128]> transpose_232 = transpose(perm = transpose_232_perm_0, x = q_h_47_cast_fp16)[name = string("transpose_700")];
            tensor<fp16, [1, 4, 304, 304]> scores_45_cast_fp16 = matmul(transpose_x = scores_45_transpose_x_0, transpose_y = scores_45_transpose_y_0, x = transpose_232, y = transpose_233)[name = string("scores_45_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1436_cast_fp16 = softmax(axis = var_29, x = scores_45_cast_fp16)[name = string("op_1436_cast_fp16")];
            bool x_121_transpose_x_0 = const()[name = string("x_121_transpose_x_0"), val = bool(false)];
            bool x_121_transpose_y_0 = const()[name = string("x_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = var_1409_cast_fp16)[name = string("transpose_703")];
            tensor<fp16, [1, 4, 304, 128]> x_121_cast_fp16 = matmul(transpose_x = x_121_transpose_x_0, transpose_y = x_121_transpose_y_0, x = var_1436_cast_fp16, y = value_23_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> var_1440_perm_0 = const()[name = string("op_1440_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1442 = const()[name = string("op_1442"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1440_cast_fp16 = transpose(perm = var_1440_perm_0, x = x_121_cast_fp16)[name = string("transpose_698")];
            tensor<fp16, [1, 304, 512]> input_349_cast_fp16 = reshape(shape = var_1442, x = var_1440_cast_fp16)[name = string("input_349_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_10_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_10_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71568128)))];
            tensor<fp16, [512]> m_encoder_encoders_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_10_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72092480)))];
            tensor<fp16, [1, 304, 512]> linear_45_cast_fp16 = linear(bias = m_encoder_encoders_10_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_10_self_attn_linear_out_weight_to_fp16, x = input_349_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_351_cast_fp16 = add(x = linear_45_cast_fp16, y = input_345_cast_fp16)[name = string("input_351_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_353_cast_fp16 = add(x = input_337_cast_fp16, y = input_351_cast_fp16)[name = string("input_353_cast_fp16")];
            tensor<int32, [1]> output_47_axes_0 = const()[name = string("output_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_150_to_fp16 = const()[name = string("const_150_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72093568)))];
            tensor<fp16, [512]> const_151_to_fp16 = const()[name = string("const_151_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72094656)))];
            tensor<fp16, [1, 304, 512]> output_47_cast_fp16 = layer_norm(axes = output_47_axes_0, beta = const_151_to_fp16, epsilon = var_22_to_fp16, gamma = const_150_to_fp16, x = input_353_cast_fp16)[name = string("output_47_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_10_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_10_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72095744)))];
            tensor<fp16, [2048]> m_encoder_encoders_10_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_10_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74192960)))];
            tensor<fp16, [1, 304, 2048]> linear_46_cast_fp16 = linear(bias = m_encoder_encoders_10_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_10_feed_forward_w_1_weight_to_fp16, x = output_47_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_361_cast_fp16 = relu(x = linear_46_cast_fp16)[name = string("input_361_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_10_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_10_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74197120)))];
            tensor<fp16, [512]> m_encoder_encoders_10_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_10_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76294336)))];
            tensor<fp16, [1, 304, 512]> linear_47_cast_fp16 = linear(bias = m_encoder_encoders_10_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_10_feed_forward_w_2_weight_to_fp16, x = input_361_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_367_cast_fp16 = add(x = input_353_cast_fp16, y = linear_47_cast_fp16)[name = string("input_367_cast_fp16")];
            tensor<int32, [1]> output_49_axes_0 = const()[name = string("output_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_152_to_fp16 = const()[name = string("const_152_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76295424)))];
            tensor<fp16, [512]> const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76296512)))];
            tensor<fp16, [1, 304, 512]> output_49_cast_fp16 = layer_norm(axes = output_49_axes_0, beta = const_153_to_fp16, epsilon = var_22_to_fp16, gamma = const_152_to_fp16, x = input_367_cast_fp16)[name = string("output_49_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_11_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_11_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76297600)))];
            tensor<fp16, [1536]> m_encoder_encoders_11_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_11_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77870528)))];
            tensor<fp16, [1, 304, 1536]> linear_48_cast_fp16 = linear(bias = m_encoder_encoders_11_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_11_self_attn_linear_q_k_v_weight_to_fp16, x = output_49_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [3]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_1500_axis_0 = const()[name = string("op_1500_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_1500_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_1500_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_1500_cast_fp16_2 = split(axis = var_1500_axis_0, split_sizes = tile_12, x = linear_48_cast_fp16)[name = string("op_1500_cast_fp16")];
            tensor<int32, [4]> var_1504 = const()[name = string("op_1504"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1505_cast_fp16 = reshape(shape = var_1504, x = var_1500_cast_fp16_0)[name = string("op_1505_cast_fp16")];
            tensor<int32, [4]> var_1507 = const()[name = string("op_1507"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1508_cast_fp16 = reshape(shape = var_1507, x = var_1500_cast_fp16_1)[name = string("op_1508_cast_fp16")];
            tensor<int32, [4]> var_1510 = const()[name = string("op_1510"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1511_cast_fp16 = reshape(shape = var_1510, x = var_1500_cast_fp16_2)[name = string("op_1511_cast_fp16")];
            tensor<int32, [4]> value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_371_perm_0 = const()[name = string("input_371_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_373_pad_0 = const()[name = string("input_373_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_373_mode_0 = const()[name = string("input_373_mode_0"), val = string("constant")];
            fp16 const_159_to_fp16 = const()[name = string("const_159_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_371_cast_fp16 = transpose(perm = input_371_perm_0, x = var_1500_cast_fp16_2)[name = string("transpose_696")];
            tensor<fp16, [1, 512, 314]> input_373_cast_fp16 = pad(constant_val = const_159_to_fp16, mode = input_373_mode_0, pad = input_373_pad_0, x = input_371_cast_fp16)[name = string("input_373_cast_fp16")];
            string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")];
            int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_11_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_11_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77873664)))];
            tensor<fp16, [1, 512, 304]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = m_encoder_encoders_11_self_attn_fsmn_block_weight_to_fp16, x = input_373_cast_fp16)[name = string("x_125_cast_fp16")];
            tensor<int32, [3]> x_127_perm_0 = const()[name = string("x_127_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_127_cast_fp16 = transpose(perm = x_127_perm_0, x = x_125_cast_fp16)[name = string("transpose_695")];
            tensor<fp16, [1, 304, 512]> input_375_cast_fp16 = add(x = x_127_cast_fp16, y = var_1500_cast_fp16_2)[name = string("input_375_cast_fp16")];
            fp16 var_1530_to_fp16 = const()[name = string("op_1530_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_51_cast_fp16 = mul(x = var_1505_cast_fp16, y = var_1530_to_fp16)[name = string("q_h_51_cast_fp16")];
            bool scores_49_transpose_x_0 = const()[name = string("scores_49_transpose_x_0"), val = bool(false)];
            bool scores_49_transpose_y_0 = const()[name = string("scores_49_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_234_perm_0 = const()[name = string("transpose_234_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_235_perm_0 = const()[name = string("transpose_235_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_235 = transpose(perm = transpose_235_perm_0, x = var_1508_cast_fp16)[name = string("transpose_693")];
            tensor<fp16, [1, 4, 304, 128]> transpose_234 = transpose(perm = transpose_234_perm_0, x = q_h_51_cast_fp16)[name = string("transpose_694")];
            tensor<fp16, [1, 4, 304, 304]> scores_49_cast_fp16 = matmul(transpose_x = scores_49_transpose_x_0, transpose_y = scores_49_transpose_y_0, x = transpose_234, y = transpose_235)[name = string("scores_49_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1538_cast_fp16 = softmax(axis = var_29, x = scores_49_cast_fp16)[name = string("op_1538_cast_fp16")];
            bool x_131_transpose_x_0 = const()[name = string("x_131_transpose_x_0"), val = bool(false)];
            bool x_131_transpose_y_0 = const()[name = string("x_131_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = var_1511_cast_fp16)[name = string("transpose_697")];
            tensor<fp16, [1, 4, 304, 128]> x_131_cast_fp16 = matmul(transpose_x = x_131_transpose_x_0, transpose_y = x_131_transpose_y_0, x = var_1538_cast_fp16, y = value_25_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> var_1542_perm_0 = const()[name = string("op_1542_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1544 = const()[name = string("op_1544"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1542_cast_fp16 = transpose(perm = var_1542_perm_0, x = x_131_cast_fp16)[name = string("transpose_692")];
            tensor<fp16, [1, 304, 512]> input_379_cast_fp16 = reshape(shape = var_1544, x = var_1542_cast_fp16)[name = string("input_379_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_11_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_11_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77884992)))];
            tensor<fp16, [512]> m_encoder_encoders_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_11_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78409344)))];
            tensor<fp16, [1, 304, 512]> linear_49_cast_fp16 = linear(bias = m_encoder_encoders_11_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_11_self_attn_linear_out_weight_to_fp16, x = input_379_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_381_cast_fp16 = add(x = linear_49_cast_fp16, y = input_375_cast_fp16)[name = string("input_381_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_383_cast_fp16 = add(x = input_367_cast_fp16, y = input_381_cast_fp16)[name = string("input_383_cast_fp16")];
            tensor<int32, [1]> output_51_axes_0 = const()[name = string("output_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_161_to_fp16 = const()[name = string("const_161_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78410432)))];
            tensor<fp16, [512]> const_162_to_fp16 = const()[name = string("const_162_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78411520)))];
            tensor<fp16, [1, 304, 512]> output_51_cast_fp16 = layer_norm(axes = output_51_axes_0, beta = const_162_to_fp16, epsilon = var_22_to_fp16, gamma = const_161_to_fp16, x = input_383_cast_fp16)[name = string("output_51_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_11_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_11_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78412608)))];
            tensor<fp16, [2048]> m_encoder_encoders_11_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_11_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80509824)))];
            tensor<fp16, [1, 304, 2048]> linear_50_cast_fp16 = linear(bias = m_encoder_encoders_11_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_11_feed_forward_w_1_weight_to_fp16, x = output_51_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_391_cast_fp16 = relu(x = linear_50_cast_fp16)[name = string("input_391_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_11_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_11_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80513984)))];
            tensor<fp16, [512]> m_encoder_encoders_11_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_11_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82611200)))];
            tensor<fp16, [1, 304, 512]> linear_51_cast_fp16 = linear(bias = m_encoder_encoders_11_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_11_feed_forward_w_2_weight_to_fp16, x = input_391_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_397_cast_fp16 = add(x = input_383_cast_fp16, y = linear_51_cast_fp16)[name = string("input_397_cast_fp16")];
            tensor<int32, [1]> output_53_axes_0 = const()[name = string("output_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_163_to_fp16 = const()[name = string("const_163_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82612288)))];
            tensor<fp16, [512]> const_164_to_fp16 = const()[name = string("const_164_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82613376)))];
            tensor<fp16, [1, 304, 512]> output_53_cast_fp16 = layer_norm(axes = output_53_axes_0, beta = const_164_to_fp16, epsilon = var_22_to_fp16, gamma = const_163_to_fp16, x = input_397_cast_fp16)[name = string("output_53_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_12_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_12_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82614464)))];
            tensor<fp16, [1536]> m_encoder_encoders_12_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_12_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84187392)))];
            tensor<fp16, [1, 304, 1536]> linear_52_cast_fp16 = linear(bias = m_encoder_encoders_12_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_12_self_attn_linear_q_k_v_weight_to_fp16, x = output_53_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<int32, [3]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_1602_axis_0 = const()[name = string("op_1602_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_1602_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_1602_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_1602_cast_fp16_2 = split(axis = var_1602_axis_0, split_sizes = tile_13, x = linear_52_cast_fp16)[name = string("op_1602_cast_fp16")];
            tensor<int32, [4]> var_1606 = const()[name = string("op_1606"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1607_cast_fp16 = reshape(shape = var_1606, x = var_1602_cast_fp16_0)[name = string("op_1607_cast_fp16")];
            tensor<int32, [4]> var_1609 = const()[name = string("op_1609"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1610_cast_fp16 = reshape(shape = var_1609, x = var_1602_cast_fp16_1)[name = string("op_1610_cast_fp16")];
            tensor<int32, [4]> var_1612 = const()[name = string("op_1612"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1613_cast_fp16 = reshape(shape = var_1612, x = var_1602_cast_fp16_2)[name = string("op_1613_cast_fp16")];
            tensor<int32, [4]> value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_401_perm_0 = const()[name = string("input_401_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_403_pad_0 = const()[name = string("input_403_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_403_mode_0 = const()[name = string("input_403_mode_0"), val = string("constant")];
            fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_401_cast_fp16 = transpose(perm = input_401_perm_0, x = var_1602_cast_fp16_2)[name = string("transpose_690")];
            tensor<fp16, [1, 512, 314]> input_403_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = input_403_mode_0, pad = input_403_pad_0, x = input_401_cast_fp16)[name = string("input_403_cast_fp16")];
            string x_135_pad_type_0 = const()[name = string("x_135_pad_type_0"), val = string("valid")];
            int32 x_135_groups_0 = const()[name = string("x_135_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_135_strides_0 = const()[name = string("x_135_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_135_pad_0 = const()[name = string("x_135_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_135_dilations_0 = const()[name = string("x_135_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_12_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_12_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84190528)))];
            tensor<fp16, [1, 512, 304]> x_135_cast_fp16 = conv(dilations = x_135_dilations_0, groups = x_135_groups_0, pad = x_135_pad_0, pad_type = x_135_pad_type_0, strides = x_135_strides_0, weight = m_encoder_encoders_12_self_attn_fsmn_block_weight_to_fp16, x = input_403_cast_fp16)[name = string("x_135_cast_fp16")];
            tensor<int32, [3]> x_137_perm_0 = const()[name = string("x_137_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_137_cast_fp16 = transpose(perm = x_137_perm_0, x = x_135_cast_fp16)[name = string("transpose_689")];
            tensor<fp16, [1, 304, 512]> input_405_cast_fp16 = add(x = x_137_cast_fp16, y = var_1602_cast_fp16_2)[name = string("input_405_cast_fp16")];
            fp16 var_1632_to_fp16 = const()[name = string("op_1632_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_55_cast_fp16 = mul(x = var_1607_cast_fp16, y = var_1632_to_fp16)[name = string("q_h_55_cast_fp16")];
            bool scores_53_transpose_x_0 = const()[name = string("scores_53_transpose_x_0"), val = bool(false)];
            bool scores_53_transpose_y_0 = const()[name = string("scores_53_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_236_perm_0 = const()[name = string("transpose_236_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_237_perm_0 = const()[name = string("transpose_237_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_237 = transpose(perm = transpose_237_perm_0, x = var_1610_cast_fp16)[name = string("transpose_687")];
            tensor<fp16, [1, 4, 304, 128]> transpose_236 = transpose(perm = transpose_236_perm_0, x = q_h_55_cast_fp16)[name = string("transpose_688")];
            tensor<fp16, [1, 4, 304, 304]> scores_53_cast_fp16 = matmul(transpose_x = scores_53_transpose_x_0, transpose_y = scores_53_transpose_y_0, x = transpose_236, y = transpose_237)[name = string("scores_53_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1640_cast_fp16 = softmax(axis = var_29, x = scores_53_cast_fp16)[name = string("op_1640_cast_fp16")];
            bool x_141_transpose_x_0 = const()[name = string("x_141_transpose_x_0"), val = bool(false)];
            bool x_141_transpose_y_0 = const()[name = string("x_141_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = var_1613_cast_fp16)[name = string("transpose_691")];
            tensor<fp16, [1, 4, 304, 128]> x_141_cast_fp16 = matmul(transpose_x = x_141_transpose_x_0, transpose_y = x_141_transpose_y_0, x = var_1640_cast_fp16, y = value_27_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [4]> var_1644_perm_0 = const()[name = string("op_1644_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1646 = const()[name = string("op_1646"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1644_cast_fp16 = transpose(perm = var_1644_perm_0, x = x_141_cast_fp16)[name = string("transpose_686")];
            tensor<fp16, [1, 304, 512]> input_409_cast_fp16 = reshape(shape = var_1646, x = var_1644_cast_fp16)[name = string("input_409_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_12_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_12_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84201856)))];
            tensor<fp16, [512]> m_encoder_encoders_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_12_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84726208)))];
            tensor<fp16, [1, 304, 512]> linear_53_cast_fp16 = linear(bias = m_encoder_encoders_12_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_12_self_attn_linear_out_weight_to_fp16, x = input_409_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_411_cast_fp16 = add(x = linear_53_cast_fp16, y = input_405_cast_fp16)[name = string("input_411_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_413_cast_fp16 = add(x = input_397_cast_fp16, y = input_411_cast_fp16)[name = string("input_413_cast_fp16")];
            tensor<int32, [1]> output_55_axes_0 = const()[name = string("output_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84727296)))];
            tensor<fp16, [512]> const_173_to_fp16 = const()[name = string("const_173_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84728384)))];
            tensor<fp16, [1, 304, 512]> output_55_cast_fp16 = layer_norm(axes = output_55_axes_0, beta = const_173_to_fp16, epsilon = var_22_to_fp16, gamma = const_172_to_fp16, x = input_413_cast_fp16)[name = string("output_55_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_12_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_12_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84729472)))];
            tensor<fp16, [2048]> m_encoder_encoders_12_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_12_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86826688)))];
            tensor<fp16, [1, 304, 2048]> linear_54_cast_fp16 = linear(bias = m_encoder_encoders_12_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_12_feed_forward_w_1_weight_to_fp16, x = output_55_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_421_cast_fp16 = relu(x = linear_54_cast_fp16)[name = string("input_421_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_12_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_12_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86830848)))];
            tensor<fp16, [512]> m_encoder_encoders_12_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_12_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88928064)))];
            tensor<fp16, [1, 304, 512]> linear_55_cast_fp16 = linear(bias = m_encoder_encoders_12_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_12_feed_forward_w_2_weight_to_fp16, x = input_421_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_427_cast_fp16 = add(x = input_413_cast_fp16, y = linear_55_cast_fp16)[name = string("input_427_cast_fp16")];
            tensor<int32, [1]> output_57_axes_0 = const()[name = string("output_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_174_to_fp16 = const()[name = string("const_174_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88929152)))];
            tensor<fp16, [512]> const_175_to_fp16 = const()[name = string("const_175_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88930240)))];
            tensor<fp16, [1, 304, 512]> output_57_cast_fp16 = layer_norm(axes = output_57_axes_0, beta = const_175_to_fp16, epsilon = var_22_to_fp16, gamma = const_174_to_fp16, x = input_427_cast_fp16)[name = string("output_57_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_13_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_13_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88931328)))];
            tensor<fp16, [1536]> m_encoder_encoders_13_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_13_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90504256)))];
            tensor<fp16, [1, 304, 1536]> linear_56_cast_fp16 = linear(bias = m_encoder_encoders_13_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_13_self_attn_linear_q_k_v_weight_to_fp16, x = output_57_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [3]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_1704_axis_0 = const()[name = string("op_1704_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_1704_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_1704_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_1704_cast_fp16_2 = split(axis = var_1704_axis_0, split_sizes = tile_14, x = linear_56_cast_fp16)[name = string("op_1704_cast_fp16")];
            tensor<int32, [4]> var_1708 = const()[name = string("op_1708"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1709_cast_fp16 = reshape(shape = var_1708, x = var_1704_cast_fp16_0)[name = string("op_1709_cast_fp16")];
            tensor<int32, [4]> var_1711 = const()[name = string("op_1711"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1712_cast_fp16 = reshape(shape = var_1711, x = var_1704_cast_fp16_1)[name = string("op_1712_cast_fp16")];
            tensor<int32, [4]> var_1714 = const()[name = string("op_1714"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1715_cast_fp16 = reshape(shape = var_1714, x = var_1704_cast_fp16_2)[name = string("op_1715_cast_fp16")];
            tensor<int32, [4]> value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_431_perm_0 = const()[name = string("input_431_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_433_pad_0 = const()[name = string("input_433_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_433_mode_0 = const()[name = string("input_433_mode_0"), val = string("constant")];
            fp16 const_181_to_fp16 = const()[name = string("const_181_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_431_cast_fp16 = transpose(perm = input_431_perm_0, x = var_1704_cast_fp16_2)[name = string("transpose_684")];
            tensor<fp16, [1, 512, 314]> input_433_cast_fp16 = pad(constant_val = const_181_to_fp16, mode = input_433_mode_0, pad = input_433_pad_0, x = input_431_cast_fp16)[name = string("input_433_cast_fp16")];
            string x_145_pad_type_0 = const()[name = string("x_145_pad_type_0"), val = string("valid")];
            int32 x_145_groups_0 = const()[name = string("x_145_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_145_strides_0 = const()[name = string("x_145_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_145_pad_0 = const()[name = string("x_145_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_145_dilations_0 = const()[name = string("x_145_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_13_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_13_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90507392)))];
            tensor<fp16, [1, 512, 304]> x_145_cast_fp16 = conv(dilations = x_145_dilations_0, groups = x_145_groups_0, pad = x_145_pad_0, pad_type = x_145_pad_type_0, strides = x_145_strides_0, weight = m_encoder_encoders_13_self_attn_fsmn_block_weight_to_fp16, x = input_433_cast_fp16)[name = string("x_145_cast_fp16")];
            tensor<int32, [3]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = x_145_cast_fp16)[name = string("transpose_683")];
            tensor<fp16, [1, 304, 512]> input_435_cast_fp16 = add(x = x_147_cast_fp16, y = var_1704_cast_fp16_2)[name = string("input_435_cast_fp16")];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_59_cast_fp16 = mul(x = var_1709_cast_fp16, y = var_1734_to_fp16)[name = string("q_h_59_cast_fp16")];
            bool scores_57_transpose_x_0 = const()[name = string("scores_57_transpose_x_0"), val = bool(false)];
            bool scores_57_transpose_y_0 = const()[name = string("scores_57_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_238_perm_0 = const()[name = string("transpose_238_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_239_perm_0 = const()[name = string("transpose_239_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_239 = transpose(perm = transpose_239_perm_0, x = var_1712_cast_fp16)[name = string("transpose_681")];
            tensor<fp16, [1, 4, 304, 128]> transpose_238 = transpose(perm = transpose_238_perm_0, x = q_h_59_cast_fp16)[name = string("transpose_682")];
            tensor<fp16, [1, 4, 304, 304]> scores_57_cast_fp16 = matmul(transpose_x = scores_57_transpose_x_0, transpose_y = scores_57_transpose_y_0, x = transpose_238, y = transpose_239)[name = string("scores_57_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1742_cast_fp16 = softmax(axis = var_29, x = scores_57_cast_fp16)[name = string("op_1742_cast_fp16")];
            bool x_151_transpose_x_0 = const()[name = string("x_151_transpose_x_0"), val = bool(false)];
            bool x_151_transpose_y_0 = const()[name = string("x_151_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = var_1715_cast_fp16)[name = string("transpose_685")];
            tensor<fp16, [1, 4, 304, 128]> x_151_cast_fp16 = matmul(transpose_x = x_151_transpose_x_0, transpose_y = x_151_transpose_y_0, x = var_1742_cast_fp16, y = value_29_cast_fp16)[name = string("x_151_cast_fp16")];
            tensor<int32, [4]> var_1746_perm_0 = const()[name = string("op_1746_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1748 = const()[name = string("op_1748"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1746_cast_fp16 = transpose(perm = var_1746_perm_0, x = x_151_cast_fp16)[name = string("transpose_680")];
            tensor<fp16, [1, 304, 512]> input_439_cast_fp16 = reshape(shape = var_1748, x = var_1746_cast_fp16)[name = string("input_439_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_13_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_13_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90518720)))];
            tensor<fp16, [512]> m_encoder_encoders_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_13_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91043072)))];
            tensor<fp16, [1, 304, 512]> linear_57_cast_fp16 = linear(bias = m_encoder_encoders_13_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_13_self_attn_linear_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_441_cast_fp16 = add(x = linear_57_cast_fp16, y = input_435_cast_fp16)[name = string("input_441_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_443_cast_fp16 = add(x = input_427_cast_fp16, y = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
            tensor<int32, [1]> output_59_axes_0 = const()[name = string("output_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91044160)))];
            tensor<fp16, [512]> const_184_to_fp16 = const()[name = string("const_184_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91045248)))];
            tensor<fp16, [1, 304, 512]> output_59_cast_fp16 = layer_norm(axes = output_59_axes_0, beta = const_184_to_fp16, epsilon = var_22_to_fp16, gamma = const_183_to_fp16, x = input_443_cast_fp16)[name = string("output_59_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_13_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_13_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91046336)))];
            tensor<fp16, [2048]> m_encoder_encoders_13_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_13_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93143552)))];
            tensor<fp16, [1, 304, 2048]> linear_58_cast_fp16 = linear(bias = m_encoder_encoders_13_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_13_feed_forward_w_1_weight_to_fp16, x = output_59_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_451_cast_fp16 = relu(x = linear_58_cast_fp16)[name = string("input_451_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_13_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_13_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93147712)))];
            tensor<fp16, [512]> m_encoder_encoders_13_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_13_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95244928)))];
            tensor<fp16, [1, 304, 512]> linear_59_cast_fp16 = linear(bias = m_encoder_encoders_13_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_13_feed_forward_w_2_weight_to_fp16, x = input_451_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_457_cast_fp16 = add(x = input_443_cast_fp16, y = linear_59_cast_fp16)[name = string("input_457_cast_fp16")];
            tensor<int32, [1]> output_61_axes_0 = const()[name = string("output_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_185_to_fp16 = const()[name = string("const_185_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95246016)))];
            tensor<fp16, [512]> const_186_to_fp16 = const()[name = string("const_186_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95247104)))];
            tensor<fp16, [1, 304, 512]> output_61_cast_fp16 = layer_norm(axes = output_61_axes_0, beta = const_186_to_fp16, epsilon = var_22_to_fp16, gamma = const_185_to_fp16, x = input_457_cast_fp16)[name = string("output_61_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_14_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_14_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95248192)))];
            tensor<fp16, [1536]> m_encoder_encoders_14_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_14_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96821120)))];
            tensor<fp16, [1, 304, 1536]> linear_60_cast_fp16 = linear(bias = m_encoder_encoders_14_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_14_self_attn_linear_q_k_v_weight_to_fp16, x = output_61_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<int32, [3]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_1806_axis_0 = const()[name = string("op_1806_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_1806_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_1806_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_1806_cast_fp16_2 = split(axis = var_1806_axis_0, split_sizes = tile_15, x = linear_60_cast_fp16)[name = string("op_1806_cast_fp16")];
            tensor<int32, [4]> var_1810 = const()[name = string("op_1810"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1811_cast_fp16 = reshape(shape = var_1810, x = var_1806_cast_fp16_0)[name = string("op_1811_cast_fp16")];
            tensor<int32, [4]> var_1813 = const()[name = string("op_1813"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1814_cast_fp16 = reshape(shape = var_1813, x = var_1806_cast_fp16_1)[name = string("op_1814_cast_fp16")];
            tensor<int32, [4]> var_1816 = const()[name = string("op_1816"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1817_cast_fp16 = reshape(shape = var_1816, x = var_1806_cast_fp16_2)[name = string("op_1817_cast_fp16")];
            tensor<int32, [4]> value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_461_perm_0 = const()[name = string("input_461_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_463_pad_0 = const()[name = string("input_463_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_463_mode_0 = const()[name = string("input_463_mode_0"), val = string("constant")];
            fp16 const_192_to_fp16 = const()[name = string("const_192_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_461_cast_fp16 = transpose(perm = input_461_perm_0, x = var_1806_cast_fp16_2)[name = string("transpose_678")];
            tensor<fp16, [1, 512, 314]> input_463_cast_fp16 = pad(constant_val = const_192_to_fp16, mode = input_463_mode_0, pad = input_463_pad_0, x = input_461_cast_fp16)[name = string("input_463_cast_fp16")];
            string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")];
            int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_14_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_14_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96824256)))];
            tensor<fp16, [1, 512, 304]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = m_encoder_encoders_14_self_attn_fsmn_block_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<int32, [3]> x_157_perm_0 = const()[name = string("x_157_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_157_cast_fp16 = transpose(perm = x_157_perm_0, x = x_155_cast_fp16)[name = string("transpose_677")];
            tensor<fp16, [1, 304, 512]> input_465_cast_fp16 = add(x = x_157_cast_fp16, y = var_1806_cast_fp16_2)[name = string("input_465_cast_fp16")];
            fp16 var_1836_to_fp16 = const()[name = string("op_1836_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_63_cast_fp16 = mul(x = var_1811_cast_fp16, y = var_1836_to_fp16)[name = string("q_h_63_cast_fp16")];
            bool scores_61_transpose_x_0 = const()[name = string("scores_61_transpose_x_0"), val = bool(false)];
            bool scores_61_transpose_y_0 = const()[name = string("scores_61_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_240_perm_0 = const()[name = string("transpose_240_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_241_perm_0 = const()[name = string("transpose_241_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_241 = transpose(perm = transpose_241_perm_0, x = var_1814_cast_fp16)[name = string("transpose_675")];
            tensor<fp16, [1, 4, 304, 128]> transpose_240 = transpose(perm = transpose_240_perm_0, x = q_h_63_cast_fp16)[name = string("transpose_676")];
            tensor<fp16, [1, 4, 304, 304]> scores_61_cast_fp16 = matmul(transpose_x = scores_61_transpose_x_0, transpose_y = scores_61_transpose_y_0, x = transpose_240, y = transpose_241)[name = string("scores_61_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1844_cast_fp16 = softmax(axis = var_29, x = scores_61_cast_fp16)[name = string("op_1844_cast_fp16")];
            bool x_161_transpose_x_0 = const()[name = string("x_161_transpose_x_0"), val = bool(false)];
            bool x_161_transpose_y_0 = const()[name = string("x_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = var_1817_cast_fp16)[name = string("transpose_679")];
            tensor<fp16, [1, 4, 304, 128]> x_161_cast_fp16 = matmul(transpose_x = x_161_transpose_x_0, transpose_y = x_161_transpose_y_0, x = var_1844_cast_fp16, y = value_31_cast_fp16)[name = string("x_161_cast_fp16")];
            tensor<int32, [4]> var_1848_perm_0 = const()[name = string("op_1848_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1850 = const()[name = string("op_1850"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1848_cast_fp16 = transpose(perm = var_1848_perm_0, x = x_161_cast_fp16)[name = string("transpose_674")];
            tensor<fp16, [1, 304, 512]> input_469_cast_fp16 = reshape(shape = var_1850, x = var_1848_cast_fp16)[name = string("input_469_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_14_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_14_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96835584)))];
            tensor<fp16, [512]> m_encoder_encoders_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_14_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97359936)))];
            tensor<fp16, [1, 304, 512]> linear_61_cast_fp16 = linear(bias = m_encoder_encoders_14_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_14_self_attn_linear_out_weight_to_fp16, x = input_469_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_471_cast_fp16 = add(x = linear_61_cast_fp16, y = input_465_cast_fp16)[name = string("input_471_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_473_cast_fp16 = add(x = input_457_cast_fp16, y = input_471_cast_fp16)[name = string("input_473_cast_fp16")];
            tensor<int32, [1]> output_63_axes_0 = const()[name = string("output_63_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_194_to_fp16 = const()[name = string("const_194_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97361024)))];
            tensor<fp16, [512]> const_195_to_fp16 = const()[name = string("const_195_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97362112)))];
            tensor<fp16, [1, 304, 512]> output_63_cast_fp16 = layer_norm(axes = output_63_axes_0, beta = const_195_to_fp16, epsilon = var_22_to_fp16, gamma = const_194_to_fp16, x = input_473_cast_fp16)[name = string("output_63_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_14_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_14_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97363200)))];
            tensor<fp16, [2048]> m_encoder_encoders_14_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_14_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99460416)))];
            tensor<fp16, [1, 304, 2048]> linear_62_cast_fp16 = linear(bias = m_encoder_encoders_14_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_14_feed_forward_w_1_weight_to_fp16, x = output_63_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_481_cast_fp16 = relu(x = linear_62_cast_fp16)[name = string("input_481_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_14_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_14_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99464576)))];
            tensor<fp16, [512]> m_encoder_encoders_14_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_14_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101561792)))];
            tensor<fp16, [1, 304, 512]> linear_63_cast_fp16 = linear(bias = m_encoder_encoders_14_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_14_feed_forward_w_2_weight_to_fp16, x = input_481_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_487_cast_fp16 = add(x = input_473_cast_fp16, y = linear_63_cast_fp16)[name = string("input_487_cast_fp16")];
            tensor<int32, [1]> output_65_axes_0 = const()[name = string("output_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101562880)))];
            tensor<fp16, [512]> const_197_to_fp16 = const()[name = string("const_197_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101563968)))];
            tensor<fp16, [1, 304, 512]> output_65_cast_fp16 = layer_norm(axes = output_65_axes_0, beta = const_197_to_fp16, epsilon = var_22_to_fp16, gamma = const_196_to_fp16, x = input_487_cast_fp16)[name = string("output_65_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_15_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_15_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101565056)))];
            tensor<fp16, [1536]> m_encoder_encoders_15_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_15_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103137984)))];
            tensor<fp16, [1, 304, 1536]> linear_64_cast_fp16 = linear(bias = m_encoder_encoders_15_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_15_self_attn_linear_q_k_v_weight_to_fp16, x = output_65_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [3]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_1908_axis_0 = const()[name = string("op_1908_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_1908_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_1908_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_1908_cast_fp16_2 = split(axis = var_1908_axis_0, split_sizes = tile_16, x = linear_64_cast_fp16)[name = string("op_1908_cast_fp16")];
            tensor<int32, [4]> var_1912 = const()[name = string("op_1912"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1913_cast_fp16 = reshape(shape = var_1912, x = var_1908_cast_fp16_0)[name = string("op_1913_cast_fp16")];
            tensor<int32, [4]> var_1915 = const()[name = string("op_1915"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1916_cast_fp16 = reshape(shape = var_1915, x = var_1908_cast_fp16_1)[name = string("op_1916_cast_fp16")];
            tensor<int32, [4]> var_1918 = const()[name = string("op_1918"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_1919_cast_fp16 = reshape(shape = var_1918, x = var_1908_cast_fp16_2)[name = string("op_1919_cast_fp16")];
            tensor<int32, [4]> value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_491_perm_0 = const()[name = string("input_491_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_493_pad_0 = const()[name = string("input_493_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_493_mode_0 = const()[name = string("input_493_mode_0"), val = string("constant")];
            fp16 const_203_to_fp16 = const()[name = string("const_203_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_491_cast_fp16 = transpose(perm = input_491_perm_0, x = var_1908_cast_fp16_2)[name = string("transpose_672")];
            tensor<fp16, [1, 512, 314]> input_493_cast_fp16 = pad(constant_val = const_203_to_fp16, mode = input_493_mode_0, pad = input_493_pad_0, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")];
            string x_165_pad_type_0 = const()[name = string("x_165_pad_type_0"), val = string("valid")];
            int32 x_165_groups_0 = const()[name = string("x_165_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_165_strides_0 = const()[name = string("x_165_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_165_dilations_0 = const()[name = string("x_165_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_15_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_15_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103141120)))];
            tensor<fp16, [1, 512, 304]> x_165_cast_fp16 = conv(dilations = x_165_dilations_0, groups = x_165_groups_0, pad = x_165_pad_0, pad_type = x_165_pad_type_0, strides = x_165_strides_0, weight = m_encoder_encoders_15_self_attn_fsmn_block_weight_to_fp16, x = input_493_cast_fp16)[name = string("x_165_cast_fp16")];
            tensor<int32, [3]> x_167_perm_0 = const()[name = string("x_167_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_167_cast_fp16 = transpose(perm = x_167_perm_0, x = x_165_cast_fp16)[name = string("transpose_671")];
            tensor<fp16, [1, 304, 512]> input_495_cast_fp16 = add(x = x_167_cast_fp16, y = var_1908_cast_fp16_2)[name = string("input_495_cast_fp16")];
            fp16 var_1938_to_fp16 = const()[name = string("op_1938_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_67_cast_fp16 = mul(x = var_1913_cast_fp16, y = var_1938_to_fp16)[name = string("q_h_67_cast_fp16")];
            bool scores_65_transpose_x_0 = const()[name = string("scores_65_transpose_x_0"), val = bool(false)];
            bool scores_65_transpose_y_0 = const()[name = string("scores_65_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_242_perm_0 = const()[name = string("transpose_242_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_243_perm_0 = const()[name = string("transpose_243_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_243 = transpose(perm = transpose_243_perm_0, x = var_1916_cast_fp16)[name = string("transpose_669")];
            tensor<fp16, [1, 4, 304, 128]> transpose_242 = transpose(perm = transpose_242_perm_0, x = q_h_67_cast_fp16)[name = string("transpose_670")];
            tensor<fp16, [1, 4, 304, 304]> scores_65_cast_fp16 = matmul(transpose_x = scores_65_transpose_x_0, transpose_y = scores_65_transpose_y_0, x = transpose_242, y = transpose_243)[name = string("scores_65_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_1946_cast_fp16 = softmax(axis = var_29, x = scores_65_cast_fp16)[name = string("op_1946_cast_fp16")];
            bool x_171_transpose_x_0 = const()[name = string("x_171_transpose_x_0"), val = bool(false)];
            bool x_171_transpose_y_0 = const()[name = string("x_171_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = var_1919_cast_fp16)[name = string("transpose_673")];
            tensor<fp16, [1, 4, 304, 128]> x_171_cast_fp16 = matmul(transpose_x = x_171_transpose_x_0, transpose_y = x_171_transpose_y_0, x = var_1946_cast_fp16, y = value_33_cast_fp16)[name = string("x_171_cast_fp16")];
            tensor<int32, [4]> var_1950_perm_0 = const()[name = string("op_1950_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1952 = const()[name = string("op_1952"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_1950_cast_fp16 = transpose(perm = var_1950_perm_0, x = x_171_cast_fp16)[name = string("transpose_668")];
            tensor<fp16, [1, 304, 512]> input_499_cast_fp16 = reshape(shape = var_1952, x = var_1950_cast_fp16)[name = string("input_499_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_15_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_15_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103152448)))];
            tensor<fp16, [512]> m_encoder_encoders_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_15_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103676800)))];
            tensor<fp16, [1, 304, 512]> linear_65_cast_fp16 = linear(bias = m_encoder_encoders_15_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_15_self_attn_linear_out_weight_to_fp16, x = input_499_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_501_cast_fp16 = add(x = linear_65_cast_fp16, y = input_495_cast_fp16)[name = string("input_501_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_503_cast_fp16 = add(x = input_487_cast_fp16, y = input_501_cast_fp16)[name = string("input_503_cast_fp16")];
            tensor<int32, [1]> output_67_axes_0 = const()[name = string("output_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_205_to_fp16 = const()[name = string("const_205_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103677888)))];
            tensor<fp16, [512]> const_206_to_fp16 = const()[name = string("const_206_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103678976)))];
            tensor<fp16, [1, 304, 512]> output_67_cast_fp16 = layer_norm(axes = output_67_axes_0, beta = const_206_to_fp16, epsilon = var_22_to_fp16, gamma = const_205_to_fp16, x = input_503_cast_fp16)[name = string("output_67_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_15_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_15_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103680064)))];
            tensor<fp16, [2048]> m_encoder_encoders_15_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_15_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105777280)))];
            tensor<fp16, [1, 304, 2048]> linear_66_cast_fp16 = linear(bias = m_encoder_encoders_15_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_15_feed_forward_w_1_weight_to_fp16, x = output_67_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_511_cast_fp16 = relu(x = linear_66_cast_fp16)[name = string("input_511_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_15_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_15_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105781440)))];
            tensor<fp16, [512]> m_encoder_encoders_15_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_15_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107878656)))];
            tensor<fp16, [1, 304, 512]> linear_67_cast_fp16 = linear(bias = m_encoder_encoders_15_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_15_feed_forward_w_2_weight_to_fp16, x = input_511_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_517_cast_fp16 = add(x = input_503_cast_fp16, y = linear_67_cast_fp16)[name = string("input_517_cast_fp16")];
            tensor<int32, [1]> output_69_axes_0 = const()[name = string("output_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107879744)))];
            tensor<fp16, [512]> const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107880832)))];
            tensor<fp16, [1, 304, 512]> output_69_cast_fp16 = layer_norm(axes = output_69_axes_0, beta = const_208_to_fp16, epsilon = var_22_to_fp16, gamma = const_207_to_fp16, x = input_517_cast_fp16)[name = string("output_69_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_16_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_16_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107881920)))];
            tensor<fp16, [1536]> m_encoder_encoders_16_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_16_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109454848)))];
            tensor<fp16, [1, 304, 1536]> linear_68_cast_fp16 = linear(bias = m_encoder_encoders_16_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_16_self_attn_linear_q_k_v_weight_to_fp16, x = output_69_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [3]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2010_axis_0 = const()[name = string("op_2010_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2010_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2010_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2010_cast_fp16_2 = split(axis = var_2010_axis_0, split_sizes = tile_17, x = linear_68_cast_fp16)[name = string("op_2010_cast_fp16")];
            tensor<int32, [4]> var_2014 = const()[name = string("op_2014"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2015_cast_fp16 = reshape(shape = var_2014, x = var_2010_cast_fp16_0)[name = string("op_2015_cast_fp16")];
            tensor<int32, [4]> var_2017 = const()[name = string("op_2017"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2018_cast_fp16 = reshape(shape = var_2017, x = var_2010_cast_fp16_1)[name = string("op_2018_cast_fp16")];
            tensor<int32, [4]> var_2020 = const()[name = string("op_2020"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2021_cast_fp16 = reshape(shape = var_2020, x = var_2010_cast_fp16_2)[name = string("op_2021_cast_fp16")];
            tensor<int32, [4]> value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_521_perm_0 = const()[name = string("input_521_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_523_pad_0 = const()[name = string("input_523_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_523_mode_0 = const()[name = string("input_523_mode_0"), val = string("constant")];
            fp16 const_214_to_fp16 = const()[name = string("const_214_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_521_cast_fp16 = transpose(perm = input_521_perm_0, x = var_2010_cast_fp16_2)[name = string("transpose_666")];
            tensor<fp16, [1, 512, 314]> input_523_cast_fp16 = pad(constant_val = const_214_to_fp16, mode = input_523_mode_0, pad = input_523_pad_0, x = input_521_cast_fp16)[name = string("input_523_cast_fp16")];
            string x_175_pad_type_0 = const()[name = string("x_175_pad_type_0"), val = string("valid")];
            int32 x_175_groups_0 = const()[name = string("x_175_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_175_strides_0 = const()[name = string("x_175_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_175_pad_0 = const()[name = string("x_175_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_175_dilations_0 = const()[name = string("x_175_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_16_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_16_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109457984)))];
            tensor<fp16, [1, 512, 304]> x_175_cast_fp16 = conv(dilations = x_175_dilations_0, groups = x_175_groups_0, pad = x_175_pad_0, pad_type = x_175_pad_type_0, strides = x_175_strides_0, weight = m_encoder_encoders_16_self_attn_fsmn_block_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_175_cast_fp16")];
            tensor<int32, [3]> x_177_perm_0 = const()[name = string("x_177_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_177_cast_fp16 = transpose(perm = x_177_perm_0, x = x_175_cast_fp16)[name = string("transpose_665")];
            tensor<fp16, [1, 304, 512]> input_525_cast_fp16 = add(x = x_177_cast_fp16, y = var_2010_cast_fp16_2)[name = string("input_525_cast_fp16")];
            fp16 var_2040_to_fp16 = const()[name = string("op_2040_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_71_cast_fp16 = mul(x = var_2015_cast_fp16, y = var_2040_to_fp16)[name = string("q_h_71_cast_fp16")];
            bool scores_69_transpose_x_0 = const()[name = string("scores_69_transpose_x_0"), val = bool(false)];
            bool scores_69_transpose_y_0 = const()[name = string("scores_69_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_244_perm_0 = const()[name = string("transpose_244_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_245_perm_0 = const()[name = string("transpose_245_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_245 = transpose(perm = transpose_245_perm_0, x = var_2018_cast_fp16)[name = string("transpose_663")];
            tensor<fp16, [1, 4, 304, 128]> transpose_244 = transpose(perm = transpose_244_perm_0, x = q_h_71_cast_fp16)[name = string("transpose_664")];
            tensor<fp16, [1, 4, 304, 304]> scores_69_cast_fp16 = matmul(transpose_x = scores_69_transpose_x_0, transpose_y = scores_69_transpose_y_0, x = transpose_244, y = transpose_245)[name = string("scores_69_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2048_cast_fp16 = softmax(axis = var_29, x = scores_69_cast_fp16)[name = string("op_2048_cast_fp16")];
            bool x_181_transpose_x_0 = const()[name = string("x_181_transpose_x_0"), val = bool(false)];
            bool x_181_transpose_y_0 = const()[name = string("x_181_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = var_2021_cast_fp16)[name = string("transpose_667")];
            tensor<fp16, [1, 4, 304, 128]> x_181_cast_fp16 = matmul(transpose_x = x_181_transpose_x_0, transpose_y = x_181_transpose_y_0, x = var_2048_cast_fp16, y = value_35_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<int32, [4]> var_2052_perm_0 = const()[name = string("op_2052_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2054 = const()[name = string("op_2054"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2052_cast_fp16 = transpose(perm = var_2052_perm_0, x = x_181_cast_fp16)[name = string("transpose_662")];
            tensor<fp16, [1, 304, 512]> input_529_cast_fp16 = reshape(shape = var_2054, x = var_2052_cast_fp16)[name = string("input_529_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_16_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_16_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109469312)))];
            tensor<fp16, [512]> m_encoder_encoders_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_16_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109993664)))];
            tensor<fp16, [1, 304, 512]> linear_69_cast_fp16 = linear(bias = m_encoder_encoders_16_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_16_self_attn_linear_out_weight_to_fp16, x = input_529_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_531_cast_fp16 = add(x = linear_69_cast_fp16, y = input_525_cast_fp16)[name = string("input_531_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_533_cast_fp16 = add(x = input_517_cast_fp16, y = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
            tensor<int32, [1]> output_71_axes_0 = const()[name = string("output_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_216_to_fp16 = const()[name = string("const_216_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109994752)))];
            tensor<fp16, [512]> const_217_to_fp16 = const()[name = string("const_217_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109995840)))];
            tensor<fp16, [1, 304, 512]> output_71_cast_fp16 = layer_norm(axes = output_71_axes_0, beta = const_217_to_fp16, epsilon = var_22_to_fp16, gamma = const_216_to_fp16, x = input_533_cast_fp16)[name = string("output_71_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_16_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_16_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109996928)))];
            tensor<fp16, [2048]> m_encoder_encoders_16_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_16_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112094144)))];
            tensor<fp16, [1, 304, 2048]> linear_70_cast_fp16 = linear(bias = m_encoder_encoders_16_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_16_feed_forward_w_1_weight_to_fp16, x = output_71_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_541_cast_fp16 = relu(x = linear_70_cast_fp16)[name = string("input_541_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_16_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_16_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112098304)))];
            tensor<fp16, [512]> m_encoder_encoders_16_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_16_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114195520)))];
            tensor<fp16, [1, 304, 512]> linear_71_cast_fp16 = linear(bias = m_encoder_encoders_16_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_16_feed_forward_w_2_weight_to_fp16, x = input_541_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_547_cast_fp16 = add(x = input_533_cast_fp16, y = linear_71_cast_fp16)[name = string("input_547_cast_fp16")];
            tensor<int32, [1]> output_73_axes_0 = const()[name = string("output_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_218_to_fp16 = const()[name = string("const_218_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114196608)))];
            tensor<fp16, [512]> const_219_to_fp16 = const()[name = string("const_219_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114197696)))];
            tensor<fp16, [1, 304, 512]> output_73_cast_fp16 = layer_norm(axes = output_73_axes_0, beta = const_219_to_fp16, epsilon = var_22_to_fp16, gamma = const_218_to_fp16, x = input_547_cast_fp16)[name = string("output_73_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_17_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_17_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114198784)))];
            tensor<fp16, [1536]> m_encoder_encoders_17_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_17_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115771712)))];
            tensor<fp16, [1, 304, 1536]> linear_72_cast_fp16 = linear(bias = m_encoder_encoders_17_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_17_self_attn_linear_q_k_v_weight_to_fp16, x = output_73_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [3]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2112_axis_0 = const()[name = string("op_2112_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2112_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2112_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2112_cast_fp16_2 = split(axis = var_2112_axis_0, split_sizes = tile_18, x = linear_72_cast_fp16)[name = string("op_2112_cast_fp16")];
            tensor<int32, [4]> var_2116 = const()[name = string("op_2116"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2117_cast_fp16 = reshape(shape = var_2116, x = var_2112_cast_fp16_0)[name = string("op_2117_cast_fp16")];
            tensor<int32, [4]> var_2119 = const()[name = string("op_2119"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2120_cast_fp16 = reshape(shape = var_2119, x = var_2112_cast_fp16_1)[name = string("op_2120_cast_fp16")];
            tensor<int32, [4]> var_2122 = const()[name = string("op_2122"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2123_cast_fp16 = reshape(shape = var_2122, x = var_2112_cast_fp16_2)[name = string("op_2123_cast_fp16")];
            tensor<int32, [4]> value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_551_perm_0 = const()[name = string("input_551_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_553_pad_0 = const()[name = string("input_553_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_553_mode_0 = const()[name = string("input_553_mode_0"), val = string("constant")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_551_cast_fp16 = transpose(perm = input_551_perm_0, x = var_2112_cast_fp16_2)[name = string("transpose_660")];
            tensor<fp16, [1, 512, 314]> input_553_cast_fp16 = pad(constant_val = const_225_to_fp16, mode = input_553_mode_0, pad = input_553_pad_0, x = input_551_cast_fp16)[name = string("input_553_cast_fp16")];
            string x_185_pad_type_0 = const()[name = string("x_185_pad_type_0"), val = string("valid")];
            int32 x_185_groups_0 = const()[name = string("x_185_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_185_strides_0 = const()[name = string("x_185_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_185_pad_0 = const()[name = string("x_185_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_185_dilations_0 = const()[name = string("x_185_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_17_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_17_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115774848)))];
            tensor<fp16, [1, 512, 304]> x_185_cast_fp16 = conv(dilations = x_185_dilations_0, groups = x_185_groups_0, pad = x_185_pad_0, pad_type = x_185_pad_type_0, strides = x_185_strides_0, weight = m_encoder_encoders_17_self_attn_fsmn_block_weight_to_fp16, x = input_553_cast_fp16)[name = string("x_185_cast_fp16")];
            tensor<int32, [3]> x_187_perm_0 = const()[name = string("x_187_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_187_cast_fp16 = transpose(perm = x_187_perm_0, x = x_185_cast_fp16)[name = string("transpose_659")];
            tensor<fp16, [1, 304, 512]> input_555_cast_fp16 = add(x = x_187_cast_fp16, y = var_2112_cast_fp16_2)[name = string("input_555_cast_fp16")];
            fp16 var_2142_to_fp16 = const()[name = string("op_2142_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_75_cast_fp16 = mul(x = var_2117_cast_fp16, y = var_2142_to_fp16)[name = string("q_h_75_cast_fp16")];
            bool scores_73_transpose_x_0 = const()[name = string("scores_73_transpose_x_0"), val = bool(false)];
            bool scores_73_transpose_y_0 = const()[name = string("scores_73_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_246_perm_0 = const()[name = string("transpose_246_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_247_perm_0 = const()[name = string("transpose_247_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_247 = transpose(perm = transpose_247_perm_0, x = var_2120_cast_fp16)[name = string("transpose_657")];
            tensor<fp16, [1, 4, 304, 128]> transpose_246 = transpose(perm = transpose_246_perm_0, x = q_h_75_cast_fp16)[name = string("transpose_658")];
            tensor<fp16, [1, 4, 304, 304]> scores_73_cast_fp16 = matmul(transpose_x = scores_73_transpose_x_0, transpose_y = scores_73_transpose_y_0, x = transpose_246, y = transpose_247)[name = string("scores_73_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2150_cast_fp16 = softmax(axis = var_29, x = scores_73_cast_fp16)[name = string("op_2150_cast_fp16")];
            bool x_191_transpose_x_0 = const()[name = string("x_191_transpose_x_0"), val = bool(false)];
            bool x_191_transpose_y_0 = const()[name = string("x_191_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = var_2123_cast_fp16)[name = string("transpose_661")];
            tensor<fp16, [1, 4, 304, 128]> x_191_cast_fp16 = matmul(transpose_x = x_191_transpose_x_0, transpose_y = x_191_transpose_y_0, x = var_2150_cast_fp16, y = value_37_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<int32, [4]> var_2154_perm_0 = const()[name = string("op_2154_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2156 = const()[name = string("op_2156"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2154_cast_fp16 = transpose(perm = var_2154_perm_0, x = x_191_cast_fp16)[name = string("transpose_656")];
            tensor<fp16, [1, 304, 512]> input_559_cast_fp16 = reshape(shape = var_2156, x = var_2154_cast_fp16)[name = string("input_559_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_17_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_17_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115786176)))];
            tensor<fp16, [512]> m_encoder_encoders_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_17_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116310528)))];
            tensor<fp16, [1, 304, 512]> linear_73_cast_fp16 = linear(bias = m_encoder_encoders_17_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_17_self_attn_linear_out_weight_to_fp16, x = input_559_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_561_cast_fp16 = add(x = linear_73_cast_fp16, y = input_555_cast_fp16)[name = string("input_561_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_563_cast_fp16 = add(x = input_547_cast_fp16, y = input_561_cast_fp16)[name = string("input_563_cast_fp16")];
            tensor<int32, [1]> output_75_axes_0 = const()[name = string("output_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_227_to_fp16 = const()[name = string("const_227_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116311616)))];
            tensor<fp16, [512]> const_228_to_fp16 = const()[name = string("const_228_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116312704)))];
            tensor<fp16, [1, 304, 512]> output_75_cast_fp16 = layer_norm(axes = output_75_axes_0, beta = const_228_to_fp16, epsilon = var_22_to_fp16, gamma = const_227_to_fp16, x = input_563_cast_fp16)[name = string("output_75_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_17_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_17_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116313792)))];
            tensor<fp16, [2048]> m_encoder_encoders_17_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_17_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118411008)))];
            tensor<fp16, [1, 304, 2048]> linear_74_cast_fp16 = linear(bias = m_encoder_encoders_17_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_17_feed_forward_w_1_weight_to_fp16, x = output_75_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_571_cast_fp16 = relu(x = linear_74_cast_fp16)[name = string("input_571_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_17_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_17_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118415168)))];
            tensor<fp16, [512]> m_encoder_encoders_17_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_17_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120512384)))];
            tensor<fp16, [1, 304, 512]> linear_75_cast_fp16 = linear(bias = m_encoder_encoders_17_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_17_feed_forward_w_2_weight_to_fp16, x = input_571_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_577_cast_fp16 = add(x = input_563_cast_fp16, y = linear_75_cast_fp16)[name = string("input_577_cast_fp16")];
            tensor<int32, [1]> output_77_axes_0 = const()[name = string("output_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_229_to_fp16 = const()[name = string("const_229_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120513472)))];
            tensor<fp16, [512]> const_230_to_fp16 = const()[name = string("const_230_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120514560)))];
            tensor<fp16, [1, 304, 512]> output_77_cast_fp16 = layer_norm(axes = output_77_axes_0, beta = const_230_to_fp16, epsilon = var_22_to_fp16, gamma = const_229_to_fp16, x = input_577_cast_fp16)[name = string("output_77_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_18_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_18_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120515648)))];
            tensor<fp16, [1536]> m_encoder_encoders_18_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_18_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122088576)))];
            tensor<fp16, [1, 304, 1536]> linear_76_cast_fp16 = linear(bias = m_encoder_encoders_18_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_18_self_attn_linear_q_k_v_weight_to_fp16, x = output_77_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<int32, [3]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2214_axis_0 = const()[name = string("op_2214_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2214_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2214_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2214_cast_fp16_2 = split(axis = var_2214_axis_0, split_sizes = tile_19, x = linear_76_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [4]> var_2218 = const()[name = string("op_2218"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2219_cast_fp16 = reshape(shape = var_2218, x = var_2214_cast_fp16_0)[name = string("op_2219_cast_fp16")];
            tensor<int32, [4]> var_2221 = const()[name = string("op_2221"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2222_cast_fp16 = reshape(shape = var_2221, x = var_2214_cast_fp16_1)[name = string("op_2222_cast_fp16")];
            tensor<int32, [4]> var_2224 = const()[name = string("op_2224"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2225_cast_fp16 = reshape(shape = var_2224, x = var_2214_cast_fp16_2)[name = string("op_2225_cast_fp16")];
            tensor<int32, [4]> value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_583_pad_0 = const()[name = string("input_583_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_583_mode_0 = const()[name = string("input_583_mode_0"), val = string("constant")];
            fp16 const_236_to_fp16 = const()[name = string("const_236_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = var_2214_cast_fp16_2)[name = string("transpose_654")];
            tensor<fp16, [1, 512, 314]> input_583_cast_fp16 = pad(constant_val = const_236_to_fp16, mode = input_583_mode_0, pad = input_583_pad_0, x = input_581_cast_fp16)[name = string("input_583_cast_fp16")];
            string x_195_pad_type_0 = const()[name = string("x_195_pad_type_0"), val = string("valid")];
            int32 x_195_groups_0 = const()[name = string("x_195_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_195_strides_0 = const()[name = string("x_195_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_195_pad_0 = const()[name = string("x_195_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_195_dilations_0 = const()[name = string("x_195_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_18_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_18_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122091712)))];
            tensor<fp16, [1, 512, 304]> x_195_cast_fp16 = conv(dilations = x_195_dilations_0, groups = x_195_groups_0, pad = x_195_pad_0, pad_type = x_195_pad_type_0, strides = x_195_strides_0, weight = m_encoder_encoders_18_self_attn_fsmn_block_weight_to_fp16, x = input_583_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<int32, [3]> x_197_perm_0 = const()[name = string("x_197_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_197_cast_fp16 = transpose(perm = x_197_perm_0, x = x_195_cast_fp16)[name = string("transpose_653")];
            tensor<fp16, [1, 304, 512]> input_585_cast_fp16 = add(x = x_197_cast_fp16, y = var_2214_cast_fp16_2)[name = string("input_585_cast_fp16")];
            fp16 var_2244_to_fp16 = const()[name = string("op_2244_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_79_cast_fp16 = mul(x = var_2219_cast_fp16, y = var_2244_to_fp16)[name = string("q_h_79_cast_fp16")];
            bool scores_77_transpose_x_0 = const()[name = string("scores_77_transpose_x_0"), val = bool(false)];
            bool scores_77_transpose_y_0 = const()[name = string("scores_77_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_248_perm_0 = const()[name = string("transpose_248_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_249_perm_0 = const()[name = string("transpose_249_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_249 = transpose(perm = transpose_249_perm_0, x = var_2222_cast_fp16)[name = string("transpose_651")];
            tensor<fp16, [1, 4, 304, 128]> transpose_248 = transpose(perm = transpose_248_perm_0, x = q_h_79_cast_fp16)[name = string("transpose_652")];
            tensor<fp16, [1, 4, 304, 304]> scores_77_cast_fp16 = matmul(transpose_x = scores_77_transpose_x_0, transpose_y = scores_77_transpose_y_0, x = transpose_248, y = transpose_249)[name = string("scores_77_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2252_cast_fp16 = softmax(axis = var_29, x = scores_77_cast_fp16)[name = string("op_2252_cast_fp16")];
            bool x_201_transpose_x_0 = const()[name = string("x_201_transpose_x_0"), val = bool(false)];
            bool x_201_transpose_y_0 = const()[name = string("x_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = var_2225_cast_fp16)[name = string("transpose_655")];
            tensor<fp16, [1, 4, 304, 128]> x_201_cast_fp16 = matmul(transpose_x = x_201_transpose_x_0, transpose_y = x_201_transpose_y_0, x = var_2252_cast_fp16, y = value_39_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<int32, [4]> var_2256_perm_0 = const()[name = string("op_2256_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2258 = const()[name = string("op_2258"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2256_cast_fp16 = transpose(perm = var_2256_perm_0, x = x_201_cast_fp16)[name = string("transpose_650")];
            tensor<fp16, [1, 304, 512]> input_589_cast_fp16 = reshape(shape = var_2258, x = var_2256_cast_fp16)[name = string("input_589_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_18_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_18_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122103040)))];
            tensor<fp16, [512]> m_encoder_encoders_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_18_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122627392)))];
            tensor<fp16, [1, 304, 512]> linear_77_cast_fp16 = linear(bias = m_encoder_encoders_18_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_18_self_attn_linear_out_weight_to_fp16, x = input_589_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_591_cast_fp16 = add(x = linear_77_cast_fp16, y = input_585_cast_fp16)[name = string("input_591_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_593_cast_fp16 = add(x = input_577_cast_fp16, y = input_591_cast_fp16)[name = string("input_593_cast_fp16")];
            tensor<int32, [1]> output_79_axes_0 = const()[name = string("output_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_238_to_fp16 = const()[name = string("const_238_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122628480)))];
            tensor<fp16, [512]> const_239_to_fp16 = const()[name = string("const_239_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122629568)))];
            tensor<fp16, [1, 304, 512]> output_79_cast_fp16 = layer_norm(axes = output_79_axes_0, beta = const_239_to_fp16, epsilon = var_22_to_fp16, gamma = const_238_to_fp16, x = input_593_cast_fp16)[name = string("output_79_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_18_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_18_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122630656)))];
            tensor<fp16, [2048]> m_encoder_encoders_18_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_18_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124727872)))];
            tensor<fp16, [1, 304, 2048]> linear_78_cast_fp16 = linear(bias = m_encoder_encoders_18_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_18_feed_forward_w_1_weight_to_fp16, x = output_79_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_601_cast_fp16 = relu(x = linear_78_cast_fp16)[name = string("input_601_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_18_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_18_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124732032)))];
            tensor<fp16, [512]> m_encoder_encoders_18_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_18_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126829248)))];
            tensor<fp16, [1, 304, 512]> linear_79_cast_fp16 = linear(bias = m_encoder_encoders_18_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_18_feed_forward_w_2_weight_to_fp16, x = input_601_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_607_cast_fp16 = add(x = input_593_cast_fp16, y = linear_79_cast_fp16)[name = string("input_607_cast_fp16")];
            tensor<int32, [1]> output_81_axes_0 = const()[name = string("output_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_240_to_fp16 = const()[name = string("const_240_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126830336)))];
            tensor<fp16, [512]> const_241_to_fp16 = const()[name = string("const_241_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126831424)))];
            tensor<fp16, [1, 304, 512]> output_81_cast_fp16 = layer_norm(axes = output_81_axes_0, beta = const_241_to_fp16, epsilon = var_22_to_fp16, gamma = const_240_to_fp16, x = input_607_cast_fp16)[name = string("output_81_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_19_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_19_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126832512)))];
            tensor<fp16, [1536]> m_encoder_encoders_19_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_19_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128405440)))];
            tensor<fp16, [1, 304, 1536]> linear_80_cast_fp16 = linear(bias = m_encoder_encoders_19_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_19_self_attn_linear_q_k_v_weight_to_fp16, x = output_81_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [3]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2316_axis_0 = const()[name = string("op_2316_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2316_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2316_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2316_cast_fp16_2 = split(axis = var_2316_axis_0, split_sizes = tile_20, x = linear_80_cast_fp16)[name = string("op_2316_cast_fp16")];
            tensor<int32, [4]> var_2320 = const()[name = string("op_2320"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2321_cast_fp16 = reshape(shape = var_2320, x = var_2316_cast_fp16_0)[name = string("op_2321_cast_fp16")];
            tensor<int32, [4]> var_2323 = const()[name = string("op_2323"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2324_cast_fp16 = reshape(shape = var_2323, x = var_2316_cast_fp16_1)[name = string("op_2324_cast_fp16")];
            tensor<int32, [4]> var_2326 = const()[name = string("op_2326"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2327_cast_fp16 = reshape(shape = var_2326, x = var_2316_cast_fp16_2)[name = string("op_2327_cast_fp16")];
            tensor<int32, [4]> value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_611_perm_0 = const()[name = string("input_611_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_613_pad_0 = const()[name = string("input_613_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_613_mode_0 = const()[name = string("input_613_mode_0"), val = string("constant")];
            fp16 const_247_to_fp16 = const()[name = string("const_247_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_611_cast_fp16 = transpose(perm = input_611_perm_0, x = var_2316_cast_fp16_2)[name = string("transpose_648")];
            tensor<fp16, [1, 512, 314]> input_613_cast_fp16 = pad(constant_val = const_247_to_fp16, mode = input_613_mode_0, pad = input_613_pad_0, x = input_611_cast_fp16)[name = string("input_613_cast_fp16")];
            string x_205_pad_type_0 = const()[name = string("x_205_pad_type_0"), val = string("valid")];
            int32 x_205_groups_0 = const()[name = string("x_205_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_205_strides_0 = const()[name = string("x_205_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_205_pad_0 = const()[name = string("x_205_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_205_dilations_0 = const()[name = string("x_205_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_19_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_19_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128408576)))];
            tensor<fp16, [1, 512, 304]> x_205_cast_fp16 = conv(dilations = x_205_dilations_0, groups = x_205_groups_0, pad = x_205_pad_0, pad_type = x_205_pad_type_0, strides = x_205_strides_0, weight = m_encoder_encoders_19_self_attn_fsmn_block_weight_to_fp16, x = input_613_cast_fp16)[name = string("x_205_cast_fp16")];
            tensor<int32, [3]> x_207_perm_0 = const()[name = string("x_207_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_207_cast_fp16 = transpose(perm = x_207_perm_0, x = x_205_cast_fp16)[name = string("transpose_647")];
            tensor<fp16, [1, 304, 512]> input_615_cast_fp16 = add(x = x_207_cast_fp16, y = var_2316_cast_fp16_2)[name = string("input_615_cast_fp16")];
            fp16 var_2346_to_fp16 = const()[name = string("op_2346_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_83_cast_fp16 = mul(x = var_2321_cast_fp16, y = var_2346_to_fp16)[name = string("q_h_83_cast_fp16")];
            bool scores_81_transpose_x_0 = const()[name = string("scores_81_transpose_x_0"), val = bool(false)];
            bool scores_81_transpose_y_0 = const()[name = string("scores_81_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_250_perm_0 = const()[name = string("transpose_250_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_251_perm_0 = const()[name = string("transpose_251_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_251 = transpose(perm = transpose_251_perm_0, x = var_2324_cast_fp16)[name = string("transpose_645")];
            tensor<fp16, [1, 4, 304, 128]> transpose_250 = transpose(perm = transpose_250_perm_0, x = q_h_83_cast_fp16)[name = string("transpose_646")];
            tensor<fp16, [1, 4, 304, 304]> scores_81_cast_fp16 = matmul(transpose_x = scores_81_transpose_x_0, transpose_y = scores_81_transpose_y_0, x = transpose_250, y = transpose_251)[name = string("scores_81_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2354_cast_fp16 = softmax(axis = var_29, x = scores_81_cast_fp16)[name = string("op_2354_cast_fp16")];
            bool x_211_transpose_x_0 = const()[name = string("x_211_transpose_x_0"), val = bool(false)];
            bool x_211_transpose_y_0 = const()[name = string("x_211_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = var_2327_cast_fp16)[name = string("transpose_649")];
            tensor<fp16, [1, 4, 304, 128]> x_211_cast_fp16 = matmul(transpose_x = x_211_transpose_x_0, transpose_y = x_211_transpose_y_0, x = var_2354_cast_fp16, y = value_41_cast_fp16)[name = string("x_211_cast_fp16")];
            tensor<int32, [4]> var_2358_perm_0 = const()[name = string("op_2358_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2360 = const()[name = string("op_2360"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2358_cast_fp16 = transpose(perm = var_2358_perm_0, x = x_211_cast_fp16)[name = string("transpose_644")];
            tensor<fp16, [1, 304, 512]> input_619_cast_fp16 = reshape(shape = var_2360, x = var_2358_cast_fp16)[name = string("input_619_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_19_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_19_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128419904)))];
            tensor<fp16, [512]> m_encoder_encoders_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_19_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128944256)))];
            tensor<fp16, [1, 304, 512]> linear_81_cast_fp16 = linear(bias = m_encoder_encoders_19_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_19_self_attn_linear_out_weight_to_fp16, x = input_619_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_621_cast_fp16 = add(x = linear_81_cast_fp16, y = input_615_cast_fp16)[name = string("input_621_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_623_cast_fp16 = add(x = input_607_cast_fp16, y = input_621_cast_fp16)[name = string("input_623_cast_fp16")];
            tensor<int32, [1]> output_83_axes_0 = const()[name = string("output_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_249_to_fp16 = const()[name = string("const_249_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128945344)))];
            tensor<fp16, [512]> const_250_to_fp16 = const()[name = string("const_250_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128946432)))];
            tensor<fp16, [1, 304, 512]> output_83_cast_fp16 = layer_norm(axes = output_83_axes_0, beta = const_250_to_fp16, epsilon = var_22_to_fp16, gamma = const_249_to_fp16, x = input_623_cast_fp16)[name = string("output_83_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_19_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_19_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128947520)))];
            tensor<fp16, [2048]> m_encoder_encoders_19_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_19_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131044736)))];
            tensor<fp16, [1, 304, 2048]> linear_82_cast_fp16 = linear(bias = m_encoder_encoders_19_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_19_feed_forward_w_1_weight_to_fp16, x = output_83_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_631_cast_fp16 = relu(x = linear_82_cast_fp16)[name = string("input_631_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_19_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_19_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131048896)))];
            tensor<fp16, [512]> m_encoder_encoders_19_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_19_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133146112)))];
            tensor<fp16, [1, 304, 512]> linear_83_cast_fp16 = linear(bias = m_encoder_encoders_19_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_19_feed_forward_w_2_weight_to_fp16, x = input_631_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_637_cast_fp16 = add(x = input_623_cast_fp16, y = linear_83_cast_fp16)[name = string("input_637_cast_fp16")];
            tensor<int32, [1]> output_85_axes_0 = const()[name = string("output_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_251_to_fp16 = const()[name = string("const_251_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133147200)))];
            tensor<fp16, [512]> const_252_to_fp16 = const()[name = string("const_252_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133148288)))];
            tensor<fp16, [1, 304, 512]> output_85_cast_fp16 = layer_norm(axes = output_85_axes_0, beta = const_252_to_fp16, epsilon = var_22_to_fp16, gamma = const_251_to_fp16, x = input_637_cast_fp16)[name = string("output_85_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_20_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_20_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133149376)))];
            tensor<fp16, [1536]> m_encoder_encoders_20_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_20_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134722304)))];
            tensor<fp16, [1, 304, 1536]> linear_84_cast_fp16 = linear(bias = m_encoder_encoders_20_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_20_self_attn_linear_q_k_v_weight_to_fp16, x = output_85_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [3]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2418_axis_0 = const()[name = string("op_2418_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2418_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2418_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2418_cast_fp16_2 = split(axis = var_2418_axis_0, split_sizes = tile_21, x = linear_84_cast_fp16)[name = string("op_2418_cast_fp16")];
            tensor<int32, [4]> var_2422 = const()[name = string("op_2422"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2423_cast_fp16 = reshape(shape = var_2422, x = var_2418_cast_fp16_0)[name = string("op_2423_cast_fp16")];
            tensor<int32, [4]> var_2425 = const()[name = string("op_2425"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2426_cast_fp16 = reshape(shape = var_2425, x = var_2418_cast_fp16_1)[name = string("op_2426_cast_fp16")];
            tensor<int32, [4]> var_2428 = const()[name = string("op_2428"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2429_cast_fp16 = reshape(shape = var_2428, x = var_2418_cast_fp16_2)[name = string("op_2429_cast_fp16")];
            tensor<int32, [4]> value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_641_perm_0 = const()[name = string("input_641_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_643_pad_0 = const()[name = string("input_643_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_643_mode_0 = const()[name = string("input_643_mode_0"), val = string("constant")];
            fp16 const_258_to_fp16 = const()[name = string("const_258_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_641_cast_fp16 = transpose(perm = input_641_perm_0, x = var_2418_cast_fp16_2)[name = string("transpose_642")];
            tensor<fp16, [1, 512, 314]> input_643_cast_fp16 = pad(constant_val = const_258_to_fp16, mode = input_643_mode_0, pad = input_643_pad_0, x = input_641_cast_fp16)[name = string("input_643_cast_fp16")];
            string x_215_pad_type_0 = const()[name = string("x_215_pad_type_0"), val = string("valid")];
            int32 x_215_groups_0 = const()[name = string("x_215_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_215_strides_0 = const()[name = string("x_215_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_215_pad_0 = const()[name = string("x_215_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_215_dilations_0 = const()[name = string("x_215_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_20_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_20_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134725440)))];
            tensor<fp16, [1, 512, 304]> x_215_cast_fp16 = conv(dilations = x_215_dilations_0, groups = x_215_groups_0, pad = x_215_pad_0, pad_type = x_215_pad_type_0, strides = x_215_strides_0, weight = m_encoder_encoders_20_self_attn_fsmn_block_weight_to_fp16, x = input_643_cast_fp16)[name = string("x_215_cast_fp16")];
            tensor<int32, [3]> x_217_perm_0 = const()[name = string("x_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_217_cast_fp16 = transpose(perm = x_217_perm_0, x = x_215_cast_fp16)[name = string("transpose_641")];
            tensor<fp16, [1, 304, 512]> input_645_cast_fp16 = add(x = x_217_cast_fp16, y = var_2418_cast_fp16_2)[name = string("input_645_cast_fp16")];
            fp16 var_2448_to_fp16 = const()[name = string("op_2448_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_87_cast_fp16 = mul(x = var_2423_cast_fp16, y = var_2448_to_fp16)[name = string("q_h_87_cast_fp16")];
            bool scores_85_transpose_x_0 = const()[name = string("scores_85_transpose_x_0"), val = bool(false)];
            bool scores_85_transpose_y_0 = const()[name = string("scores_85_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_252_perm_0 = const()[name = string("transpose_252_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_253_perm_0 = const()[name = string("transpose_253_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_253 = transpose(perm = transpose_253_perm_0, x = var_2426_cast_fp16)[name = string("transpose_639")];
            tensor<fp16, [1, 4, 304, 128]> transpose_252 = transpose(perm = transpose_252_perm_0, x = q_h_87_cast_fp16)[name = string("transpose_640")];
            tensor<fp16, [1, 4, 304, 304]> scores_85_cast_fp16 = matmul(transpose_x = scores_85_transpose_x_0, transpose_y = scores_85_transpose_y_0, x = transpose_252, y = transpose_253)[name = string("scores_85_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2456_cast_fp16 = softmax(axis = var_29, x = scores_85_cast_fp16)[name = string("op_2456_cast_fp16")];
            bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)];
            bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = var_2429_cast_fp16)[name = string("transpose_643")];
            tensor<fp16, [1, 4, 304, 128]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = var_2456_cast_fp16, y = value_43_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<int32, [4]> var_2460_perm_0 = const()[name = string("op_2460_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2462 = const()[name = string("op_2462"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2460_cast_fp16 = transpose(perm = var_2460_perm_0, x = x_221_cast_fp16)[name = string("transpose_638")];
            tensor<fp16, [1, 304, 512]> input_649_cast_fp16 = reshape(shape = var_2462, x = var_2460_cast_fp16)[name = string("input_649_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_20_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_20_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134736768)))];
            tensor<fp16, [512]> m_encoder_encoders_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_20_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135261120)))];
            tensor<fp16, [1, 304, 512]> linear_85_cast_fp16 = linear(bias = m_encoder_encoders_20_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_20_self_attn_linear_out_weight_to_fp16, x = input_649_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_651_cast_fp16 = add(x = linear_85_cast_fp16, y = input_645_cast_fp16)[name = string("input_651_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_653_cast_fp16 = add(x = input_637_cast_fp16, y = input_651_cast_fp16)[name = string("input_653_cast_fp16")];
            tensor<int32, [1]> output_87_axes_0 = const()[name = string("output_87_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_260_to_fp16 = const()[name = string("const_260_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135262208)))];
            tensor<fp16, [512]> const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135263296)))];
            tensor<fp16, [1, 304, 512]> output_87_cast_fp16 = layer_norm(axes = output_87_axes_0, beta = const_261_to_fp16, epsilon = var_22_to_fp16, gamma = const_260_to_fp16, x = input_653_cast_fp16)[name = string("output_87_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_20_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_20_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135264384)))];
            tensor<fp16, [2048]> m_encoder_encoders_20_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_20_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137361600)))];
            tensor<fp16, [1, 304, 2048]> linear_86_cast_fp16 = linear(bias = m_encoder_encoders_20_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_20_feed_forward_w_1_weight_to_fp16, x = output_87_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_661_cast_fp16 = relu(x = linear_86_cast_fp16)[name = string("input_661_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_20_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_20_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137365760)))];
            tensor<fp16, [512]> m_encoder_encoders_20_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_20_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139462976)))];
            tensor<fp16, [1, 304, 512]> linear_87_cast_fp16 = linear(bias = m_encoder_encoders_20_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_20_feed_forward_w_2_weight_to_fp16, x = input_661_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_667_cast_fp16 = add(x = input_653_cast_fp16, y = linear_87_cast_fp16)[name = string("input_667_cast_fp16")];
            tensor<int32, [1]> output_89_axes_0 = const()[name = string("output_89_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139464064)))];
            tensor<fp16, [512]> const_263_to_fp16 = const()[name = string("const_263_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139465152)))];
            tensor<fp16, [1, 304, 512]> output_89_cast_fp16 = layer_norm(axes = output_89_axes_0, beta = const_263_to_fp16, epsilon = var_22_to_fp16, gamma = const_262_to_fp16, x = input_667_cast_fp16)[name = string("output_89_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_21_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_21_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139466240)))];
            tensor<fp16, [1536]> m_encoder_encoders_21_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_21_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141039168)))];
            tensor<fp16, [1, 304, 1536]> linear_88_cast_fp16 = linear(bias = m_encoder_encoders_21_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_21_self_attn_linear_q_k_v_weight_to_fp16, x = output_89_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<int32, [3]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2520_axis_0 = const()[name = string("op_2520_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2520_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2520_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2520_cast_fp16_2 = split(axis = var_2520_axis_0, split_sizes = tile_22, x = linear_88_cast_fp16)[name = string("op_2520_cast_fp16")];
            tensor<int32, [4]> var_2524 = const()[name = string("op_2524"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2525_cast_fp16 = reshape(shape = var_2524, x = var_2520_cast_fp16_0)[name = string("op_2525_cast_fp16")];
            tensor<int32, [4]> var_2527 = const()[name = string("op_2527"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2528_cast_fp16 = reshape(shape = var_2527, x = var_2520_cast_fp16_1)[name = string("op_2528_cast_fp16")];
            tensor<int32, [4]> var_2530 = const()[name = string("op_2530"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2531_cast_fp16 = reshape(shape = var_2530, x = var_2520_cast_fp16_2)[name = string("op_2531_cast_fp16")];
            tensor<int32, [4]> value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_671_perm_0 = const()[name = string("input_671_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_673_pad_0 = const()[name = string("input_673_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_673_mode_0 = const()[name = string("input_673_mode_0"), val = string("constant")];
            fp16 const_269_to_fp16 = const()[name = string("const_269_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_671_cast_fp16 = transpose(perm = input_671_perm_0, x = var_2520_cast_fp16_2)[name = string("transpose_636")];
            tensor<fp16, [1, 512, 314]> input_673_cast_fp16 = pad(constant_val = const_269_to_fp16, mode = input_673_mode_0, pad = input_673_pad_0, x = input_671_cast_fp16)[name = string("input_673_cast_fp16")];
            string x_225_pad_type_0 = const()[name = string("x_225_pad_type_0"), val = string("valid")];
            int32 x_225_groups_0 = const()[name = string("x_225_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_225_strides_0 = const()[name = string("x_225_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_225_pad_0 = const()[name = string("x_225_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_225_dilations_0 = const()[name = string("x_225_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_21_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_21_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141042304)))];
            tensor<fp16, [1, 512, 304]> x_225_cast_fp16 = conv(dilations = x_225_dilations_0, groups = x_225_groups_0, pad = x_225_pad_0, pad_type = x_225_pad_type_0, strides = x_225_strides_0, weight = m_encoder_encoders_21_self_attn_fsmn_block_weight_to_fp16, x = input_673_cast_fp16)[name = string("x_225_cast_fp16")];
            tensor<int32, [3]> x_227_perm_0 = const()[name = string("x_227_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_227_cast_fp16 = transpose(perm = x_227_perm_0, x = x_225_cast_fp16)[name = string("transpose_635")];
            tensor<fp16, [1, 304, 512]> input_675_cast_fp16 = add(x = x_227_cast_fp16, y = var_2520_cast_fp16_2)[name = string("input_675_cast_fp16")];
            fp16 var_2550_to_fp16 = const()[name = string("op_2550_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_91_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2550_to_fp16)[name = string("q_h_91_cast_fp16")];
            bool scores_89_transpose_x_0 = const()[name = string("scores_89_transpose_x_0"), val = bool(false)];
            bool scores_89_transpose_y_0 = const()[name = string("scores_89_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_254_perm_0 = const()[name = string("transpose_254_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_255_perm_0 = const()[name = string("transpose_255_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_255 = transpose(perm = transpose_255_perm_0, x = var_2528_cast_fp16)[name = string("transpose_633")];
            tensor<fp16, [1, 4, 304, 128]> transpose_254 = transpose(perm = transpose_254_perm_0, x = q_h_91_cast_fp16)[name = string("transpose_634")];
            tensor<fp16, [1, 4, 304, 304]> scores_89_cast_fp16 = matmul(transpose_x = scores_89_transpose_x_0, transpose_y = scores_89_transpose_y_0, x = transpose_254, y = transpose_255)[name = string("scores_89_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2558_cast_fp16 = softmax(axis = var_29, x = scores_89_cast_fp16)[name = string("op_2558_cast_fp16")];
            bool x_231_transpose_x_0 = const()[name = string("x_231_transpose_x_0"), val = bool(false)];
            bool x_231_transpose_y_0 = const()[name = string("x_231_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = var_2531_cast_fp16)[name = string("transpose_637")];
            tensor<fp16, [1, 4, 304, 128]> x_231_cast_fp16 = matmul(transpose_x = x_231_transpose_x_0, transpose_y = x_231_transpose_y_0, x = var_2558_cast_fp16, y = value_45_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [4]> var_2562_perm_0 = const()[name = string("op_2562_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2564 = const()[name = string("op_2564"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2562_cast_fp16 = transpose(perm = var_2562_perm_0, x = x_231_cast_fp16)[name = string("transpose_632")];
            tensor<fp16, [1, 304, 512]> input_679_cast_fp16 = reshape(shape = var_2564, x = var_2562_cast_fp16)[name = string("input_679_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_21_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141053632)))];
            tensor<fp16, [512]> m_encoder_encoders_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_21_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141577984)))];
            tensor<fp16, [1, 304, 512]> linear_89_cast_fp16 = linear(bias = m_encoder_encoders_21_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_21_self_attn_linear_out_weight_to_fp16, x = input_679_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_681_cast_fp16 = add(x = linear_89_cast_fp16, y = input_675_cast_fp16)[name = string("input_681_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_683_cast_fp16 = add(x = input_667_cast_fp16, y = input_681_cast_fp16)[name = string("input_683_cast_fp16")];
            tensor<int32, [1]> output_91_axes_0 = const()[name = string("output_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_271_to_fp16 = const()[name = string("const_271_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141579072)))];
            tensor<fp16, [512]> const_272_to_fp16 = const()[name = string("const_272_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141580160)))];
            tensor<fp16, [1, 304, 512]> output_91_cast_fp16 = layer_norm(axes = output_91_axes_0, beta = const_272_to_fp16, epsilon = var_22_to_fp16, gamma = const_271_to_fp16, x = input_683_cast_fp16)[name = string("output_91_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_21_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_21_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141581248)))];
            tensor<fp16, [2048]> m_encoder_encoders_21_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_21_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143678464)))];
            tensor<fp16, [1, 304, 2048]> linear_90_cast_fp16 = linear(bias = m_encoder_encoders_21_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_21_feed_forward_w_1_weight_to_fp16, x = output_91_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_691_cast_fp16 = relu(x = linear_90_cast_fp16)[name = string("input_691_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_21_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_21_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143682624)))];
            tensor<fp16, [512]> m_encoder_encoders_21_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_21_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145779840)))];
            tensor<fp16, [1, 304, 512]> linear_91_cast_fp16 = linear(bias = m_encoder_encoders_21_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_21_feed_forward_w_2_weight_to_fp16, x = input_691_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_697_cast_fp16 = add(x = input_683_cast_fp16, y = linear_91_cast_fp16)[name = string("input_697_cast_fp16")];
            tensor<int32, [1]> output_93_axes_0 = const()[name = string("output_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_273_to_fp16 = const()[name = string("const_273_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145780928)))];
            tensor<fp16, [512]> const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145782016)))];
            tensor<fp16, [1, 304, 512]> output_93_cast_fp16 = layer_norm(axes = output_93_axes_0, beta = const_274_to_fp16, epsilon = var_22_to_fp16, gamma = const_273_to_fp16, x = input_697_cast_fp16)[name = string("output_93_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_22_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_22_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145783104)))];
            tensor<fp16, [1536]> m_encoder_encoders_22_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_22_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147356032)))];
            tensor<fp16, [1, 304, 1536]> linear_92_cast_fp16 = linear(bias = m_encoder_encoders_22_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_22_self_attn_linear_q_k_v_weight_to_fp16, x = output_93_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [3]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2622_axis_0 = const()[name = string("op_2622_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2622_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2622_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2622_cast_fp16_2 = split(axis = var_2622_axis_0, split_sizes = tile_23, x = linear_92_cast_fp16)[name = string("op_2622_cast_fp16")];
            tensor<int32, [4]> var_2626 = const()[name = string("op_2626"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2627_cast_fp16 = reshape(shape = var_2626, x = var_2622_cast_fp16_0)[name = string("op_2627_cast_fp16")];
            tensor<int32, [4]> var_2629 = const()[name = string("op_2629"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2630_cast_fp16 = reshape(shape = var_2629, x = var_2622_cast_fp16_1)[name = string("op_2630_cast_fp16")];
            tensor<int32, [4]> var_2632 = const()[name = string("op_2632"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2633_cast_fp16 = reshape(shape = var_2632, x = var_2622_cast_fp16_2)[name = string("op_2633_cast_fp16")];
            tensor<int32, [4]> value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_701_perm_0 = const()[name = string("input_701_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_703_pad_0 = const()[name = string("input_703_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_703_mode_0 = const()[name = string("input_703_mode_0"), val = string("constant")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_701_cast_fp16 = transpose(perm = input_701_perm_0, x = var_2622_cast_fp16_2)[name = string("transpose_630")];
            tensor<fp16, [1, 512, 314]> input_703_cast_fp16 = pad(constant_val = const_280_to_fp16, mode = input_703_mode_0, pad = input_703_pad_0, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")];
            string x_235_pad_type_0 = const()[name = string("x_235_pad_type_0"), val = string("valid")];
            int32 x_235_groups_0 = const()[name = string("x_235_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_235_strides_0 = const()[name = string("x_235_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_235_pad_0 = const()[name = string("x_235_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_235_dilations_0 = const()[name = string("x_235_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_22_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_22_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147359168)))];
            tensor<fp16, [1, 512, 304]> x_235_cast_fp16 = conv(dilations = x_235_dilations_0, groups = x_235_groups_0, pad = x_235_pad_0, pad_type = x_235_pad_type_0, strides = x_235_strides_0, weight = m_encoder_encoders_22_self_attn_fsmn_block_weight_to_fp16, x = input_703_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<int32, [3]> x_237_perm_0 = const()[name = string("x_237_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_237_cast_fp16 = transpose(perm = x_237_perm_0, x = x_235_cast_fp16)[name = string("transpose_629")];
            tensor<fp16, [1, 304, 512]> input_705_cast_fp16 = add(x = x_237_cast_fp16, y = var_2622_cast_fp16_2)[name = string("input_705_cast_fp16")];
            fp16 var_2652_to_fp16 = const()[name = string("op_2652_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_95_cast_fp16 = mul(x = var_2627_cast_fp16, y = var_2652_to_fp16)[name = string("q_h_95_cast_fp16")];
            bool scores_93_transpose_x_0 = const()[name = string("scores_93_transpose_x_0"), val = bool(false)];
            bool scores_93_transpose_y_0 = const()[name = string("scores_93_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_256_perm_0 = const()[name = string("transpose_256_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_257_perm_0 = const()[name = string("transpose_257_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_257 = transpose(perm = transpose_257_perm_0, x = var_2630_cast_fp16)[name = string("transpose_627")];
            tensor<fp16, [1, 4, 304, 128]> transpose_256 = transpose(perm = transpose_256_perm_0, x = q_h_95_cast_fp16)[name = string("transpose_628")];
            tensor<fp16, [1, 4, 304, 304]> scores_93_cast_fp16 = matmul(transpose_x = scores_93_transpose_x_0, transpose_y = scores_93_transpose_y_0, x = transpose_256, y = transpose_257)[name = string("scores_93_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2660_cast_fp16 = softmax(axis = var_29, x = scores_93_cast_fp16)[name = string("op_2660_cast_fp16")];
            bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)];
            bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = var_2633_cast_fp16)[name = string("transpose_631")];
            tensor<fp16, [1, 4, 304, 128]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = var_2660_cast_fp16, y = value_47_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<int32, [4]> var_2664_perm_0 = const()[name = string("op_2664_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2666 = const()[name = string("op_2666"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2664_cast_fp16 = transpose(perm = var_2664_perm_0, x = x_241_cast_fp16)[name = string("transpose_626")];
            tensor<fp16, [1, 304, 512]> input_709_cast_fp16 = reshape(shape = var_2666, x = var_2664_cast_fp16)[name = string("input_709_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_22_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147370496)))];
            tensor<fp16, [512]> m_encoder_encoders_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_22_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147894848)))];
            tensor<fp16, [1, 304, 512]> linear_93_cast_fp16 = linear(bias = m_encoder_encoders_22_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_22_self_attn_linear_out_weight_to_fp16, x = input_709_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_711_cast_fp16 = add(x = linear_93_cast_fp16, y = input_705_cast_fp16)[name = string("input_711_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_713_cast_fp16 = add(x = input_697_cast_fp16, y = input_711_cast_fp16)[name = string("input_713_cast_fp16")];
            tensor<int32, [1]> output_95_axes_0 = const()[name = string("output_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_282_to_fp16 = const()[name = string("const_282_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147895936)))];
            tensor<fp16, [512]> const_283_to_fp16 = const()[name = string("const_283_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147897024)))];
            tensor<fp16, [1, 304, 512]> output_95_cast_fp16 = layer_norm(axes = output_95_axes_0, beta = const_283_to_fp16, epsilon = var_22_to_fp16, gamma = const_282_to_fp16, x = input_713_cast_fp16)[name = string("output_95_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_22_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_22_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147898112)))];
            tensor<fp16, [2048]> m_encoder_encoders_22_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_22_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149995328)))];
            tensor<fp16, [1, 304, 2048]> linear_94_cast_fp16 = linear(bias = m_encoder_encoders_22_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_22_feed_forward_w_1_weight_to_fp16, x = output_95_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_721_cast_fp16 = relu(x = linear_94_cast_fp16)[name = string("input_721_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_22_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_22_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149999488)))];
            tensor<fp16, [512]> m_encoder_encoders_22_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_22_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152096704)))];
            tensor<fp16, [1, 304, 512]> linear_95_cast_fp16 = linear(bias = m_encoder_encoders_22_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_22_feed_forward_w_2_weight_to_fp16, x = input_721_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_727_cast_fp16 = add(x = input_713_cast_fp16, y = linear_95_cast_fp16)[name = string("input_727_cast_fp16")];
            tensor<int32, [1]> output_97_axes_0 = const()[name = string("output_97_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_284_to_fp16 = const()[name = string("const_284_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152097792)))];
            tensor<fp16, [512]> const_285_to_fp16 = const()[name = string("const_285_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152098880)))];
            tensor<fp16, [1, 304, 512]> output_97_cast_fp16 = layer_norm(axes = output_97_axes_0, beta = const_285_to_fp16, epsilon = var_22_to_fp16, gamma = const_284_to_fp16, x = input_727_cast_fp16)[name = string("output_97_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_23_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_23_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152099968)))];
            tensor<fp16, [1536]> m_encoder_encoders_23_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_23_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153672896)))];
            tensor<fp16, [1, 304, 1536]> linear_96_cast_fp16 = linear(bias = m_encoder_encoders_23_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_23_self_attn_linear_q_k_v_weight_to_fp16, x = output_97_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<int32, [3]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2724_axis_0 = const()[name = string("op_2724_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2724_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2724_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2724_cast_fp16_2 = split(axis = var_2724_axis_0, split_sizes = tile_24, x = linear_96_cast_fp16)[name = string("op_2724_cast_fp16")];
            tensor<int32, [4]> var_2728 = const()[name = string("op_2728"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2729_cast_fp16 = reshape(shape = var_2728, x = var_2724_cast_fp16_0)[name = string("op_2729_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2732_cast_fp16 = reshape(shape = var_2731, x = var_2724_cast_fp16_1)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> var_2734 = const()[name = string("op_2734"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2735_cast_fp16 = reshape(shape = var_2734, x = var_2724_cast_fp16_2)[name = string("op_2735_cast_fp16")];
            tensor<int32, [4]> value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_733_pad_0 = const()[name = string("input_733_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_733_mode_0 = const()[name = string("input_733_mode_0"), val = string("constant")];
            fp16 const_291_to_fp16 = const()[name = string("const_291_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = var_2724_cast_fp16_2)[name = string("transpose_624")];
            tensor<fp16, [1, 512, 314]> input_733_cast_fp16 = pad(constant_val = const_291_to_fp16, mode = input_733_mode_0, pad = input_733_pad_0, x = input_731_cast_fp16)[name = string("input_733_cast_fp16")];
            string x_245_pad_type_0 = const()[name = string("x_245_pad_type_0"), val = string("valid")];
            int32 x_245_groups_0 = const()[name = string("x_245_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_245_strides_0 = const()[name = string("x_245_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_245_pad_0 = const()[name = string("x_245_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_245_dilations_0 = const()[name = string("x_245_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_23_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_23_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153676032)))];
            tensor<fp16, [1, 512, 304]> x_245_cast_fp16 = conv(dilations = x_245_dilations_0, groups = x_245_groups_0, pad = x_245_pad_0, pad_type = x_245_pad_type_0, strides = x_245_strides_0, weight = m_encoder_encoders_23_self_attn_fsmn_block_weight_to_fp16, x = input_733_cast_fp16)[name = string("x_245_cast_fp16")];
            tensor<int32, [3]> x_247_perm_0 = const()[name = string("x_247_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_247_cast_fp16 = transpose(perm = x_247_perm_0, x = x_245_cast_fp16)[name = string("transpose_623")];
            tensor<fp16, [1, 304, 512]> input_735_cast_fp16 = add(x = x_247_cast_fp16, y = var_2724_cast_fp16_2)[name = string("input_735_cast_fp16")];
            fp16 var_2754_to_fp16 = const()[name = string("op_2754_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_99_cast_fp16 = mul(x = var_2729_cast_fp16, y = var_2754_to_fp16)[name = string("q_h_99_cast_fp16")];
            bool scores_97_transpose_x_0 = const()[name = string("scores_97_transpose_x_0"), val = bool(false)];
            bool scores_97_transpose_y_0 = const()[name = string("scores_97_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_258_perm_0 = const()[name = string("transpose_258_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_259_perm_0 = const()[name = string("transpose_259_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_259 = transpose(perm = transpose_259_perm_0, x = var_2732_cast_fp16)[name = string("transpose_621")];
            tensor<fp16, [1, 4, 304, 128]> transpose_258 = transpose(perm = transpose_258_perm_0, x = q_h_99_cast_fp16)[name = string("transpose_622")];
            tensor<fp16, [1, 4, 304, 304]> scores_97_cast_fp16 = matmul(transpose_x = scores_97_transpose_x_0, transpose_y = scores_97_transpose_y_0, x = transpose_258, y = transpose_259)[name = string("scores_97_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2762_cast_fp16 = softmax(axis = var_29, x = scores_97_cast_fp16)[name = string("op_2762_cast_fp16")];
            bool x_251_transpose_x_0 = const()[name = string("x_251_transpose_x_0"), val = bool(false)];
            bool x_251_transpose_y_0 = const()[name = string("x_251_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = var_2735_cast_fp16)[name = string("transpose_625")];
            tensor<fp16, [1, 4, 304, 128]> x_251_cast_fp16 = matmul(transpose_x = x_251_transpose_x_0, transpose_y = x_251_transpose_y_0, x = var_2762_cast_fp16, y = value_49_cast_fp16)[name = string("x_251_cast_fp16")];
            tensor<int32, [4]> var_2766_perm_0 = const()[name = string("op_2766_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2768 = const()[name = string("op_2768"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2766_cast_fp16 = transpose(perm = var_2766_perm_0, x = x_251_cast_fp16)[name = string("transpose_620")];
            tensor<fp16, [1, 304, 512]> input_739_cast_fp16 = reshape(shape = var_2768, x = var_2766_cast_fp16)[name = string("input_739_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_23_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153687360)))];
            tensor<fp16, [512]> m_encoder_encoders_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_23_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154211712)))];
            tensor<fp16, [1, 304, 512]> linear_97_cast_fp16 = linear(bias = m_encoder_encoders_23_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_23_self_attn_linear_out_weight_to_fp16, x = input_739_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_741_cast_fp16 = add(x = linear_97_cast_fp16, y = input_735_cast_fp16)[name = string("input_741_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_743_cast_fp16 = add(x = input_727_cast_fp16, y = input_741_cast_fp16)[name = string("input_743_cast_fp16")];
            tensor<int32, [1]> output_99_axes_0 = const()[name = string("output_99_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_293_to_fp16 = const()[name = string("const_293_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154212800)))];
            tensor<fp16, [512]> const_294_to_fp16 = const()[name = string("const_294_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154213888)))];
            tensor<fp16, [1, 304, 512]> output_99_cast_fp16 = layer_norm(axes = output_99_axes_0, beta = const_294_to_fp16, epsilon = var_22_to_fp16, gamma = const_293_to_fp16, x = input_743_cast_fp16)[name = string("output_99_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_23_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_23_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154214976)))];
            tensor<fp16, [2048]> m_encoder_encoders_23_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_23_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156312192)))];
            tensor<fp16, [1, 304, 2048]> linear_98_cast_fp16 = linear(bias = m_encoder_encoders_23_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_23_feed_forward_w_1_weight_to_fp16, x = output_99_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_751_cast_fp16 = relu(x = linear_98_cast_fp16)[name = string("input_751_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_23_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_23_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156316352)))];
            tensor<fp16, [512]> m_encoder_encoders_23_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_23_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158413568)))];
            tensor<fp16, [1, 304, 512]> linear_99_cast_fp16 = linear(bias = m_encoder_encoders_23_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_23_feed_forward_w_2_weight_to_fp16, x = input_751_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_757_cast_fp16 = add(x = input_743_cast_fp16, y = linear_99_cast_fp16)[name = string("input_757_cast_fp16")];
            tensor<int32, [1]> output_101_axes_0 = const()[name = string("output_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_295_to_fp16 = const()[name = string("const_295_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158414656)))];
            tensor<fp16, [512]> const_296_to_fp16 = const()[name = string("const_296_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158415744)))];
            tensor<fp16, [1, 304, 512]> output_101_cast_fp16 = layer_norm(axes = output_101_axes_0, beta = const_296_to_fp16, epsilon = var_22_to_fp16, gamma = const_295_to_fp16, x = input_757_cast_fp16)[name = string("output_101_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_24_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_24_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158416832)))];
            tensor<fp16, [1536]> m_encoder_encoders_24_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_24_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159989760)))];
            tensor<fp16, [1, 304, 1536]> linear_100_cast_fp16 = linear(bias = m_encoder_encoders_24_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_24_self_attn_linear_q_k_v_weight_to_fp16, x = output_101_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [3]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2826_axis_0 = const()[name = string("op_2826_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2826_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2826_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2826_cast_fp16_2 = split(axis = var_2826_axis_0, split_sizes = tile_25, x = linear_100_cast_fp16)[name = string("op_2826_cast_fp16")];
            tensor<int32, [4]> var_2830 = const()[name = string("op_2830"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2831_cast_fp16 = reshape(shape = var_2830, x = var_2826_cast_fp16_0)[name = string("op_2831_cast_fp16")];
            tensor<int32, [4]> var_2833 = const()[name = string("op_2833"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2834_cast_fp16 = reshape(shape = var_2833, x = var_2826_cast_fp16_1)[name = string("op_2834_cast_fp16")];
            tensor<int32, [4]> var_2836 = const()[name = string("op_2836"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2837_cast_fp16 = reshape(shape = var_2836, x = var_2826_cast_fp16_2)[name = string("op_2837_cast_fp16")];
            tensor<int32, [4]> value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_761_perm_0 = const()[name = string("input_761_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_763_pad_0 = const()[name = string("input_763_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_763_mode_0 = const()[name = string("input_763_mode_0"), val = string("constant")];
            fp16 const_302_to_fp16 = const()[name = string("const_302_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_761_cast_fp16 = transpose(perm = input_761_perm_0, x = var_2826_cast_fp16_2)[name = string("transpose_618")];
            tensor<fp16, [1, 512, 314]> input_763_cast_fp16 = pad(constant_val = const_302_to_fp16, mode = input_763_mode_0, pad = input_763_pad_0, x = input_761_cast_fp16)[name = string("input_763_cast_fp16")];
            string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")];
            int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_24_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_24_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159992896)))];
            tensor<fp16, [1, 512, 304]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = m_encoder_encoders_24_self_attn_fsmn_block_weight_to_fp16, x = input_763_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<int32, [3]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = x_255_cast_fp16)[name = string("transpose_617")];
            tensor<fp16, [1, 304, 512]> input_765_cast_fp16 = add(x = x_257_cast_fp16, y = var_2826_cast_fp16_2)[name = string("input_765_cast_fp16")];
            fp16 var_2856_to_fp16 = const()[name = string("op_2856_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_103_cast_fp16 = mul(x = var_2831_cast_fp16, y = var_2856_to_fp16)[name = string("q_h_103_cast_fp16")];
            bool scores_101_transpose_x_0 = const()[name = string("scores_101_transpose_x_0"), val = bool(false)];
            bool scores_101_transpose_y_0 = const()[name = string("scores_101_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_260_perm_0 = const()[name = string("transpose_260_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_261_perm_0 = const()[name = string("transpose_261_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_261 = transpose(perm = transpose_261_perm_0, x = var_2834_cast_fp16)[name = string("transpose_615")];
            tensor<fp16, [1, 4, 304, 128]> transpose_260 = transpose(perm = transpose_260_perm_0, x = q_h_103_cast_fp16)[name = string("transpose_616")];
            tensor<fp16, [1, 4, 304, 304]> scores_101_cast_fp16 = matmul(transpose_x = scores_101_transpose_x_0, transpose_y = scores_101_transpose_y_0, x = transpose_260, y = transpose_261)[name = string("scores_101_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2864_cast_fp16 = softmax(axis = var_29, x = scores_101_cast_fp16)[name = string("op_2864_cast_fp16")];
            bool x_261_transpose_x_0 = const()[name = string("x_261_transpose_x_0"), val = bool(false)];
            bool x_261_transpose_y_0 = const()[name = string("x_261_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = var_2837_cast_fp16)[name = string("transpose_619")];
            tensor<fp16, [1, 4, 304, 128]> x_261_cast_fp16 = matmul(transpose_x = x_261_transpose_x_0, transpose_y = x_261_transpose_y_0, x = var_2864_cast_fp16, y = value_51_cast_fp16)[name = string("x_261_cast_fp16")];
            tensor<int32, [4]> var_2868_perm_0 = const()[name = string("op_2868_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2870 = const()[name = string("op_2870"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2868_cast_fp16 = transpose(perm = var_2868_perm_0, x = x_261_cast_fp16)[name = string("transpose_614")];
            tensor<fp16, [1, 304, 512]> input_769_cast_fp16 = reshape(shape = var_2870, x = var_2868_cast_fp16)[name = string("input_769_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_24_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_24_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160004224)))];
            tensor<fp16, [512]> m_encoder_encoders_24_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_24_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160528576)))];
            tensor<fp16, [1, 304, 512]> linear_101_cast_fp16 = linear(bias = m_encoder_encoders_24_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_24_self_attn_linear_out_weight_to_fp16, x = input_769_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_771_cast_fp16 = add(x = linear_101_cast_fp16, y = input_765_cast_fp16)[name = string("input_771_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_773_cast_fp16 = add(x = input_757_cast_fp16, y = input_771_cast_fp16)[name = string("input_773_cast_fp16")];
            tensor<int32, [1]> output_103_axes_0 = const()[name = string("output_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_304_to_fp16 = const()[name = string("const_304_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160529664)))];
            tensor<fp16, [512]> const_305_to_fp16 = const()[name = string("const_305_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160530752)))];
            tensor<fp16, [1, 304, 512]> output_103_cast_fp16 = layer_norm(axes = output_103_axes_0, beta = const_305_to_fp16, epsilon = var_22_to_fp16, gamma = const_304_to_fp16, x = input_773_cast_fp16)[name = string("output_103_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_24_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_24_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160531840)))];
            tensor<fp16, [2048]> m_encoder_encoders_24_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_24_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162629056)))];
            tensor<fp16, [1, 304, 2048]> linear_102_cast_fp16 = linear(bias = m_encoder_encoders_24_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_24_feed_forward_w_1_weight_to_fp16, x = output_103_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_781_cast_fp16 = relu(x = linear_102_cast_fp16)[name = string("input_781_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_24_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_24_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162633216)))];
            tensor<fp16, [512]> m_encoder_encoders_24_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_24_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164730432)))];
            tensor<fp16, [1, 304, 512]> linear_103_cast_fp16 = linear(bias = m_encoder_encoders_24_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_24_feed_forward_w_2_weight_to_fp16, x = input_781_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_787_cast_fp16 = add(x = input_773_cast_fp16, y = linear_103_cast_fp16)[name = string("input_787_cast_fp16")];
            tensor<int32, [1]> output_105_axes_0 = const()[name = string("output_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_306_to_fp16 = const()[name = string("const_306_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164731520)))];
            tensor<fp16, [512]> const_307_to_fp16 = const()[name = string("const_307_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164732608)))];
            tensor<fp16, [1, 304, 512]> output_105_cast_fp16 = layer_norm(axes = output_105_axes_0, beta = const_307_to_fp16, epsilon = var_22_to_fp16, gamma = const_306_to_fp16, x = input_787_cast_fp16)[name = string("output_105_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_25_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_25_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164733696)))];
            tensor<fp16, [1536]> m_encoder_encoders_25_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_25_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166306624)))];
            tensor<fp16, [1, 304, 1536]> linear_104_cast_fp16 = linear(bias = m_encoder_encoders_25_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_25_self_attn_linear_q_k_v_weight_to_fp16, x = output_105_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<int32, [3]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_2928_axis_0 = const()[name = string("op_2928_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_2928_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_2928_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_2928_cast_fp16_2 = split(axis = var_2928_axis_0, split_sizes = tile_26, x = linear_104_cast_fp16)[name = string("op_2928_cast_fp16")];
            tensor<int32, [4]> var_2932 = const()[name = string("op_2932"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2933_cast_fp16 = reshape(shape = var_2932, x = var_2928_cast_fp16_0)[name = string("op_2933_cast_fp16")];
            tensor<int32, [4]> var_2935 = const()[name = string("op_2935"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2936_cast_fp16 = reshape(shape = var_2935, x = var_2928_cast_fp16_1)[name = string("op_2936_cast_fp16")];
            tensor<int32, [4]> var_2938 = const()[name = string("op_2938"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_2939_cast_fp16 = reshape(shape = var_2938, x = var_2928_cast_fp16_2)[name = string("op_2939_cast_fp16")];
            tensor<int32, [4]> value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_791_perm_0 = const()[name = string("input_791_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_793_pad_0 = const()[name = string("input_793_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_793_mode_0 = const()[name = string("input_793_mode_0"), val = string("constant")];
            fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_791_cast_fp16 = transpose(perm = input_791_perm_0, x = var_2928_cast_fp16_2)[name = string("transpose_612")];
            tensor<fp16, [1, 512, 314]> input_793_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = input_793_mode_0, pad = input_793_pad_0, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")];
            string x_265_pad_type_0 = const()[name = string("x_265_pad_type_0"), val = string("valid")];
            int32 x_265_groups_0 = const()[name = string("x_265_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_265_strides_0 = const()[name = string("x_265_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_265_pad_0 = const()[name = string("x_265_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_265_dilations_0 = const()[name = string("x_265_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_25_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_25_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166309760)))];
            tensor<fp16, [1, 512, 304]> x_265_cast_fp16 = conv(dilations = x_265_dilations_0, groups = x_265_groups_0, pad = x_265_pad_0, pad_type = x_265_pad_type_0, strides = x_265_strides_0, weight = m_encoder_encoders_25_self_attn_fsmn_block_weight_to_fp16, x = input_793_cast_fp16)[name = string("x_265_cast_fp16")];
            tensor<int32, [3]> x_267_perm_0 = const()[name = string("x_267_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_267_cast_fp16 = transpose(perm = x_267_perm_0, x = x_265_cast_fp16)[name = string("transpose_611")];
            tensor<fp16, [1, 304, 512]> input_795_cast_fp16 = add(x = x_267_cast_fp16, y = var_2928_cast_fp16_2)[name = string("input_795_cast_fp16")];
            fp16 var_2958_to_fp16 = const()[name = string("op_2958_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_107_cast_fp16 = mul(x = var_2933_cast_fp16, y = var_2958_to_fp16)[name = string("q_h_107_cast_fp16")];
            bool scores_105_transpose_x_0 = const()[name = string("scores_105_transpose_x_0"), val = bool(false)];
            bool scores_105_transpose_y_0 = const()[name = string("scores_105_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_262_perm_0 = const()[name = string("transpose_262_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_263_perm_0 = const()[name = string("transpose_263_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_263 = transpose(perm = transpose_263_perm_0, x = var_2936_cast_fp16)[name = string("transpose_609")];
            tensor<fp16, [1, 4, 304, 128]> transpose_262 = transpose(perm = transpose_262_perm_0, x = q_h_107_cast_fp16)[name = string("transpose_610")];
            tensor<fp16, [1, 4, 304, 304]> scores_105_cast_fp16 = matmul(transpose_x = scores_105_transpose_x_0, transpose_y = scores_105_transpose_y_0, x = transpose_262, y = transpose_263)[name = string("scores_105_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_2966_cast_fp16 = softmax(axis = var_29, x = scores_105_cast_fp16)[name = string("op_2966_cast_fp16")];
            bool x_271_transpose_x_0 = const()[name = string("x_271_transpose_x_0"), val = bool(false)];
            bool x_271_transpose_y_0 = const()[name = string("x_271_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = var_2939_cast_fp16)[name = string("transpose_613")];
            tensor<fp16, [1, 4, 304, 128]> x_271_cast_fp16 = matmul(transpose_x = x_271_transpose_x_0, transpose_y = x_271_transpose_y_0, x = var_2966_cast_fp16, y = value_53_cast_fp16)[name = string("x_271_cast_fp16")];
            tensor<int32, [4]> var_2970_perm_0 = const()[name = string("op_2970_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2972 = const()[name = string("op_2972"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_2970_cast_fp16 = transpose(perm = var_2970_perm_0, x = x_271_cast_fp16)[name = string("transpose_608")];
            tensor<fp16, [1, 304, 512]> input_799_cast_fp16 = reshape(shape = var_2972, x = var_2970_cast_fp16)[name = string("input_799_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_25_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_25_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166321088)))];
            tensor<fp16, [512]> m_encoder_encoders_25_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_25_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166845440)))];
            tensor<fp16, [1, 304, 512]> linear_105_cast_fp16 = linear(bias = m_encoder_encoders_25_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_25_self_attn_linear_out_weight_to_fp16, x = input_799_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_801_cast_fp16 = add(x = linear_105_cast_fp16, y = input_795_cast_fp16)[name = string("input_801_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_803_cast_fp16 = add(x = input_787_cast_fp16, y = input_801_cast_fp16)[name = string("input_803_cast_fp16")];
            tensor<int32, [1]> output_107_axes_0 = const()[name = string("output_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166846528)))];
            tensor<fp16, [512]> const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166847616)))];
            tensor<fp16, [1, 304, 512]> output_107_cast_fp16 = layer_norm(axes = output_107_axes_0, beta = const_316_to_fp16, epsilon = var_22_to_fp16, gamma = const_315_to_fp16, x = input_803_cast_fp16)[name = string("output_107_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_25_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_25_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166848704)))];
            tensor<fp16, [2048]> m_encoder_encoders_25_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_25_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168945920)))];
            tensor<fp16, [1, 304, 2048]> linear_106_cast_fp16 = linear(bias = m_encoder_encoders_25_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_25_feed_forward_w_1_weight_to_fp16, x = output_107_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_811_cast_fp16 = relu(x = linear_106_cast_fp16)[name = string("input_811_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_25_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_25_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168950080)))];
            tensor<fp16, [512]> m_encoder_encoders_25_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_25_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171047296)))];
            tensor<fp16, [1, 304, 512]> linear_107_cast_fp16 = linear(bias = m_encoder_encoders_25_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_25_feed_forward_w_2_weight_to_fp16, x = input_811_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_817_cast_fp16 = add(x = input_803_cast_fp16, y = linear_107_cast_fp16)[name = string("input_817_cast_fp16")];
            tensor<int32, [1]> output_109_axes_0 = const()[name = string("output_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_317_to_fp16 = const()[name = string("const_317_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171048384)))];
            tensor<fp16, [512]> const_318_to_fp16 = const()[name = string("const_318_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171049472)))];
            tensor<fp16, [1, 304, 512]> output_109_cast_fp16 = layer_norm(axes = output_109_axes_0, beta = const_318_to_fp16, epsilon = var_22_to_fp16, gamma = const_317_to_fp16, x = input_817_cast_fp16)[name = string("output_109_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_26_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_26_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171050560)))];
            tensor<fp16, [1536]> m_encoder_encoders_26_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_26_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172623488)))];
            tensor<fp16, [1, 304, 1536]> linear_108_cast_fp16 = linear(bias = m_encoder_encoders_26_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_26_self_attn_linear_q_k_v_weight_to_fp16, x = output_109_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<int32, [3]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3030_axis_0 = const()[name = string("op_3030_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3030_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3030_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3030_cast_fp16_2 = split(axis = var_3030_axis_0, split_sizes = tile_27, x = linear_108_cast_fp16)[name = string("op_3030_cast_fp16")];
            tensor<int32, [4]> var_3034 = const()[name = string("op_3034"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3035_cast_fp16 = reshape(shape = var_3034, x = var_3030_cast_fp16_0)[name = string("op_3035_cast_fp16")];
            tensor<int32, [4]> var_3037 = const()[name = string("op_3037"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3038_cast_fp16 = reshape(shape = var_3037, x = var_3030_cast_fp16_1)[name = string("op_3038_cast_fp16")];
            tensor<int32, [4]> var_3040 = const()[name = string("op_3040"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3041_cast_fp16 = reshape(shape = var_3040, x = var_3030_cast_fp16_2)[name = string("op_3041_cast_fp16")];
            tensor<int32, [4]> value_55_perm_0 = const()[name = string("value_55_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_821_perm_0 = const()[name = string("input_821_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_823_pad_0 = const()[name = string("input_823_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_823_mode_0 = const()[name = string("input_823_mode_0"), val = string("constant")];
            fp16 const_324_to_fp16 = const()[name = string("const_324_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_821_cast_fp16 = transpose(perm = input_821_perm_0, x = var_3030_cast_fp16_2)[name = string("transpose_606")];
            tensor<fp16, [1, 512, 314]> input_823_cast_fp16 = pad(constant_val = const_324_to_fp16, mode = input_823_mode_0, pad = input_823_pad_0, x = input_821_cast_fp16)[name = string("input_823_cast_fp16")];
            string x_275_pad_type_0 = const()[name = string("x_275_pad_type_0"), val = string("valid")];
            int32 x_275_groups_0 = const()[name = string("x_275_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_275_strides_0 = const()[name = string("x_275_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_275_pad_0 = const()[name = string("x_275_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_275_dilations_0 = const()[name = string("x_275_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_26_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_26_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172626624)))];
            tensor<fp16, [1, 512, 304]> x_275_cast_fp16 = conv(dilations = x_275_dilations_0, groups = x_275_groups_0, pad = x_275_pad_0, pad_type = x_275_pad_type_0, strides = x_275_strides_0, weight = m_encoder_encoders_26_self_attn_fsmn_block_weight_to_fp16, x = input_823_cast_fp16)[name = string("x_275_cast_fp16")];
            tensor<int32, [3]> x_277_perm_0 = const()[name = string("x_277_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_277_cast_fp16 = transpose(perm = x_277_perm_0, x = x_275_cast_fp16)[name = string("transpose_605")];
            tensor<fp16, [1, 304, 512]> input_825_cast_fp16 = add(x = x_277_cast_fp16, y = var_3030_cast_fp16_2)[name = string("input_825_cast_fp16")];
            fp16 var_3060_to_fp16 = const()[name = string("op_3060_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_111_cast_fp16 = mul(x = var_3035_cast_fp16, y = var_3060_to_fp16)[name = string("q_h_111_cast_fp16")];
            bool scores_109_transpose_x_0 = const()[name = string("scores_109_transpose_x_0"), val = bool(false)];
            bool scores_109_transpose_y_0 = const()[name = string("scores_109_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_264_perm_0 = const()[name = string("transpose_264_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_265_perm_0 = const()[name = string("transpose_265_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_265 = transpose(perm = transpose_265_perm_0, x = var_3038_cast_fp16)[name = string("transpose_603")];
            tensor<fp16, [1, 4, 304, 128]> transpose_264 = transpose(perm = transpose_264_perm_0, x = q_h_111_cast_fp16)[name = string("transpose_604")];
            tensor<fp16, [1, 4, 304, 304]> scores_109_cast_fp16 = matmul(transpose_x = scores_109_transpose_x_0, transpose_y = scores_109_transpose_y_0, x = transpose_264, y = transpose_265)[name = string("scores_109_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3068_cast_fp16 = softmax(axis = var_29, x = scores_109_cast_fp16)[name = string("op_3068_cast_fp16")];
            bool x_281_transpose_x_0 = const()[name = string("x_281_transpose_x_0"), val = bool(false)];
            bool x_281_transpose_y_0 = const()[name = string("x_281_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_55_cast_fp16 = transpose(perm = value_55_perm_0, x = var_3041_cast_fp16)[name = string("transpose_607")];
            tensor<fp16, [1, 4, 304, 128]> x_281_cast_fp16 = matmul(transpose_x = x_281_transpose_x_0, transpose_y = x_281_transpose_y_0, x = var_3068_cast_fp16, y = value_55_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<int32, [4]> var_3072_perm_0 = const()[name = string("op_3072_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3074 = const()[name = string("op_3074"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3072_cast_fp16 = transpose(perm = var_3072_perm_0, x = x_281_cast_fp16)[name = string("transpose_602")];
            tensor<fp16, [1, 304, 512]> input_829_cast_fp16 = reshape(shape = var_3074, x = var_3072_cast_fp16)[name = string("input_829_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_26_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_26_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172637952)))];
            tensor<fp16, [512]> m_encoder_encoders_26_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_26_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173162304)))];
            tensor<fp16, [1, 304, 512]> linear_109_cast_fp16 = linear(bias = m_encoder_encoders_26_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_26_self_attn_linear_out_weight_to_fp16, x = input_829_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_831_cast_fp16 = add(x = linear_109_cast_fp16, y = input_825_cast_fp16)[name = string("input_831_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_833_cast_fp16 = add(x = input_817_cast_fp16, y = input_831_cast_fp16)[name = string("input_833_cast_fp16")];
            tensor<int32, [1]> output_111_axes_0 = const()[name = string("output_111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173163392)))];
            tensor<fp16, [512]> const_327_to_fp16 = const()[name = string("const_327_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173164480)))];
            tensor<fp16, [1, 304, 512]> output_111_cast_fp16 = layer_norm(axes = output_111_axes_0, beta = const_327_to_fp16, epsilon = var_22_to_fp16, gamma = const_326_to_fp16, x = input_833_cast_fp16)[name = string("output_111_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_26_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_26_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173165568)))];
            tensor<fp16, [2048]> m_encoder_encoders_26_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_26_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175262784)))];
            tensor<fp16, [1, 304, 2048]> linear_110_cast_fp16 = linear(bias = m_encoder_encoders_26_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_26_feed_forward_w_1_weight_to_fp16, x = output_111_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_841_cast_fp16 = relu(x = linear_110_cast_fp16)[name = string("input_841_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_26_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_26_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175266944)))];
            tensor<fp16, [512]> m_encoder_encoders_26_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_26_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177364160)))];
            tensor<fp16, [1, 304, 512]> linear_111_cast_fp16 = linear(bias = m_encoder_encoders_26_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_26_feed_forward_w_2_weight_to_fp16, x = input_841_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_847_cast_fp16 = add(x = input_833_cast_fp16, y = linear_111_cast_fp16)[name = string("input_847_cast_fp16")];
            tensor<int32, [1]> output_113_axes_0 = const()[name = string("output_113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_328_to_fp16 = const()[name = string("const_328_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177365248)))];
            tensor<fp16, [512]> const_329_to_fp16 = const()[name = string("const_329_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177366336)))];
            tensor<fp16, [1, 304, 512]> output_113_cast_fp16 = layer_norm(axes = output_113_axes_0, beta = const_329_to_fp16, epsilon = var_22_to_fp16, gamma = const_328_to_fp16, x = input_847_cast_fp16)[name = string("output_113_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_27_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_27_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177367424)))];
            tensor<fp16, [1536]> m_encoder_encoders_27_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_27_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178940352)))];
            tensor<fp16, [1, 304, 1536]> linear_112_cast_fp16 = linear(bias = m_encoder_encoders_27_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_27_self_attn_linear_q_k_v_weight_to_fp16, x = output_113_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [3]> tile_28 = const()[name = string("tile_28"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3132_axis_0 = const()[name = string("op_3132_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3132_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3132_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3132_cast_fp16_2 = split(axis = var_3132_axis_0, split_sizes = tile_28, x = linear_112_cast_fp16)[name = string("op_3132_cast_fp16")];
            tensor<int32, [4]> var_3136 = const()[name = string("op_3136"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3137_cast_fp16 = reshape(shape = var_3136, x = var_3132_cast_fp16_0)[name = string("op_3137_cast_fp16")];
            tensor<int32, [4]> var_3139 = const()[name = string("op_3139"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3140_cast_fp16 = reshape(shape = var_3139, x = var_3132_cast_fp16_1)[name = string("op_3140_cast_fp16")];
            tensor<int32, [4]> var_3142 = const()[name = string("op_3142"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3143_cast_fp16 = reshape(shape = var_3142, x = var_3132_cast_fp16_2)[name = string("op_3143_cast_fp16")];
            tensor<int32, [4]> value_57_perm_0 = const()[name = string("value_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_851_perm_0 = const()[name = string("input_851_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_853_pad_0 = const()[name = string("input_853_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_853_mode_0 = const()[name = string("input_853_mode_0"), val = string("constant")];
            fp16 const_335_to_fp16 = const()[name = string("const_335_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_851_cast_fp16 = transpose(perm = input_851_perm_0, x = var_3132_cast_fp16_2)[name = string("transpose_600")];
            tensor<fp16, [1, 512, 314]> input_853_cast_fp16 = pad(constant_val = const_335_to_fp16, mode = input_853_mode_0, pad = input_853_pad_0, x = input_851_cast_fp16)[name = string("input_853_cast_fp16")];
            string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")];
            int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_27_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_27_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178943488)))];
            tensor<fp16, [1, 512, 304]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = m_encoder_encoders_27_self_attn_fsmn_block_weight_to_fp16, x = input_853_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [3]> x_287_perm_0 = const()[name = string("x_287_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_287_cast_fp16 = transpose(perm = x_287_perm_0, x = x_285_cast_fp16)[name = string("transpose_599")];
            tensor<fp16, [1, 304, 512]> input_855_cast_fp16 = add(x = x_287_cast_fp16, y = var_3132_cast_fp16_2)[name = string("input_855_cast_fp16")];
            fp16 var_3162_to_fp16 = const()[name = string("op_3162_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_115_cast_fp16 = mul(x = var_3137_cast_fp16, y = var_3162_to_fp16)[name = string("q_h_115_cast_fp16")];
            bool scores_113_transpose_x_0 = const()[name = string("scores_113_transpose_x_0"), val = bool(false)];
            bool scores_113_transpose_y_0 = const()[name = string("scores_113_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_266_perm_0 = const()[name = string("transpose_266_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_267_perm_0 = const()[name = string("transpose_267_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_267 = transpose(perm = transpose_267_perm_0, x = var_3140_cast_fp16)[name = string("transpose_597")];
            tensor<fp16, [1, 4, 304, 128]> transpose_266 = transpose(perm = transpose_266_perm_0, x = q_h_115_cast_fp16)[name = string("transpose_598")];
            tensor<fp16, [1, 4, 304, 304]> scores_113_cast_fp16 = matmul(transpose_x = scores_113_transpose_x_0, transpose_y = scores_113_transpose_y_0, x = transpose_266, y = transpose_267)[name = string("scores_113_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3170_cast_fp16 = softmax(axis = var_29, x = scores_113_cast_fp16)[name = string("op_3170_cast_fp16")];
            bool x_291_transpose_x_0 = const()[name = string("x_291_transpose_x_0"), val = bool(false)];
            bool x_291_transpose_y_0 = const()[name = string("x_291_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_57_cast_fp16 = transpose(perm = value_57_perm_0, x = var_3143_cast_fp16)[name = string("transpose_601")];
            tensor<fp16, [1, 4, 304, 128]> x_291_cast_fp16 = matmul(transpose_x = x_291_transpose_x_0, transpose_y = x_291_transpose_y_0, x = var_3170_cast_fp16, y = value_57_cast_fp16)[name = string("x_291_cast_fp16")];
            tensor<int32, [4]> var_3174_perm_0 = const()[name = string("op_3174_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3176 = const()[name = string("op_3176"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3174_cast_fp16 = transpose(perm = var_3174_perm_0, x = x_291_cast_fp16)[name = string("transpose_596")];
            tensor<fp16, [1, 304, 512]> input_859_cast_fp16 = reshape(shape = var_3176, x = var_3174_cast_fp16)[name = string("input_859_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_27_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_27_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178954816)))];
            tensor<fp16, [512]> m_encoder_encoders_27_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_27_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179479168)))];
            tensor<fp16, [1, 304, 512]> linear_113_cast_fp16 = linear(bias = m_encoder_encoders_27_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_27_self_attn_linear_out_weight_to_fp16, x = input_859_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_861_cast_fp16 = add(x = linear_113_cast_fp16, y = input_855_cast_fp16)[name = string("input_861_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_863_cast_fp16 = add(x = input_847_cast_fp16, y = input_861_cast_fp16)[name = string("input_863_cast_fp16")];
            tensor<int32, [1]> output_115_axes_0 = const()[name = string("output_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_337_to_fp16 = const()[name = string("const_337_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179480256)))];
            tensor<fp16, [512]> const_338_to_fp16 = const()[name = string("const_338_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179481344)))];
            tensor<fp16, [1, 304, 512]> output_115_cast_fp16 = layer_norm(axes = output_115_axes_0, beta = const_338_to_fp16, epsilon = var_22_to_fp16, gamma = const_337_to_fp16, x = input_863_cast_fp16)[name = string("output_115_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_27_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_27_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179482432)))];
            tensor<fp16, [2048]> m_encoder_encoders_27_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_27_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181579648)))];
            tensor<fp16, [1, 304, 2048]> linear_114_cast_fp16 = linear(bias = m_encoder_encoders_27_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_27_feed_forward_w_1_weight_to_fp16, x = output_115_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_871_cast_fp16 = relu(x = linear_114_cast_fp16)[name = string("input_871_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_27_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_27_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181583808)))];
            tensor<fp16, [512]> m_encoder_encoders_27_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_27_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183681024)))];
            tensor<fp16, [1, 304, 512]> linear_115_cast_fp16 = linear(bias = m_encoder_encoders_27_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_27_feed_forward_w_2_weight_to_fp16, x = input_871_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_877_cast_fp16 = add(x = input_863_cast_fp16, y = linear_115_cast_fp16)[name = string("input_877_cast_fp16")];
            tensor<int32, [1]> output_117_axes_0 = const()[name = string("output_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183682112)))];
            tensor<fp16, [512]> const_340_to_fp16 = const()[name = string("const_340_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183683200)))];
            tensor<fp16, [1, 304, 512]> output_117_cast_fp16 = layer_norm(axes = output_117_axes_0, beta = const_340_to_fp16, epsilon = var_22_to_fp16, gamma = const_339_to_fp16, x = input_877_cast_fp16)[name = string("output_117_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_28_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_28_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183684288)))];
            tensor<fp16, [1536]> m_encoder_encoders_28_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_28_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185257216)))];
            tensor<fp16, [1, 304, 1536]> linear_116_cast_fp16 = linear(bias = m_encoder_encoders_28_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_28_self_attn_linear_q_k_v_weight_to_fp16, x = output_117_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<int32, [3]> tile_29 = const()[name = string("tile_29"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3234_axis_0 = const()[name = string("op_3234_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3234_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3234_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3234_cast_fp16_2 = split(axis = var_3234_axis_0, split_sizes = tile_29, x = linear_116_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [4]> var_3238 = const()[name = string("op_3238"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3239_cast_fp16 = reshape(shape = var_3238, x = var_3234_cast_fp16_0)[name = string("op_3239_cast_fp16")];
            tensor<int32, [4]> var_3241 = const()[name = string("op_3241"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3242_cast_fp16 = reshape(shape = var_3241, x = var_3234_cast_fp16_1)[name = string("op_3242_cast_fp16")];
            tensor<int32, [4]> var_3244 = const()[name = string("op_3244"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3245_cast_fp16 = reshape(shape = var_3244, x = var_3234_cast_fp16_2)[name = string("op_3245_cast_fp16")];
            tensor<int32, [4]> value_59_perm_0 = const()[name = string("value_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_883_mode_0 = const()[name = string("input_883_mode_0"), val = string("constant")];
            fp16 const_346_to_fp16 = const()[name = string("const_346_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = var_3234_cast_fp16_2)[name = string("transpose_594")];
            tensor<fp16, [1, 512, 314]> input_883_cast_fp16 = pad(constant_val = const_346_to_fp16, mode = input_883_mode_0, pad = input_883_pad_0, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")];
            string x_295_pad_type_0 = const()[name = string("x_295_pad_type_0"), val = string("valid")];
            int32 x_295_groups_0 = const()[name = string("x_295_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_295_strides_0 = const()[name = string("x_295_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_295_dilations_0 = const()[name = string("x_295_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_28_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_28_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185260352)))];
            tensor<fp16, [1, 512, 304]> x_295_cast_fp16 = conv(dilations = x_295_dilations_0, groups = x_295_groups_0, pad = x_295_pad_0, pad_type = x_295_pad_type_0, strides = x_295_strides_0, weight = m_encoder_encoders_28_self_attn_fsmn_block_weight_to_fp16, x = input_883_cast_fp16)[name = string("x_295_cast_fp16")];
            tensor<int32, [3]> x_297_perm_0 = const()[name = string("x_297_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_297_cast_fp16 = transpose(perm = x_297_perm_0, x = x_295_cast_fp16)[name = string("transpose_593")];
            tensor<fp16, [1, 304, 512]> input_885_cast_fp16 = add(x = x_297_cast_fp16, y = var_3234_cast_fp16_2)[name = string("input_885_cast_fp16")];
            fp16 var_3264_to_fp16 = const()[name = string("op_3264_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_119_cast_fp16 = mul(x = var_3239_cast_fp16, y = var_3264_to_fp16)[name = string("q_h_119_cast_fp16")];
            bool scores_117_transpose_x_0 = const()[name = string("scores_117_transpose_x_0"), val = bool(false)];
            bool scores_117_transpose_y_0 = const()[name = string("scores_117_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_268_perm_0 = const()[name = string("transpose_268_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_269_perm_0 = const()[name = string("transpose_269_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_269 = transpose(perm = transpose_269_perm_0, x = var_3242_cast_fp16)[name = string("transpose_591")];
            tensor<fp16, [1, 4, 304, 128]> transpose_268 = transpose(perm = transpose_268_perm_0, x = q_h_119_cast_fp16)[name = string("transpose_592")];
            tensor<fp16, [1, 4, 304, 304]> scores_117_cast_fp16 = matmul(transpose_x = scores_117_transpose_x_0, transpose_y = scores_117_transpose_y_0, x = transpose_268, y = transpose_269)[name = string("scores_117_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3272_cast_fp16 = softmax(axis = var_29, x = scores_117_cast_fp16)[name = string("op_3272_cast_fp16")];
            bool x_301_transpose_x_0 = const()[name = string("x_301_transpose_x_0"), val = bool(false)];
            bool x_301_transpose_y_0 = const()[name = string("x_301_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_59_cast_fp16 = transpose(perm = value_59_perm_0, x = var_3245_cast_fp16)[name = string("transpose_595")];
            tensor<fp16, [1, 4, 304, 128]> x_301_cast_fp16 = matmul(transpose_x = x_301_transpose_x_0, transpose_y = x_301_transpose_y_0, x = var_3272_cast_fp16, y = value_59_cast_fp16)[name = string("x_301_cast_fp16")];
            tensor<int32, [4]> var_3276_perm_0 = const()[name = string("op_3276_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3278 = const()[name = string("op_3278"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3276_cast_fp16 = transpose(perm = var_3276_perm_0, x = x_301_cast_fp16)[name = string("transpose_590")];
            tensor<fp16, [1, 304, 512]> input_889_cast_fp16 = reshape(shape = var_3278, x = var_3276_cast_fp16)[name = string("input_889_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_28_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_28_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185271680)))];
            tensor<fp16, [512]> m_encoder_encoders_28_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_28_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185796032)))];
            tensor<fp16, [1, 304, 512]> linear_117_cast_fp16 = linear(bias = m_encoder_encoders_28_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_28_self_attn_linear_out_weight_to_fp16, x = input_889_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_891_cast_fp16 = add(x = linear_117_cast_fp16, y = input_885_cast_fp16)[name = string("input_891_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_893_cast_fp16 = add(x = input_877_cast_fp16, y = input_891_cast_fp16)[name = string("input_893_cast_fp16")];
            tensor<int32, [1]> output_119_axes_0 = const()[name = string("output_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_348_to_fp16 = const()[name = string("const_348_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185797120)))];
            tensor<fp16, [512]> const_349_to_fp16 = const()[name = string("const_349_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185798208)))];
            tensor<fp16, [1, 304, 512]> output_119_cast_fp16 = layer_norm(axes = output_119_axes_0, beta = const_349_to_fp16, epsilon = var_22_to_fp16, gamma = const_348_to_fp16, x = input_893_cast_fp16)[name = string("output_119_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_28_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_28_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185799296)))];
            tensor<fp16, [2048]> m_encoder_encoders_28_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_28_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187896512)))];
            tensor<fp16, [1, 304, 2048]> linear_118_cast_fp16 = linear(bias = m_encoder_encoders_28_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_28_feed_forward_w_1_weight_to_fp16, x = output_119_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_901_cast_fp16 = relu(x = linear_118_cast_fp16)[name = string("input_901_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_28_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_28_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187900672)))];
            tensor<fp16, [512]> m_encoder_encoders_28_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_28_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189997888)))];
            tensor<fp16, [1, 304, 512]> linear_119_cast_fp16 = linear(bias = m_encoder_encoders_28_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_28_feed_forward_w_2_weight_to_fp16, x = input_901_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_907_cast_fp16 = add(x = input_893_cast_fp16, y = linear_119_cast_fp16)[name = string("input_907_cast_fp16")];
            tensor<int32, [1]> output_121_axes_0 = const()[name = string("output_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_350_to_fp16 = const()[name = string("const_350_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189998976)))];
            tensor<fp16, [512]> const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190000064)))];
            tensor<fp16, [1, 304, 512]> output_121_cast_fp16 = layer_norm(axes = output_121_axes_0, beta = const_351_to_fp16, epsilon = var_22_to_fp16, gamma = const_350_to_fp16, x = input_907_cast_fp16)[name = string("output_121_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_29_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_29_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190001152)))];
            tensor<fp16, [1536]> m_encoder_encoders_29_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_29_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191574080)))];
            tensor<fp16, [1, 304, 1536]> linear_120_cast_fp16 = linear(bias = m_encoder_encoders_29_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_29_self_attn_linear_q_k_v_weight_to_fp16, x = output_121_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [3]> tile_30 = const()[name = string("tile_30"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3336_axis_0 = const()[name = string("op_3336_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3336_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3336_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3336_cast_fp16_2 = split(axis = var_3336_axis_0, split_sizes = tile_30, x = linear_120_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> var_3340 = const()[name = string("op_3340"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3341_cast_fp16 = reshape(shape = var_3340, x = var_3336_cast_fp16_0)[name = string("op_3341_cast_fp16")];
            tensor<int32, [4]> var_3343 = const()[name = string("op_3343"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3344_cast_fp16 = reshape(shape = var_3343, x = var_3336_cast_fp16_1)[name = string("op_3344_cast_fp16")];
            tensor<int32, [4]> var_3346 = const()[name = string("op_3346"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3347_cast_fp16 = reshape(shape = var_3346, x = var_3336_cast_fp16_2)[name = string("op_3347_cast_fp16")];
            tensor<int32, [4]> value_61_perm_0 = const()[name = string("value_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_911_perm_0 = const()[name = string("input_911_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_913_pad_0 = const()[name = string("input_913_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_913_mode_0 = const()[name = string("input_913_mode_0"), val = string("constant")];
            fp16 const_357_to_fp16 = const()[name = string("const_357_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_911_cast_fp16 = transpose(perm = input_911_perm_0, x = var_3336_cast_fp16_2)[name = string("transpose_588")];
            tensor<fp16, [1, 512, 314]> input_913_cast_fp16 = pad(constant_val = const_357_to_fp16, mode = input_913_mode_0, pad = input_913_pad_0, x = input_911_cast_fp16)[name = string("input_913_cast_fp16")];
            string x_305_pad_type_0 = const()[name = string("x_305_pad_type_0"), val = string("valid")];
            int32 x_305_groups_0 = const()[name = string("x_305_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_305_strides_0 = const()[name = string("x_305_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_305_pad_0 = const()[name = string("x_305_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_305_dilations_0 = const()[name = string("x_305_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_29_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_29_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191577216)))];
            tensor<fp16, [1, 512, 304]> x_305_cast_fp16 = conv(dilations = x_305_dilations_0, groups = x_305_groups_0, pad = x_305_pad_0, pad_type = x_305_pad_type_0, strides = x_305_strides_0, weight = m_encoder_encoders_29_self_attn_fsmn_block_weight_to_fp16, x = input_913_cast_fp16)[name = string("x_305_cast_fp16")];
            tensor<int32, [3]> x_307_perm_0 = const()[name = string("x_307_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_307_cast_fp16 = transpose(perm = x_307_perm_0, x = x_305_cast_fp16)[name = string("transpose_587")];
            tensor<fp16, [1, 304, 512]> input_915_cast_fp16 = add(x = x_307_cast_fp16, y = var_3336_cast_fp16_2)[name = string("input_915_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_123_cast_fp16 = mul(x = var_3341_cast_fp16, y = var_3366_to_fp16)[name = string("q_h_123_cast_fp16")];
            bool scores_121_transpose_x_0 = const()[name = string("scores_121_transpose_x_0"), val = bool(false)];
            bool scores_121_transpose_y_0 = const()[name = string("scores_121_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_270_perm_0 = const()[name = string("transpose_270_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_271_perm_0 = const()[name = string("transpose_271_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_271 = transpose(perm = transpose_271_perm_0, x = var_3344_cast_fp16)[name = string("transpose_585")];
            tensor<fp16, [1, 4, 304, 128]> transpose_270 = transpose(perm = transpose_270_perm_0, x = q_h_123_cast_fp16)[name = string("transpose_586")];
            tensor<fp16, [1, 4, 304, 304]> scores_121_cast_fp16 = matmul(transpose_x = scores_121_transpose_x_0, transpose_y = scores_121_transpose_y_0, x = transpose_270, y = transpose_271)[name = string("scores_121_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3374_cast_fp16 = softmax(axis = var_29, x = scores_121_cast_fp16)[name = string("op_3374_cast_fp16")];
            bool x_311_transpose_x_0 = const()[name = string("x_311_transpose_x_0"), val = bool(false)];
            bool x_311_transpose_y_0 = const()[name = string("x_311_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_61_cast_fp16 = transpose(perm = value_61_perm_0, x = var_3347_cast_fp16)[name = string("transpose_589")];
            tensor<fp16, [1, 4, 304, 128]> x_311_cast_fp16 = matmul(transpose_x = x_311_transpose_x_0, transpose_y = x_311_transpose_y_0, x = var_3374_cast_fp16, y = value_61_cast_fp16)[name = string("x_311_cast_fp16")];
            tensor<int32, [4]> var_3378_perm_0 = const()[name = string("op_3378_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3380 = const()[name = string("op_3380"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3378_cast_fp16 = transpose(perm = var_3378_perm_0, x = x_311_cast_fp16)[name = string("transpose_584")];
            tensor<fp16, [1, 304, 512]> input_919_cast_fp16 = reshape(shape = var_3380, x = var_3378_cast_fp16)[name = string("input_919_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_29_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_29_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191588544)))];
            tensor<fp16, [512]> m_encoder_encoders_29_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_29_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192112896)))];
            tensor<fp16, [1, 304, 512]> linear_121_cast_fp16 = linear(bias = m_encoder_encoders_29_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_29_self_attn_linear_out_weight_to_fp16, x = input_919_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_921_cast_fp16 = add(x = linear_121_cast_fp16, y = input_915_cast_fp16)[name = string("input_921_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_923_cast_fp16 = add(x = input_907_cast_fp16, y = input_921_cast_fp16)[name = string("input_923_cast_fp16")];
            tensor<int32, [1]> output_123_axes_0 = const()[name = string("output_123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_359_to_fp16 = const()[name = string("const_359_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192113984)))];
            tensor<fp16, [512]> const_360_to_fp16 = const()[name = string("const_360_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192115072)))];
            tensor<fp16, [1, 304, 512]> output_123_cast_fp16 = layer_norm(axes = output_123_axes_0, beta = const_360_to_fp16, epsilon = var_22_to_fp16, gamma = const_359_to_fp16, x = input_923_cast_fp16)[name = string("output_123_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_29_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_29_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192116160)))];
            tensor<fp16, [2048]> m_encoder_encoders_29_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_29_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194213376)))];
            tensor<fp16, [1, 304, 2048]> linear_122_cast_fp16 = linear(bias = m_encoder_encoders_29_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_29_feed_forward_w_1_weight_to_fp16, x = output_123_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_931_cast_fp16 = relu(x = linear_122_cast_fp16)[name = string("input_931_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_29_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_29_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194217536)))];
            tensor<fp16, [512]> m_encoder_encoders_29_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_29_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196314752)))];
            tensor<fp16, [1, 304, 512]> linear_123_cast_fp16 = linear(bias = m_encoder_encoders_29_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_29_feed_forward_w_2_weight_to_fp16, x = input_931_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_937_cast_fp16 = add(x = input_923_cast_fp16, y = linear_123_cast_fp16)[name = string("input_937_cast_fp16")];
            tensor<int32, [1]> output_125_axes_0 = const()[name = string("output_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_361_to_fp16 = const()[name = string("const_361_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196315840)))];
            tensor<fp16, [512]> const_362_to_fp16 = const()[name = string("const_362_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196316928)))];
            tensor<fp16, [1, 304, 512]> output_125_cast_fp16 = layer_norm(axes = output_125_axes_0, beta = const_362_to_fp16, epsilon = var_22_to_fp16, gamma = const_361_to_fp16, x = input_937_cast_fp16)[name = string("output_125_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_30_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_30_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196318016)))];
            tensor<fp16, [1536]> m_encoder_encoders_30_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_30_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197890944)))];
            tensor<fp16, [1, 304, 1536]> linear_124_cast_fp16 = linear(bias = m_encoder_encoders_30_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_30_self_attn_linear_q_k_v_weight_to_fp16, x = output_125_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<int32, [3]> tile_31 = const()[name = string("tile_31"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3438_axis_0 = const()[name = string("op_3438_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3438_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3438_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3438_cast_fp16_2 = split(axis = var_3438_axis_0, split_sizes = tile_31, x = linear_124_cast_fp16)[name = string("op_3438_cast_fp16")];
            tensor<int32, [4]> var_3442 = const()[name = string("op_3442"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3443_cast_fp16 = reshape(shape = var_3442, x = var_3438_cast_fp16_0)[name = string("op_3443_cast_fp16")];
            tensor<int32, [4]> var_3445 = const()[name = string("op_3445"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3446_cast_fp16 = reshape(shape = var_3445, x = var_3438_cast_fp16_1)[name = string("op_3446_cast_fp16")];
            tensor<int32, [4]> var_3448 = const()[name = string("op_3448"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3449_cast_fp16 = reshape(shape = var_3448, x = var_3438_cast_fp16_2)[name = string("op_3449_cast_fp16")];
            tensor<int32, [4]> value_63_perm_0 = const()[name = string("value_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_943_pad_0 = const()[name = string("input_943_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_943_mode_0 = const()[name = string("input_943_mode_0"), val = string("constant")];
            fp16 const_368_to_fp16 = const()[name = string("const_368_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = var_3438_cast_fp16_2)[name = string("transpose_582")];
            tensor<fp16, [1, 512, 314]> input_943_cast_fp16 = pad(constant_val = const_368_to_fp16, mode = input_943_mode_0, pad = input_943_pad_0, x = input_941_cast_fp16)[name = string("input_943_cast_fp16")];
            string x_315_pad_type_0 = const()[name = string("x_315_pad_type_0"), val = string("valid")];
            int32 x_315_groups_0 = const()[name = string("x_315_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_315_strides_0 = const()[name = string("x_315_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_315_pad_0 = const()[name = string("x_315_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_315_dilations_0 = const()[name = string("x_315_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_30_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_30_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197894080)))];
            tensor<fp16, [1, 512, 304]> x_315_cast_fp16 = conv(dilations = x_315_dilations_0, groups = x_315_groups_0, pad = x_315_pad_0, pad_type = x_315_pad_type_0, strides = x_315_strides_0, weight = m_encoder_encoders_30_self_attn_fsmn_block_weight_to_fp16, x = input_943_cast_fp16)[name = string("x_315_cast_fp16")];
            tensor<int32, [3]> x_317_perm_0 = const()[name = string("x_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_317_cast_fp16 = transpose(perm = x_317_perm_0, x = x_315_cast_fp16)[name = string("transpose_581")];
            tensor<fp16, [1, 304, 512]> input_945_cast_fp16 = add(x = x_317_cast_fp16, y = var_3438_cast_fp16_2)[name = string("input_945_cast_fp16")];
            fp16 var_3468_to_fp16 = const()[name = string("op_3468_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_127_cast_fp16 = mul(x = var_3443_cast_fp16, y = var_3468_to_fp16)[name = string("q_h_127_cast_fp16")];
            bool scores_125_transpose_x_0 = const()[name = string("scores_125_transpose_x_0"), val = bool(false)];
            bool scores_125_transpose_y_0 = const()[name = string("scores_125_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_272_perm_0 = const()[name = string("transpose_272_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_273_perm_0 = const()[name = string("transpose_273_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_273 = transpose(perm = transpose_273_perm_0, x = var_3446_cast_fp16)[name = string("transpose_579")];
            tensor<fp16, [1, 4, 304, 128]> transpose_272 = transpose(perm = transpose_272_perm_0, x = q_h_127_cast_fp16)[name = string("transpose_580")];
            tensor<fp16, [1, 4, 304, 304]> scores_125_cast_fp16 = matmul(transpose_x = scores_125_transpose_x_0, transpose_y = scores_125_transpose_y_0, x = transpose_272, y = transpose_273)[name = string("scores_125_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3476_cast_fp16 = softmax(axis = var_29, x = scores_125_cast_fp16)[name = string("op_3476_cast_fp16")];
            bool x_321_transpose_x_0 = const()[name = string("x_321_transpose_x_0"), val = bool(false)];
            bool x_321_transpose_y_0 = const()[name = string("x_321_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_63_cast_fp16 = transpose(perm = value_63_perm_0, x = var_3449_cast_fp16)[name = string("transpose_583")];
            tensor<fp16, [1, 4, 304, 128]> x_321_cast_fp16 = matmul(transpose_x = x_321_transpose_x_0, transpose_y = x_321_transpose_y_0, x = var_3476_cast_fp16, y = value_63_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [4]> var_3480_perm_0 = const()[name = string("op_3480_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3482 = const()[name = string("op_3482"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3480_cast_fp16 = transpose(perm = var_3480_perm_0, x = x_321_cast_fp16)[name = string("transpose_578")];
            tensor<fp16, [1, 304, 512]> input_949_cast_fp16 = reshape(shape = var_3482, x = var_3480_cast_fp16)[name = string("input_949_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_30_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_30_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197905408)))];
            tensor<fp16, [512]> m_encoder_encoders_30_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_30_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198429760)))];
            tensor<fp16, [1, 304, 512]> linear_125_cast_fp16 = linear(bias = m_encoder_encoders_30_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_30_self_attn_linear_out_weight_to_fp16, x = input_949_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_951_cast_fp16 = add(x = linear_125_cast_fp16, y = input_945_cast_fp16)[name = string("input_951_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_953_cast_fp16 = add(x = input_937_cast_fp16, y = input_951_cast_fp16)[name = string("input_953_cast_fp16")];
            tensor<int32, [1]> output_127_axes_0 = const()[name = string("output_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198430848)))];
            tensor<fp16, [512]> const_371_to_fp16 = const()[name = string("const_371_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198431936)))];
            tensor<fp16, [1, 304, 512]> output_127_cast_fp16 = layer_norm(axes = output_127_axes_0, beta = const_371_to_fp16, epsilon = var_22_to_fp16, gamma = const_370_to_fp16, x = input_953_cast_fp16)[name = string("output_127_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_30_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_30_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198433024)))];
            tensor<fp16, [2048]> m_encoder_encoders_30_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_30_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200530240)))];
            tensor<fp16, [1, 304, 2048]> linear_126_cast_fp16 = linear(bias = m_encoder_encoders_30_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_30_feed_forward_w_1_weight_to_fp16, x = output_127_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_961_cast_fp16 = relu(x = linear_126_cast_fp16)[name = string("input_961_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_30_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_30_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200534400)))];
            tensor<fp16, [512]> m_encoder_encoders_30_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_30_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202631616)))];
            tensor<fp16, [1, 304, 512]> linear_127_cast_fp16 = linear(bias = m_encoder_encoders_30_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_30_feed_forward_w_2_weight_to_fp16, x = input_961_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_967_cast_fp16 = add(x = input_953_cast_fp16, y = linear_127_cast_fp16)[name = string("input_967_cast_fp16")];
            tensor<int32, [1]> output_129_axes_0 = const()[name = string("output_129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_372_to_fp16 = const()[name = string("const_372_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202632704)))];
            tensor<fp16, [512]> const_373_to_fp16 = const()[name = string("const_373_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202633792)))];
            tensor<fp16, [1, 304, 512]> output_129_cast_fp16 = layer_norm(axes = output_129_axes_0, beta = const_373_to_fp16, epsilon = var_22_to_fp16, gamma = const_372_to_fp16, x = input_967_cast_fp16)[name = string("output_129_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_31_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_31_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202634880)))];
            tensor<fp16, [1536]> m_encoder_encoders_31_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_31_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204207808)))];
            tensor<fp16, [1, 304, 1536]> linear_128_cast_fp16 = linear(bias = m_encoder_encoders_31_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_31_self_attn_linear_q_k_v_weight_to_fp16, x = output_129_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [3]> tile_32 = const()[name = string("tile_32"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3540_axis_0 = const()[name = string("op_3540_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3540_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3540_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3540_cast_fp16_2 = split(axis = var_3540_axis_0, split_sizes = tile_32, x = linear_128_cast_fp16)[name = string("op_3540_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3545_cast_fp16 = reshape(shape = var_3544, x = var_3540_cast_fp16_0)[name = string("op_3545_cast_fp16")];
            tensor<int32, [4]> var_3547 = const()[name = string("op_3547"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3548_cast_fp16 = reshape(shape = var_3547, x = var_3540_cast_fp16_1)[name = string("op_3548_cast_fp16")];
            tensor<int32, [4]> var_3550 = const()[name = string("op_3550"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3551_cast_fp16 = reshape(shape = var_3550, x = var_3540_cast_fp16_2)[name = string("op_3551_cast_fp16")];
            tensor<int32, [4]> value_65_perm_0 = const()[name = string("value_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_971_perm_0 = const()[name = string("input_971_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_973_pad_0 = const()[name = string("input_973_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_973_mode_0 = const()[name = string("input_973_mode_0"), val = string("constant")];
            fp16 const_379_to_fp16 = const()[name = string("const_379_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_971_cast_fp16 = transpose(perm = input_971_perm_0, x = var_3540_cast_fp16_2)[name = string("transpose_576")];
            tensor<fp16, [1, 512, 314]> input_973_cast_fp16 = pad(constant_val = const_379_to_fp16, mode = input_973_mode_0, pad = input_973_pad_0, x = input_971_cast_fp16)[name = string("input_973_cast_fp16")];
            string x_325_pad_type_0 = const()[name = string("x_325_pad_type_0"), val = string("valid")];
            int32 x_325_groups_0 = const()[name = string("x_325_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_325_strides_0 = const()[name = string("x_325_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_325_pad_0 = const()[name = string("x_325_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_325_dilations_0 = const()[name = string("x_325_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_31_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_31_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204210944)))];
            tensor<fp16, [1, 512, 304]> x_325_cast_fp16 = conv(dilations = x_325_dilations_0, groups = x_325_groups_0, pad = x_325_pad_0, pad_type = x_325_pad_type_0, strides = x_325_strides_0, weight = m_encoder_encoders_31_self_attn_fsmn_block_weight_to_fp16, x = input_973_cast_fp16)[name = string("x_325_cast_fp16")];
            tensor<int32, [3]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = x_325_cast_fp16)[name = string("transpose_575")];
            tensor<fp16, [1, 304, 512]> input_975_cast_fp16 = add(x = x_327_cast_fp16, y = var_3540_cast_fp16_2)[name = string("input_975_cast_fp16")];
            fp16 var_3570_to_fp16 = const()[name = string("op_3570_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_131_cast_fp16 = mul(x = var_3545_cast_fp16, y = var_3570_to_fp16)[name = string("q_h_131_cast_fp16")];
            bool scores_129_transpose_x_0 = const()[name = string("scores_129_transpose_x_0"), val = bool(false)];
            bool scores_129_transpose_y_0 = const()[name = string("scores_129_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_274_perm_0 = const()[name = string("transpose_274_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_275_perm_0 = const()[name = string("transpose_275_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_275 = transpose(perm = transpose_275_perm_0, x = var_3548_cast_fp16)[name = string("transpose_573")];
            tensor<fp16, [1, 4, 304, 128]> transpose_274 = transpose(perm = transpose_274_perm_0, x = q_h_131_cast_fp16)[name = string("transpose_574")];
            tensor<fp16, [1, 4, 304, 304]> scores_129_cast_fp16 = matmul(transpose_x = scores_129_transpose_x_0, transpose_y = scores_129_transpose_y_0, x = transpose_274, y = transpose_275)[name = string("scores_129_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3578_cast_fp16 = softmax(axis = var_29, x = scores_129_cast_fp16)[name = string("op_3578_cast_fp16")];
            bool x_331_transpose_x_0 = const()[name = string("x_331_transpose_x_0"), val = bool(false)];
            bool x_331_transpose_y_0 = const()[name = string("x_331_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_65_cast_fp16 = transpose(perm = value_65_perm_0, x = var_3551_cast_fp16)[name = string("transpose_577")];
            tensor<fp16, [1, 4, 304, 128]> x_331_cast_fp16 = matmul(transpose_x = x_331_transpose_x_0, transpose_y = x_331_transpose_y_0, x = var_3578_cast_fp16, y = value_65_cast_fp16)[name = string("x_331_cast_fp16")];
            tensor<int32, [4]> var_3582_perm_0 = const()[name = string("op_3582_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3584 = const()[name = string("op_3584"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3582_cast_fp16 = transpose(perm = var_3582_perm_0, x = x_331_cast_fp16)[name = string("transpose_572")];
            tensor<fp16, [1, 304, 512]> input_979_cast_fp16 = reshape(shape = var_3584, x = var_3582_cast_fp16)[name = string("input_979_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_31_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_31_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204222272)))];
            tensor<fp16, [512]> m_encoder_encoders_31_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_31_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204746624)))];
            tensor<fp16, [1, 304, 512]> linear_129_cast_fp16 = linear(bias = m_encoder_encoders_31_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_31_self_attn_linear_out_weight_to_fp16, x = input_979_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_981_cast_fp16 = add(x = linear_129_cast_fp16, y = input_975_cast_fp16)[name = string("input_981_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_983_cast_fp16 = add(x = input_967_cast_fp16, y = input_981_cast_fp16)[name = string("input_983_cast_fp16")];
            tensor<int32, [1]> output_131_axes_0 = const()[name = string("output_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_381_to_fp16 = const()[name = string("const_381_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204747712)))];
            tensor<fp16, [512]> const_382_to_fp16 = const()[name = string("const_382_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204748800)))];
            tensor<fp16, [1, 304, 512]> output_131_cast_fp16 = layer_norm(axes = output_131_axes_0, beta = const_382_to_fp16, epsilon = var_22_to_fp16, gamma = const_381_to_fp16, x = input_983_cast_fp16)[name = string("output_131_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_31_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_31_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204749888)))];
            tensor<fp16, [2048]> m_encoder_encoders_31_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_31_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206847104)))];
            tensor<fp16, [1, 304, 2048]> linear_130_cast_fp16 = linear(bias = m_encoder_encoders_31_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_31_feed_forward_w_1_weight_to_fp16, x = output_131_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_991_cast_fp16 = relu(x = linear_130_cast_fp16)[name = string("input_991_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_31_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_31_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206851264)))];
            tensor<fp16, [512]> m_encoder_encoders_31_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_31_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208948480)))];
            tensor<fp16, [1, 304, 512]> linear_131_cast_fp16 = linear(bias = m_encoder_encoders_31_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_31_feed_forward_w_2_weight_to_fp16, x = input_991_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_997_cast_fp16 = add(x = input_983_cast_fp16, y = linear_131_cast_fp16)[name = string("input_997_cast_fp16")];
            tensor<int32, [1]> output_133_axes_0 = const()[name = string("output_133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_383_to_fp16 = const()[name = string("const_383_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208949568)))];
            tensor<fp16, [512]> const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208950656)))];
            tensor<fp16, [1, 304, 512]> output_133_cast_fp16 = layer_norm(axes = output_133_axes_0, beta = const_384_to_fp16, epsilon = var_22_to_fp16, gamma = const_383_to_fp16, x = input_997_cast_fp16)[name = string("output_133_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_32_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_32_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208951744)))];
            tensor<fp16, [1536]> m_encoder_encoders_32_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_32_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210524672)))];
            tensor<fp16, [1, 304, 1536]> linear_132_cast_fp16 = linear(bias = m_encoder_encoders_32_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_32_self_attn_linear_q_k_v_weight_to_fp16, x = output_133_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<int32, [3]> tile_33 = const()[name = string("tile_33"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3642_axis_0 = const()[name = string("op_3642_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3642_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3642_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3642_cast_fp16_2 = split(axis = var_3642_axis_0, split_sizes = tile_33, x = linear_132_cast_fp16)[name = string("op_3642_cast_fp16")];
            tensor<int32, [4]> var_3646 = const()[name = string("op_3646"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3647_cast_fp16 = reshape(shape = var_3646, x = var_3642_cast_fp16_0)[name = string("op_3647_cast_fp16")];
            tensor<int32, [4]> var_3649 = const()[name = string("op_3649"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3650_cast_fp16 = reshape(shape = var_3649, x = var_3642_cast_fp16_1)[name = string("op_3650_cast_fp16")];
            tensor<int32, [4]> var_3652 = const()[name = string("op_3652"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3653_cast_fp16 = reshape(shape = var_3652, x = var_3642_cast_fp16_2)[name = string("op_3653_cast_fp16")];
            tensor<int32, [4]> value_67_perm_0 = const()[name = string("value_67_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1001_perm_0 = const()[name = string("input_1001_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1003_pad_0 = const()[name = string("input_1003_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1003_mode_0 = const()[name = string("input_1003_mode_0"), val = string("constant")];
            fp16 const_390_to_fp16 = const()[name = string("const_390_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1001_cast_fp16 = transpose(perm = input_1001_perm_0, x = var_3642_cast_fp16_2)[name = string("transpose_570")];
            tensor<fp16, [1, 512, 314]> input_1003_cast_fp16 = pad(constant_val = const_390_to_fp16, mode = input_1003_mode_0, pad = input_1003_pad_0, x = input_1001_cast_fp16)[name = string("input_1003_cast_fp16")];
            string x_335_pad_type_0 = const()[name = string("x_335_pad_type_0"), val = string("valid")];
            int32 x_335_groups_0 = const()[name = string("x_335_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_335_strides_0 = const()[name = string("x_335_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_335_pad_0 = const()[name = string("x_335_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_335_dilations_0 = const()[name = string("x_335_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_32_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_32_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210527808)))];
            tensor<fp16, [1, 512, 304]> x_335_cast_fp16 = conv(dilations = x_335_dilations_0, groups = x_335_groups_0, pad = x_335_pad_0, pad_type = x_335_pad_type_0, strides = x_335_strides_0, weight = m_encoder_encoders_32_self_attn_fsmn_block_weight_to_fp16, x = input_1003_cast_fp16)[name = string("x_335_cast_fp16")];
            tensor<int32, [3]> x_337_perm_0 = const()[name = string("x_337_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_337_cast_fp16 = transpose(perm = x_337_perm_0, x = x_335_cast_fp16)[name = string("transpose_569")];
            tensor<fp16, [1, 304, 512]> input_1005_cast_fp16 = add(x = x_337_cast_fp16, y = var_3642_cast_fp16_2)[name = string("input_1005_cast_fp16")];
            fp16 var_3672_to_fp16 = const()[name = string("op_3672_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_135_cast_fp16 = mul(x = var_3647_cast_fp16, y = var_3672_to_fp16)[name = string("q_h_135_cast_fp16")];
            bool scores_133_transpose_x_0 = const()[name = string("scores_133_transpose_x_0"), val = bool(false)];
            bool scores_133_transpose_y_0 = const()[name = string("scores_133_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_276_perm_0 = const()[name = string("transpose_276_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_277_perm_0 = const()[name = string("transpose_277_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_277 = transpose(perm = transpose_277_perm_0, x = var_3650_cast_fp16)[name = string("transpose_567")];
            tensor<fp16, [1, 4, 304, 128]> transpose_276 = transpose(perm = transpose_276_perm_0, x = q_h_135_cast_fp16)[name = string("transpose_568")];
            tensor<fp16, [1, 4, 304, 304]> scores_133_cast_fp16 = matmul(transpose_x = scores_133_transpose_x_0, transpose_y = scores_133_transpose_y_0, x = transpose_276, y = transpose_277)[name = string("scores_133_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3680_cast_fp16 = softmax(axis = var_29, x = scores_133_cast_fp16)[name = string("op_3680_cast_fp16")];
            bool x_341_transpose_x_0 = const()[name = string("x_341_transpose_x_0"), val = bool(false)];
            bool x_341_transpose_y_0 = const()[name = string("x_341_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_67_cast_fp16 = transpose(perm = value_67_perm_0, x = var_3653_cast_fp16)[name = string("transpose_571")];
            tensor<fp16, [1, 4, 304, 128]> x_341_cast_fp16 = matmul(transpose_x = x_341_transpose_x_0, transpose_y = x_341_transpose_y_0, x = var_3680_cast_fp16, y = value_67_cast_fp16)[name = string("x_341_cast_fp16")];
            tensor<int32, [4]> var_3684_perm_0 = const()[name = string("op_3684_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3686 = const()[name = string("op_3686"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3684_cast_fp16 = transpose(perm = var_3684_perm_0, x = x_341_cast_fp16)[name = string("transpose_566")];
            tensor<fp16, [1, 304, 512]> input_1009_cast_fp16 = reshape(shape = var_3686, x = var_3684_cast_fp16)[name = string("input_1009_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_32_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_32_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210539136)))];
            tensor<fp16, [512]> m_encoder_encoders_32_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_32_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211063488)))];
            tensor<fp16, [1, 304, 512]> linear_133_cast_fp16 = linear(bias = m_encoder_encoders_32_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_32_self_attn_linear_out_weight_to_fp16, x = input_1009_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1011_cast_fp16 = add(x = linear_133_cast_fp16, y = input_1005_cast_fp16)[name = string("input_1011_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1013_cast_fp16 = add(x = input_997_cast_fp16, y = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")];
            tensor<int32, [1]> output_135_axes_0 = const()[name = string("output_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_392_to_fp16 = const()[name = string("const_392_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211064576)))];
            tensor<fp16, [512]> const_393_to_fp16 = const()[name = string("const_393_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211065664)))];
            tensor<fp16, [1, 304, 512]> output_135_cast_fp16 = layer_norm(axes = output_135_axes_0, beta = const_393_to_fp16, epsilon = var_22_to_fp16, gamma = const_392_to_fp16, x = input_1013_cast_fp16)[name = string("output_135_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_32_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_32_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211066752)))];
            tensor<fp16, [2048]> m_encoder_encoders_32_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_32_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213163968)))];
            tensor<fp16, [1, 304, 2048]> linear_134_cast_fp16 = linear(bias = m_encoder_encoders_32_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_32_feed_forward_w_1_weight_to_fp16, x = output_135_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1021_cast_fp16 = relu(x = linear_134_cast_fp16)[name = string("input_1021_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_32_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_32_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213168128)))];
            tensor<fp16, [512]> m_encoder_encoders_32_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_32_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215265344)))];
            tensor<fp16, [1, 304, 512]> linear_135_cast_fp16 = linear(bias = m_encoder_encoders_32_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_32_feed_forward_w_2_weight_to_fp16, x = input_1021_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1027_cast_fp16 = add(x = input_1013_cast_fp16, y = linear_135_cast_fp16)[name = string("input_1027_cast_fp16")];
            tensor<int32, [1]> output_137_axes_0 = const()[name = string("output_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_394_to_fp16 = const()[name = string("const_394_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215266432)))];
            tensor<fp16, [512]> const_395_to_fp16 = const()[name = string("const_395_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215267520)))];
            tensor<fp16, [1, 304, 512]> output_137_cast_fp16 = layer_norm(axes = output_137_axes_0, beta = const_395_to_fp16, epsilon = var_22_to_fp16, gamma = const_394_to_fp16, x = input_1027_cast_fp16)[name = string("output_137_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_33_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_33_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215268608)))];
            tensor<fp16, [1536]> m_encoder_encoders_33_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_33_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216841536)))];
            tensor<fp16, [1, 304, 1536]> linear_136_cast_fp16 = linear(bias = m_encoder_encoders_33_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_33_self_attn_linear_q_k_v_weight_to_fp16, x = output_137_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<int32, [3]> tile_34 = const()[name = string("tile_34"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3744_axis_0 = const()[name = string("op_3744_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3744_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3744_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3744_cast_fp16_2 = split(axis = var_3744_axis_0, split_sizes = tile_34, x = linear_136_cast_fp16)[name = string("op_3744_cast_fp16")];
            tensor<int32, [4]> var_3748 = const()[name = string("op_3748"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3749_cast_fp16 = reshape(shape = var_3748, x = var_3744_cast_fp16_0)[name = string("op_3749_cast_fp16")];
            tensor<int32, [4]> var_3751 = const()[name = string("op_3751"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3752_cast_fp16 = reshape(shape = var_3751, x = var_3744_cast_fp16_1)[name = string("op_3752_cast_fp16")];
            tensor<int32, [4]> var_3754 = const()[name = string("op_3754"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3755_cast_fp16 = reshape(shape = var_3754, x = var_3744_cast_fp16_2)[name = string("op_3755_cast_fp16")];
            tensor<int32, [4]> value_69_perm_0 = const()[name = string("value_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1031_perm_0 = const()[name = string("input_1031_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1033_pad_0 = const()[name = string("input_1033_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1033_mode_0 = const()[name = string("input_1033_mode_0"), val = string("constant")];
            fp16 const_401_to_fp16 = const()[name = string("const_401_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1031_cast_fp16 = transpose(perm = input_1031_perm_0, x = var_3744_cast_fp16_2)[name = string("transpose_564")];
            tensor<fp16, [1, 512, 314]> input_1033_cast_fp16 = pad(constant_val = const_401_to_fp16, mode = input_1033_mode_0, pad = input_1033_pad_0, x = input_1031_cast_fp16)[name = string("input_1033_cast_fp16")];
            string x_345_pad_type_0 = const()[name = string("x_345_pad_type_0"), val = string("valid")];
            int32 x_345_groups_0 = const()[name = string("x_345_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_345_strides_0 = const()[name = string("x_345_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_345_pad_0 = const()[name = string("x_345_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_345_dilations_0 = const()[name = string("x_345_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_33_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_33_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216844672)))];
            tensor<fp16, [1, 512, 304]> x_345_cast_fp16 = conv(dilations = x_345_dilations_0, groups = x_345_groups_0, pad = x_345_pad_0, pad_type = x_345_pad_type_0, strides = x_345_strides_0, weight = m_encoder_encoders_33_self_attn_fsmn_block_weight_to_fp16, x = input_1033_cast_fp16)[name = string("x_345_cast_fp16")];
            tensor<int32, [3]> x_347_perm_0 = const()[name = string("x_347_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_347_cast_fp16 = transpose(perm = x_347_perm_0, x = x_345_cast_fp16)[name = string("transpose_563")];
            tensor<fp16, [1, 304, 512]> input_1035_cast_fp16 = add(x = x_347_cast_fp16, y = var_3744_cast_fp16_2)[name = string("input_1035_cast_fp16")];
            fp16 var_3774_to_fp16 = const()[name = string("op_3774_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_139_cast_fp16 = mul(x = var_3749_cast_fp16, y = var_3774_to_fp16)[name = string("q_h_139_cast_fp16")];
            bool scores_137_transpose_x_0 = const()[name = string("scores_137_transpose_x_0"), val = bool(false)];
            bool scores_137_transpose_y_0 = const()[name = string("scores_137_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_278_perm_0 = const()[name = string("transpose_278_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_279_perm_0 = const()[name = string("transpose_279_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_279 = transpose(perm = transpose_279_perm_0, x = var_3752_cast_fp16)[name = string("transpose_561")];
            tensor<fp16, [1, 4, 304, 128]> transpose_278 = transpose(perm = transpose_278_perm_0, x = q_h_139_cast_fp16)[name = string("transpose_562")];
            tensor<fp16, [1, 4, 304, 304]> scores_137_cast_fp16 = matmul(transpose_x = scores_137_transpose_x_0, transpose_y = scores_137_transpose_y_0, x = transpose_278, y = transpose_279)[name = string("scores_137_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3782_cast_fp16 = softmax(axis = var_29, x = scores_137_cast_fp16)[name = string("op_3782_cast_fp16")];
            bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)];
            bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_69_cast_fp16 = transpose(perm = value_69_perm_0, x = var_3755_cast_fp16)[name = string("transpose_565")];
            tensor<fp16, [1, 4, 304, 128]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = var_3782_cast_fp16, y = value_69_cast_fp16)[name = string("x_351_cast_fp16")];
            tensor<int32, [4]> var_3786_perm_0 = const()[name = string("op_3786_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3788 = const()[name = string("op_3788"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3786_cast_fp16 = transpose(perm = var_3786_perm_0, x = x_351_cast_fp16)[name = string("transpose_560")];
            tensor<fp16, [1, 304, 512]> input_1039_cast_fp16 = reshape(shape = var_3788, x = var_3786_cast_fp16)[name = string("input_1039_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_33_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_33_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216856000)))];
            tensor<fp16, [512]> m_encoder_encoders_33_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_33_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217380352)))];
            tensor<fp16, [1, 304, 512]> linear_137_cast_fp16 = linear(bias = m_encoder_encoders_33_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_33_self_attn_linear_out_weight_to_fp16, x = input_1039_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1041_cast_fp16 = add(x = linear_137_cast_fp16, y = input_1035_cast_fp16)[name = string("input_1041_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1043_cast_fp16 = add(x = input_1027_cast_fp16, y = input_1041_cast_fp16)[name = string("input_1043_cast_fp16")];
            tensor<int32, [1]> output_139_axes_0 = const()[name = string("output_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_403_to_fp16 = const()[name = string("const_403_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217381440)))];
            tensor<fp16, [512]> const_404_to_fp16 = const()[name = string("const_404_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217382528)))];
            tensor<fp16, [1, 304, 512]> output_139_cast_fp16 = layer_norm(axes = output_139_axes_0, beta = const_404_to_fp16, epsilon = var_22_to_fp16, gamma = const_403_to_fp16, x = input_1043_cast_fp16)[name = string("output_139_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_33_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_33_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217383616)))];
            tensor<fp16, [2048]> m_encoder_encoders_33_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_33_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219480832)))];
            tensor<fp16, [1, 304, 2048]> linear_138_cast_fp16 = linear(bias = m_encoder_encoders_33_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_33_feed_forward_w_1_weight_to_fp16, x = output_139_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1051_cast_fp16 = relu(x = linear_138_cast_fp16)[name = string("input_1051_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_33_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_33_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219484992)))];
            tensor<fp16, [512]> m_encoder_encoders_33_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_33_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221582208)))];
            tensor<fp16, [1, 304, 512]> linear_139_cast_fp16 = linear(bias = m_encoder_encoders_33_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_33_feed_forward_w_2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1057_cast_fp16 = add(x = input_1043_cast_fp16, y = linear_139_cast_fp16)[name = string("input_1057_cast_fp16")];
            tensor<int32, [1]> output_141_axes_0 = const()[name = string("output_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221583296)))];
            tensor<fp16, [512]> const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221584384)))];
            tensor<fp16, [1, 304, 512]> output_141_cast_fp16 = layer_norm(axes = output_141_axes_0, beta = const_406_to_fp16, epsilon = var_22_to_fp16, gamma = const_405_to_fp16, x = input_1057_cast_fp16)[name = string("output_141_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_34_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_34_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221585472)))];
            tensor<fp16, [1536]> m_encoder_encoders_34_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_34_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223158400)))];
            tensor<fp16, [1, 304, 1536]> linear_140_cast_fp16 = linear(bias = m_encoder_encoders_34_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_34_self_attn_linear_q_k_v_weight_to_fp16, x = output_141_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [3]> tile_35 = const()[name = string("tile_35"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3846_axis_0 = const()[name = string("op_3846_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3846_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3846_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3846_cast_fp16_2 = split(axis = var_3846_axis_0, split_sizes = tile_35, x = linear_140_cast_fp16)[name = string("op_3846_cast_fp16")];
            tensor<int32, [4]> var_3850 = const()[name = string("op_3850"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3851_cast_fp16 = reshape(shape = var_3850, x = var_3846_cast_fp16_0)[name = string("op_3851_cast_fp16")];
            tensor<int32, [4]> var_3853 = const()[name = string("op_3853"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3854_cast_fp16 = reshape(shape = var_3853, x = var_3846_cast_fp16_1)[name = string("op_3854_cast_fp16")];
            tensor<int32, [4]> var_3856 = const()[name = string("op_3856"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3857_cast_fp16 = reshape(shape = var_3856, x = var_3846_cast_fp16_2)[name = string("op_3857_cast_fp16")];
            tensor<int32, [4]> value_71_perm_0 = const()[name = string("value_71_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1061_perm_0 = const()[name = string("input_1061_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1063_pad_0 = const()[name = string("input_1063_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1063_mode_0 = const()[name = string("input_1063_mode_0"), val = string("constant")];
            fp16 const_412_to_fp16 = const()[name = string("const_412_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1061_cast_fp16 = transpose(perm = input_1061_perm_0, x = var_3846_cast_fp16_2)[name = string("transpose_558")];
            tensor<fp16, [1, 512, 314]> input_1063_cast_fp16 = pad(constant_val = const_412_to_fp16, mode = input_1063_mode_0, pad = input_1063_pad_0, x = input_1061_cast_fp16)[name = string("input_1063_cast_fp16")];
            string x_355_pad_type_0 = const()[name = string("x_355_pad_type_0"), val = string("valid")];
            int32 x_355_groups_0 = const()[name = string("x_355_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_355_strides_0 = const()[name = string("x_355_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_355_pad_0 = const()[name = string("x_355_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_355_dilations_0 = const()[name = string("x_355_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_34_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_34_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223161536)))];
            tensor<fp16, [1, 512, 304]> x_355_cast_fp16 = conv(dilations = x_355_dilations_0, groups = x_355_groups_0, pad = x_355_pad_0, pad_type = x_355_pad_type_0, strides = x_355_strides_0, weight = m_encoder_encoders_34_self_attn_fsmn_block_weight_to_fp16, x = input_1063_cast_fp16)[name = string("x_355_cast_fp16")];
            tensor<int32, [3]> x_357_perm_0 = const()[name = string("x_357_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_357_cast_fp16 = transpose(perm = x_357_perm_0, x = x_355_cast_fp16)[name = string("transpose_557")];
            tensor<fp16, [1, 304, 512]> input_1065_cast_fp16 = add(x = x_357_cast_fp16, y = var_3846_cast_fp16_2)[name = string("input_1065_cast_fp16")];
            fp16 var_3876_to_fp16 = const()[name = string("op_3876_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_143_cast_fp16 = mul(x = var_3851_cast_fp16, y = var_3876_to_fp16)[name = string("q_h_143_cast_fp16")];
            bool scores_141_transpose_x_0 = const()[name = string("scores_141_transpose_x_0"), val = bool(false)];
            bool scores_141_transpose_y_0 = const()[name = string("scores_141_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_280_perm_0 = const()[name = string("transpose_280_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_281_perm_0 = const()[name = string("transpose_281_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_281 = transpose(perm = transpose_281_perm_0, x = var_3854_cast_fp16)[name = string("transpose_555")];
            tensor<fp16, [1, 4, 304, 128]> transpose_280 = transpose(perm = transpose_280_perm_0, x = q_h_143_cast_fp16)[name = string("transpose_556")];
            tensor<fp16, [1, 4, 304, 304]> scores_141_cast_fp16 = matmul(transpose_x = scores_141_transpose_x_0, transpose_y = scores_141_transpose_y_0, x = transpose_280, y = transpose_281)[name = string("scores_141_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3884_cast_fp16 = softmax(axis = var_29, x = scores_141_cast_fp16)[name = string("op_3884_cast_fp16")];
            bool x_361_transpose_x_0 = const()[name = string("x_361_transpose_x_0"), val = bool(false)];
            bool x_361_transpose_y_0 = const()[name = string("x_361_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_71_cast_fp16 = transpose(perm = value_71_perm_0, x = var_3857_cast_fp16)[name = string("transpose_559")];
            tensor<fp16, [1, 4, 304, 128]> x_361_cast_fp16 = matmul(transpose_x = x_361_transpose_x_0, transpose_y = x_361_transpose_y_0, x = var_3884_cast_fp16, y = value_71_cast_fp16)[name = string("x_361_cast_fp16")];
            tensor<int32, [4]> var_3888_perm_0 = const()[name = string("op_3888_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3890 = const()[name = string("op_3890"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3888_cast_fp16 = transpose(perm = var_3888_perm_0, x = x_361_cast_fp16)[name = string("transpose_554")];
            tensor<fp16, [1, 304, 512]> input_1069_cast_fp16 = reshape(shape = var_3890, x = var_3888_cast_fp16)[name = string("input_1069_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_34_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_34_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223172864)))];
            tensor<fp16, [512]> m_encoder_encoders_34_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_34_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223697216)))];
            tensor<fp16, [1, 304, 512]> linear_141_cast_fp16 = linear(bias = m_encoder_encoders_34_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_34_self_attn_linear_out_weight_to_fp16, x = input_1069_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1071_cast_fp16 = add(x = linear_141_cast_fp16, y = input_1065_cast_fp16)[name = string("input_1071_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1073_cast_fp16 = add(x = input_1057_cast_fp16, y = input_1071_cast_fp16)[name = string("input_1073_cast_fp16")];
            tensor<int32, [1]> output_143_axes_0 = const()[name = string("output_143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_414_to_fp16 = const()[name = string("const_414_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223698304)))];
            tensor<fp16, [512]> const_415_to_fp16 = const()[name = string("const_415_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223699392)))];
            tensor<fp16, [1, 304, 512]> output_143_cast_fp16 = layer_norm(axes = output_143_axes_0, beta = const_415_to_fp16, epsilon = var_22_to_fp16, gamma = const_414_to_fp16, x = input_1073_cast_fp16)[name = string("output_143_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_34_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_34_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223700480)))];
            tensor<fp16, [2048]> m_encoder_encoders_34_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_34_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225797696)))];
            tensor<fp16, [1, 304, 2048]> linear_142_cast_fp16 = linear(bias = m_encoder_encoders_34_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_34_feed_forward_w_1_weight_to_fp16, x = output_143_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1081_cast_fp16 = relu(x = linear_142_cast_fp16)[name = string("input_1081_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_34_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_34_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225801856)))];
            tensor<fp16, [512]> m_encoder_encoders_34_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_34_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227899072)))];
            tensor<fp16, [1, 304, 512]> linear_143_cast_fp16 = linear(bias = m_encoder_encoders_34_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_34_feed_forward_w_2_weight_to_fp16, x = input_1081_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1087_cast_fp16 = add(x = input_1073_cast_fp16, y = linear_143_cast_fp16)[name = string("input_1087_cast_fp16")];
            tensor<int32, [1]> output_145_axes_0 = const()[name = string("output_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_416_to_fp16 = const()[name = string("const_416_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227900160)))];
            tensor<fp16, [512]> const_417_to_fp16 = const()[name = string("const_417_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227901248)))];
            tensor<fp16, [1, 304, 512]> output_145_cast_fp16 = layer_norm(axes = output_145_axes_0, beta = const_417_to_fp16, epsilon = var_22_to_fp16, gamma = const_416_to_fp16, x = input_1087_cast_fp16)[name = string("output_145_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_35_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_35_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227902336)))];
            tensor<fp16, [1536]> m_encoder_encoders_35_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_35_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229475264)))];
            tensor<fp16, [1, 304, 1536]> linear_144_cast_fp16 = linear(bias = m_encoder_encoders_35_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_35_self_attn_linear_q_k_v_weight_to_fp16, x = output_145_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<int32, [3]> tile_36 = const()[name = string("tile_36"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_3948_axis_0 = const()[name = string("op_3948_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_3948_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_3948_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_3948_cast_fp16_2 = split(axis = var_3948_axis_0, split_sizes = tile_36, x = linear_144_cast_fp16)[name = string("op_3948_cast_fp16")];
            tensor<int32, [4]> var_3952 = const()[name = string("op_3952"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3953_cast_fp16 = reshape(shape = var_3952, x = var_3948_cast_fp16_0)[name = string("op_3953_cast_fp16")];
            tensor<int32, [4]> var_3955 = const()[name = string("op_3955"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3956_cast_fp16 = reshape(shape = var_3955, x = var_3948_cast_fp16_1)[name = string("op_3956_cast_fp16")];
            tensor<int32, [4]> var_3958 = const()[name = string("op_3958"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_3959_cast_fp16 = reshape(shape = var_3958, x = var_3948_cast_fp16_2)[name = string("op_3959_cast_fp16")];
            tensor<int32, [4]> value_73_perm_0 = const()[name = string("value_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1091_perm_0 = const()[name = string("input_1091_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1093_pad_0 = const()[name = string("input_1093_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1093_mode_0 = const()[name = string("input_1093_mode_0"), val = string("constant")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1091_cast_fp16 = transpose(perm = input_1091_perm_0, x = var_3948_cast_fp16_2)[name = string("transpose_552")];
            tensor<fp16, [1, 512, 314]> input_1093_cast_fp16 = pad(constant_val = const_423_to_fp16, mode = input_1093_mode_0, pad = input_1093_pad_0, x = input_1091_cast_fp16)[name = string("input_1093_cast_fp16")];
            string x_365_pad_type_0 = const()[name = string("x_365_pad_type_0"), val = string("valid")];
            int32 x_365_groups_0 = const()[name = string("x_365_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_365_strides_0 = const()[name = string("x_365_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_365_pad_0 = const()[name = string("x_365_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_365_dilations_0 = const()[name = string("x_365_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_35_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_35_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229478400)))];
            tensor<fp16, [1, 512, 304]> x_365_cast_fp16 = conv(dilations = x_365_dilations_0, groups = x_365_groups_0, pad = x_365_pad_0, pad_type = x_365_pad_type_0, strides = x_365_strides_0, weight = m_encoder_encoders_35_self_attn_fsmn_block_weight_to_fp16, x = input_1093_cast_fp16)[name = string("x_365_cast_fp16")];
            tensor<int32, [3]> x_367_perm_0 = const()[name = string("x_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_367_cast_fp16 = transpose(perm = x_367_perm_0, x = x_365_cast_fp16)[name = string("transpose_551")];
            tensor<fp16, [1, 304, 512]> input_1095_cast_fp16 = add(x = x_367_cast_fp16, y = var_3948_cast_fp16_2)[name = string("input_1095_cast_fp16")];
            fp16 var_3978_to_fp16 = const()[name = string("op_3978_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_147_cast_fp16 = mul(x = var_3953_cast_fp16, y = var_3978_to_fp16)[name = string("q_h_147_cast_fp16")];
            bool scores_145_transpose_x_0 = const()[name = string("scores_145_transpose_x_0"), val = bool(false)];
            bool scores_145_transpose_y_0 = const()[name = string("scores_145_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_282_perm_0 = const()[name = string("transpose_282_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_283_perm_0 = const()[name = string("transpose_283_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_283 = transpose(perm = transpose_283_perm_0, x = var_3956_cast_fp16)[name = string("transpose_549")];
            tensor<fp16, [1, 4, 304, 128]> transpose_282 = transpose(perm = transpose_282_perm_0, x = q_h_147_cast_fp16)[name = string("transpose_550")];
            tensor<fp16, [1, 4, 304, 304]> scores_145_cast_fp16 = matmul(transpose_x = scores_145_transpose_x_0, transpose_y = scores_145_transpose_y_0, x = transpose_282, y = transpose_283)[name = string("scores_145_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_3986_cast_fp16 = softmax(axis = var_29, x = scores_145_cast_fp16)[name = string("op_3986_cast_fp16")];
            bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)];
            bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_73_cast_fp16 = transpose(perm = value_73_perm_0, x = var_3959_cast_fp16)[name = string("transpose_553")];
            tensor<fp16, [1, 4, 304, 128]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = var_3986_cast_fp16, y = value_73_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<int32, [4]> var_3990_perm_0 = const()[name = string("op_3990_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3992 = const()[name = string("op_3992"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_3990_cast_fp16 = transpose(perm = var_3990_perm_0, x = x_371_cast_fp16)[name = string("transpose_548")];
            tensor<fp16, [1, 304, 512]> input_1099_cast_fp16 = reshape(shape = var_3992, x = var_3990_cast_fp16)[name = string("input_1099_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_35_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_35_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229489728)))];
            tensor<fp16, [512]> m_encoder_encoders_35_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_35_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230014080)))];
            tensor<fp16, [1, 304, 512]> linear_145_cast_fp16 = linear(bias = m_encoder_encoders_35_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_35_self_attn_linear_out_weight_to_fp16, x = input_1099_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1101_cast_fp16 = add(x = linear_145_cast_fp16, y = input_1095_cast_fp16)[name = string("input_1101_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")];
            tensor<int32, [1]> output_147_axes_0 = const()[name = string("output_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_425_to_fp16 = const()[name = string("const_425_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230015168)))];
            tensor<fp16, [512]> const_426_to_fp16 = const()[name = string("const_426_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230016256)))];
            tensor<fp16, [1, 304, 512]> output_147_cast_fp16 = layer_norm(axes = output_147_axes_0, beta = const_426_to_fp16, epsilon = var_22_to_fp16, gamma = const_425_to_fp16, x = input_1103_cast_fp16)[name = string("output_147_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_35_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_35_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230017344)))];
            tensor<fp16, [2048]> m_encoder_encoders_35_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_35_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232114560)))];
            tensor<fp16, [1, 304, 2048]> linear_146_cast_fp16 = linear(bias = m_encoder_encoders_35_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_35_feed_forward_w_1_weight_to_fp16, x = output_147_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1111_cast_fp16 = relu(x = linear_146_cast_fp16)[name = string("input_1111_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_35_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_35_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232118720)))];
            tensor<fp16, [512]> m_encoder_encoders_35_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_35_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234215936)))];
            tensor<fp16, [1, 304, 512]> linear_147_cast_fp16 = linear(bias = m_encoder_encoders_35_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_35_feed_forward_w_2_weight_to_fp16, x = input_1111_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1117_cast_fp16 = add(x = input_1103_cast_fp16, y = linear_147_cast_fp16)[name = string("input_1117_cast_fp16")];
            tensor<int32, [1]> output_149_axes_0 = const()[name = string("output_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_427_to_fp16 = const()[name = string("const_427_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234217024)))];
            tensor<fp16, [512]> const_428_to_fp16 = const()[name = string("const_428_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234218112)))];
            tensor<fp16, [1, 304, 512]> output_149_cast_fp16 = layer_norm(axes = output_149_axes_0, beta = const_428_to_fp16, epsilon = var_22_to_fp16, gamma = const_427_to_fp16, x = input_1117_cast_fp16)[name = string("output_149_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_36_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_36_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234219200)))];
            tensor<fp16, [1536]> m_encoder_encoders_36_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_36_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235792128)))];
            tensor<fp16, [1, 304, 1536]> linear_148_cast_fp16 = linear(bias = m_encoder_encoders_36_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_36_self_attn_linear_q_k_v_weight_to_fp16, x = output_149_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [3]> tile_37 = const()[name = string("tile_37"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4050_axis_0 = const()[name = string("op_4050_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4050_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4050_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4050_cast_fp16_2 = split(axis = var_4050_axis_0, split_sizes = tile_37, x = linear_148_cast_fp16)[name = string("op_4050_cast_fp16")];
            tensor<int32, [4]> var_4054 = const()[name = string("op_4054"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4055_cast_fp16 = reshape(shape = var_4054, x = var_4050_cast_fp16_0)[name = string("op_4055_cast_fp16")];
            tensor<int32, [4]> var_4057 = const()[name = string("op_4057"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4058_cast_fp16 = reshape(shape = var_4057, x = var_4050_cast_fp16_1)[name = string("op_4058_cast_fp16")];
            tensor<int32, [4]> var_4060 = const()[name = string("op_4060"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4061_cast_fp16 = reshape(shape = var_4060, x = var_4050_cast_fp16_2)[name = string("op_4061_cast_fp16")];
            tensor<int32, [4]> value_75_perm_0 = const()[name = string("value_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1121_perm_0 = const()[name = string("input_1121_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1123_pad_0 = const()[name = string("input_1123_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1123_mode_0 = const()[name = string("input_1123_mode_0"), val = string("constant")];
            fp16 const_434_to_fp16 = const()[name = string("const_434_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1121_cast_fp16 = transpose(perm = input_1121_perm_0, x = var_4050_cast_fp16_2)[name = string("transpose_546")];
            tensor<fp16, [1, 512, 314]> input_1123_cast_fp16 = pad(constant_val = const_434_to_fp16, mode = input_1123_mode_0, pad = input_1123_pad_0, x = input_1121_cast_fp16)[name = string("input_1123_cast_fp16")];
            string x_375_pad_type_0 = const()[name = string("x_375_pad_type_0"), val = string("valid")];
            int32 x_375_groups_0 = const()[name = string("x_375_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_375_strides_0 = const()[name = string("x_375_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_375_pad_0 = const()[name = string("x_375_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_375_dilations_0 = const()[name = string("x_375_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_36_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_36_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235795264)))];
            tensor<fp16, [1, 512, 304]> x_375_cast_fp16 = conv(dilations = x_375_dilations_0, groups = x_375_groups_0, pad = x_375_pad_0, pad_type = x_375_pad_type_0, strides = x_375_strides_0, weight = m_encoder_encoders_36_self_attn_fsmn_block_weight_to_fp16, x = input_1123_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [3]> x_377_perm_0 = const()[name = string("x_377_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_377_cast_fp16 = transpose(perm = x_377_perm_0, x = x_375_cast_fp16)[name = string("transpose_545")];
            tensor<fp16, [1, 304, 512]> input_1125_cast_fp16 = add(x = x_377_cast_fp16, y = var_4050_cast_fp16_2)[name = string("input_1125_cast_fp16")];
            fp16 var_4080_to_fp16 = const()[name = string("op_4080_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_151_cast_fp16 = mul(x = var_4055_cast_fp16, y = var_4080_to_fp16)[name = string("q_h_151_cast_fp16")];
            bool scores_149_transpose_x_0 = const()[name = string("scores_149_transpose_x_0"), val = bool(false)];
            bool scores_149_transpose_y_0 = const()[name = string("scores_149_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_284_perm_0 = const()[name = string("transpose_284_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_285_perm_0 = const()[name = string("transpose_285_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_285 = transpose(perm = transpose_285_perm_0, x = var_4058_cast_fp16)[name = string("transpose_543")];
            tensor<fp16, [1, 4, 304, 128]> transpose_284 = transpose(perm = transpose_284_perm_0, x = q_h_151_cast_fp16)[name = string("transpose_544")];
            tensor<fp16, [1, 4, 304, 304]> scores_149_cast_fp16 = matmul(transpose_x = scores_149_transpose_x_0, transpose_y = scores_149_transpose_y_0, x = transpose_284, y = transpose_285)[name = string("scores_149_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_4088_cast_fp16 = softmax(axis = var_29, x = scores_149_cast_fp16)[name = string("op_4088_cast_fp16")];
            bool x_381_transpose_x_0 = const()[name = string("x_381_transpose_x_0"), val = bool(false)];
            bool x_381_transpose_y_0 = const()[name = string("x_381_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_75_cast_fp16 = transpose(perm = value_75_perm_0, x = var_4061_cast_fp16)[name = string("transpose_547")];
            tensor<fp16, [1, 4, 304, 128]> x_381_cast_fp16 = matmul(transpose_x = x_381_transpose_x_0, transpose_y = x_381_transpose_y_0, x = var_4088_cast_fp16, y = value_75_cast_fp16)[name = string("x_381_cast_fp16")];
            tensor<int32, [4]> var_4092_perm_0 = const()[name = string("op_4092_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4094 = const()[name = string("op_4094"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_4092_cast_fp16 = transpose(perm = var_4092_perm_0, x = x_381_cast_fp16)[name = string("transpose_542")];
            tensor<fp16, [1, 304, 512]> input_1129_cast_fp16 = reshape(shape = var_4094, x = var_4092_cast_fp16)[name = string("input_1129_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_36_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_36_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235806592)))];
            tensor<fp16, [512]> m_encoder_encoders_36_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_36_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236330944)))];
            tensor<fp16, [1, 304, 512]> linear_149_cast_fp16 = linear(bias = m_encoder_encoders_36_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_36_self_attn_linear_out_weight_to_fp16, x = input_1129_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1131_cast_fp16 = add(x = linear_149_cast_fp16, y = input_1125_cast_fp16)[name = string("input_1131_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1133_cast_fp16 = add(x = input_1117_cast_fp16, y = input_1131_cast_fp16)[name = string("input_1133_cast_fp16")];
            tensor<int32, [1]> output_151_axes_0 = const()[name = string("output_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_436_to_fp16 = const()[name = string("const_436_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236332032)))];
            tensor<fp16, [512]> const_437_to_fp16 = const()[name = string("const_437_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236333120)))];
            tensor<fp16, [1, 304, 512]> output_151_cast_fp16 = layer_norm(axes = output_151_axes_0, beta = const_437_to_fp16, epsilon = var_22_to_fp16, gamma = const_436_to_fp16, x = input_1133_cast_fp16)[name = string("output_151_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_36_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_36_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236334208)))];
            tensor<fp16, [2048]> m_encoder_encoders_36_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_36_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238431424)))];
            tensor<fp16, [1, 304, 2048]> linear_150_cast_fp16 = linear(bias = m_encoder_encoders_36_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_36_feed_forward_w_1_weight_to_fp16, x = output_151_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1141_cast_fp16 = relu(x = linear_150_cast_fp16)[name = string("input_1141_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_36_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_36_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238435584)))];
            tensor<fp16, [512]> m_encoder_encoders_36_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_36_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240532800)))];
            tensor<fp16, [1, 304, 512]> linear_151_cast_fp16 = linear(bias = m_encoder_encoders_36_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_36_feed_forward_w_2_weight_to_fp16, x = input_1141_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1147_cast_fp16 = add(x = input_1133_cast_fp16, y = linear_151_cast_fp16)[name = string("input_1147_cast_fp16")];
            tensor<int32, [1]> output_153_axes_0 = const()[name = string("output_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_438_to_fp16 = const()[name = string("const_438_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240533888)))];
            tensor<fp16, [512]> const_439_to_fp16 = const()[name = string("const_439_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240534976)))];
            tensor<fp16, [1, 304, 512]> output_153_cast_fp16 = layer_norm(axes = output_153_axes_0, beta = const_439_to_fp16, epsilon = var_22_to_fp16, gamma = const_438_to_fp16, x = input_1147_cast_fp16)[name = string("output_153_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_37_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_37_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240536064)))];
            tensor<fp16, [1536]> m_encoder_encoders_37_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_37_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242108992)))];
            tensor<fp16, [1, 304, 1536]> linear_152_cast_fp16 = linear(bias = m_encoder_encoders_37_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_37_self_attn_linear_q_k_v_weight_to_fp16, x = output_153_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<int32, [3]> tile_38 = const()[name = string("tile_38"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4152_axis_0 = const()[name = string("op_4152_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4152_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4152_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4152_cast_fp16_2 = split(axis = var_4152_axis_0, split_sizes = tile_38, x = linear_152_cast_fp16)[name = string("op_4152_cast_fp16")];
            tensor<int32, [4]> var_4156 = const()[name = string("op_4156"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4157_cast_fp16 = reshape(shape = var_4156, x = var_4152_cast_fp16_0)[name = string("op_4157_cast_fp16")];
            tensor<int32, [4]> var_4159 = const()[name = string("op_4159"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4160_cast_fp16 = reshape(shape = var_4159, x = var_4152_cast_fp16_1)[name = string("op_4160_cast_fp16")];
            tensor<int32, [4]> var_4162 = const()[name = string("op_4162"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4163_cast_fp16 = reshape(shape = var_4162, x = var_4152_cast_fp16_2)[name = string("op_4163_cast_fp16")];
            tensor<int32, [4]> value_77_perm_0 = const()[name = string("value_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1151_perm_0 = const()[name = string("input_1151_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1153_pad_0 = const()[name = string("input_1153_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1153_mode_0 = const()[name = string("input_1153_mode_0"), val = string("constant")];
            fp16 const_445_to_fp16 = const()[name = string("const_445_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1151_cast_fp16 = transpose(perm = input_1151_perm_0, x = var_4152_cast_fp16_2)[name = string("transpose_540")];
            tensor<fp16, [1, 512, 314]> input_1153_cast_fp16 = pad(constant_val = const_445_to_fp16, mode = input_1153_mode_0, pad = input_1153_pad_0, x = input_1151_cast_fp16)[name = string("input_1153_cast_fp16")];
            string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")];
            int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_37_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_37_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242112128)))];
            tensor<fp16, [1, 512, 304]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = m_encoder_encoders_37_self_attn_fsmn_block_weight_to_fp16, x = input_1153_cast_fp16)[name = string("x_385_cast_fp16")];
            tensor<int32, [3]> x_387_perm_0 = const()[name = string("x_387_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_387_cast_fp16 = transpose(perm = x_387_perm_0, x = x_385_cast_fp16)[name = string("transpose_539")];
            tensor<fp16, [1, 304, 512]> input_1155_cast_fp16 = add(x = x_387_cast_fp16, y = var_4152_cast_fp16_2)[name = string("input_1155_cast_fp16")];
            fp16 var_4182_to_fp16 = const()[name = string("op_4182_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_155_cast_fp16 = mul(x = var_4157_cast_fp16, y = var_4182_to_fp16)[name = string("q_h_155_cast_fp16")];
            bool scores_153_transpose_x_0 = const()[name = string("scores_153_transpose_x_0"), val = bool(false)];
            bool scores_153_transpose_y_0 = const()[name = string("scores_153_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_286_perm_0 = const()[name = string("transpose_286_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_287_perm_0 = const()[name = string("transpose_287_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_287 = transpose(perm = transpose_287_perm_0, x = var_4160_cast_fp16)[name = string("transpose_537")];
            tensor<fp16, [1, 4, 304, 128]> transpose_286 = transpose(perm = transpose_286_perm_0, x = q_h_155_cast_fp16)[name = string("transpose_538")];
            tensor<fp16, [1, 4, 304, 304]> scores_153_cast_fp16 = matmul(transpose_x = scores_153_transpose_x_0, transpose_y = scores_153_transpose_y_0, x = transpose_286, y = transpose_287)[name = string("scores_153_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_4190_cast_fp16 = softmax(axis = var_29, x = scores_153_cast_fp16)[name = string("op_4190_cast_fp16")];
            bool x_391_transpose_x_0 = const()[name = string("x_391_transpose_x_0"), val = bool(false)];
            bool x_391_transpose_y_0 = const()[name = string("x_391_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_77_cast_fp16 = transpose(perm = value_77_perm_0, x = var_4163_cast_fp16)[name = string("transpose_541")];
            tensor<fp16, [1, 4, 304, 128]> x_391_cast_fp16 = matmul(transpose_x = x_391_transpose_x_0, transpose_y = x_391_transpose_y_0, x = var_4190_cast_fp16, y = value_77_cast_fp16)[name = string("x_391_cast_fp16")];
            tensor<int32, [4]> var_4194_perm_0 = const()[name = string("op_4194_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4196 = const()[name = string("op_4196"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_4194_cast_fp16 = transpose(perm = var_4194_perm_0, x = x_391_cast_fp16)[name = string("transpose_536")];
            tensor<fp16, [1, 304, 512]> input_1159_cast_fp16 = reshape(shape = var_4196, x = var_4194_cast_fp16)[name = string("input_1159_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_37_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_37_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242123456)))];
            tensor<fp16, [512]> m_encoder_encoders_37_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_37_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242647808)))];
            tensor<fp16, [1, 304, 512]> linear_153_cast_fp16 = linear(bias = m_encoder_encoders_37_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_37_self_attn_linear_out_weight_to_fp16, x = input_1159_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1161_cast_fp16 = add(x = linear_153_cast_fp16, y = input_1155_cast_fp16)[name = string("input_1161_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1163_cast_fp16 = add(x = input_1147_cast_fp16, y = input_1161_cast_fp16)[name = string("input_1163_cast_fp16")];
            tensor<int32, [1]> output_155_axes_0 = const()[name = string("output_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_447_to_fp16 = const()[name = string("const_447_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242648896)))];
            tensor<fp16, [512]> const_448_to_fp16 = const()[name = string("const_448_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242649984)))];
            tensor<fp16, [1, 304, 512]> output_155_cast_fp16 = layer_norm(axes = output_155_axes_0, beta = const_448_to_fp16, epsilon = var_22_to_fp16, gamma = const_447_to_fp16, x = input_1163_cast_fp16)[name = string("output_155_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_37_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_37_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242651072)))];
            tensor<fp16, [2048]> m_encoder_encoders_37_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_37_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244748288)))];
            tensor<fp16, [1, 304, 2048]> linear_154_cast_fp16 = linear(bias = m_encoder_encoders_37_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_37_feed_forward_w_1_weight_to_fp16, x = output_155_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1171_cast_fp16 = relu(x = linear_154_cast_fp16)[name = string("input_1171_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_37_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_37_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244752448)))];
            tensor<fp16, [512]> m_encoder_encoders_37_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_37_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246849664)))];
            tensor<fp16, [1, 304, 512]> linear_155_cast_fp16 = linear(bias = m_encoder_encoders_37_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_37_feed_forward_w_2_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1177_cast_fp16 = add(x = input_1163_cast_fp16, y = linear_155_cast_fp16)[name = string("input_1177_cast_fp16")];
            tensor<int32, [1]> output_157_axes_0 = const()[name = string("output_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_449_to_fp16 = const()[name = string("const_449_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246850752)))];
            tensor<fp16, [512]> const_450_to_fp16 = const()[name = string("const_450_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246851840)))];
            tensor<fp16, [1, 304, 512]> output_157_cast_fp16 = layer_norm(axes = output_157_axes_0, beta = const_450_to_fp16, epsilon = var_22_to_fp16, gamma = const_449_to_fp16, x = input_1177_cast_fp16)[name = string("output_157_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_38_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_38_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246852928)))];
            tensor<fp16, [1536]> m_encoder_encoders_38_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_38_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248425856)))];
            tensor<fp16, [1, 304, 1536]> linear_156_cast_fp16 = linear(bias = m_encoder_encoders_38_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_38_self_attn_linear_q_k_v_weight_to_fp16, x = output_157_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [3]> tile_39 = const()[name = string("tile_39"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4254_axis_0 = const()[name = string("op_4254_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4254_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4254_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4254_cast_fp16_2 = split(axis = var_4254_axis_0, split_sizes = tile_39, x = linear_156_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<int32, [4]> var_4258 = const()[name = string("op_4258"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4259_cast_fp16 = reshape(shape = var_4258, x = var_4254_cast_fp16_0)[name = string("op_4259_cast_fp16")];
            tensor<int32, [4]> var_4261 = const()[name = string("op_4261"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4262_cast_fp16 = reshape(shape = var_4261, x = var_4254_cast_fp16_1)[name = string("op_4262_cast_fp16")];
            tensor<int32, [4]> var_4264 = const()[name = string("op_4264"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4265_cast_fp16 = reshape(shape = var_4264, x = var_4254_cast_fp16_2)[name = string("op_4265_cast_fp16")];
            tensor<int32, [4]> value_79_perm_0 = const()[name = string("value_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1181_perm_0 = const()[name = string("input_1181_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1183_pad_0 = const()[name = string("input_1183_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1183_mode_0 = const()[name = string("input_1183_mode_0"), val = string("constant")];
            fp16 const_456_to_fp16 = const()[name = string("const_456_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1181_cast_fp16 = transpose(perm = input_1181_perm_0, x = var_4254_cast_fp16_2)[name = string("transpose_534")];
            tensor<fp16, [1, 512, 314]> input_1183_cast_fp16 = pad(constant_val = const_456_to_fp16, mode = input_1183_mode_0, pad = input_1183_pad_0, x = input_1181_cast_fp16)[name = string("input_1183_cast_fp16")];
            string x_395_pad_type_0 = const()[name = string("x_395_pad_type_0"), val = string("valid")];
            int32 x_395_groups_0 = const()[name = string("x_395_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_395_strides_0 = const()[name = string("x_395_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_395_pad_0 = const()[name = string("x_395_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_395_dilations_0 = const()[name = string("x_395_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_38_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_38_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248428992)))];
            tensor<fp16, [1, 512, 304]> x_395_cast_fp16 = conv(dilations = x_395_dilations_0, groups = x_395_groups_0, pad = x_395_pad_0, pad_type = x_395_pad_type_0, strides = x_395_strides_0, weight = m_encoder_encoders_38_self_attn_fsmn_block_weight_to_fp16, x = input_1183_cast_fp16)[name = string("x_395_cast_fp16")];
            tensor<int32, [3]> x_397_perm_0 = const()[name = string("x_397_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_397_cast_fp16 = transpose(perm = x_397_perm_0, x = x_395_cast_fp16)[name = string("transpose_533")];
            tensor<fp16, [1, 304, 512]> input_1185_cast_fp16 = add(x = x_397_cast_fp16, y = var_4254_cast_fp16_2)[name = string("input_1185_cast_fp16")];
            fp16 var_4284_to_fp16 = const()[name = string("op_4284_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_159_cast_fp16 = mul(x = var_4259_cast_fp16, y = var_4284_to_fp16)[name = string("q_h_159_cast_fp16")];
            bool scores_157_transpose_x_0 = const()[name = string("scores_157_transpose_x_0"), val = bool(false)];
            bool scores_157_transpose_y_0 = const()[name = string("scores_157_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_288_perm_0 = const()[name = string("transpose_288_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_289_perm_0 = const()[name = string("transpose_289_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_289 = transpose(perm = transpose_289_perm_0, x = var_4262_cast_fp16)[name = string("transpose_531")];
            tensor<fp16, [1, 4, 304, 128]> transpose_288 = transpose(perm = transpose_288_perm_0, x = q_h_159_cast_fp16)[name = string("transpose_532")];
            tensor<fp16, [1, 4, 304, 304]> scores_157_cast_fp16 = matmul(transpose_x = scores_157_transpose_x_0, transpose_y = scores_157_transpose_y_0, x = transpose_288, y = transpose_289)[name = string("scores_157_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_4292_cast_fp16 = softmax(axis = var_29, x = scores_157_cast_fp16)[name = string("op_4292_cast_fp16")];
            bool x_401_transpose_x_0 = const()[name = string("x_401_transpose_x_0"), val = bool(false)];
            bool x_401_transpose_y_0 = const()[name = string("x_401_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_79_cast_fp16 = transpose(perm = value_79_perm_0, x = var_4265_cast_fp16)[name = string("transpose_535")];
            tensor<fp16, [1, 4, 304, 128]> x_401_cast_fp16 = matmul(transpose_x = x_401_transpose_x_0, transpose_y = x_401_transpose_y_0, x = var_4292_cast_fp16, y = value_79_cast_fp16)[name = string("x_401_cast_fp16")];
            tensor<int32, [4]> var_4296_perm_0 = const()[name = string("op_4296_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4298 = const()[name = string("op_4298"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_4296_cast_fp16 = transpose(perm = var_4296_perm_0, x = x_401_cast_fp16)[name = string("transpose_530")];
            tensor<fp16, [1, 304, 512]> input_1189_cast_fp16 = reshape(shape = var_4298, x = var_4296_cast_fp16)[name = string("input_1189_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_38_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_38_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248440320)))];
            tensor<fp16, [512]> m_encoder_encoders_38_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_38_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248964672)))];
            tensor<fp16, [1, 304, 512]> linear_157_cast_fp16 = linear(bias = m_encoder_encoders_38_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_38_self_attn_linear_out_weight_to_fp16, x = input_1189_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1191_cast_fp16 = add(x = linear_157_cast_fp16, y = input_1185_cast_fp16)[name = string("input_1191_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1193_cast_fp16 = add(x = input_1177_cast_fp16, y = input_1191_cast_fp16)[name = string("input_1193_cast_fp16")];
            tensor<int32, [1]> output_159_axes_0 = const()[name = string("output_159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_458_to_fp16 = const()[name = string("const_458_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248965760)))];
            tensor<fp16, [512]> const_459_to_fp16 = const()[name = string("const_459_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248966848)))];
            tensor<fp16, [1, 304, 512]> output_159_cast_fp16 = layer_norm(axes = output_159_axes_0, beta = const_459_to_fp16, epsilon = var_22_to_fp16, gamma = const_458_to_fp16, x = input_1193_cast_fp16)[name = string("output_159_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_38_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_38_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248967936)))];
            tensor<fp16, [2048]> m_encoder_encoders_38_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_38_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251065152)))];
            tensor<fp16, [1, 304, 2048]> linear_158_cast_fp16 = linear(bias = m_encoder_encoders_38_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_38_feed_forward_w_1_weight_to_fp16, x = output_159_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1201_cast_fp16 = relu(x = linear_158_cast_fp16)[name = string("input_1201_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_38_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_38_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251069312)))];
            tensor<fp16, [512]> m_encoder_encoders_38_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_38_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253166528)))];
            tensor<fp16, [1, 304, 512]> linear_159_cast_fp16 = linear(bias = m_encoder_encoders_38_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_38_feed_forward_w_2_weight_to_fp16, x = input_1201_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1207_cast_fp16 = add(x = input_1193_cast_fp16, y = linear_159_cast_fp16)[name = string("input_1207_cast_fp16")];
            tensor<int32, [1]> output_161_axes_0 = const()[name = string("output_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_460_to_fp16 = const()[name = string("const_460_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253167616)))];
            tensor<fp16, [512]> const_461_to_fp16 = const()[name = string("const_461_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253168704)))];
            tensor<fp16, [1, 304, 512]> output_161_cast_fp16 = layer_norm(axes = output_161_axes_0, beta = const_461_to_fp16, epsilon = var_22_to_fp16, gamma = const_460_to_fp16, x = input_1207_cast_fp16)[name = string("output_161_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_39_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_39_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253169792)))];
            tensor<fp16, [1536]> m_encoder_encoders_39_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_39_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254742720)))];
            tensor<fp16, [1, 304, 1536]> linear_160_cast_fp16 = linear(bias = m_encoder_encoders_39_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_39_self_attn_linear_q_k_v_weight_to_fp16, x = output_161_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<int32, [3]> tile_40 = const()[name = string("tile_40"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4356_axis_0 = const()[name = string("op_4356_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4356_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4356_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4356_cast_fp16_2 = split(axis = var_4356_axis_0, split_sizes = tile_40, x = linear_160_cast_fp16)[name = string("op_4356_cast_fp16")];
            tensor<int32, [4]> var_4360 = const()[name = string("op_4360"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4361_cast_fp16 = reshape(shape = var_4360, x = var_4356_cast_fp16_0)[name = string("op_4361_cast_fp16")];
            tensor<int32, [4]> var_4363 = const()[name = string("op_4363"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4364_cast_fp16 = reshape(shape = var_4363, x = var_4356_cast_fp16_1)[name = string("op_4364_cast_fp16")];
            tensor<int32, [4]> var_4366 = const()[name = string("op_4366"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4367_cast_fp16 = reshape(shape = var_4366, x = var_4356_cast_fp16_2)[name = string("op_4367_cast_fp16")];
            tensor<int32, [4]> value_81_perm_0 = const()[name = string("value_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1211_perm_0 = const()[name = string("input_1211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1213_pad_0 = const()[name = string("input_1213_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1213_mode_0 = const()[name = string("input_1213_mode_0"), val = string("constant")];
            fp16 const_467_to_fp16 = const()[name = string("const_467_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1211_cast_fp16 = transpose(perm = input_1211_perm_0, x = var_4356_cast_fp16_2)[name = string("transpose_528")];
            tensor<fp16, [1, 512, 314]> input_1213_cast_fp16 = pad(constant_val = const_467_to_fp16, mode = input_1213_mode_0, pad = input_1213_pad_0, x = input_1211_cast_fp16)[name = string("input_1213_cast_fp16")];
            string x_405_pad_type_0 = const()[name = string("x_405_pad_type_0"), val = string("valid")];
            int32 x_405_groups_0 = const()[name = string("x_405_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_405_strides_0 = const()[name = string("x_405_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_405_pad_0 = const()[name = string("x_405_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_405_dilations_0 = const()[name = string("x_405_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_39_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_39_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254745856)))];
            tensor<fp16, [1, 512, 304]> x_405_cast_fp16 = conv(dilations = x_405_dilations_0, groups = x_405_groups_0, pad = x_405_pad_0, pad_type = x_405_pad_type_0, strides = x_405_strides_0, weight = m_encoder_encoders_39_self_attn_fsmn_block_weight_to_fp16, x = input_1213_cast_fp16)[name = string("x_405_cast_fp16")];
            tensor<int32, [3]> x_407_perm_0 = const()[name = string("x_407_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_407_cast_fp16 = transpose(perm = x_407_perm_0, x = x_405_cast_fp16)[name = string("transpose_527")];
            tensor<fp16, [1, 304, 512]> input_1215_cast_fp16 = add(x = x_407_cast_fp16, y = var_4356_cast_fp16_2)[name = string("input_1215_cast_fp16")];
            fp16 var_4386_to_fp16 = const()[name = string("op_4386_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_163_cast_fp16 = mul(x = var_4361_cast_fp16, y = var_4386_to_fp16)[name = string("q_h_163_cast_fp16")];
            bool scores_161_transpose_x_0 = const()[name = string("scores_161_transpose_x_0"), val = bool(false)];
            bool scores_161_transpose_y_0 = const()[name = string("scores_161_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_290_perm_0 = const()[name = string("transpose_290_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_291_perm_0 = const()[name = string("transpose_291_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_291 = transpose(perm = transpose_291_perm_0, x = var_4364_cast_fp16)[name = string("transpose_525")];
            tensor<fp16, [1, 4, 304, 128]> transpose_290 = transpose(perm = transpose_290_perm_0, x = q_h_163_cast_fp16)[name = string("transpose_526")];
            tensor<fp16, [1, 4, 304, 304]> scores_161_cast_fp16 = matmul(transpose_x = scores_161_transpose_x_0, transpose_y = scores_161_transpose_y_0, x = transpose_290, y = transpose_291)[name = string("scores_161_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_4394_cast_fp16 = softmax(axis = var_29, x = scores_161_cast_fp16)[name = string("op_4394_cast_fp16")];
            bool x_411_transpose_x_0 = const()[name = string("x_411_transpose_x_0"), val = bool(false)];
            bool x_411_transpose_y_0 = const()[name = string("x_411_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_81_cast_fp16 = transpose(perm = value_81_perm_0, x = var_4367_cast_fp16)[name = string("transpose_529")];
            tensor<fp16, [1, 4, 304, 128]> x_411_cast_fp16 = matmul(transpose_x = x_411_transpose_x_0, transpose_y = x_411_transpose_y_0, x = var_4394_cast_fp16, y = value_81_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<int32, [4]> var_4398_perm_0 = const()[name = string("op_4398_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4400 = const()[name = string("op_4400"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_4398_cast_fp16 = transpose(perm = var_4398_perm_0, x = x_411_cast_fp16)[name = string("transpose_524")];
            tensor<fp16, [1, 304, 512]> input_1219_cast_fp16 = reshape(shape = var_4400, x = var_4398_cast_fp16)[name = string("input_1219_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_39_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_39_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254757184)))];
            tensor<fp16, [512]> m_encoder_encoders_39_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_39_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255281536)))];
            tensor<fp16, [1, 304, 512]> linear_161_cast_fp16 = linear(bias = m_encoder_encoders_39_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_39_self_attn_linear_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1221_cast_fp16 = add(x = linear_161_cast_fp16, y = input_1215_cast_fp16)[name = string("input_1221_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1223_cast_fp16 = add(x = input_1207_cast_fp16, y = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")];
            tensor<int32, [1]> output_163_axes_0 = const()[name = string("output_163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_469_to_fp16 = const()[name = string("const_469_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255282624)))];
            tensor<fp16, [512]> const_470_to_fp16 = const()[name = string("const_470_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255283712)))];
            tensor<fp16, [1, 304, 512]> output_163_cast_fp16 = layer_norm(axes = output_163_axes_0, beta = const_470_to_fp16, epsilon = var_22_to_fp16, gamma = const_469_to_fp16, x = input_1223_cast_fp16)[name = string("output_163_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_39_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_39_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255284800)))];
            tensor<fp16, [2048]> m_encoder_encoders_39_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_39_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257382016)))];
            tensor<fp16, [1, 304, 2048]> linear_162_cast_fp16 = linear(bias = m_encoder_encoders_39_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_39_feed_forward_w_1_weight_to_fp16, x = output_163_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1231_cast_fp16 = relu(x = linear_162_cast_fp16)[name = string("input_1231_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_39_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_39_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257386176)))];
            tensor<fp16, [512]> m_encoder_encoders_39_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_39_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259483392)))];
            tensor<fp16, [1, 304, 512]> linear_163_cast_fp16 = linear(bias = m_encoder_encoders_39_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_39_feed_forward_w_2_weight_to_fp16, x = input_1231_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1237_cast_fp16 = add(x = input_1223_cast_fp16, y = linear_163_cast_fp16)[name = string("input_1237_cast_fp16")];
            tensor<int32, [1]> output_165_axes_0 = const()[name = string("output_165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_471_to_fp16 = const()[name = string("const_471_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259484480)))];
            tensor<fp16, [512]> const_472_to_fp16 = const()[name = string("const_472_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259485568)))];
            tensor<fp16, [1, 304, 512]> output_165_cast_fp16 = layer_norm(axes = output_165_axes_0, beta = const_472_to_fp16, epsilon = var_22_to_fp16, gamma = const_471_to_fp16, x = input_1237_cast_fp16)[name = string("output_165_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_40_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_40_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259486656)))];
            tensor<fp16, [1536]> m_encoder_encoders_40_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_40_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261059584)))];
            tensor<fp16, [1, 304, 1536]> linear_164_cast_fp16 = linear(bias = m_encoder_encoders_40_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_40_self_attn_linear_q_k_v_weight_to_fp16, x = output_165_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<int32, [3]> tile_41 = const()[name = string("tile_41"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4458_axis_0 = const()[name = string("op_4458_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4458_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4458_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4458_cast_fp16_2 = split(axis = var_4458_axis_0, split_sizes = tile_41, x = linear_164_cast_fp16)[name = string("op_4458_cast_fp16")];
            tensor<int32, [4]> var_4462 = const()[name = string("op_4462"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4463_cast_fp16 = reshape(shape = var_4462, x = var_4458_cast_fp16_0)[name = string("op_4463_cast_fp16")];
            tensor<int32, [4]> var_4465 = const()[name = string("op_4465"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4466_cast_fp16 = reshape(shape = var_4465, x = var_4458_cast_fp16_1)[name = string("op_4466_cast_fp16")];
            tensor<int32, [4]> var_4468 = const()[name = string("op_4468"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4469_cast_fp16 = reshape(shape = var_4468, x = var_4458_cast_fp16_2)[name = string("op_4469_cast_fp16")];
            tensor<int32, [4]> value_83_perm_0 = const()[name = string("value_83_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1241_perm_0 = const()[name = string("input_1241_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1243_pad_0 = const()[name = string("input_1243_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1243_mode_0 = const()[name = string("input_1243_mode_0"), val = string("constant")];
            fp16 const_478_to_fp16 = const()[name = string("const_478_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1241_cast_fp16 = transpose(perm = input_1241_perm_0, x = var_4458_cast_fp16_2)[name = string("transpose_522")];
            tensor<fp16, [1, 512, 314]> input_1243_cast_fp16 = pad(constant_val = const_478_to_fp16, mode = input_1243_mode_0, pad = input_1243_pad_0, x = input_1241_cast_fp16)[name = string("input_1243_cast_fp16")];
            string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")];
            int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_40_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_40_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261062720)))];
            tensor<fp16, [1, 512, 304]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = m_encoder_encoders_40_self_attn_fsmn_block_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_415_cast_fp16")];
            tensor<int32, [3]> x_417_perm_0 = const()[name = string("x_417_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_417_cast_fp16 = transpose(perm = x_417_perm_0, x = x_415_cast_fp16)[name = string("transpose_521")];
            tensor<fp16, [1, 304, 512]> input_1245_cast_fp16 = add(x = x_417_cast_fp16, y = var_4458_cast_fp16_2)[name = string("input_1245_cast_fp16")];
            fp16 var_4488_to_fp16 = const()[name = string("op_4488_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_167_cast_fp16 = mul(x = var_4463_cast_fp16, y = var_4488_to_fp16)[name = string("q_h_167_cast_fp16")];
            bool scores_165_transpose_x_0 = const()[name = string("scores_165_transpose_x_0"), val = bool(false)];
            bool scores_165_transpose_y_0 = const()[name = string("scores_165_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_292_perm_0 = const()[name = string("transpose_292_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_293_perm_0 = const()[name = string("transpose_293_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_293 = transpose(perm = transpose_293_perm_0, x = var_4466_cast_fp16)[name = string("transpose_519")];
            tensor<fp16, [1, 4, 304, 128]> transpose_292 = transpose(perm = transpose_292_perm_0, x = q_h_167_cast_fp16)[name = string("transpose_520")];
            tensor<fp16, [1, 4, 304, 304]> scores_165_cast_fp16 = matmul(transpose_x = scores_165_transpose_x_0, transpose_y = scores_165_transpose_y_0, x = transpose_292, y = transpose_293)[name = string("scores_165_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_4496_cast_fp16 = softmax(axis = var_29, x = scores_165_cast_fp16)[name = string("op_4496_cast_fp16")];
            bool x_421_transpose_x_0 = const()[name = string("x_421_transpose_x_0"), val = bool(false)];
            bool x_421_transpose_y_0 = const()[name = string("x_421_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_83_cast_fp16 = transpose(perm = value_83_perm_0, x = var_4469_cast_fp16)[name = string("transpose_523")];
            tensor<fp16, [1, 4, 304, 128]> x_421_cast_fp16 = matmul(transpose_x = x_421_transpose_x_0, transpose_y = x_421_transpose_y_0, x = var_4496_cast_fp16, y = value_83_cast_fp16)[name = string("x_421_cast_fp16")];
            tensor<int32, [4]> var_4500_perm_0 = const()[name = string("op_4500_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4502 = const()[name = string("op_4502"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_4500_cast_fp16 = transpose(perm = var_4500_perm_0, x = x_421_cast_fp16)[name = string("transpose_518")];
            tensor<fp16, [1, 304, 512]> input_1249_cast_fp16 = reshape(shape = var_4502, x = var_4500_cast_fp16)[name = string("input_1249_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_40_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_40_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261074048)))];
            tensor<fp16, [512]> m_encoder_encoders_40_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_40_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261598400)))];
            tensor<fp16, [1, 304, 512]> linear_165_cast_fp16 = linear(bias = m_encoder_encoders_40_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_40_self_attn_linear_out_weight_to_fp16, x = input_1249_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1251_cast_fp16 = add(x = linear_165_cast_fp16, y = input_1245_cast_fp16)[name = string("input_1251_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1253_cast_fp16 = add(x = input_1237_cast_fp16, y = input_1251_cast_fp16)[name = string("input_1253_cast_fp16")];
            tensor<int32, [1]> output_167_axes_0 = const()[name = string("output_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_480_to_fp16 = const()[name = string("const_480_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261599488)))];
            tensor<fp16, [512]> const_481_to_fp16 = const()[name = string("const_481_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261600576)))];
            tensor<fp16, [1, 304, 512]> output_167_cast_fp16 = layer_norm(axes = output_167_axes_0, beta = const_481_to_fp16, epsilon = var_22_to_fp16, gamma = const_480_to_fp16, x = input_1253_cast_fp16)[name = string("output_167_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_40_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_40_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261601664)))];
            tensor<fp16, [2048]> m_encoder_encoders_40_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_40_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263698880)))];
            tensor<fp16, [1, 304, 2048]> linear_166_cast_fp16 = linear(bias = m_encoder_encoders_40_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_40_feed_forward_w_1_weight_to_fp16, x = output_167_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1261_cast_fp16 = relu(x = linear_166_cast_fp16)[name = string("input_1261_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_40_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_40_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263703040)))];
            tensor<fp16, [512]> m_encoder_encoders_40_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_40_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265800256)))];
            tensor<fp16, [1, 304, 512]> linear_167_cast_fp16 = linear(bias = m_encoder_encoders_40_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_40_feed_forward_w_2_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1267_cast_fp16 = add(x = input_1253_cast_fp16, y = linear_167_cast_fp16)[name = string("input_1267_cast_fp16")];
            tensor<int32, [1]> output_169_axes_0 = const()[name = string("output_169_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_482_to_fp16 = const()[name = string("const_482_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265801344)))];
            tensor<fp16, [512]> const_483_to_fp16 = const()[name = string("const_483_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265802432)))];
            tensor<fp16, [1, 304, 512]> output_169_cast_fp16 = layer_norm(axes = output_169_axes_0, beta = const_483_to_fp16, epsilon = var_22_to_fp16, gamma = const_482_to_fp16, x = input_1267_cast_fp16)[name = string("output_169_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_41_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_41_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265803520)))];
            tensor<fp16, [1536]> m_encoder_encoders_41_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_41_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267376448)))];
            tensor<fp16, [1, 304, 1536]> linear_168_cast_fp16 = linear(bias = m_encoder_encoders_41_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_41_self_attn_linear_q_k_v_weight_to_fp16, x = output_169_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<int32, [3]> tile_42 = const()[name = string("tile_42"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4560_axis_0 = const()[name = string("op_4560_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4560_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4560_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4560_cast_fp16_2 = split(axis = var_4560_axis_0, split_sizes = tile_42, x = linear_168_cast_fp16)[name = string("op_4560_cast_fp16")];
            tensor<int32, [4]> var_4564 = const()[name = string("op_4564"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4565_cast_fp16 = reshape(shape = var_4564, x = var_4560_cast_fp16_0)[name = string("op_4565_cast_fp16")];
            tensor<int32, [4]> var_4567 = const()[name = string("op_4567"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4568_cast_fp16 = reshape(shape = var_4567, x = var_4560_cast_fp16_1)[name = string("op_4568_cast_fp16")];
            tensor<int32, [4]> var_4570 = const()[name = string("op_4570"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4571_cast_fp16 = reshape(shape = var_4570, x = var_4560_cast_fp16_2)[name = string("op_4571_cast_fp16")];
            tensor<int32, [4]> value_85_perm_0 = const()[name = string("value_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1271_perm_0 = const()[name = string("input_1271_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1273_pad_0 = const()[name = string("input_1273_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1273_mode_0 = const()[name = string("input_1273_mode_0"), val = string("constant")];
            fp16 const_489_to_fp16 = const()[name = string("const_489_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1271_cast_fp16 = transpose(perm = input_1271_perm_0, x = var_4560_cast_fp16_2)[name = string("transpose_516")];
            tensor<fp16, [1, 512, 314]> input_1273_cast_fp16 = pad(constant_val = const_489_to_fp16, mode = input_1273_mode_0, pad = input_1273_pad_0, x = input_1271_cast_fp16)[name = string("input_1273_cast_fp16")];
            string x_425_pad_type_0 = const()[name = string("x_425_pad_type_0"), val = string("valid")];
            int32 x_425_groups_0 = const()[name = string("x_425_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_425_strides_0 = const()[name = string("x_425_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_425_dilations_0 = const()[name = string("x_425_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_41_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_41_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267379584)))];
            tensor<fp16, [1, 512, 304]> x_425_cast_fp16 = conv(dilations = x_425_dilations_0, groups = x_425_groups_0, pad = x_425_pad_0, pad_type = x_425_pad_type_0, strides = x_425_strides_0, weight = m_encoder_encoders_41_self_attn_fsmn_block_weight_to_fp16, x = input_1273_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<int32, [3]> x_427_perm_0 = const()[name = string("x_427_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_427_cast_fp16 = transpose(perm = x_427_perm_0, x = x_425_cast_fp16)[name = string("transpose_515")];
            tensor<fp16, [1, 304, 512]> input_1275_cast_fp16 = add(x = x_427_cast_fp16, y = var_4560_cast_fp16_2)[name = string("input_1275_cast_fp16")];
            fp16 var_4590_to_fp16 = const()[name = string("op_4590_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_171_cast_fp16 = mul(x = var_4565_cast_fp16, y = var_4590_to_fp16)[name = string("q_h_171_cast_fp16")];
            bool scores_169_transpose_x_0 = const()[name = string("scores_169_transpose_x_0"), val = bool(false)];
            bool scores_169_transpose_y_0 = const()[name = string("scores_169_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_294_perm_0 = const()[name = string("transpose_294_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_295_perm_0 = const()[name = string("transpose_295_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_295 = transpose(perm = transpose_295_perm_0, x = var_4568_cast_fp16)[name = string("transpose_513")];
            tensor<fp16, [1, 4, 304, 128]> transpose_294 = transpose(perm = transpose_294_perm_0, x = q_h_171_cast_fp16)[name = string("transpose_514")];
            tensor<fp16, [1, 4, 304, 304]> scores_169_cast_fp16 = matmul(transpose_x = scores_169_transpose_x_0, transpose_y = scores_169_transpose_y_0, x = transpose_294, y = transpose_295)[name = string("scores_169_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_4598_cast_fp16 = softmax(axis = var_29, x = scores_169_cast_fp16)[name = string("op_4598_cast_fp16")];
            bool x_431_transpose_x_0 = const()[name = string("x_431_transpose_x_0"), val = bool(false)];
            bool x_431_transpose_y_0 = const()[name = string("x_431_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_85_cast_fp16 = transpose(perm = value_85_perm_0, x = var_4571_cast_fp16)[name = string("transpose_517")];
            tensor<fp16, [1, 4, 304, 128]> x_431_cast_fp16 = matmul(transpose_x = x_431_transpose_x_0, transpose_y = x_431_transpose_y_0, x = var_4598_cast_fp16, y = value_85_cast_fp16)[name = string("x_431_cast_fp16")];
            tensor<int32, [4]> var_4602_perm_0 = const()[name = string("op_4602_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4604 = const()[name = string("op_4604"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_4602_cast_fp16 = transpose(perm = var_4602_perm_0, x = x_431_cast_fp16)[name = string("transpose_512")];
            tensor<fp16, [1, 304, 512]> input_1279_cast_fp16 = reshape(shape = var_4604, x = var_4602_cast_fp16)[name = string("input_1279_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_41_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_41_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267390912)))];
            tensor<fp16, [512]> m_encoder_encoders_41_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_41_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267915264)))];
            tensor<fp16, [1, 304, 512]> linear_169_cast_fp16 = linear(bias = m_encoder_encoders_41_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_41_self_attn_linear_out_weight_to_fp16, x = input_1279_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1281_cast_fp16 = add(x = linear_169_cast_fp16, y = input_1275_cast_fp16)[name = string("input_1281_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1283_cast_fp16 = add(x = input_1267_cast_fp16, y = input_1281_cast_fp16)[name = string("input_1283_cast_fp16")];
            tensor<int32, [1]> output_171_axes_0 = const()[name = string("output_171_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_491_to_fp16 = const()[name = string("const_491_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267916352)))];
            tensor<fp16, [512]> const_492_to_fp16 = const()[name = string("const_492_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267917440)))];
            tensor<fp16, [1, 304, 512]> output_171_cast_fp16 = layer_norm(axes = output_171_axes_0, beta = const_492_to_fp16, epsilon = var_22_to_fp16, gamma = const_491_to_fp16, x = input_1283_cast_fp16)[name = string("output_171_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_41_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_41_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267918528)))];
            tensor<fp16, [2048]> m_encoder_encoders_41_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_41_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270015744)))];
            tensor<fp16, [1, 304, 2048]> linear_170_cast_fp16 = linear(bias = m_encoder_encoders_41_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_41_feed_forward_w_1_weight_to_fp16, x = output_171_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1291_cast_fp16 = relu(x = linear_170_cast_fp16)[name = string("input_1291_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_41_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_41_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270019904)))];
            tensor<fp16, [512]> m_encoder_encoders_41_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_41_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272117120)))];
            tensor<fp16, [1, 304, 512]> linear_171_cast_fp16 = linear(bias = m_encoder_encoders_41_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_41_feed_forward_w_2_weight_to_fp16, x = input_1291_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1297_cast_fp16 = add(x = input_1283_cast_fp16, y = linear_171_cast_fp16)[name = string("input_1297_cast_fp16")];
            tensor<int32, [1]> output_173_axes_0 = const()[name = string("output_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_493_to_fp16 = const()[name = string("const_493_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272118208)))];
            tensor<fp16, [512]> const_494_to_fp16 = const()[name = string("const_494_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272119296)))];
            tensor<fp16, [1, 304, 512]> output_173_cast_fp16 = layer_norm(axes = output_173_axes_0, beta = const_494_to_fp16, epsilon = var_22_to_fp16, gamma = const_493_to_fp16, x = input_1297_cast_fp16)[name = string("output_173_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_42_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_42_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272120384)))];
            tensor<fp16, [1536]> m_encoder_encoders_42_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_42_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273693312)))];
            tensor<fp16, [1, 304, 1536]> linear_172_cast_fp16 = linear(bias = m_encoder_encoders_42_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_42_self_attn_linear_q_k_v_weight_to_fp16, x = output_173_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [3]> tile_43 = const()[name = string("tile_43"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4662_axis_0 = const()[name = string("op_4662_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4662_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4662_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4662_cast_fp16_2 = split(axis = var_4662_axis_0, split_sizes = tile_43, x = linear_172_cast_fp16)[name = string("op_4662_cast_fp16")];
            tensor<int32, [4]> var_4666 = const()[name = string("op_4666"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4667_cast_fp16 = reshape(shape = var_4666, x = var_4662_cast_fp16_0)[name = string("op_4667_cast_fp16")];
            tensor<int32, [4]> var_4669 = const()[name = string("op_4669"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4670_cast_fp16 = reshape(shape = var_4669, x = var_4662_cast_fp16_1)[name = string("op_4670_cast_fp16")];
            tensor<int32, [4]> var_4672 = const()[name = string("op_4672"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4673_cast_fp16 = reshape(shape = var_4672, x = var_4662_cast_fp16_2)[name = string("op_4673_cast_fp16")];
            tensor<int32, [4]> value_87_perm_0 = const()[name = string("value_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1301_perm_0 = const()[name = string("input_1301_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1303_pad_0 = const()[name = string("input_1303_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1303_mode_0 = const()[name = string("input_1303_mode_0"), val = string("constant")];
            fp16 const_500_to_fp16 = const()[name = string("const_500_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1301_cast_fp16 = transpose(perm = input_1301_perm_0, x = var_4662_cast_fp16_2)[name = string("transpose_510")];
            tensor<fp16, [1, 512, 314]> input_1303_cast_fp16 = pad(constant_val = const_500_to_fp16, mode = input_1303_mode_0, pad = input_1303_pad_0, x = input_1301_cast_fp16)[name = string("input_1303_cast_fp16")];
            string x_435_pad_type_0 = const()[name = string("x_435_pad_type_0"), val = string("valid")];
            int32 x_435_groups_0 = const()[name = string("x_435_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_435_strides_0 = const()[name = string("x_435_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_435_pad_0 = const()[name = string("x_435_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_435_dilations_0 = const()[name = string("x_435_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_42_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_42_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273696448)))];
            tensor<fp16, [1, 512, 304]> x_435_cast_fp16 = conv(dilations = x_435_dilations_0, groups = x_435_groups_0, pad = x_435_pad_0, pad_type = x_435_pad_type_0, strides = x_435_strides_0, weight = m_encoder_encoders_42_self_attn_fsmn_block_weight_to_fp16, x = input_1303_cast_fp16)[name = string("x_435_cast_fp16")];
            tensor<int32, [3]> x_437_perm_0 = const()[name = string("x_437_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_437_cast_fp16 = transpose(perm = x_437_perm_0, x = x_435_cast_fp16)[name = string("transpose_509")];
            tensor<fp16, [1, 304, 512]> input_1305_cast_fp16 = add(x = x_437_cast_fp16, y = var_4662_cast_fp16_2)[name = string("input_1305_cast_fp16")];
            fp16 var_4692_to_fp16 = const()[name = string("op_4692_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_175_cast_fp16 = mul(x = var_4667_cast_fp16, y = var_4692_to_fp16)[name = string("q_h_175_cast_fp16")];
            bool scores_173_transpose_x_0 = const()[name = string("scores_173_transpose_x_0"), val = bool(false)];
            bool scores_173_transpose_y_0 = const()[name = string("scores_173_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_296_perm_0 = const()[name = string("transpose_296_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_297_perm_0 = const()[name = string("transpose_297_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_297 = transpose(perm = transpose_297_perm_0, x = var_4670_cast_fp16)[name = string("transpose_507")];
            tensor<fp16, [1, 4, 304, 128]> transpose_296 = transpose(perm = transpose_296_perm_0, x = q_h_175_cast_fp16)[name = string("transpose_508")];
            tensor<fp16, [1, 4, 304, 304]> scores_173_cast_fp16 = matmul(transpose_x = scores_173_transpose_x_0, transpose_y = scores_173_transpose_y_0, x = transpose_296, y = transpose_297)[name = string("scores_173_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_4700_cast_fp16 = softmax(axis = var_29, x = scores_173_cast_fp16)[name = string("op_4700_cast_fp16")];
            bool x_441_transpose_x_0 = const()[name = string("x_441_transpose_x_0"), val = bool(false)];
            bool x_441_transpose_y_0 = const()[name = string("x_441_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_87_cast_fp16 = transpose(perm = value_87_perm_0, x = var_4673_cast_fp16)[name = string("transpose_511")];
            tensor<fp16, [1, 4, 304, 128]> x_441_cast_fp16 = matmul(transpose_x = x_441_transpose_x_0, transpose_y = x_441_transpose_y_0, x = var_4700_cast_fp16, y = value_87_cast_fp16)[name = string("x_441_cast_fp16")];
            tensor<int32, [4]> var_4704_perm_0 = const()[name = string("op_4704_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4706 = const()[name = string("op_4706"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_4704_cast_fp16 = transpose(perm = var_4704_perm_0, x = x_441_cast_fp16)[name = string("transpose_506")];
            tensor<fp16, [1, 304, 512]> input_1309_cast_fp16 = reshape(shape = var_4706, x = var_4704_cast_fp16)[name = string("input_1309_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_42_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_42_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273707776)))];
            tensor<fp16, [512]> m_encoder_encoders_42_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_42_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274232128)))];
            tensor<fp16, [1, 304, 512]> linear_173_cast_fp16 = linear(bias = m_encoder_encoders_42_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_42_self_attn_linear_out_weight_to_fp16, x = input_1309_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1311_cast_fp16 = add(x = linear_173_cast_fp16, y = input_1305_cast_fp16)[name = string("input_1311_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1313_cast_fp16 = add(x = input_1297_cast_fp16, y = input_1311_cast_fp16)[name = string("input_1313_cast_fp16")];
            tensor<int32, [1]> output_175_axes_0 = const()[name = string("output_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_502_to_fp16 = const()[name = string("const_502_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274233216)))];
            tensor<fp16, [512]> const_503_to_fp16 = const()[name = string("const_503_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274234304)))];
            tensor<fp16, [1, 304, 512]> output_175_cast_fp16 = layer_norm(axes = output_175_axes_0, beta = const_503_to_fp16, epsilon = var_22_to_fp16, gamma = const_502_to_fp16, x = input_1313_cast_fp16)[name = string("output_175_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_42_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_42_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274235392)))];
            tensor<fp16, [2048]> m_encoder_encoders_42_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_42_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276332608)))];
            tensor<fp16, [1, 304, 2048]> linear_174_cast_fp16 = linear(bias = m_encoder_encoders_42_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_42_feed_forward_w_1_weight_to_fp16, x = output_175_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1321_cast_fp16 = relu(x = linear_174_cast_fp16)[name = string("input_1321_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_42_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_42_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276336768)))];
            tensor<fp16, [512]> m_encoder_encoders_42_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_42_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278433984)))];
            tensor<fp16, [1, 304, 512]> linear_175_cast_fp16 = linear(bias = m_encoder_encoders_42_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_42_feed_forward_w_2_weight_to_fp16, x = input_1321_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1327_cast_fp16 = add(x = input_1313_cast_fp16, y = linear_175_cast_fp16)[name = string("input_1327_cast_fp16")];
            tensor<int32, [1]> output_177_axes_0 = const()[name = string("output_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_504_to_fp16 = const()[name = string("const_504_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278435072)))];
            tensor<fp16, [512]> const_505_to_fp16 = const()[name = string("const_505_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278436160)))];
            tensor<fp16, [1, 304, 512]> output_177_cast_fp16 = layer_norm(axes = output_177_axes_0, beta = const_505_to_fp16, epsilon = var_22_to_fp16, gamma = const_504_to_fp16, x = input_1327_cast_fp16)[name = string("output_177_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_43_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_43_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278437248)))];
            tensor<fp16, [1536]> m_encoder_encoders_43_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_43_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280010176)))];
            tensor<fp16, [1, 304, 1536]> linear_176_cast_fp16 = linear(bias = m_encoder_encoders_43_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_43_self_attn_linear_q_k_v_weight_to_fp16, x = output_177_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<int32, [3]> tile_44 = const()[name = string("tile_44"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4764_axis_0 = const()[name = string("op_4764_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4764_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4764_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4764_cast_fp16_2 = split(axis = var_4764_axis_0, split_sizes = tile_44, x = linear_176_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<int32, [4]> var_4768 = const()[name = string("op_4768"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4769_cast_fp16 = reshape(shape = var_4768, x = var_4764_cast_fp16_0)[name = string("op_4769_cast_fp16")];
            tensor<int32, [4]> var_4771 = const()[name = string("op_4771"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4772_cast_fp16 = reshape(shape = var_4771, x = var_4764_cast_fp16_1)[name = string("op_4772_cast_fp16")];
            tensor<int32, [4]> var_4774 = const()[name = string("op_4774"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4775_cast_fp16 = reshape(shape = var_4774, x = var_4764_cast_fp16_2)[name = string("op_4775_cast_fp16")];
            tensor<int32, [4]> value_89_perm_0 = const()[name = string("value_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1331_perm_0 = const()[name = string("input_1331_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1333_pad_0 = const()[name = string("input_1333_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1333_mode_0 = const()[name = string("input_1333_mode_0"), val = string("constant")];
            fp16 const_511_to_fp16 = const()[name = string("const_511_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1331_cast_fp16 = transpose(perm = input_1331_perm_0, x = var_4764_cast_fp16_2)[name = string("transpose_504")];
            tensor<fp16, [1, 512, 314]> input_1333_cast_fp16 = pad(constant_val = const_511_to_fp16, mode = input_1333_mode_0, pad = input_1333_pad_0, x = input_1331_cast_fp16)[name = string("input_1333_cast_fp16")];
            string x_445_pad_type_0 = const()[name = string("x_445_pad_type_0"), val = string("valid")];
            int32 x_445_groups_0 = const()[name = string("x_445_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_445_strides_0 = const()[name = string("x_445_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_445_pad_0 = const()[name = string("x_445_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_445_dilations_0 = const()[name = string("x_445_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_43_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_43_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280013312)))];
            tensor<fp16, [1, 512, 304]> x_445_cast_fp16 = conv(dilations = x_445_dilations_0, groups = x_445_groups_0, pad = x_445_pad_0, pad_type = x_445_pad_type_0, strides = x_445_strides_0, weight = m_encoder_encoders_43_self_attn_fsmn_block_weight_to_fp16, x = input_1333_cast_fp16)[name = string("x_445_cast_fp16")];
            tensor<int32, [3]> x_447_perm_0 = const()[name = string("x_447_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_447_cast_fp16 = transpose(perm = x_447_perm_0, x = x_445_cast_fp16)[name = string("transpose_503")];
            tensor<fp16, [1, 304, 512]> input_1335_cast_fp16 = add(x = x_447_cast_fp16, y = var_4764_cast_fp16_2)[name = string("input_1335_cast_fp16")];
            fp16 var_4794_to_fp16 = const()[name = string("op_4794_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_179_cast_fp16 = mul(x = var_4769_cast_fp16, y = var_4794_to_fp16)[name = string("q_h_179_cast_fp16")];
            bool scores_177_transpose_x_0 = const()[name = string("scores_177_transpose_x_0"), val = bool(false)];
            bool scores_177_transpose_y_0 = const()[name = string("scores_177_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_298_perm_0 = const()[name = string("transpose_298_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_299_perm_0 = const()[name = string("transpose_299_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_299 = transpose(perm = transpose_299_perm_0, x = var_4772_cast_fp16)[name = string("transpose_501")];
            tensor<fp16, [1, 4, 304, 128]> transpose_298 = transpose(perm = transpose_298_perm_0, x = q_h_179_cast_fp16)[name = string("transpose_502")];
            tensor<fp16, [1, 4, 304, 304]> scores_177_cast_fp16 = matmul(transpose_x = scores_177_transpose_x_0, transpose_y = scores_177_transpose_y_0, x = transpose_298, y = transpose_299)[name = string("scores_177_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_4802_cast_fp16 = softmax(axis = var_29, x = scores_177_cast_fp16)[name = string("op_4802_cast_fp16")];
            bool x_451_transpose_x_0 = const()[name = string("x_451_transpose_x_0"), val = bool(false)];
            bool x_451_transpose_y_0 = const()[name = string("x_451_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_89_cast_fp16 = transpose(perm = value_89_perm_0, x = var_4775_cast_fp16)[name = string("transpose_505")];
            tensor<fp16, [1, 4, 304, 128]> x_451_cast_fp16 = matmul(transpose_x = x_451_transpose_x_0, transpose_y = x_451_transpose_y_0, x = var_4802_cast_fp16, y = value_89_cast_fp16)[name = string("x_451_cast_fp16")];
            tensor<int32, [4]> var_4806_perm_0 = const()[name = string("op_4806_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4808 = const()[name = string("op_4808"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_4806_cast_fp16 = transpose(perm = var_4806_perm_0, x = x_451_cast_fp16)[name = string("transpose_500")];
            tensor<fp16, [1, 304, 512]> input_1339_cast_fp16 = reshape(shape = var_4808, x = var_4806_cast_fp16)[name = string("input_1339_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_43_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_43_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280024640)))];
            tensor<fp16, [512]> m_encoder_encoders_43_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_43_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280548992)))];
            tensor<fp16, [1, 304, 512]> linear_177_cast_fp16 = linear(bias = m_encoder_encoders_43_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_43_self_attn_linear_out_weight_to_fp16, x = input_1339_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1341_cast_fp16 = add(x = linear_177_cast_fp16, y = input_1335_cast_fp16)[name = string("input_1341_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1343_cast_fp16 = add(x = input_1327_cast_fp16, y = input_1341_cast_fp16)[name = string("input_1343_cast_fp16")];
            tensor<int32, [1]> output_179_axes_0 = const()[name = string("output_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_513_to_fp16 = const()[name = string("const_513_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280550080)))];
            tensor<fp16, [512]> const_514_to_fp16 = const()[name = string("const_514_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280551168)))];
            tensor<fp16, [1, 304, 512]> output_179_cast_fp16 = layer_norm(axes = output_179_axes_0, beta = const_514_to_fp16, epsilon = var_22_to_fp16, gamma = const_513_to_fp16, x = input_1343_cast_fp16)[name = string("output_179_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_43_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_43_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280552256)))];
            tensor<fp16, [2048]> m_encoder_encoders_43_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_43_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282649472)))];
            tensor<fp16, [1, 304, 2048]> linear_178_cast_fp16 = linear(bias = m_encoder_encoders_43_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_43_feed_forward_w_1_weight_to_fp16, x = output_179_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1351_cast_fp16 = relu(x = linear_178_cast_fp16)[name = string("input_1351_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_43_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_43_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282653632)))];
            tensor<fp16, [512]> m_encoder_encoders_43_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_43_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284750848)))];
            tensor<fp16, [1, 304, 512]> linear_179_cast_fp16 = linear(bias = m_encoder_encoders_43_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_43_feed_forward_w_2_weight_to_fp16, x = input_1351_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1357_cast_fp16 = add(x = input_1343_cast_fp16, y = linear_179_cast_fp16)[name = string("input_1357_cast_fp16")];
            tensor<int32, [1]> output_181_axes_0 = const()[name = string("output_181_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_515_to_fp16 = const()[name = string("const_515_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284751936)))];
            tensor<fp16, [512]> const_516_to_fp16 = const()[name = string("const_516_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284753024)))];
            tensor<fp16, [1, 304, 512]> output_181_cast_fp16 = layer_norm(axes = output_181_axes_0, beta = const_516_to_fp16, epsilon = var_22_to_fp16, gamma = const_515_to_fp16, x = input_1357_cast_fp16)[name = string("output_181_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_44_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_44_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284754112)))];
            tensor<fp16, [1536]> m_encoder_encoders_44_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_44_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286327040)))];
            tensor<fp16, [1, 304, 1536]> linear_180_cast_fp16 = linear(bias = m_encoder_encoders_44_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_44_self_attn_linear_q_k_v_weight_to_fp16, x = output_181_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<int32, [3]> tile_45 = const()[name = string("tile_45"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4866_axis_0 = const()[name = string("op_4866_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4866_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4866_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4866_cast_fp16_2 = split(axis = var_4866_axis_0, split_sizes = tile_45, x = linear_180_cast_fp16)[name = string("op_4866_cast_fp16")];
            tensor<int32, [4]> var_4870 = const()[name = string("op_4870"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4871_cast_fp16 = reshape(shape = var_4870, x = var_4866_cast_fp16_0)[name = string("op_4871_cast_fp16")];
            tensor<int32, [4]> var_4873 = const()[name = string("op_4873"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4874_cast_fp16 = reshape(shape = var_4873, x = var_4866_cast_fp16_1)[name = string("op_4874_cast_fp16")];
            tensor<int32, [4]> var_4876 = const()[name = string("op_4876"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4877_cast_fp16 = reshape(shape = var_4876, x = var_4866_cast_fp16_2)[name = string("op_4877_cast_fp16")];
            tensor<int32, [4]> value_91_perm_0 = const()[name = string("value_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1361_perm_0 = const()[name = string("input_1361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1363_pad_0 = const()[name = string("input_1363_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1363_mode_0 = const()[name = string("input_1363_mode_0"), val = string("constant")];
            fp16 const_522_to_fp16 = const()[name = string("const_522_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1361_cast_fp16 = transpose(perm = input_1361_perm_0, x = var_4866_cast_fp16_2)[name = string("transpose_498")];
            tensor<fp16, [1, 512, 314]> input_1363_cast_fp16 = pad(constant_val = const_522_to_fp16, mode = input_1363_mode_0, pad = input_1363_pad_0, x = input_1361_cast_fp16)[name = string("input_1363_cast_fp16")];
            string x_455_pad_type_0 = const()[name = string("x_455_pad_type_0"), val = string("valid")];
            int32 x_455_groups_0 = const()[name = string("x_455_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_455_strides_0 = const()[name = string("x_455_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_455_pad_0 = const()[name = string("x_455_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_455_dilations_0 = const()[name = string("x_455_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_44_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_44_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286330176)))];
            tensor<fp16, [1, 512, 304]> x_455_cast_fp16 = conv(dilations = x_455_dilations_0, groups = x_455_groups_0, pad = x_455_pad_0, pad_type = x_455_pad_type_0, strides = x_455_strides_0, weight = m_encoder_encoders_44_self_attn_fsmn_block_weight_to_fp16, x = input_1363_cast_fp16)[name = string("x_455_cast_fp16")];
            tensor<int32, [3]> x_457_perm_0 = const()[name = string("x_457_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_457_cast_fp16 = transpose(perm = x_457_perm_0, x = x_455_cast_fp16)[name = string("transpose_497")];
            tensor<fp16, [1, 304, 512]> input_1365_cast_fp16 = add(x = x_457_cast_fp16, y = var_4866_cast_fp16_2)[name = string("input_1365_cast_fp16")];
            fp16 var_4896_to_fp16 = const()[name = string("op_4896_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_183_cast_fp16 = mul(x = var_4871_cast_fp16, y = var_4896_to_fp16)[name = string("q_h_183_cast_fp16")];
            bool scores_181_transpose_x_0 = const()[name = string("scores_181_transpose_x_0"), val = bool(false)];
            bool scores_181_transpose_y_0 = const()[name = string("scores_181_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_300_perm_0 = const()[name = string("transpose_300_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_301_perm_0 = const()[name = string("transpose_301_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_301 = transpose(perm = transpose_301_perm_0, x = var_4874_cast_fp16)[name = string("transpose_495")];
            tensor<fp16, [1, 4, 304, 128]> transpose_300 = transpose(perm = transpose_300_perm_0, x = q_h_183_cast_fp16)[name = string("transpose_496")];
            tensor<fp16, [1, 4, 304, 304]> scores_181_cast_fp16 = matmul(transpose_x = scores_181_transpose_x_0, transpose_y = scores_181_transpose_y_0, x = transpose_300, y = transpose_301)[name = string("scores_181_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_4904_cast_fp16 = softmax(axis = var_29, x = scores_181_cast_fp16)[name = string("op_4904_cast_fp16")];
            bool x_461_transpose_x_0 = const()[name = string("x_461_transpose_x_0"), val = bool(false)];
            bool x_461_transpose_y_0 = const()[name = string("x_461_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_91_cast_fp16 = transpose(perm = value_91_perm_0, x = var_4877_cast_fp16)[name = string("transpose_499")];
            tensor<fp16, [1, 4, 304, 128]> x_461_cast_fp16 = matmul(transpose_x = x_461_transpose_x_0, transpose_y = x_461_transpose_y_0, x = var_4904_cast_fp16, y = value_91_cast_fp16)[name = string("x_461_cast_fp16")];
            tensor<int32, [4]> var_4908_perm_0 = const()[name = string("op_4908_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4910 = const()[name = string("op_4910"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_4908_cast_fp16 = transpose(perm = var_4908_perm_0, x = x_461_cast_fp16)[name = string("transpose_494")];
            tensor<fp16, [1, 304, 512]> input_1369_cast_fp16 = reshape(shape = var_4910, x = var_4908_cast_fp16)[name = string("input_1369_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_44_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_44_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286341504)))];
            tensor<fp16, [512]> m_encoder_encoders_44_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_44_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286865856)))];
            tensor<fp16, [1, 304, 512]> linear_181_cast_fp16 = linear(bias = m_encoder_encoders_44_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_44_self_attn_linear_out_weight_to_fp16, x = input_1369_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1371_cast_fp16 = add(x = linear_181_cast_fp16, y = input_1365_cast_fp16)[name = string("input_1371_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1373_cast_fp16 = add(x = input_1357_cast_fp16, y = input_1371_cast_fp16)[name = string("input_1373_cast_fp16")];
            tensor<int32, [1]> output_183_axes_0 = const()[name = string("output_183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_524_to_fp16 = const()[name = string("const_524_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286866944)))];
            tensor<fp16, [512]> const_525_to_fp16 = const()[name = string("const_525_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286868032)))];
            tensor<fp16, [1, 304, 512]> output_183_cast_fp16 = layer_norm(axes = output_183_axes_0, beta = const_525_to_fp16, epsilon = var_22_to_fp16, gamma = const_524_to_fp16, x = input_1373_cast_fp16)[name = string("output_183_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_44_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_44_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286869120)))];
            tensor<fp16, [2048]> m_encoder_encoders_44_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_44_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288966336)))];
            tensor<fp16, [1, 304, 2048]> linear_182_cast_fp16 = linear(bias = m_encoder_encoders_44_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_44_feed_forward_w_1_weight_to_fp16, x = output_183_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1381_cast_fp16 = relu(x = linear_182_cast_fp16)[name = string("input_1381_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_44_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_44_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288970496)))];
            tensor<fp16, [512]> m_encoder_encoders_44_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_44_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291067712)))];
            tensor<fp16, [1, 304, 512]> linear_183_cast_fp16 = linear(bias = m_encoder_encoders_44_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_44_feed_forward_w_2_weight_to_fp16, x = input_1381_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1387_cast_fp16 = add(x = input_1373_cast_fp16, y = linear_183_cast_fp16)[name = string("input_1387_cast_fp16")];
            tensor<int32, [1]> output_185_axes_0 = const()[name = string("output_185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_526_to_fp16 = const()[name = string("const_526_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291068800)))];
            tensor<fp16, [512]> const_527_to_fp16 = const()[name = string("const_527_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291069888)))];
            tensor<fp16, [1, 304, 512]> output_185_cast_fp16 = layer_norm(axes = output_185_axes_0, beta = const_527_to_fp16, epsilon = var_22_to_fp16, gamma = const_526_to_fp16, x = input_1387_cast_fp16)[name = string("output_185_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_45_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_45_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291070976)))];
            tensor<fp16, [1536]> m_encoder_encoders_45_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_45_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292643904)))];
            tensor<fp16, [1, 304, 1536]> linear_184_cast_fp16 = linear(bias = m_encoder_encoders_45_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_45_self_attn_linear_q_k_v_weight_to_fp16, x = output_185_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<int32, [3]> tile_46 = const()[name = string("tile_46"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_4968_axis_0 = const()[name = string("op_4968_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_4968_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_4968_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_4968_cast_fp16_2 = split(axis = var_4968_axis_0, split_sizes = tile_46, x = linear_184_cast_fp16)[name = string("op_4968_cast_fp16")];
            tensor<int32, [4]> var_4972 = const()[name = string("op_4972"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4973_cast_fp16 = reshape(shape = var_4972, x = var_4968_cast_fp16_0)[name = string("op_4973_cast_fp16")];
            tensor<int32, [4]> var_4975 = const()[name = string("op_4975"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4976_cast_fp16 = reshape(shape = var_4975, x = var_4968_cast_fp16_1)[name = string("op_4976_cast_fp16")];
            tensor<int32, [4]> var_4978 = const()[name = string("op_4978"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_4979_cast_fp16 = reshape(shape = var_4978, x = var_4968_cast_fp16_2)[name = string("op_4979_cast_fp16")];
            tensor<int32, [4]> value_93_perm_0 = const()[name = string("value_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1391_perm_0 = const()[name = string("input_1391_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1393_pad_0 = const()[name = string("input_1393_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1393_mode_0 = const()[name = string("input_1393_mode_0"), val = string("constant")];
            fp16 const_533_to_fp16 = const()[name = string("const_533_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1391_cast_fp16 = transpose(perm = input_1391_perm_0, x = var_4968_cast_fp16_2)[name = string("transpose_492")];
            tensor<fp16, [1, 512, 314]> input_1393_cast_fp16 = pad(constant_val = const_533_to_fp16, mode = input_1393_mode_0, pad = input_1393_pad_0, x = input_1391_cast_fp16)[name = string("input_1393_cast_fp16")];
            string x_465_pad_type_0 = const()[name = string("x_465_pad_type_0"), val = string("valid")];
            int32 x_465_groups_0 = const()[name = string("x_465_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_465_strides_0 = const()[name = string("x_465_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_465_pad_0 = const()[name = string("x_465_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_465_dilations_0 = const()[name = string("x_465_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_45_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_45_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292647040)))];
            tensor<fp16, [1, 512, 304]> x_465_cast_fp16 = conv(dilations = x_465_dilations_0, groups = x_465_groups_0, pad = x_465_pad_0, pad_type = x_465_pad_type_0, strides = x_465_strides_0, weight = m_encoder_encoders_45_self_attn_fsmn_block_weight_to_fp16, x = input_1393_cast_fp16)[name = string("x_465_cast_fp16")];
            tensor<int32, [3]> x_467_perm_0 = const()[name = string("x_467_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_467_cast_fp16 = transpose(perm = x_467_perm_0, x = x_465_cast_fp16)[name = string("transpose_491")];
            tensor<fp16, [1, 304, 512]> input_1395_cast_fp16 = add(x = x_467_cast_fp16, y = var_4968_cast_fp16_2)[name = string("input_1395_cast_fp16")];
            fp16 var_4998_to_fp16 = const()[name = string("op_4998_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_187_cast_fp16 = mul(x = var_4973_cast_fp16, y = var_4998_to_fp16)[name = string("q_h_187_cast_fp16")];
            bool scores_185_transpose_x_0 = const()[name = string("scores_185_transpose_x_0"), val = bool(false)];
            bool scores_185_transpose_y_0 = const()[name = string("scores_185_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_302_perm_0 = const()[name = string("transpose_302_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_303_perm_0 = const()[name = string("transpose_303_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_303 = transpose(perm = transpose_303_perm_0, x = var_4976_cast_fp16)[name = string("transpose_489")];
            tensor<fp16, [1, 4, 304, 128]> transpose_302 = transpose(perm = transpose_302_perm_0, x = q_h_187_cast_fp16)[name = string("transpose_490")];
            tensor<fp16, [1, 4, 304, 304]> scores_185_cast_fp16 = matmul(transpose_x = scores_185_transpose_x_0, transpose_y = scores_185_transpose_y_0, x = transpose_302, y = transpose_303)[name = string("scores_185_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5006_cast_fp16 = softmax(axis = var_29, x = scores_185_cast_fp16)[name = string("op_5006_cast_fp16")];
            bool x_471_transpose_x_0 = const()[name = string("x_471_transpose_x_0"), val = bool(false)];
            bool x_471_transpose_y_0 = const()[name = string("x_471_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_93_cast_fp16 = transpose(perm = value_93_perm_0, x = var_4979_cast_fp16)[name = string("transpose_493")];
            tensor<fp16, [1, 4, 304, 128]> x_471_cast_fp16 = matmul(transpose_x = x_471_transpose_x_0, transpose_y = x_471_transpose_y_0, x = var_5006_cast_fp16, y = value_93_cast_fp16)[name = string("x_471_cast_fp16")];
            tensor<int32, [4]> var_5010_perm_0 = const()[name = string("op_5010_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5012 = const()[name = string("op_5012"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5010_cast_fp16 = transpose(perm = var_5010_perm_0, x = x_471_cast_fp16)[name = string("transpose_488")];
            tensor<fp16, [1, 304, 512]> input_1399_cast_fp16 = reshape(shape = var_5012, x = var_5010_cast_fp16)[name = string("input_1399_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_45_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_45_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292658368)))];
            tensor<fp16, [512]> m_encoder_encoders_45_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_45_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293182720)))];
            tensor<fp16, [1, 304, 512]> linear_185_cast_fp16 = linear(bias = m_encoder_encoders_45_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_45_self_attn_linear_out_weight_to_fp16, x = input_1399_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1401_cast_fp16 = add(x = linear_185_cast_fp16, y = input_1395_cast_fp16)[name = string("input_1401_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1403_cast_fp16 = add(x = input_1387_cast_fp16, y = input_1401_cast_fp16)[name = string("input_1403_cast_fp16")];
            tensor<int32, [1]> output_187_axes_0 = const()[name = string("output_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_535_to_fp16 = const()[name = string("const_535_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293183808)))];
            tensor<fp16, [512]> const_536_to_fp16 = const()[name = string("const_536_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293184896)))];
            tensor<fp16, [1, 304, 512]> output_187_cast_fp16 = layer_norm(axes = output_187_axes_0, beta = const_536_to_fp16, epsilon = var_22_to_fp16, gamma = const_535_to_fp16, x = input_1403_cast_fp16)[name = string("output_187_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_45_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_45_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293185984)))];
            tensor<fp16, [2048]> m_encoder_encoders_45_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_45_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295283200)))];
            tensor<fp16, [1, 304, 2048]> linear_186_cast_fp16 = linear(bias = m_encoder_encoders_45_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_45_feed_forward_w_1_weight_to_fp16, x = output_187_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1411_cast_fp16 = relu(x = linear_186_cast_fp16)[name = string("input_1411_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_45_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_45_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295287360)))];
            tensor<fp16, [512]> m_encoder_encoders_45_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_45_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297384576)))];
            tensor<fp16, [1, 304, 512]> linear_187_cast_fp16 = linear(bias = m_encoder_encoders_45_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_45_feed_forward_w_2_weight_to_fp16, x = input_1411_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1417_cast_fp16 = add(x = input_1403_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1417_cast_fp16")];
            tensor<int32, [1]> output_189_axes_0 = const()[name = string("output_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_537_to_fp16 = const()[name = string("const_537_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297385664)))];
            tensor<fp16, [512]> const_538_to_fp16 = const()[name = string("const_538_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297386752)))];
            tensor<fp16, [1, 304, 512]> output_189_cast_fp16 = layer_norm(axes = output_189_axes_0, beta = const_538_to_fp16, epsilon = var_22_to_fp16, gamma = const_537_to_fp16, x = input_1417_cast_fp16)[name = string("output_189_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_46_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_46_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297387840)))];
            tensor<fp16, [1536]> m_encoder_encoders_46_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_46_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298960768)))];
            tensor<fp16, [1, 304, 1536]> linear_188_cast_fp16 = linear(bias = m_encoder_encoders_46_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_46_self_attn_linear_q_k_v_weight_to_fp16, x = output_189_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<int32, [3]> tile_47 = const()[name = string("tile_47"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_5070_axis_0 = const()[name = string("op_5070_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_5070_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_5070_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_5070_cast_fp16_2 = split(axis = var_5070_axis_0, split_sizes = tile_47, x = linear_188_cast_fp16)[name = string("op_5070_cast_fp16")];
            tensor<int32, [4]> var_5074 = const()[name = string("op_5074"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5075_cast_fp16 = reshape(shape = var_5074, x = var_5070_cast_fp16_0)[name = string("op_5075_cast_fp16")];
            tensor<int32, [4]> var_5077 = const()[name = string("op_5077"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5078_cast_fp16 = reshape(shape = var_5077, x = var_5070_cast_fp16_1)[name = string("op_5078_cast_fp16")];
            tensor<int32, [4]> var_5080 = const()[name = string("op_5080"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5081_cast_fp16 = reshape(shape = var_5080, x = var_5070_cast_fp16_2)[name = string("op_5081_cast_fp16")];
            tensor<int32, [4]> value_95_perm_0 = const()[name = string("value_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1421_perm_0 = const()[name = string("input_1421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1423_pad_0 = const()[name = string("input_1423_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1423_mode_0 = const()[name = string("input_1423_mode_0"), val = string("constant")];
            fp16 const_544_to_fp16 = const()[name = string("const_544_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1421_cast_fp16 = transpose(perm = input_1421_perm_0, x = var_5070_cast_fp16_2)[name = string("transpose_486")];
            tensor<fp16, [1, 512, 314]> input_1423_cast_fp16 = pad(constant_val = const_544_to_fp16, mode = input_1423_mode_0, pad = input_1423_pad_0, x = input_1421_cast_fp16)[name = string("input_1423_cast_fp16")];
            string x_475_pad_type_0 = const()[name = string("x_475_pad_type_0"), val = string("valid")];
            int32 x_475_groups_0 = const()[name = string("x_475_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_475_strides_0 = const()[name = string("x_475_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_475_pad_0 = const()[name = string("x_475_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_475_dilations_0 = const()[name = string("x_475_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_46_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_46_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298963904)))];
            tensor<fp16, [1, 512, 304]> x_475_cast_fp16 = conv(dilations = x_475_dilations_0, groups = x_475_groups_0, pad = x_475_pad_0, pad_type = x_475_pad_type_0, strides = x_475_strides_0, weight = m_encoder_encoders_46_self_attn_fsmn_block_weight_to_fp16, x = input_1423_cast_fp16)[name = string("x_475_cast_fp16")];
            tensor<int32, [3]> x_477_perm_0 = const()[name = string("x_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_477_cast_fp16 = transpose(perm = x_477_perm_0, x = x_475_cast_fp16)[name = string("transpose_485")];
            tensor<fp16, [1, 304, 512]> input_1425_cast_fp16 = add(x = x_477_cast_fp16, y = var_5070_cast_fp16_2)[name = string("input_1425_cast_fp16")];
            fp16 var_5100_to_fp16 = const()[name = string("op_5100_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_191_cast_fp16 = mul(x = var_5075_cast_fp16, y = var_5100_to_fp16)[name = string("q_h_191_cast_fp16")];
            bool scores_189_transpose_x_0 = const()[name = string("scores_189_transpose_x_0"), val = bool(false)];
            bool scores_189_transpose_y_0 = const()[name = string("scores_189_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_304_perm_0 = const()[name = string("transpose_304_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_305_perm_0 = const()[name = string("transpose_305_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_305 = transpose(perm = transpose_305_perm_0, x = var_5078_cast_fp16)[name = string("transpose_483")];
            tensor<fp16, [1, 4, 304, 128]> transpose_304 = transpose(perm = transpose_304_perm_0, x = q_h_191_cast_fp16)[name = string("transpose_484")];
            tensor<fp16, [1, 4, 304, 304]> scores_189_cast_fp16 = matmul(transpose_x = scores_189_transpose_x_0, transpose_y = scores_189_transpose_y_0, x = transpose_304, y = transpose_305)[name = string("scores_189_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5108_cast_fp16 = softmax(axis = var_29, x = scores_189_cast_fp16)[name = string("op_5108_cast_fp16")];
            bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)];
            bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_95_cast_fp16 = transpose(perm = value_95_perm_0, x = var_5081_cast_fp16)[name = string("transpose_487")];
            tensor<fp16, [1, 4, 304, 128]> x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = var_5108_cast_fp16, y = value_95_cast_fp16)[name = string("x_481_cast_fp16")];
            tensor<int32, [4]> var_5112_perm_0 = const()[name = string("op_5112_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5114 = const()[name = string("op_5114"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5112_cast_fp16 = transpose(perm = var_5112_perm_0, x = x_481_cast_fp16)[name = string("transpose_482")];
            tensor<fp16, [1, 304, 512]> input_1429_cast_fp16 = reshape(shape = var_5114, x = var_5112_cast_fp16)[name = string("input_1429_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_46_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_46_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298975232)))];
            tensor<fp16, [512]> m_encoder_encoders_46_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_46_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299499584)))];
            tensor<fp16, [1, 304, 512]> linear_189_cast_fp16 = linear(bias = m_encoder_encoders_46_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_46_self_attn_linear_out_weight_to_fp16, x = input_1429_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1431_cast_fp16 = add(x = linear_189_cast_fp16, y = input_1425_cast_fp16)[name = string("input_1431_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1433_cast_fp16 = add(x = input_1417_cast_fp16, y = input_1431_cast_fp16)[name = string("input_1433_cast_fp16")];
            tensor<int32, [1]> output_191_axes_0 = const()[name = string("output_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_546_to_fp16 = const()[name = string("const_546_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299500672)))];
            tensor<fp16, [512]> const_547_to_fp16 = const()[name = string("const_547_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299501760)))];
            tensor<fp16, [1, 304, 512]> output_191_cast_fp16 = layer_norm(axes = output_191_axes_0, beta = const_547_to_fp16, epsilon = var_22_to_fp16, gamma = const_546_to_fp16, x = input_1433_cast_fp16)[name = string("output_191_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_46_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_46_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299502848)))];
            tensor<fp16, [2048]> m_encoder_encoders_46_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_46_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301600064)))];
            tensor<fp16, [1, 304, 2048]> linear_190_cast_fp16 = linear(bias = m_encoder_encoders_46_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_46_feed_forward_w_1_weight_to_fp16, x = output_191_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1441_cast_fp16 = relu(x = linear_190_cast_fp16)[name = string("input_1441_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_46_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_46_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301604224)))];
            tensor<fp16, [512]> m_encoder_encoders_46_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_46_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303701440)))];
            tensor<fp16, [1, 304, 512]> linear_191_cast_fp16 = linear(bias = m_encoder_encoders_46_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_46_feed_forward_w_2_weight_to_fp16, x = input_1441_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1447_cast_fp16 = add(x = input_1433_cast_fp16, y = linear_191_cast_fp16)[name = string("input_1447_cast_fp16")];
            tensor<int32, [1]> output_193_axes_0 = const()[name = string("output_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_548_to_fp16 = const()[name = string("const_548_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303702528)))];
            tensor<fp16, [512]> const_549_to_fp16 = const()[name = string("const_549_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303703616)))];
            tensor<fp16, [1, 304, 512]> output_193_cast_fp16 = layer_norm(axes = output_193_axes_0, beta = const_549_to_fp16, epsilon = var_22_to_fp16, gamma = const_548_to_fp16, x = input_1447_cast_fp16)[name = string("output_193_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_47_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_47_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303704704)))];
            tensor<fp16, [1536]> m_encoder_encoders_47_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_47_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305277632)))];
            tensor<fp16, [1, 304, 1536]> linear_192_cast_fp16 = linear(bias = m_encoder_encoders_47_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_47_self_attn_linear_q_k_v_weight_to_fp16, x = output_193_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [3]> tile_48 = const()[name = string("tile_48"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_5172_axis_0 = const()[name = string("op_5172_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_5172_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_5172_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_5172_cast_fp16_2 = split(axis = var_5172_axis_0, split_sizes = tile_48, x = linear_192_cast_fp16)[name = string("op_5172_cast_fp16")];
            tensor<int32, [4]> var_5176 = const()[name = string("op_5176"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5177_cast_fp16 = reshape(shape = var_5176, x = var_5172_cast_fp16_0)[name = string("op_5177_cast_fp16")];
            tensor<int32, [4]> var_5179 = const()[name = string("op_5179"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5180_cast_fp16 = reshape(shape = var_5179, x = var_5172_cast_fp16_1)[name = string("op_5180_cast_fp16")];
            tensor<int32, [4]> var_5182 = const()[name = string("op_5182"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5183_cast_fp16 = reshape(shape = var_5182, x = var_5172_cast_fp16_2)[name = string("op_5183_cast_fp16")];
            tensor<int32, [4]> value_97_perm_0 = const()[name = string("value_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1451_perm_0 = const()[name = string("input_1451_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1453_pad_0 = const()[name = string("input_1453_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1453_mode_0 = const()[name = string("input_1453_mode_0"), val = string("constant")];
            fp16 const_555_to_fp16 = const()[name = string("const_555_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1451_cast_fp16 = transpose(perm = input_1451_perm_0, x = var_5172_cast_fp16_2)[name = string("transpose_480")];
            tensor<fp16, [1, 512, 314]> input_1453_cast_fp16 = pad(constant_val = const_555_to_fp16, mode = input_1453_mode_0, pad = input_1453_pad_0, x = input_1451_cast_fp16)[name = string("input_1453_cast_fp16")];
            string x_485_pad_type_0 = const()[name = string("x_485_pad_type_0"), val = string("valid")];
            int32 x_485_groups_0 = const()[name = string("x_485_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_485_strides_0 = const()[name = string("x_485_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_485_pad_0 = const()[name = string("x_485_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_485_dilations_0 = const()[name = string("x_485_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_47_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_47_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305280768)))];
            tensor<fp16, [1, 512, 304]> x_485_cast_fp16 = conv(dilations = x_485_dilations_0, groups = x_485_groups_0, pad = x_485_pad_0, pad_type = x_485_pad_type_0, strides = x_485_strides_0, weight = m_encoder_encoders_47_self_attn_fsmn_block_weight_to_fp16, x = input_1453_cast_fp16)[name = string("x_485_cast_fp16")];
            tensor<int32, [3]> x_487_perm_0 = const()[name = string("x_487_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_487_cast_fp16 = transpose(perm = x_487_perm_0, x = x_485_cast_fp16)[name = string("transpose_479")];
            tensor<fp16, [1, 304, 512]> input_1455_cast_fp16 = add(x = x_487_cast_fp16, y = var_5172_cast_fp16_2)[name = string("input_1455_cast_fp16")];
            fp16 var_5202_to_fp16 = const()[name = string("op_5202_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_195_cast_fp16 = mul(x = var_5177_cast_fp16, y = var_5202_to_fp16)[name = string("q_h_195_cast_fp16")];
            bool scores_193_transpose_x_0 = const()[name = string("scores_193_transpose_x_0"), val = bool(false)];
            bool scores_193_transpose_y_0 = const()[name = string("scores_193_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_306_perm_0 = const()[name = string("transpose_306_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_307_perm_0 = const()[name = string("transpose_307_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_307 = transpose(perm = transpose_307_perm_0, x = var_5180_cast_fp16)[name = string("transpose_477")];
            tensor<fp16, [1, 4, 304, 128]> transpose_306 = transpose(perm = transpose_306_perm_0, x = q_h_195_cast_fp16)[name = string("transpose_478")];
            tensor<fp16, [1, 4, 304, 304]> scores_193_cast_fp16 = matmul(transpose_x = scores_193_transpose_x_0, transpose_y = scores_193_transpose_y_0, x = transpose_306, y = transpose_307)[name = string("scores_193_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5210_cast_fp16 = softmax(axis = var_29, x = scores_193_cast_fp16)[name = string("op_5210_cast_fp16")];
            bool x_491_transpose_x_0 = const()[name = string("x_491_transpose_x_0"), val = bool(false)];
            bool x_491_transpose_y_0 = const()[name = string("x_491_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_97_cast_fp16 = transpose(perm = value_97_perm_0, x = var_5183_cast_fp16)[name = string("transpose_481")];
            tensor<fp16, [1, 4, 304, 128]> x_491_cast_fp16 = matmul(transpose_x = x_491_transpose_x_0, transpose_y = x_491_transpose_y_0, x = var_5210_cast_fp16, y = value_97_cast_fp16)[name = string("x_491_cast_fp16")];
            tensor<int32, [4]> var_5214_perm_0 = const()[name = string("op_5214_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5216 = const()[name = string("op_5216"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5214_cast_fp16 = transpose(perm = var_5214_perm_0, x = x_491_cast_fp16)[name = string("transpose_476")];
            tensor<fp16, [1, 304, 512]> input_1459_cast_fp16 = reshape(shape = var_5216, x = var_5214_cast_fp16)[name = string("input_1459_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_47_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_47_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305292096)))];
            tensor<fp16, [512]> m_encoder_encoders_47_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_47_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305816448)))];
            tensor<fp16, [1, 304, 512]> linear_193_cast_fp16 = linear(bias = m_encoder_encoders_47_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_47_self_attn_linear_out_weight_to_fp16, x = input_1459_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1461_cast_fp16 = add(x = linear_193_cast_fp16, y = input_1455_cast_fp16)[name = string("input_1461_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1463_cast_fp16 = add(x = input_1447_cast_fp16, y = input_1461_cast_fp16)[name = string("input_1463_cast_fp16")];
            tensor<int32, [1]> output_195_axes_0 = const()[name = string("output_195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_557_to_fp16 = const()[name = string("const_557_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305817536)))];
            tensor<fp16, [512]> const_558_to_fp16 = const()[name = string("const_558_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305818624)))];
            tensor<fp16, [1, 304, 512]> output_195_cast_fp16 = layer_norm(axes = output_195_axes_0, beta = const_558_to_fp16, epsilon = var_22_to_fp16, gamma = const_557_to_fp16, x = input_1463_cast_fp16)[name = string("output_195_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_47_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_47_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305819712)))];
            tensor<fp16, [2048]> m_encoder_encoders_47_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_47_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307916928)))];
            tensor<fp16, [1, 304, 2048]> linear_194_cast_fp16 = linear(bias = m_encoder_encoders_47_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_47_feed_forward_w_1_weight_to_fp16, x = output_195_cast_fp16)[name = string("linear_194_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1471_cast_fp16 = relu(x = linear_194_cast_fp16)[name = string("input_1471_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_47_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_47_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307921088)))];
            tensor<fp16, [512]> m_encoder_encoders_47_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_47_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310018304)))];
            tensor<fp16, [1, 304, 512]> linear_195_cast_fp16 = linear(bias = m_encoder_encoders_47_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_47_feed_forward_w_2_weight_to_fp16, x = input_1471_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1477_cast_fp16 = add(x = input_1463_cast_fp16, y = linear_195_cast_fp16)[name = string("input_1477_cast_fp16")];
            tensor<int32, [1]> output_197_axes_0 = const()[name = string("output_197_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_559_to_fp16 = const()[name = string("const_559_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310019392)))];
            tensor<fp16, [512]> const_560_to_fp16 = const()[name = string("const_560_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310020480)))];
            tensor<fp16, [1, 304, 512]> output_197_cast_fp16 = layer_norm(axes = output_197_axes_0, beta = const_560_to_fp16, epsilon = var_22_to_fp16, gamma = const_559_to_fp16, x = input_1477_cast_fp16)[name = string("output_197_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_encoders_48_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_encoders_48_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310021568)))];
            tensor<fp16, [1536]> m_encoder_encoders_48_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_encoders_48_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311594496)))];
            tensor<fp16, [1, 304, 1536]> linear_196_cast_fp16 = linear(bias = m_encoder_encoders_48_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_encoders_48_self_attn_linear_q_k_v_weight_to_fp16, x = output_197_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<int32, [3]> tile_49 = const()[name = string("tile_49"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_5274_axis_0 = const()[name = string("op_5274_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_5274_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_5274_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_5274_cast_fp16_2 = split(axis = var_5274_axis_0, split_sizes = tile_49, x = linear_196_cast_fp16)[name = string("op_5274_cast_fp16")];
            tensor<int32, [4]> var_5278 = const()[name = string("op_5278"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5279_cast_fp16 = reshape(shape = var_5278, x = var_5274_cast_fp16_0)[name = string("op_5279_cast_fp16")];
            tensor<int32, [4]> var_5281 = const()[name = string("op_5281"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5282_cast_fp16 = reshape(shape = var_5281, x = var_5274_cast_fp16_1)[name = string("op_5282_cast_fp16")];
            tensor<int32, [4]> var_5284 = const()[name = string("op_5284"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5285_cast_fp16 = reshape(shape = var_5284, x = var_5274_cast_fp16_2)[name = string("op_5285_cast_fp16")];
            tensor<int32, [4]> value_99_perm_0 = const()[name = string("value_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1481_perm_0 = const()[name = string("input_1481_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1483_pad_0 = const()[name = string("input_1483_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1483_mode_0 = const()[name = string("input_1483_mode_0"), val = string("constant")];
            fp16 const_566_to_fp16 = const()[name = string("const_566_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1481_cast_fp16 = transpose(perm = input_1481_perm_0, x = var_5274_cast_fp16_2)[name = string("transpose_474")];
            tensor<fp16, [1, 512, 314]> input_1483_cast_fp16 = pad(constant_val = const_566_to_fp16, mode = input_1483_mode_0, pad = input_1483_pad_0, x = input_1481_cast_fp16)[name = string("input_1483_cast_fp16")];
            string x_495_pad_type_0 = const()[name = string("x_495_pad_type_0"), val = string("valid")];
            int32 x_495_groups_0 = const()[name = string("x_495_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_495_strides_0 = const()[name = string("x_495_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_495_pad_0 = const()[name = string("x_495_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_495_dilations_0 = const()[name = string("x_495_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_encoders_48_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_encoders_48_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311597632)))];
            tensor<fp16, [1, 512, 304]> x_495_cast_fp16 = conv(dilations = x_495_dilations_0, groups = x_495_groups_0, pad = x_495_pad_0, pad_type = x_495_pad_type_0, strides = x_495_strides_0, weight = m_encoder_encoders_48_self_attn_fsmn_block_weight_to_fp16, x = input_1483_cast_fp16)[name = string("x_495_cast_fp16")];
            tensor<int32, [3]> x_497_perm_0 = const()[name = string("x_497_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_497_cast_fp16 = transpose(perm = x_497_perm_0, x = x_495_cast_fp16)[name = string("transpose_473")];
            tensor<fp16, [1, 304, 512]> input_1485_cast_fp16 = add(x = x_497_cast_fp16, y = var_5274_cast_fp16_2)[name = string("input_1485_cast_fp16")];
            fp16 var_5304_to_fp16 = const()[name = string("op_5304_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_199_cast_fp16 = mul(x = var_5279_cast_fp16, y = var_5304_to_fp16)[name = string("q_h_199_cast_fp16")];
            bool scores_197_transpose_x_0 = const()[name = string("scores_197_transpose_x_0"), val = bool(false)];
            bool scores_197_transpose_y_0 = const()[name = string("scores_197_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_308_perm_0 = const()[name = string("transpose_308_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_309_perm_0 = const()[name = string("transpose_309_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_309 = transpose(perm = transpose_309_perm_0, x = var_5282_cast_fp16)[name = string("transpose_471")];
            tensor<fp16, [1, 4, 304, 128]> transpose_308 = transpose(perm = transpose_308_perm_0, x = q_h_199_cast_fp16)[name = string("transpose_472")];
            tensor<fp16, [1, 4, 304, 304]> scores_197_cast_fp16 = matmul(transpose_x = scores_197_transpose_x_0, transpose_y = scores_197_transpose_y_0, x = transpose_308, y = transpose_309)[name = string("scores_197_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5312_cast_fp16 = softmax(axis = var_29, x = scores_197_cast_fp16)[name = string("op_5312_cast_fp16")];
            bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)];
            bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_99_cast_fp16 = transpose(perm = value_99_perm_0, x = var_5285_cast_fp16)[name = string("transpose_475")];
            tensor<fp16, [1, 4, 304, 128]> x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = var_5312_cast_fp16, y = value_99_cast_fp16)[name = string("x_501_cast_fp16")];
            tensor<int32, [4]> var_5316_perm_0 = const()[name = string("op_5316_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5318 = const()[name = string("op_5318"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5316_cast_fp16 = transpose(perm = var_5316_perm_0, x = x_501_cast_fp16)[name = string("transpose_470")];
            tensor<fp16, [1, 304, 512]> input_1489_cast_fp16 = reshape(shape = var_5318, x = var_5316_cast_fp16)[name = string("input_1489_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_encoders_48_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_encoders_48_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311608960)))];
            tensor<fp16, [512]> m_encoder_encoders_48_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_encoders_48_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312133312)))];
            tensor<fp16, [1, 304, 512]> linear_197_cast_fp16 = linear(bias = m_encoder_encoders_48_self_attn_linear_out_bias_to_fp16, weight = m_encoder_encoders_48_self_attn_linear_out_weight_to_fp16, x = input_1489_cast_fp16)[name = string("linear_197_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1491_cast_fp16 = add(x = linear_197_cast_fp16, y = input_1485_cast_fp16)[name = string("input_1491_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1493_cast_fp16 = add(x = input_1477_cast_fp16, y = input_1491_cast_fp16)[name = string("input_1493_cast_fp16")];
            tensor<int32, [1]> output_199_axes_0 = const()[name = string("output_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_568_to_fp16 = const()[name = string("const_568_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312134400)))];
            tensor<fp16, [512]> const_569_to_fp16 = const()[name = string("const_569_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312135488)))];
            tensor<fp16, [1, 304, 512]> output_199_cast_fp16 = layer_norm(axes = output_199_axes_0, beta = const_569_to_fp16, epsilon = var_22_to_fp16, gamma = const_568_to_fp16, x = input_1493_cast_fp16)[name = string("output_199_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_encoders_48_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_encoders_48_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312136576)))];
            tensor<fp16, [2048]> m_encoder_encoders_48_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_encoders_48_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314233792)))];
            tensor<fp16, [1, 304, 2048]> linear_198_cast_fp16 = linear(bias = m_encoder_encoders_48_feed_forward_w_1_bias_to_fp16, weight = m_encoder_encoders_48_feed_forward_w_1_weight_to_fp16, x = output_199_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1501_cast_fp16 = relu(x = linear_198_cast_fp16)[name = string("input_1501_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_encoders_48_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_encoders_48_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314237952)))];
            tensor<fp16, [512]> m_encoder_encoders_48_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_encoders_48_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316335168)))];
            tensor<fp16, [1, 304, 512]> linear_199_cast_fp16 = linear(bias = m_encoder_encoders_48_feed_forward_w_2_bias_to_fp16, weight = m_encoder_encoders_48_feed_forward_w_2_weight_to_fp16, x = input_1501_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1507_cast_fp16 = add(x = input_1493_cast_fp16, y = linear_199_cast_fp16)[name = string("input_1507_cast_fp16")];
            tensor<int32, [1]> output_201_axes_0 = const()[name = string("output_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_570_to_fp16 = const()[name = string("const_570_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316336256)))];
            tensor<fp16, [512]> const_571_to_fp16 = const()[name = string("const_571_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316337344)))];
            tensor<fp16, [1, 304, 512]> output_201_cast_fp16 = layer_norm(axes = output_201_axes_0, beta = const_571_to_fp16, epsilon = var_22_to_fp16, gamma = const_570_to_fp16, x = input_1507_cast_fp16)[name = string("output_201_cast_fp16")];
            tensor<int32, [1]> encoder_out_lens = const()[name = string("const_572"), val = tensor<int32, [1]>([304])];
            tensor<int32, [1]> output_203_axes_0 = const()[name = string("output_203_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_573_to_fp16 = const()[name = string("const_573_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316338432)))];
            tensor<fp16, [512]> const_574_to_fp16 = const()[name = string("const_574_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316339520)))];
            tensor<fp16, [1, 304, 512]> output_203_cast_fp16 = layer_norm(axes = output_203_axes_0, beta = const_574_to_fp16, epsilon = var_22_to_fp16, gamma = const_573_to_fp16, x = output_201_cast_fp16)[name = string("output_203_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_0_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_0_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316340608)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_0_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_0_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317913536)))];
            tensor<fp16, [1, 304, 1536]> linear_200_cast_fp16 = linear(bias = m_encoder_tp_encoders_0_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_0_self_attn_linear_q_k_v_weight_to_fp16, x = output_203_cast_fp16)[name = string("linear_200_cast_fp16")];
            tensor<int32, [3]> tile_50 = const()[name = string("tile_50"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_5388_axis_0 = const()[name = string("op_5388_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_5388_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_5388_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_5388_cast_fp16_2 = split(axis = var_5388_axis_0, split_sizes = tile_50, x = linear_200_cast_fp16)[name = string("op_5388_cast_fp16")];
            tensor<int32, [4]> var_5392 = const()[name = string("op_5392"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5393_cast_fp16 = reshape(shape = var_5392, x = var_5388_cast_fp16_0)[name = string("op_5393_cast_fp16")];
            tensor<int32, [4]> var_5395 = const()[name = string("op_5395"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5396_cast_fp16 = reshape(shape = var_5395, x = var_5388_cast_fp16_1)[name = string("op_5396_cast_fp16")];
            tensor<int32, [4]> var_5398 = const()[name = string("op_5398"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5399_cast_fp16 = reshape(shape = var_5398, x = var_5388_cast_fp16_2)[name = string("op_5399_cast_fp16")];
            tensor<int32, [4]> value_101_perm_0 = const()[name = string("value_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1515_perm_0 = const()[name = string("input_1515_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1517_pad_0 = const()[name = string("input_1517_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1517_mode_0 = const()[name = string("input_1517_mode_0"), val = string("constant")];
            fp16 const_580_to_fp16 = const()[name = string("const_580_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1515_cast_fp16 = transpose(perm = input_1515_perm_0, x = var_5388_cast_fp16_2)[name = string("transpose_468")];
            tensor<fp16, [1, 512, 314]> input_1517_cast_fp16 = pad(constant_val = const_580_to_fp16, mode = input_1517_mode_0, pad = input_1517_pad_0, x = input_1515_cast_fp16)[name = string("input_1517_cast_fp16")];
            string x_505_pad_type_0 = const()[name = string("x_505_pad_type_0"), val = string("valid")];
            int32 x_505_groups_0 = const()[name = string("x_505_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_505_strides_0 = const()[name = string("x_505_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_505_pad_0 = const()[name = string("x_505_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_505_dilations_0 = const()[name = string("x_505_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_0_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_0_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317916672)))];
            tensor<fp16, [1, 512, 304]> x_505_cast_fp16 = conv(dilations = x_505_dilations_0, groups = x_505_groups_0, pad = x_505_pad_0, pad_type = x_505_pad_type_0, strides = x_505_strides_0, weight = m_encoder_tp_encoders_0_self_attn_fsmn_block_weight_to_fp16, x = input_1517_cast_fp16)[name = string("x_505_cast_fp16")];
            tensor<int32, [3]> x_507_perm_0 = const()[name = string("x_507_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_507_cast_fp16 = transpose(perm = x_507_perm_0, x = x_505_cast_fp16)[name = string("transpose_467")];
            tensor<fp16, [1, 304, 512]> input_1519_cast_fp16 = add(x = x_507_cast_fp16, y = var_5388_cast_fp16_2)[name = string("input_1519_cast_fp16")];
            fp16 var_5418_to_fp16 = const()[name = string("op_5418_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_203_cast_fp16 = mul(x = var_5393_cast_fp16, y = var_5418_to_fp16)[name = string("q_h_203_cast_fp16")];
            bool scores_201_transpose_x_0 = const()[name = string("scores_201_transpose_x_0"), val = bool(false)];
            bool scores_201_transpose_y_0 = const()[name = string("scores_201_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_310_perm_0 = const()[name = string("transpose_310_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_311_perm_0 = const()[name = string("transpose_311_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_311 = transpose(perm = transpose_311_perm_0, x = var_5396_cast_fp16)[name = string("transpose_465")];
            tensor<fp16, [1, 4, 304, 128]> transpose_310 = transpose(perm = transpose_310_perm_0, x = q_h_203_cast_fp16)[name = string("transpose_466")];
            tensor<fp16, [1, 4, 304, 304]> scores_201_cast_fp16 = matmul(transpose_x = scores_201_transpose_x_0, transpose_y = scores_201_transpose_y_0, x = transpose_310, y = transpose_311)[name = string("scores_201_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5426_cast_fp16 = softmax(axis = var_29, x = scores_201_cast_fp16)[name = string("op_5426_cast_fp16")];
            bool x_511_transpose_x_0 = const()[name = string("x_511_transpose_x_0"), val = bool(false)];
            bool x_511_transpose_y_0 = const()[name = string("x_511_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_101_cast_fp16 = transpose(perm = value_101_perm_0, x = var_5399_cast_fp16)[name = string("transpose_469")];
            tensor<fp16, [1, 4, 304, 128]> x_511_cast_fp16 = matmul(transpose_x = x_511_transpose_x_0, transpose_y = x_511_transpose_y_0, x = var_5426_cast_fp16, y = value_101_cast_fp16)[name = string("x_511_cast_fp16")];
            tensor<int32, [4]> var_5430_perm_0 = const()[name = string("op_5430_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5432 = const()[name = string("op_5432"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5430_cast_fp16 = transpose(perm = var_5430_perm_0, x = x_511_cast_fp16)[name = string("transpose_464")];
            tensor<fp16, [1, 304, 512]> input_1523_cast_fp16 = reshape(shape = var_5432, x = var_5430_cast_fp16)[name = string("input_1523_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_0_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_0_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317928000)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318452352)))];
            tensor<fp16, [1, 304, 512]> linear_201_cast_fp16 = linear(bias = m_encoder_tp_encoders_0_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_0_self_attn_linear_out_weight_to_fp16, x = input_1523_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1525_cast_fp16 = add(x = linear_201_cast_fp16, y = input_1519_cast_fp16)[name = string("input_1525_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1527_cast_fp16 = add(x = output_201_cast_fp16, y = input_1525_cast_fp16)[name = string("input_1527_cast_fp16")];
            tensor<int32, [1]> output_205_axes_0 = const()[name = string("output_205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_582_to_fp16 = const()[name = string("const_582_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318453440)))];
            tensor<fp16, [512]> const_583_to_fp16 = const()[name = string("const_583_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318454528)))];
            tensor<fp16, [1, 304, 512]> output_205_cast_fp16 = layer_norm(axes = output_205_axes_0, beta = const_583_to_fp16, epsilon = var_22_to_fp16, gamma = const_582_to_fp16, x = input_1527_cast_fp16)[name = string("output_205_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_0_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_0_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318455616)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_0_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_0_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320552832)))];
            tensor<fp16, [1, 304, 2048]> linear_202_cast_fp16 = linear(bias = m_encoder_tp_encoders_0_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_0_feed_forward_w_1_weight_to_fp16, x = output_205_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1535_cast_fp16 = relu(x = linear_202_cast_fp16)[name = string("input_1535_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_0_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_0_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320556992)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_0_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_0_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322654208)))];
            tensor<fp16, [1, 304, 512]> linear_203_cast_fp16 = linear(bias = m_encoder_tp_encoders_0_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_0_feed_forward_w_2_weight_to_fp16, x = input_1535_cast_fp16)[name = string("linear_203_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1541_cast_fp16 = add(x = input_1527_cast_fp16, y = linear_203_cast_fp16)[name = string("input_1541_cast_fp16")];
            tensor<int32, [1]> output_207_axes_0 = const()[name = string("output_207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_584_to_fp16 = const()[name = string("const_584_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322655296)))];
            tensor<fp16, [512]> const_585_to_fp16 = const()[name = string("const_585_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322656384)))];
            tensor<fp16, [1, 304, 512]> output_207_cast_fp16 = layer_norm(axes = output_207_axes_0, beta = const_585_to_fp16, epsilon = var_22_to_fp16, gamma = const_584_to_fp16, x = input_1541_cast_fp16)[name = string("output_207_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_1_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_1_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322657472)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_1_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_1_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324230400)))];
            tensor<fp16, [1, 304, 1536]> linear_204_cast_fp16 = linear(bias = m_encoder_tp_encoders_1_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_1_self_attn_linear_q_k_v_weight_to_fp16, x = output_207_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<int32, [3]> tile_51 = const()[name = string("tile_51"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_5490_axis_0 = const()[name = string("op_5490_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_5490_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_5490_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_5490_cast_fp16_2 = split(axis = var_5490_axis_0, split_sizes = tile_51, x = linear_204_cast_fp16)[name = string("op_5490_cast_fp16")];
            tensor<int32, [4]> var_5494 = const()[name = string("op_5494"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5495_cast_fp16 = reshape(shape = var_5494, x = var_5490_cast_fp16_0)[name = string("op_5495_cast_fp16")];
            tensor<int32, [4]> var_5497 = const()[name = string("op_5497"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5498_cast_fp16 = reshape(shape = var_5497, x = var_5490_cast_fp16_1)[name = string("op_5498_cast_fp16")];
            tensor<int32, [4]> var_5500 = const()[name = string("op_5500"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5501_cast_fp16 = reshape(shape = var_5500, x = var_5490_cast_fp16_2)[name = string("op_5501_cast_fp16")];
            tensor<int32, [4]> value_103_perm_0 = const()[name = string("value_103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1545_perm_0 = const()[name = string("input_1545_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1547_pad_0 = const()[name = string("input_1547_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1547_mode_0 = const()[name = string("input_1547_mode_0"), val = string("constant")];
            fp16 const_591_to_fp16 = const()[name = string("const_591_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1545_cast_fp16 = transpose(perm = input_1545_perm_0, x = var_5490_cast_fp16_2)[name = string("transpose_462")];
            tensor<fp16, [1, 512, 314]> input_1547_cast_fp16 = pad(constant_val = const_591_to_fp16, mode = input_1547_mode_0, pad = input_1547_pad_0, x = input_1545_cast_fp16)[name = string("input_1547_cast_fp16")];
            string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")];
            int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_1_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_1_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324233536)))];
            tensor<fp16, [1, 512, 304]> x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = m_encoder_tp_encoders_1_self_attn_fsmn_block_weight_to_fp16, x = input_1547_cast_fp16)[name = string("x_515_cast_fp16")];
            tensor<int32, [3]> x_517_perm_0 = const()[name = string("x_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_517_cast_fp16 = transpose(perm = x_517_perm_0, x = x_515_cast_fp16)[name = string("transpose_461")];
            tensor<fp16, [1, 304, 512]> input_1549_cast_fp16 = add(x = x_517_cast_fp16, y = var_5490_cast_fp16_2)[name = string("input_1549_cast_fp16")];
            fp16 var_5520_to_fp16 = const()[name = string("op_5520_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_207_cast_fp16 = mul(x = var_5495_cast_fp16, y = var_5520_to_fp16)[name = string("q_h_207_cast_fp16")];
            bool scores_205_transpose_x_0 = const()[name = string("scores_205_transpose_x_0"), val = bool(false)];
            bool scores_205_transpose_y_0 = const()[name = string("scores_205_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_312_perm_0 = const()[name = string("transpose_312_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_313_perm_0 = const()[name = string("transpose_313_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_313 = transpose(perm = transpose_313_perm_0, x = var_5498_cast_fp16)[name = string("transpose_459")];
            tensor<fp16, [1, 4, 304, 128]> transpose_312 = transpose(perm = transpose_312_perm_0, x = q_h_207_cast_fp16)[name = string("transpose_460")];
            tensor<fp16, [1, 4, 304, 304]> scores_205_cast_fp16 = matmul(transpose_x = scores_205_transpose_x_0, transpose_y = scores_205_transpose_y_0, x = transpose_312, y = transpose_313)[name = string("scores_205_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5528_cast_fp16 = softmax(axis = var_29, x = scores_205_cast_fp16)[name = string("op_5528_cast_fp16")];
            bool x_521_transpose_x_0 = const()[name = string("x_521_transpose_x_0"), val = bool(false)];
            bool x_521_transpose_y_0 = const()[name = string("x_521_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_103_cast_fp16 = transpose(perm = value_103_perm_0, x = var_5501_cast_fp16)[name = string("transpose_463")];
            tensor<fp16, [1, 4, 304, 128]> x_521_cast_fp16 = matmul(transpose_x = x_521_transpose_x_0, transpose_y = x_521_transpose_y_0, x = var_5528_cast_fp16, y = value_103_cast_fp16)[name = string("x_521_cast_fp16")];
            tensor<int32, [4]> var_5532_perm_0 = const()[name = string("op_5532_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5534 = const()[name = string("op_5534"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5532_cast_fp16 = transpose(perm = var_5532_perm_0, x = x_521_cast_fp16)[name = string("transpose_458")];
            tensor<fp16, [1, 304, 512]> input_1553_cast_fp16 = reshape(shape = var_5534, x = var_5532_cast_fp16)[name = string("input_1553_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_1_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_1_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324244864)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324769216)))];
            tensor<fp16, [1, 304, 512]> linear_205_cast_fp16 = linear(bias = m_encoder_tp_encoders_1_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_1_self_attn_linear_out_weight_to_fp16, x = input_1553_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1555_cast_fp16 = add(x = linear_205_cast_fp16, y = input_1549_cast_fp16)[name = string("input_1555_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1557_cast_fp16 = add(x = input_1541_cast_fp16, y = input_1555_cast_fp16)[name = string("input_1557_cast_fp16")];
            tensor<int32, [1]> output_209_axes_0 = const()[name = string("output_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_593_to_fp16 = const()[name = string("const_593_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324770304)))];
            tensor<fp16, [512]> const_594_to_fp16 = const()[name = string("const_594_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324771392)))];
            tensor<fp16, [1, 304, 512]> output_209_cast_fp16 = layer_norm(axes = output_209_axes_0, beta = const_594_to_fp16, epsilon = var_22_to_fp16, gamma = const_593_to_fp16, x = input_1557_cast_fp16)[name = string("output_209_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_1_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_1_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324772480)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_1_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_1_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326869696)))];
            tensor<fp16, [1, 304, 2048]> linear_206_cast_fp16 = linear(bias = m_encoder_tp_encoders_1_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_1_feed_forward_w_1_weight_to_fp16, x = output_209_cast_fp16)[name = string("linear_206_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1565_cast_fp16 = relu(x = linear_206_cast_fp16)[name = string("input_1565_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_1_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_1_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326873856)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_1_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_1_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328971072)))];
            tensor<fp16, [1, 304, 512]> linear_207_cast_fp16 = linear(bias = m_encoder_tp_encoders_1_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_1_feed_forward_w_2_weight_to_fp16, x = input_1565_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1571_cast_fp16 = add(x = input_1557_cast_fp16, y = linear_207_cast_fp16)[name = string("input_1571_cast_fp16")];
            tensor<int32, [1]> output_211_axes_0 = const()[name = string("output_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_595_to_fp16 = const()[name = string("const_595_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328972160)))];
            tensor<fp16, [512]> const_596_to_fp16 = const()[name = string("const_596_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328973248)))];
            tensor<fp16, [1, 304, 512]> output_211_cast_fp16 = layer_norm(axes = output_211_axes_0, beta = const_596_to_fp16, epsilon = var_22_to_fp16, gamma = const_595_to_fp16, x = input_1571_cast_fp16)[name = string("output_211_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_2_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_2_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328974336)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_2_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_2_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330547264)))];
            tensor<fp16, [1, 304, 1536]> linear_208_cast_fp16 = linear(bias = m_encoder_tp_encoders_2_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_2_self_attn_linear_q_k_v_weight_to_fp16, x = output_211_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<int32, [3]> tile_52 = const()[name = string("tile_52"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_5592_axis_0 = const()[name = string("op_5592_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_5592_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_5592_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_5592_cast_fp16_2 = split(axis = var_5592_axis_0, split_sizes = tile_52, x = linear_208_cast_fp16)[name = string("op_5592_cast_fp16")];
            tensor<int32, [4]> var_5596 = const()[name = string("op_5596"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5597_cast_fp16 = reshape(shape = var_5596, x = var_5592_cast_fp16_0)[name = string("op_5597_cast_fp16")];
            tensor<int32, [4]> var_5599 = const()[name = string("op_5599"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5600_cast_fp16 = reshape(shape = var_5599, x = var_5592_cast_fp16_1)[name = string("op_5600_cast_fp16")];
            tensor<int32, [4]> var_5602 = const()[name = string("op_5602"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5603_cast_fp16 = reshape(shape = var_5602, x = var_5592_cast_fp16_2)[name = string("op_5603_cast_fp16")];
            tensor<int32, [4]> value_105_perm_0 = const()[name = string("value_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1575_perm_0 = const()[name = string("input_1575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1577_pad_0 = const()[name = string("input_1577_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1577_mode_0 = const()[name = string("input_1577_mode_0"), val = string("constant")];
            fp16 const_602_to_fp16 = const()[name = string("const_602_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1575_cast_fp16 = transpose(perm = input_1575_perm_0, x = var_5592_cast_fp16_2)[name = string("transpose_456")];
            tensor<fp16, [1, 512, 314]> input_1577_cast_fp16 = pad(constant_val = const_602_to_fp16, mode = input_1577_mode_0, pad = input_1577_pad_0, x = input_1575_cast_fp16)[name = string("input_1577_cast_fp16")];
            string x_525_pad_type_0 = const()[name = string("x_525_pad_type_0"), val = string("valid")];
            int32 x_525_groups_0 = const()[name = string("x_525_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_525_strides_0 = const()[name = string("x_525_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_525_pad_0 = const()[name = string("x_525_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_525_dilations_0 = const()[name = string("x_525_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_2_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_2_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330550400)))];
            tensor<fp16, [1, 512, 304]> x_525_cast_fp16 = conv(dilations = x_525_dilations_0, groups = x_525_groups_0, pad = x_525_pad_0, pad_type = x_525_pad_type_0, strides = x_525_strides_0, weight = m_encoder_tp_encoders_2_self_attn_fsmn_block_weight_to_fp16, x = input_1577_cast_fp16)[name = string("x_525_cast_fp16")];
            tensor<int32, [3]> x_527_perm_0 = const()[name = string("x_527_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_527_cast_fp16 = transpose(perm = x_527_perm_0, x = x_525_cast_fp16)[name = string("transpose_455")];
            tensor<fp16, [1, 304, 512]> input_1579_cast_fp16 = add(x = x_527_cast_fp16, y = var_5592_cast_fp16_2)[name = string("input_1579_cast_fp16")];
            fp16 var_5622_to_fp16 = const()[name = string("op_5622_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_211_cast_fp16 = mul(x = var_5597_cast_fp16, y = var_5622_to_fp16)[name = string("q_h_211_cast_fp16")];
            bool scores_209_transpose_x_0 = const()[name = string("scores_209_transpose_x_0"), val = bool(false)];
            bool scores_209_transpose_y_0 = const()[name = string("scores_209_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_314_perm_0 = const()[name = string("transpose_314_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_315_perm_0 = const()[name = string("transpose_315_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_315 = transpose(perm = transpose_315_perm_0, x = var_5600_cast_fp16)[name = string("transpose_453")];
            tensor<fp16, [1, 4, 304, 128]> transpose_314 = transpose(perm = transpose_314_perm_0, x = q_h_211_cast_fp16)[name = string("transpose_454")];
            tensor<fp16, [1, 4, 304, 304]> scores_209_cast_fp16 = matmul(transpose_x = scores_209_transpose_x_0, transpose_y = scores_209_transpose_y_0, x = transpose_314, y = transpose_315)[name = string("scores_209_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5630_cast_fp16 = softmax(axis = var_29, x = scores_209_cast_fp16)[name = string("op_5630_cast_fp16")];
            bool x_531_transpose_x_0 = const()[name = string("x_531_transpose_x_0"), val = bool(false)];
            bool x_531_transpose_y_0 = const()[name = string("x_531_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_105_cast_fp16 = transpose(perm = value_105_perm_0, x = var_5603_cast_fp16)[name = string("transpose_457")];
            tensor<fp16, [1, 4, 304, 128]> x_531_cast_fp16 = matmul(transpose_x = x_531_transpose_x_0, transpose_y = x_531_transpose_y_0, x = var_5630_cast_fp16, y = value_105_cast_fp16)[name = string("x_531_cast_fp16")];
            tensor<int32, [4]> var_5634_perm_0 = const()[name = string("op_5634_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5636 = const()[name = string("op_5636"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5634_cast_fp16 = transpose(perm = var_5634_perm_0, x = x_531_cast_fp16)[name = string("transpose_452")];
            tensor<fp16, [1, 304, 512]> input_1583_cast_fp16 = reshape(shape = var_5636, x = var_5634_cast_fp16)[name = string("input_1583_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_2_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_2_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330561728)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331086080)))];
            tensor<fp16, [1, 304, 512]> linear_209_cast_fp16 = linear(bias = m_encoder_tp_encoders_2_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_2_self_attn_linear_out_weight_to_fp16, x = input_1583_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1585_cast_fp16 = add(x = linear_209_cast_fp16, y = input_1579_cast_fp16)[name = string("input_1585_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1587_cast_fp16 = add(x = input_1571_cast_fp16, y = input_1585_cast_fp16)[name = string("input_1587_cast_fp16")];
            tensor<int32, [1]> output_213_axes_0 = const()[name = string("output_213_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_604_to_fp16 = const()[name = string("const_604_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331087168)))];
            tensor<fp16, [512]> const_605_to_fp16 = const()[name = string("const_605_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331088256)))];
            tensor<fp16, [1, 304, 512]> output_213_cast_fp16 = layer_norm(axes = output_213_axes_0, beta = const_605_to_fp16, epsilon = var_22_to_fp16, gamma = const_604_to_fp16, x = input_1587_cast_fp16)[name = string("output_213_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_2_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_2_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331089344)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_2_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_2_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333186560)))];
            tensor<fp16, [1, 304, 2048]> linear_210_cast_fp16 = linear(bias = m_encoder_tp_encoders_2_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_2_feed_forward_w_1_weight_to_fp16, x = output_213_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1595_cast_fp16 = relu(x = linear_210_cast_fp16)[name = string("input_1595_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_2_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_2_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333190720)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_2_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_2_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335287936)))];
            tensor<fp16, [1, 304, 512]> linear_211_cast_fp16 = linear(bias = m_encoder_tp_encoders_2_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_2_feed_forward_w_2_weight_to_fp16, x = input_1595_cast_fp16)[name = string("linear_211_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1601_cast_fp16 = add(x = input_1587_cast_fp16, y = linear_211_cast_fp16)[name = string("input_1601_cast_fp16")];
            tensor<int32, [1]> output_215_axes_0 = const()[name = string("output_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_606_to_fp16 = const()[name = string("const_606_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335289024)))];
            tensor<fp16, [512]> const_607_to_fp16 = const()[name = string("const_607_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335290112)))];
            tensor<fp16, [1, 304, 512]> output_215_cast_fp16 = layer_norm(axes = output_215_axes_0, beta = const_607_to_fp16, epsilon = var_22_to_fp16, gamma = const_606_to_fp16, x = input_1601_cast_fp16)[name = string("output_215_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_3_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_3_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335291200)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_3_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_3_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336864128)))];
            tensor<fp16, [1, 304, 1536]> linear_212_cast_fp16 = linear(bias = m_encoder_tp_encoders_3_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_3_self_attn_linear_q_k_v_weight_to_fp16, x = output_215_cast_fp16)[name = string("linear_212_cast_fp16")];
            tensor<int32, [3]> tile_53 = const()[name = string("tile_53"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_5694_axis_0 = const()[name = string("op_5694_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_5694_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_5694_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_5694_cast_fp16_2 = split(axis = var_5694_axis_0, split_sizes = tile_53, x = linear_212_cast_fp16)[name = string("op_5694_cast_fp16")];
            tensor<int32, [4]> var_5698 = const()[name = string("op_5698"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5699_cast_fp16 = reshape(shape = var_5698, x = var_5694_cast_fp16_0)[name = string("op_5699_cast_fp16")];
            tensor<int32, [4]> var_5701 = const()[name = string("op_5701"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5702_cast_fp16 = reshape(shape = var_5701, x = var_5694_cast_fp16_1)[name = string("op_5702_cast_fp16")];
            tensor<int32, [4]> var_5704 = const()[name = string("op_5704"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5705_cast_fp16 = reshape(shape = var_5704, x = var_5694_cast_fp16_2)[name = string("op_5705_cast_fp16")];
            tensor<int32, [4]> value_107_perm_0 = const()[name = string("value_107_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1605_perm_0 = const()[name = string("input_1605_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1607_pad_0 = const()[name = string("input_1607_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1607_mode_0 = const()[name = string("input_1607_mode_0"), val = string("constant")];
            fp16 const_613_to_fp16 = const()[name = string("const_613_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1605_cast_fp16 = transpose(perm = input_1605_perm_0, x = var_5694_cast_fp16_2)[name = string("transpose_450")];
            tensor<fp16, [1, 512, 314]> input_1607_cast_fp16 = pad(constant_val = const_613_to_fp16, mode = input_1607_mode_0, pad = input_1607_pad_0, x = input_1605_cast_fp16)[name = string("input_1607_cast_fp16")];
            string x_535_pad_type_0 = const()[name = string("x_535_pad_type_0"), val = string("valid")];
            int32 x_535_groups_0 = const()[name = string("x_535_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_535_strides_0 = const()[name = string("x_535_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_535_pad_0 = const()[name = string("x_535_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_535_dilations_0 = const()[name = string("x_535_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_3_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_3_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336867264)))];
            tensor<fp16, [1, 512, 304]> x_535_cast_fp16 = conv(dilations = x_535_dilations_0, groups = x_535_groups_0, pad = x_535_pad_0, pad_type = x_535_pad_type_0, strides = x_535_strides_0, weight = m_encoder_tp_encoders_3_self_attn_fsmn_block_weight_to_fp16, x = input_1607_cast_fp16)[name = string("x_535_cast_fp16")];
            tensor<int32, [3]> x_537_perm_0 = const()[name = string("x_537_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_537_cast_fp16 = transpose(perm = x_537_perm_0, x = x_535_cast_fp16)[name = string("transpose_449")];
            tensor<fp16, [1, 304, 512]> input_1609_cast_fp16 = add(x = x_537_cast_fp16, y = var_5694_cast_fp16_2)[name = string("input_1609_cast_fp16")];
            fp16 var_5724_to_fp16 = const()[name = string("op_5724_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_215_cast_fp16 = mul(x = var_5699_cast_fp16, y = var_5724_to_fp16)[name = string("q_h_215_cast_fp16")];
            bool scores_213_transpose_x_0 = const()[name = string("scores_213_transpose_x_0"), val = bool(false)];
            bool scores_213_transpose_y_0 = const()[name = string("scores_213_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_316_perm_0 = const()[name = string("transpose_316_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_317_perm_0 = const()[name = string("transpose_317_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_317 = transpose(perm = transpose_317_perm_0, x = var_5702_cast_fp16)[name = string("transpose_447")];
            tensor<fp16, [1, 4, 304, 128]> transpose_316 = transpose(perm = transpose_316_perm_0, x = q_h_215_cast_fp16)[name = string("transpose_448")];
            tensor<fp16, [1, 4, 304, 304]> scores_213_cast_fp16 = matmul(transpose_x = scores_213_transpose_x_0, transpose_y = scores_213_transpose_y_0, x = transpose_316, y = transpose_317)[name = string("scores_213_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5732_cast_fp16 = softmax(axis = var_29, x = scores_213_cast_fp16)[name = string("op_5732_cast_fp16")];
            bool x_541_transpose_x_0 = const()[name = string("x_541_transpose_x_0"), val = bool(false)];
            bool x_541_transpose_y_0 = const()[name = string("x_541_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_107_cast_fp16 = transpose(perm = value_107_perm_0, x = var_5705_cast_fp16)[name = string("transpose_451")];
            tensor<fp16, [1, 4, 304, 128]> x_541_cast_fp16 = matmul(transpose_x = x_541_transpose_x_0, transpose_y = x_541_transpose_y_0, x = var_5732_cast_fp16, y = value_107_cast_fp16)[name = string("x_541_cast_fp16")];
            tensor<int32, [4]> var_5736_perm_0 = const()[name = string("op_5736_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5738 = const()[name = string("op_5738"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5736_cast_fp16 = transpose(perm = var_5736_perm_0, x = x_541_cast_fp16)[name = string("transpose_446")];
            tensor<fp16, [1, 304, 512]> input_1613_cast_fp16 = reshape(shape = var_5738, x = var_5736_cast_fp16)[name = string("input_1613_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_3_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_3_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336878592)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337402944)))];
            tensor<fp16, [1, 304, 512]> linear_213_cast_fp16 = linear(bias = m_encoder_tp_encoders_3_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_3_self_attn_linear_out_weight_to_fp16, x = input_1613_cast_fp16)[name = string("linear_213_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1615_cast_fp16 = add(x = linear_213_cast_fp16, y = input_1609_cast_fp16)[name = string("input_1615_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1617_cast_fp16 = add(x = input_1601_cast_fp16, y = input_1615_cast_fp16)[name = string("input_1617_cast_fp16")];
            tensor<int32, [1]> output_217_axes_0 = const()[name = string("output_217_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_615_to_fp16 = const()[name = string("const_615_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337404032)))];
            tensor<fp16, [512]> const_616_to_fp16 = const()[name = string("const_616_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337405120)))];
            tensor<fp16, [1, 304, 512]> output_217_cast_fp16 = layer_norm(axes = output_217_axes_0, beta = const_616_to_fp16, epsilon = var_22_to_fp16, gamma = const_615_to_fp16, x = input_1617_cast_fp16)[name = string("output_217_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_3_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_3_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337406208)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_3_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_3_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339503424)))];
            tensor<fp16, [1, 304, 2048]> linear_214_cast_fp16 = linear(bias = m_encoder_tp_encoders_3_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_3_feed_forward_w_1_weight_to_fp16, x = output_217_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1625_cast_fp16 = relu(x = linear_214_cast_fp16)[name = string("input_1625_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_3_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_3_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339507584)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_3_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_3_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341604800)))];
            tensor<fp16, [1, 304, 512]> linear_215_cast_fp16 = linear(bias = m_encoder_tp_encoders_3_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_3_feed_forward_w_2_weight_to_fp16, x = input_1625_cast_fp16)[name = string("linear_215_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1631_cast_fp16 = add(x = input_1617_cast_fp16, y = linear_215_cast_fp16)[name = string("input_1631_cast_fp16")];
            tensor<int32, [1]> output_219_axes_0 = const()[name = string("output_219_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_617_to_fp16 = const()[name = string("const_617_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341605888)))];
            tensor<fp16, [512]> const_618_to_fp16 = const()[name = string("const_618_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341606976)))];
            tensor<fp16, [1, 304, 512]> output_219_cast_fp16 = layer_norm(axes = output_219_axes_0, beta = const_618_to_fp16, epsilon = var_22_to_fp16, gamma = const_617_to_fp16, x = input_1631_cast_fp16)[name = string("output_219_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_4_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_4_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341608064)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_4_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_4_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343180992)))];
            tensor<fp16, [1, 304, 1536]> linear_216_cast_fp16 = linear(bias = m_encoder_tp_encoders_4_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_4_self_attn_linear_q_k_v_weight_to_fp16, x = output_219_cast_fp16)[name = string("linear_216_cast_fp16")];
            tensor<int32, [3]> tile_54 = const()[name = string("tile_54"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_5796_axis_0 = const()[name = string("op_5796_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_5796_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_5796_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_5796_cast_fp16_2 = split(axis = var_5796_axis_0, split_sizes = tile_54, x = linear_216_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<int32, [4]> var_5800 = const()[name = string("op_5800"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5801_cast_fp16 = reshape(shape = var_5800, x = var_5796_cast_fp16_0)[name = string("op_5801_cast_fp16")];
            tensor<int32, [4]> var_5803 = const()[name = string("op_5803"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5804_cast_fp16 = reshape(shape = var_5803, x = var_5796_cast_fp16_1)[name = string("op_5804_cast_fp16")];
            tensor<int32, [4]> var_5806 = const()[name = string("op_5806"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5807_cast_fp16 = reshape(shape = var_5806, x = var_5796_cast_fp16_2)[name = string("op_5807_cast_fp16")];
            tensor<int32, [4]> value_109_perm_0 = const()[name = string("value_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1635_perm_0 = const()[name = string("input_1635_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1637_pad_0 = const()[name = string("input_1637_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1637_mode_0 = const()[name = string("input_1637_mode_0"), val = string("constant")];
            fp16 const_624_to_fp16 = const()[name = string("const_624_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1635_cast_fp16 = transpose(perm = input_1635_perm_0, x = var_5796_cast_fp16_2)[name = string("transpose_444")];
            tensor<fp16, [1, 512, 314]> input_1637_cast_fp16 = pad(constant_val = const_624_to_fp16, mode = input_1637_mode_0, pad = input_1637_pad_0, x = input_1635_cast_fp16)[name = string("input_1637_cast_fp16")];
            string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")];
            int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_4_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_4_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343184128)))];
            tensor<fp16, [1, 512, 304]> x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = m_encoder_tp_encoders_4_self_attn_fsmn_block_weight_to_fp16, x = input_1637_cast_fp16)[name = string("x_545_cast_fp16")];
            tensor<int32, [3]> x_547_perm_0 = const()[name = string("x_547_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_547_cast_fp16 = transpose(perm = x_547_perm_0, x = x_545_cast_fp16)[name = string("transpose_443")];
            tensor<fp16, [1, 304, 512]> input_1639_cast_fp16 = add(x = x_547_cast_fp16, y = var_5796_cast_fp16_2)[name = string("input_1639_cast_fp16")];
            fp16 var_5826_to_fp16 = const()[name = string("op_5826_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_219_cast_fp16 = mul(x = var_5801_cast_fp16, y = var_5826_to_fp16)[name = string("q_h_219_cast_fp16")];
            bool scores_217_transpose_x_0 = const()[name = string("scores_217_transpose_x_0"), val = bool(false)];
            bool scores_217_transpose_y_0 = const()[name = string("scores_217_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_318_perm_0 = const()[name = string("transpose_318_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_319_perm_0 = const()[name = string("transpose_319_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_319 = transpose(perm = transpose_319_perm_0, x = var_5804_cast_fp16)[name = string("transpose_441")];
            tensor<fp16, [1, 4, 304, 128]> transpose_318 = transpose(perm = transpose_318_perm_0, x = q_h_219_cast_fp16)[name = string("transpose_442")];
            tensor<fp16, [1, 4, 304, 304]> scores_217_cast_fp16 = matmul(transpose_x = scores_217_transpose_x_0, transpose_y = scores_217_transpose_y_0, x = transpose_318, y = transpose_319)[name = string("scores_217_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5834_cast_fp16 = softmax(axis = var_29, x = scores_217_cast_fp16)[name = string("op_5834_cast_fp16")];
            bool x_551_transpose_x_0 = const()[name = string("x_551_transpose_x_0"), val = bool(false)];
            bool x_551_transpose_y_0 = const()[name = string("x_551_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_109_cast_fp16 = transpose(perm = value_109_perm_0, x = var_5807_cast_fp16)[name = string("transpose_445")];
            tensor<fp16, [1, 4, 304, 128]> x_551_cast_fp16 = matmul(transpose_x = x_551_transpose_x_0, transpose_y = x_551_transpose_y_0, x = var_5834_cast_fp16, y = value_109_cast_fp16)[name = string("x_551_cast_fp16")];
            tensor<int32, [4]> var_5838_perm_0 = const()[name = string("op_5838_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5840 = const()[name = string("op_5840"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5838_cast_fp16 = transpose(perm = var_5838_perm_0, x = x_551_cast_fp16)[name = string("transpose_440")];
            tensor<fp16, [1, 304, 512]> input_1643_cast_fp16 = reshape(shape = var_5840, x = var_5838_cast_fp16)[name = string("input_1643_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_4_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_4_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343195456)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343719808)))];
            tensor<fp16, [1, 304, 512]> linear_217_cast_fp16 = linear(bias = m_encoder_tp_encoders_4_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_4_self_attn_linear_out_weight_to_fp16, x = input_1643_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1645_cast_fp16 = add(x = linear_217_cast_fp16, y = input_1639_cast_fp16)[name = string("input_1645_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1647_cast_fp16 = add(x = input_1631_cast_fp16, y = input_1645_cast_fp16)[name = string("input_1647_cast_fp16")];
            tensor<int32, [1]> output_221_axes_0 = const()[name = string("output_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_626_to_fp16 = const()[name = string("const_626_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343720896)))];
            tensor<fp16, [512]> const_627_to_fp16 = const()[name = string("const_627_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343721984)))];
            tensor<fp16, [1, 304, 512]> output_221_cast_fp16 = layer_norm(axes = output_221_axes_0, beta = const_627_to_fp16, epsilon = var_22_to_fp16, gamma = const_626_to_fp16, x = input_1647_cast_fp16)[name = string("output_221_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_4_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_4_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343723072)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_4_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_4_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345820288)))];
            tensor<fp16, [1, 304, 2048]> linear_218_cast_fp16 = linear(bias = m_encoder_tp_encoders_4_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_4_feed_forward_w_1_weight_to_fp16, x = output_221_cast_fp16)[name = string("linear_218_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1655_cast_fp16 = relu(x = linear_218_cast_fp16)[name = string("input_1655_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_4_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_4_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345824448)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_4_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_4_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347921664)))];
            tensor<fp16, [1, 304, 512]> linear_219_cast_fp16 = linear(bias = m_encoder_tp_encoders_4_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_4_feed_forward_w_2_weight_to_fp16, x = input_1655_cast_fp16)[name = string("linear_219_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1661_cast_fp16 = add(x = input_1647_cast_fp16, y = linear_219_cast_fp16)[name = string("input_1661_cast_fp16")];
            tensor<int32, [1]> output_223_axes_0 = const()[name = string("output_223_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_628_to_fp16 = const()[name = string("const_628_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347922752)))];
            tensor<fp16, [512]> const_629_to_fp16 = const()[name = string("const_629_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347923840)))];
            tensor<fp16, [1, 304, 512]> output_223_cast_fp16 = layer_norm(axes = output_223_axes_0, beta = const_629_to_fp16, epsilon = var_22_to_fp16, gamma = const_628_to_fp16, x = input_1661_cast_fp16)[name = string("output_223_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_5_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_5_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347924928)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_5_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_5_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349497856)))];
            tensor<fp16, [1, 304, 1536]> linear_220_cast_fp16 = linear(bias = m_encoder_tp_encoders_5_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_5_self_attn_linear_q_k_v_weight_to_fp16, x = output_223_cast_fp16)[name = string("linear_220_cast_fp16")];
            tensor<int32, [3]> tile_55 = const()[name = string("tile_55"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_5898_axis_0 = const()[name = string("op_5898_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_5898_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_5898_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_5898_cast_fp16_2 = split(axis = var_5898_axis_0, split_sizes = tile_55, x = linear_220_cast_fp16)[name = string("op_5898_cast_fp16")];
            tensor<int32, [4]> var_5902 = const()[name = string("op_5902"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5903_cast_fp16 = reshape(shape = var_5902, x = var_5898_cast_fp16_0)[name = string("op_5903_cast_fp16")];
            tensor<int32, [4]> var_5905 = const()[name = string("op_5905"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5906_cast_fp16 = reshape(shape = var_5905, x = var_5898_cast_fp16_1)[name = string("op_5906_cast_fp16")];
            tensor<int32, [4]> var_5908 = const()[name = string("op_5908"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_5909_cast_fp16 = reshape(shape = var_5908, x = var_5898_cast_fp16_2)[name = string("op_5909_cast_fp16")];
            tensor<int32, [4]> value_111_perm_0 = const()[name = string("value_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1665_perm_0 = const()[name = string("input_1665_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1667_pad_0 = const()[name = string("input_1667_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1667_mode_0 = const()[name = string("input_1667_mode_0"), val = string("constant")];
            fp16 const_635_to_fp16 = const()[name = string("const_635_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1665_cast_fp16 = transpose(perm = input_1665_perm_0, x = var_5898_cast_fp16_2)[name = string("transpose_438")];
            tensor<fp16, [1, 512, 314]> input_1667_cast_fp16 = pad(constant_val = const_635_to_fp16, mode = input_1667_mode_0, pad = input_1667_pad_0, x = input_1665_cast_fp16)[name = string("input_1667_cast_fp16")];
            string x_555_pad_type_0 = const()[name = string("x_555_pad_type_0"), val = string("valid")];
            int32 x_555_groups_0 = const()[name = string("x_555_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_555_strides_0 = const()[name = string("x_555_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_555_dilations_0 = const()[name = string("x_555_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_5_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_5_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349500992)))];
            tensor<fp16, [1, 512, 304]> x_555_cast_fp16 = conv(dilations = x_555_dilations_0, groups = x_555_groups_0, pad = x_555_pad_0, pad_type = x_555_pad_type_0, strides = x_555_strides_0, weight = m_encoder_tp_encoders_5_self_attn_fsmn_block_weight_to_fp16, x = input_1667_cast_fp16)[name = string("x_555_cast_fp16")];
            tensor<int32, [3]> x_557_perm_0 = const()[name = string("x_557_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_557_cast_fp16 = transpose(perm = x_557_perm_0, x = x_555_cast_fp16)[name = string("transpose_437")];
            tensor<fp16, [1, 304, 512]> input_1669_cast_fp16 = add(x = x_557_cast_fp16, y = var_5898_cast_fp16_2)[name = string("input_1669_cast_fp16")];
            fp16 var_5928_to_fp16 = const()[name = string("op_5928_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_223_cast_fp16 = mul(x = var_5903_cast_fp16, y = var_5928_to_fp16)[name = string("q_h_223_cast_fp16")];
            bool scores_221_transpose_x_0 = const()[name = string("scores_221_transpose_x_0"), val = bool(false)];
            bool scores_221_transpose_y_0 = const()[name = string("scores_221_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_320_perm_0 = const()[name = string("transpose_320_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_321_perm_0 = const()[name = string("transpose_321_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_321 = transpose(perm = transpose_321_perm_0, x = var_5906_cast_fp16)[name = string("transpose_435")];
            tensor<fp16, [1, 4, 304, 128]> transpose_320 = transpose(perm = transpose_320_perm_0, x = q_h_223_cast_fp16)[name = string("transpose_436")];
            tensor<fp16, [1, 4, 304, 304]> scores_221_cast_fp16 = matmul(transpose_x = scores_221_transpose_x_0, transpose_y = scores_221_transpose_y_0, x = transpose_320, y = transpose_321)[name = string("scores_221_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_5936_cast_fp16 = softmax(axis = var_29, x = scores_221_cast_fp16)[name = string("op_5936_cast_fp16")];
            bool x_561_transpose_x_0 = const()[name = string("x_561_transpose_x_0"), val = bool(false)];
            bool x_561_transpose_y_0 = const()[name = string("x_561_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_111_cast_fp16 = transpose(perm = value_111_perm_0, x = var_5909_cast_fp16)[name = string("transpose_439")];
            tensor<fp16, [1, 4, 304, 128]> x_561_cast_fp16 = matmul(transpose_x = x_561_transpose_x_0, transpose_y = x_561_transpose_y_0, x = var_5936_cast_fp16, y = value_111_cast_fp16)[name = string("x_561_cast_fp16")];
            tensor<int32, [4]> var_5940_perm_0 = const()[name = string("op_5940_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5942 = const()[name = string("op_5942"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_5940_cast_fp16 = transpose(perm = var_5940_perm_0, x = x_561_cast_fp16)[name = string("transpose_434")];
            tensor<fp16, [1, 304, 512]> input_1673_cast_fp16 = reshape(shape = var_5942, x = var_5940_cast_fp16)[name = string("input_1673_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_5_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_5_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349512320)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350036672)))];
            tensor<fp16, [1, 304, 512]> linear_221_cast_fp16 = linear(bias = m_encoder_tp_encoders_5_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_5_self_attn_linear_out_weight_to_fp16, x = input_1673_cast_fp16)[name = string("linear_221_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1675_cast_fp16 = add(x = linear_221_cast_fp16, y = input_1669_cast_fp16)[name = string("input_1675_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1677_cast_fp16 = add(x = input_1661_cast_fp16, y = input_1675_cast_fp16)[name = string("input_1677_cast_fp16")];
            tensor<int32, [1]> output_225_axes_0 = const()[name = string("output_225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_637_to_fp16 = const()[name = string("const_637_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350037760)))];
            tensor<fp16, [512]> const_638_to_fp16 = const()[name = string("const_638_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350038848)))];
            tensor<fp16, [1, 304, 512]> output_225_cast_fp16 = layer_norm(axes = output_225_axes_0, beta = const_638_to_fp16, epsilon = var_22_to_fp16, gamma = const_637_to_fp16, x = input_1677_cast_fp16)[name = string("output_225_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_5_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_5_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350039936)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_5_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_5_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352137152)))];
            tensor<fp16, [1, 304, 2048]> linear_222_cast_fp16 = linear(bias = m_encoder_tp_encoders_5_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_5_feed_forward_w_1_weight_to_fp16, x = output_225_cast_fp16)[name = string("linear_222_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1685_cast_fp16 = relu(x = linear_222_cast_fp16)[name = string("input_1685_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_5_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_5_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352141312)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_5_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_5_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354238528)))];
            tensor<fp16, [1, 304, 512]> linear_223_cast_fp16 = linear(bias = m_encoder_tp_encoders_5_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_5_feed_forward_w_2_weight_to_fp16, x = input_1685_cast_fp16)[name = string("linear_223_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1691_cast_fp16 = add(x = input_1677_cast_fp16, y = linear_223_cast_fp16)[name = string("input_1691_cast_fp16")];
            tensor<int32, [1]> output_227_axes_0 = const()[name = string("output_227_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_639_to_fp16 = const()[name = string("const_639_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354239616)))];
            tensor<fp16, [512]> const_640_to_fp16 = const()[name = string("const_640_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354240704)))];
            tensor<fp16, [1, 304, 512]> output_227_cast_fp16 = layer_norm(axes = output_227_axes_0, beta = const_640_to_fp16, epsilon = var_22_to_fp16, gamma = const_639_to_fp16, x = input_1691_cast_fp16)[name = string("output_227_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_6_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_6_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354241792)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_6_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_6_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355814720)))];
            tensor<fp16, [1, 304, 1536]> linear_224_cast_fp16 = linear(bias = m_encoder_tp_encoders_6_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_6_self_attn_linear_q_k_v_weight_to_fp16, x = output_227_cast_fp16)[name = string("linear_224_cast_fp16")];
            tensor<int32, [3]> tile_56 = const()[name = string("tile_56"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6000_axis_0 = const()[name = string("op_6000_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6000_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6000_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6000_cast_fp16_2 = split(axis = var_6000_axis_0, split_sizes = tile_56, x = linear_224_cast_fp16)[name = string("op_6000_cast_fp16")];
            tensor<int32, [4]> var_6004 = const()[name = string("op_6004"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6005_cast_fp16 = reshape(shape = var_6004, x = var_6000_cast_fp16_0)[name = string("op_6005_cast_fp16")];
            tensor<int32, [4]> var_6007 = const()[name = string("op_6007"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6008_cast_fp16 = reshape(shape = var_6007, x = var_6000_cast_fp16_1)[name = string("op_6008_cast_fp16")];
            tensor<int32, [4]> var_6010 = const()[name = string("op_6010"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6011_cast_fp16 = reshape(shape = var_6010, x = var_6000_cast_fp16_2)[name = string("op_6011_cast_fp16")];
            tensor<int32, [4]> value_113_perm_0 = const()[name = string("value_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1695_perm_0 = const()[name = string("input_1695_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1697_pad_0 = const()[name = string("input_1697_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1697_mode_0 = const()[name = string("input_1697_mode_0"), val = string("constant")];
            fp16 const_646_to_fp16 = const()[name = string("const_646_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1695_cast_fp16 = transpose(perm = input_1695_perm_0, x = var_6000_cast_fp16_2)[name = string("transpose_432")];
            tensor<fp16, [1, 512, 314]> input_1697_cast_fp16 = pad(constant_val = const_646_to_fp16, mode = input_1697_mode_0, pad = input_1697_pad_0, x = input_1695_cast_fp16)[name = string("input_1697_cast_fp16")];
            string x_565_pad_type_0 = const()[name = string("x_565_pad_type_0"), val = string("valid")];
            int32 x_565_groups_0 = const()[name = string("x_565_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_565_strides_0 = const()[name = string("x_565_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_565_pad_0 = const()[name = string("x_565_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_565_dilations_0 = const()[name = string("x_565_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_6_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_6_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355817856)))];
            tensor<fp16, [1, 512, 304]> x_565_cast_fp16 = conv(dilations = x_565_dilations_0, groups = x_565_groups_0, pad = x_565_pad_0, pad_type = x_565_pad_type_0, strides = x_565_strides_0, weight = m_encoder_tp_encoders_6_self_attn_fsmn_block_weight_to_fp16, x = input_1697_cast_fp16)[name = string("x_565_cast_fp16")];
            tensor<int32, [3]> x_567_perm_0 = const()[name = string("x_567_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_567_cast_fp16 = transpose(perm = x_567_perm_0, x = x_565_cast_fp16)[name = string("transpose_431")];
            tensor<fp16, [1, 304, 512]> input_1699_cast_fp16 = add(x = x_567_cast_fp16, y = var_6000_cast_fp16_2)[name = string("input_1699_cast_fp16")];
            fp16 var_6030_to_fp16 = const()[name = string("op_6030_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_227_cast_fp16 = mul(x = var_6005_cast_fp16, y = var_6030_to_fp16)[name = string("q_h_227_cast_fp16")];
            bool scores_225_transpose_x_0 = const()[name = string("scores_225_transpose_x_0"), val = bool(false)];
            bool scores_225_transpose_y_0 = const()[name = string("scores_225_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_322_perm_0 = const()[name = string("transpose_322_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_323_perm_0 = const()[name = string("transpose_323_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_323 = transpose(perm = transpose_323_perm_0, x = var_6008_cast_fp16)[name = string("transpose_429")];
            tensor<fp16, [1, 4, 304, 128]> transpose_322 = transpose(perm = transpose_322_perm_0, x = q_h_227_cast_fp16)[name = string("transpose_430")];
            tensor<fp16, [1, 4, 304, 304]> scores_225_cast_fp16 = matmul(transpose_x = scores_225_transpose_x_0, transpose_y = scores_225_transpose_y_0, x = transpose_322, y = transpose_323)[name = string("scores_225_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6038_cast_fp16 = softmax(axis = var_29, x = scores_225_cast_fp16)[name = string("op_6038_cast_fp16")];
            bool x_571_transpose_x_0 = const()[name = string("x_571_transpose_x_0"), val = bool(false)];
            bool x_571_transpose_y_0 = const()[name = string("x_571_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_113_cast_fp16 = transpose(perm = value_113_perm_0, x = var_6011_cast_fp16)[name = string("transpose_433")];
            tensor<fp16, [1, 4, 304, 128]> x_571_cast_fp16 = matmul(transpose_x = x_571_transpose_x_0, transpose_y = x_571_transpose_y_0, x = var_6038_cast_fp16, y = value_113_cast_fp16)[name = string("x_571_cast_fp16")];
            tensor<int32, [4]> var_6042_perm_0 = const()[name = string("op_6042_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6044 = const()[name = string("op_6044"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6042_cast_fp16 = transpose(perm = var_6042_perm_0, x = x_571_cast_fp16)[name = string("transpose_428")];
            tensor<fp16, [1, 304, 512]> input_1703_cast_fp16 = reshape(shape = var_6044, x = var_6042_cast_fp16)[name = string("input_1703_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_6_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_6_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355829184)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_6_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356353536)))];
            tensor<fp16, [1, 304, 512]> linear_225_cast_fp16 = linear(bias = m_encoder_tp_encoders_6_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_6_self_attn_linear_out_weight_to_fp16, x = input_1703_cast_fp16)[name = string("linear_225_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1705_cast_fp16 = add(x = linear_225_cast_fp16, y = input_1699_cast_fp16)[name = string("input_1705_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1707_cast_fp16 = add(x = input_1691_cast_fp16, y = input_1705_cast_fp16)[name = string("input_1707_cast_fp16")];
            tensor<int32, [1]> output_229_axes_0 = const()[name = string("output_229_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_648_to_fp16 = const()[name = string("const_648_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356354624)))];
            tensor<fp16, [512]> const_649_to_fp16 = const()[name = string("const_649_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356355712)))];
            tensor<fp16, [1, 304, 512]> output_229_cast_fp16 = layer_norm(axes = output_229_axes_0, beta = const_649_to_fp16, epsilon = var_22_to_fp16, gamma = const_648_to_fp16, x = input_1707_cast_fp16)[name = string("output_229_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_6_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_6_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356356800)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_6_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_6_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358454016)))];
            tensor<fp16, [1, 304, 2048]> linear_226_cast_fp16 = linear(bias = m_encoder_tp_encoders_6_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_6_feed_forward_w_1_weight_to_fp16, x = output_229_cast_fp16)[name = string("linear_226_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1715_cast_fp16 = relu(x = linear_226_cast_fp16)[name = string("input_1715_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_6_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_6_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358458176)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_6_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_6_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360555392)))];
            tensor<fp16, [1, 304, 512]> linear_227_cast_fp16 = linear(bias = m_encoder_tp_encoders_6_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_6_feed_forward_w_2_weight_to_fp16, x = input_1715_cast_fp16)[name = string("linear_227_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1721_cast_fp16 = add(x = input_1707_cast_fp16, y = linear_227_cast_fp16)[name = string("input_1721_cast_fp16")];
            tensor<int32, [1]> output_231_axes_0 = const()[name = string("output_231_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_650_to_fp16 = const()[name = string("const_650_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360556480)))];
            tensor<fp16, [512]> const_651_to_fp16 = const()[name = string("const_651_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360557568)))];
            tensor<fp16, [1, 304, 512]> output_231_cast_fp16 = layer_norm(axes = output_231_axes_0, beta = const_651_to_fp16, epsilon = var_22_to_fp16, gamma = const_650_to_fp16, x = input_1721_cast_fp16)[name = string("output_231_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_7_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_7_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360558656)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_7_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_7_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362131584)))];
            tensor<fp16, [1, 304, 1536]> linear_228_cast_fp16 = linear(bias = m_encoder_tp_encoders_7_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_7_self_attn_linear_q_k_v_weight_to_fp16, x = output_231_cast_fp16)[name = string("linear_228_cast_fp16")];
            tensor<int32, [3]> tile_57 = const()[name = string("tile_57"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6102_axis_0 = const()[name = string("op_6102_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6102_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6102_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6102_cast_fp16_2 = split(axis = var_6102_axis_0, split_sizes = tile_57, x = linear_228_cast_fp16)[name = string("op_6102_cast_fp16")];
            tensor<int32, [4]> var_6106 = const()[name = string("op_6106"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6107_cast_fp16 = reshape(shape = var_6106, x = var_6102_cast_fp16_0)[name = string("op_6107_cast_fp16")];
            tensor<int32, [4]> var_6109 = const()[name = string("op_6109"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6110_cast_fp16 = reshape(shape = var_6109, x = var_6102_cast_fp16_1)[name = string("op_6110_cast_fp16")];
            tensor<int32, [4]> var_6112 = const()[name = string("op_6112"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6113_cast_fp16 = reshape(shape = var_6112, x = var_6102_cast_fp16_2)[name = string("op_6113_cast_fp16")];
            tensor<int32, [4]> value_115_perm_0 = const()[name = string("value_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1725_perm_0 = const()[name = string("input_1725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1727_pad_0 = const()[name = string("input_1727_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1727_mode_0 = const()[name = string("input_1727_mode_0"), val = string("constant")];
            fp16 const_657_to_fp16 = const()[name = string("const_657_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1725_cast_fp16 = transpose(perm = input_1725_perm_0, x = var_6102_cast_fp16_2)[name = string("transpose_426")];
            tensor<fp16, [1, 512, 314]> input_1727_cast_fp16 = pad(constant_val = const_657_to_fp16, mode = input_1727_mode_0, pad = input_1727_pad_0, x = input_1725_cast_fp16)[name = string("input_1727_cast_fp16")];
            string x_575_pad_type_0 = const()[name = string("x_575_pad_type_0"), val = string("valid")];
            int32 x_575_groups_0 = const()[name = string("x_575_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_575_strides_0 = const()[name = string("x_575_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_575_pad_0 = const()[name = string("x_575_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_575_dilations_0 = const()[name = string("x_575_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_7_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_7_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362134720)))];
            tensor<fp16, [1, 512, 304]> x_575_cast_fp16 = conv(dilations = x_575_dilations_0, groups = x_575_groups_0, pad = x_575_pad_0, pad_type = x_575_pad_type_0, strides = x_575_strides_0, weight = m_encoder_tp_encoders_7_self_attn_fsmn_block_weight_to_fp16, x = input_1727_cast_fp16)[name = string("x_575_cast_fp16")];
            tensor<int32, [3]> x_577_perm_0 = const()[name = string("x_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_577_cast_fp16 = transpose(perm = x_577_perm_0, x = x_575_cast_fp16)[name = string("transpose_425")];
            tensor<fp16, [1, 304, 512]> input_1729_cast_fp16 = add(x = x_577_cast_fp16, y = var_6102_cast_fp16_2)[name = string("input_1729_cast_fp16")];
            fp16 var_6132_to_fp16 = const()[name = string("op_6132_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_231_cast_fp16 = mul(x = var_6107_cast_fp16, y = var_6132_to_fp16)[name = string("q_h_231_cast_fp16")];
            bool scores_229_transpose_x_0 = const()[name = string("scores_229_transpose_x_0"), val = bool(false)];
            bool scores_229_transpose_y_0 = const()[name = string("scores_229_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_324_perm_0 = const()[name = string("transpose_324_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_325_perm_0 = const()[name = string("transpose_325_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_325 = transpose(perm = transpose_325_perm_0, x = var_6110_cast_fp16)[name = string("transpose_423")];
            tensor<fp16, [1, 4, 304, 128]> transpose_324 = transpose(perm = transpose_324_perm_0, x = q_h_231_cast_fp16)[name = string("transpose_424")];
            tensor<fp16, [1, 4, 304, 304]> scores_229_cast_fp16 = matmul(transpose_x = scores_229_transpose_x_0, transpose_y = scores_229_transpose_y_0, x = transpose_324, y = transpose_325)[name = string("scores_229_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6140_cast_fp16 = softmax(axis = var_29, x = scores_229_cast_fp16)[name = string("op_6140_cast_fp16")];
            bool x_581_transpose_x_0 = const()[name = string("x_581_transpose_x_0"), val = bool(false)];
            bool x_581_transpose_y_0 = const()[name = string("x_581_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_115_cast_fp16 = transpose(perm = value_115_perm_0, x = var_6113_cast_fp16)[name = string("transpose_427")];
            tensor<fp16, [1, 4, 304, 128]> x_581_cast_fp16 = matmul(transpose_x = x_581_transpose_x_0, transpose_y = x_581_transpose_y_0, x = var_6140_cast_fp16, y = value_115_cast_fp16)[name = string("x_581_cast_fp16")];
            tensor<int32, [4]> var_6144_perm_0 = const()[name = string("op_6144_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6146 = const()[name = string("op_6146"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6144_cast_fp16 = transpose(perm = var_6144_perm_0, x = x_581_cast_fp16)[name = string("transpose_422")];
            tensor<fp16, [1, 304, 512]> input_1733_cast_fp16 = reshape(shape = var_6146, x = var_6144_cast_fp16)[name = string("input_1733_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_7_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_7_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362146048)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_7_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362670400)))];
            tensor<fp16, [1, 304, 512]> linear_229_cast_fp16 = linear(bias = m_encoder_tp_encoders_7_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_7_self_attn_linear_out_weight_to_fp16, x = input_1733_cast_fp16)[name = string("linear_229_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1735_cast_fp16 = add(x = linear_229_cast_fp16, y = input_1729_cast_fp16)[name = string("input_1735_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1737_cast_fp16 = add(x = input_1721_cast_fp16, y = input_1735_cast_fp16)[name = string("input_1737_cast_fp16")];
            tensor<int32, [1]> output_233_axes_0 = const()[name = string("output_233_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_659_to_fp16 = const()[name = string("const_659_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362671488)))];
            tensor<fp16, [512]> const_660_to_fp16 = const()[name = string("const_660_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362672576)))];
            tensor<fp16, [1, 304, 512]> output_233_cast_fp16 = layer_norm(axes = output_233_axes_0, beta = const_660_to_fp16, epsilon = var_22_to_fp16, gamma = const_659_to_fp16, x = input_1737_cast_fp16)[name = string("output_233_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_7_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_7_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362673664)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_7_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_7_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364770880)))];
            tensor<fp16, [1, 304, 2048]> linear_230_cast_fp16 = linear(bias = m_encoder_tp_encoders_7_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_7_feed_forward_w_1_weight_to_fp16, x = output_233_cast_fp16)[name = string("linear_230_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1745_cast_fp16 = relu(x = linear_230_cast_fp16)[name = string("input_1745_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_7_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_7_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364775040)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_7_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_7_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366872256)))];
            tensor<fp16, [1, 304, 512]> linear_231_cast_fp16 = linear(bias = m_encoder_tp_encoders_7_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_7_feed_forward_w_2_weight_to_fp16, x = input_1745_cast_fp16)[name = string("linear_231_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1751_cast_fp16 = add(x = input_1737_cast_fp16, y = linear_231_cast_fp16)[name = string("input_1751_cast_fp16")];
            tensor<int32, [1]> output_235_axes_0 = const()[name = string("output_235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_661_to_fp16 = const()[name = string("const_661_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366873344)))];
            tensor<fp16, [512]> const_662_to_fp16 = const()[name = string("const_662_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366874432)))];
            tensor<fp16, [1, 304, 512]> output_235_cast_fp16 = layer_norm(axes = output_235_axes_0, beta = const_662_to_fp16, epsilon = var_22_to_fp16, gamma = const_661_to_fp16, x = input_1751_cast_fp16)[name = string("output_235_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_8_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_8_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366875520)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_8_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_8_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368448448)))];
            tensor<fp16, [1, 304, 1536]> linear_232_cast_fp16 = linear(bias = m_encoder_tp_encoders_8_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_8_self_attn_linear_q_k_v_weight_to_fp16, x = output_235_cast_fp16)[name = string("linear_232_cast_fp16")];
            tensor<int32, [3]> tile_58 = const()[name = string("tile_58"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6204_axis_0 = const()[name = string("op_6204_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6204_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6204_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6204_cast_fp16_2 = split(axis = var_6204_axis_0, split_sizes = tile_58, x = linear_232_cast_fp16)[name = string("op_6204_cast_fp16")];
            tensor<int32, [4]> var_6208 = const()[name = string("op_6208"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6209_cast_fp16 = reshape(shape = var_6208, x = var_6204_cast_fp16_0)[name = string("op_6209_cast_fp16")];
            tensor<int32, [4]> var_6211 = const()[name = string("op_6211"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6212_cast_fp16 = reshape(shape = var_6211, x = var_6204_cast_fp16_1)[name = string("op_6212_cast_fp16")];
            tensor<int32, [4]> var_6214 = const()[name = string("op_6214"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6215_cast_fp16 = reshape(shape = var_6214, x = var_6204_cast_fp16_2)[name = string("op_6215_cast_fp16")];
            tensor<int32, [4]> value_117_perm_0 = const()[name = string("value_117_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1755_perm_0 = const()[name = string("input_1755_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1757_pad_0 = const()[name = string("input_1757_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1757_mode_0 = const()[name = string("input_1757_mode_0"), val = string("constant")];
            fp16 const_668_to_fp16 = const()[name = string("const_668_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1755_cast_fp16 = transpose(perm = input_1755_perm_0, x = var_6204_cast_fp16_2)[name = string("transpose_420")];
            tensor<fp16, [1, 512, 314]> input_1757_cast_fp16 = pad(constant_val = const_668_to_fp16, mode = input_1757_mode_0, pad = input_1757_pad_0, x = input_1755_cast_fp16)[name = string("input_1757_cast_fp16")];
            string x_585_pad_type_0 = const()[name = string("x_585_pad_type_0"), val = string("valid")];
            int32 x_585_groups_0 = const()[name = string("x_585_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_585_strides_0 = const()[name = string("x_585_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_585_pad_0 = const()[name = string("x_585_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_585_dilations_0 = const()[name = string("x_585_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_8_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_8_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368451584)))];
            tensor<fp16, [1, 512, 304]> x_585_cast_fp16 = conv(dilations = x_585_dilations_0, groups = x_585_groups_0, pad = x_585_pad_0, pad_type = x_585_pad_type_0, strides = x_585_strides_0, weight = m_encoder_tp_encoders_8_self_attn_fsmn_block_weight_to_fp16, x = input_1757_cast_fp16)[name = string("x_585_cast_fp16")];
            tensor<int32, [3]> x_587_perm_0 = const()[name = string("x_587_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_587_cast_fp16 = transpose(perm = x_587_perm_0, x = x_585_cast_fp16)[name = string("transpose_419")];
            tensor<fp16, [1, 304, 512]> input_1759_cast_fp16 = add(x = x_587_cast_fp16, y = var_6204_cast_fp16_2)[name = string("input_1759_cast_fp16")];
            fp16 var_6234_to_fp16 = const()[name = string("op_6234_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_235_cast_fp16 = mul(x = var_6209_cast_fp16, y = var_6234_to_fp16)[name = string("q_h_235_cast_fp16")];
            bool scores_233_transpose_x_0 = const()[name = string("scores_233_transpose_x_0"), val = bool(false)];
            bool scores_233_transpose_y_0 = const()[name = string("scores_233_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_326_perm_0 = const()[name = string("transpose_326_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_327_perm_0 = const()[name = string("transpose_327_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_327 = transpose(perm = transpose_327_perm_0, x = var_6212_cast_fp16)[name = string("transpose_417")];
            tensor<fp16, [1, 4, 304, 128]> transpose_326 = transpose(perm = transpose_326_perm_0, x = q_h_235_cast_fp16)[name = string("transpose_418")];
            tensor<fp16, [1, 4, 304, 304]> scores_233_cast_fp16 = matmul(transpose_x = scores_233_transpose_x_0, transpose_y = scores_233_transpose_y_0, x = transpose_326, y = transpose_327)[name = string("scores_233_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6242_cast_fp16 = softmax(axis = var_29, x = scores_233_cast_fp16)[name = string("op_6242_cast_fp16")];
            bool x_591_transpose_x_0 = const()[name = string("x_591_transpose_x_0"), val = bool(false)];
            bool x_591_transpose_y_0 = const()[name = string("x_591_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_117_cast_fp16 = transpose(perm = value_117_perm_0, x = var_6215_cast_fp16)[name = string("transpose_421")];
            tensor<fp16, [1, 4, 304, 128]> x_591_cast_fp16 = matmul(transpose_x = x_591_transpose_x_0, transpose_y = x_591_transpose_y_0, x = var_6242_cast_fp16, y = value_117_cast_fp16)[name = string("x_591_cast_fp16")];
            tensor<int32, [4]> var_6246_perm_0 = const()[name = string("op_6246_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6248 = const()[name = string("op_6248"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6246_cast_fp16 = transpose(perm = var_6246_perm_0, x = x_591_cast_fp16)[name = string("transpose_416")];
            tensor<fp16, [1, 304, 512]> input_1763_cast_fp16 = reshape(shape = var_6248, x = var_6246_cast_fp16)[name = string("input_1763_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_8_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_8_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368462912)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_8_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368987264)))];
            tensor<fp16, [1, 304, 512]> linear_233_cast_fp16 = linear(bias = m_encoder_tp_encoders_8_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_8_self_attn_linear_out_weight_to_fp16, x = input_1763_cast_fp16)[name = string("linear_233_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1765_cast_fp16 = add(x = linear_233_cast_fp16, y = input_1759_cast_fp16)[name = string("input_1765_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1767_cast_fp16 = add(x = input_1751_cast_fp16, y = input_1765_cast_fp16)[name = string("input_1767_cast_fp16")];
            tensor<int32, [1]> output_237_axes_0 = const()[name = string("output_237_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_670_to_fp16 = const()[name = string("const_670_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368988352)))];
            tensor<fp16, [512]> const_671_to_fp16 = const()[name = string("const_671_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368989440)))];
            tensor<fp16, [1, 304, 512]> output_237_cast_fp16 = layer_norm(axes = output_237_axes_0, beta = const_671_to_fp16, epsilon = var_22_to_fp16, gamma = const_670_to_fp16, x = input_1767_cast_fp16)[name = string("output_237_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_8_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_8_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368990528)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_8_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_8_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371087744)))];
            tensor<fp16, [1, 304, 2048]> linear_234_cast_fp16 = linear(bias = m_encoder_tp_encoders_8_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_8_feed_forward_w_1_weight_to_fp16, x = output_237_cast_fp16)[name = string("linear_234_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1775_cast_fp16 = relu(x = linear_234_cast_fp16)[name = string("input_1775_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_8_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_8_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371091904)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_8_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_8_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373189120)))];
            tensor<fp16, [1, 304, 512]> linear_235_cast_fp16 = linear(bias = m_encoder_tp_encoders_8_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_8_feed_forward_w_2_weight_to_fp16, x = input_1775_cast_fp16)[name = string("linear_235_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1781_cast_fp16 = add(x = input_1767_cast_fp16, y = linear_235_cast_fp16)[name = string("input_1781_cast_fp16")];
            tensor<int32, [1]> output_239_axes_0 = const()[name = string("output_239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_672_to_fp16 = const()[name = string("const_672_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373190208)))];
            tensor<fp16, [512]> const_673_to_fp16 = const()[name = string("const_673_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373191296)))];
            tensor<fp16, [1, 304, 512]> output_239_cast_fp16 = layer_norm(axes = output_239_axes_0, beta = const_673_to_fp16, epsilon = var_22_to_fp16, gamma = const_672_to_fp16, x = input_1781_cast_fp16)[name = string("output_239_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_9_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_9_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373192384)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_9_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_9_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374765312)))];
            tensor<fp16, [1, 304, 1536]> linear_236_cast_fp16 = linear(bias = m_encoder_tp_encoders_9_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_9_self_attn_linear_q_k_v_weight_to_fp16, x = output_239_cast_fp16)[name = string("linear_236_cast_fp16")];
            tensor<int32, [3]> tile_59 = const()[name = string("tile_59"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6306_axis_0 = const()[name = string("op_6306_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6306_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6306_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6306_cast_fp16_2 = split(axis = var_6306_axis_0, split_sizes = tile_59, x = linear_236_cast_fp16)[name = string("op_6306_cast_fp16")];
            tensor<int32, [4]> var_6310 = const()[name = string("op_6310"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6311_cast_fp16 = reshape(shape = var_6310, x = var_6306_cast_fp16_0)[name = string("op_6311_cast_fp16")];
            tensor<int32, [4]> var_6313 = const()[name = string("op_6313"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6314_cast_fp16 = reshape(shape = var_6313, x = var_6306_cast_fp16_1)[name = string("op_6314_cast_fp16")];
            tensor<int32, [4]> var_6316 = const()[name = string("op_6316"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6317_cast_fp16 = reshape(shape = var_6316, x = var_6306_cast_fp16_2)[name = string("op_6317_cast_fp16")];
            tensor<int32, [4]> value_119_perm_0 = const()[name = string("value_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1785_perm_0 = const()[name = string("input_1785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1787_pad_0 = const()[name = string("input_1787_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1787_mode_0 = const()[name = string("input_1787_mode_0"), val = string("constant")];
            fp16 const_679_to_fp16 = const()[name = string("const_679_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1785_cast_fp16 = transpose(perm = input_1785_perm_0, x = var_6306_cast_fp16_2)[name = string("transpose_414")];
            tensor<fp16, [1, 512, 314]> input_1787_cast_fp16 = pad(constant_val = const_679_to_fp16, mode = input_1787_mode_0, pad = input_1787_pad_0, x = input_1785_cast_fp16)[name = string("input_1787_cast_fp16")];
            string x_595_pad_type_0 = const()[name = string("x_595_pad_type_0"), val = string("valid")];
            int32 x_595_groups_0 = const()[name = string("x_595_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_595_strides_0 = const()[name = string("x_595_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_595_pad_0 = const()[name = string("x_595_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_595_dilations_0 = const()[name = string("x_595_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_9_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_9_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374768448)))];
            tensor<fp16, [1, 512, 304]> x_595_cast_fp16 = conv(dilations = x_595_dilations_0, groups = x_595_groups_0, pad = x_595_pad_0, pad_type = x_595_pad_type_0, strides = x_595_strides_0, weight = m_encoder_tp_encoders_9_self_attn_fsmn_block_weight_to_fp16, x = input_1787_cast_fp16)[name = string("x_595_cast_fp16")];
            tensor<int32, [3]> x_597_perm_0 = const()[name = string("x_597_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_597_cast_fp16 = transpose(perm = x_597_perm_0, x = x_595_cast_fp16)[name = string("transpose_413")];
            tensor<fp16, [1, 304, 512]> input_1789_cast_fp16 = add(x = x_597_cast_fp16, y = var_6306_cast_fp16_2)[name = string("input_1789_cast_fp16")];
            fp16 var_6336_to_fp16 = const()[name = string("op_6336_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_239_cast_fp16 = mul(x = var_6311_cast_fp16, y = var_6336_to_fp16)[name = string("q_h_239_cast_fp16")];
            bool scores_237_transpose_x_0 = const()[name = string("scores_237_transpose_x_0"), val = bool(false)];
            bool scores_237_transpose_y_0 = const()[name = string("scores_237_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_328_perm_0 = const()[name = string("transpose_328_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_329_perm_0 = const()[name = string("transpose_329_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_329 = transpose(perm = transpose_329_perm_0, x = var_6314_cast_fp16)[name = string("transpose_411")];
            tensor<fp16, [1, 4, 304, 128]> transpose_328 = transpose(perm = transpose_328_perm_0, x = q_h_239_cast_fp16)[name = string("transpose_412")];
            tensor<fp16, [1, 4, 304, 304]> scores_237_cast_fp16 = matmul(transpose_x = scores_237_transpose_x_0, transpose_y = scores_237_transpose_y_0, x = transpose_328, y = transpose_329)[name = string("scores_237_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6344_cast_fp16 = softmax(axis = var_29, x = scores_237_cast_fp16)[name = string("op_6344_cast_fp16")];
            bool x_601_transpose_x_0 = const()[name = string("x_601_transpose_x_0"), val = bool(false)];
            bool x_601_transpose_y_0 = const()[name = string("x_601_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_119_cast_fp16 = transpose(perm = value_119_perm_0, x = var_6317_cast_fp16)[name = string("transpose_415")];
            tensor<fp16, [1, 4, 304, 128]> x_601_cast_fp16 = matmul(transpose_x = x_601_transpose_x_0, transpose_y = x_601_transpose_y_0, x = var_6344_cast_fp16, y = value_119_cast_fp16)[name = string("x_601_cast_fp16")];
            tensor<int32, [4]> var_6348_perm_0 = const()[name = string("op_6348_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6350 = const()[name = string("op_6350"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6348_cast_fp16 = transpose(perm = var_6348_perm_0, x = x_601_cast_fp16)[name = string("transpose_410")];
            tensor<fp16, [1, 304, 512]> input_1793_cast_fp16 = reshape(shape = var_6350, x = var_6348_cast_fp16)[name = string("input_1793_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_9_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_9_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374779776)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_9_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375304128)))];
            tensor<fp16, [1, 304, 512]> linear_237_cast_fp16 = linear(bias = m_encoder_tp_encoders_9_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_9_self_attn_linear_out_weight_to_fp16, x = input_1793_cast_fp16)[name = string("linear_237_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1795_cast_fp16 = add(x = linear_237_cast_fp16, y = input_1789_cast_fp16)[name = string("input_1795_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1797_cast_fp16 = add(x = input_1781_cast_fp16, y = input_1795_cast_fp16)[name = string("input_1797_cast_fp16")];
            tensor<int32, [1]> output_241_axes_0 = const()[name = string("output_241_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_681_to_fp16 = const()[name = string("const_681_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375305216)))];
            tensor<fp16, [512]> const_682_to_fp16 = const()[name = string("const_682_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375306304)))];
            tensor<fp16, [1, 304, 512]> output_241_cast_fp16 = layer_norm(axes = output_241_axes_0, beta = const_682_to_fp16, epsilon = var_22_to_fp16, gamma = const_681_to_fp16, x = input_1797_cast_fp16)[name = string("output_241_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_9_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_9_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375307392)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_9_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_9_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377404608)))];
            tensor<fp16, [1, 304, 2048]> linear_238_cast_fp16 = linear(bias = m_encoder_tp_encoders_9_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_9_feed_forward_w_1_weight_to_fp16, x = output_241_cast_fp16)[name = string("linear_238_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1805_cast_fp16 = relu(x = linear_238_cast_fp16)[name = string("input_1805_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_9_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_9_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377408768)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_9_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_9_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379505984)))];
            tensor<fp16, [1, 304, 512]> linear_239_cast_fp16 = linear(bias = m_encoder_tp_encoders_9_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_9_feed_forward_w_2_weight_to_fp16, x = input_1805_cast_fp16)[name = string("linear_239_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1811_cast_fp16 = add(x = input_1797_cast_fp16, y = linear_239_cast_fp16)[name = string("input_1811_cast_fp16")];
            tensor<int32, [1]> output_243_axes_0 = const()[name = string("output_243_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_683_to_fp16 = const()[name = string("const_683_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379507072)))];
            tensor<fp16, [512]> const_684_to_fp16 = const()[name = string("const_684_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379508160)))];
            tensor<fp16, [1, 304, 512]> output_243_cast_fp16 = layer_norm(axes = output_243_axes_0, beta = const_684_to_fp16, epsilon = var_22_to_fp16, gamma = const_683_to_fp16, x = input_1811_cast_fp16)[name = string("output_243_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_10_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_10_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379509248)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_10_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_10_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381082176)))];
            tensor<fp16, [1, 304, 1536]> linear_240_cast_fp16 = linear(bias = m_encoder_tp_encoders_10_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_10_self_attn_linear_q_k_v_weight_to_fp16, x = output_243_cast_fp16)[name = string("linear_240_cast_fp16")];
            tensor<int32, [3]> tile_60 = const()[name = string("tile_60"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6408_axis_0 = const()[name = string("op_6408_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6408_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6408_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6408_cast_fp16_2 = split(axis = var_6408_axis_0, split_sizes = tile_60, x = linear_240_cast_fp16)[name = string("op_6408_cast_fp16")];
            tensor<int32, [4]> var_6412 = const()[name = string("op_6412"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6413_cast_fp16 = reshape(shape = var_6412, x = var_6408_cast_fp16_0)[name = string("op_6413_cast_fp16")];
            tensor<int32, [4]> var_6415 = const()[name = string("op_6415"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6416_cast_fp16 = reshape(shape = var_6415, x = var_6408_cast_fp16_1)[name = string("op_6416_cast_fp16")];
            tensor<int32, [4]> var_6418 = const()[name = string("op_6418"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6419_cast_fp16 = reshape(shape = var_6418, x = var_6408_cast_fp16_2)[name = string("op_6419_cast_fp16")];
            tensor<int32, [4]> value_121_perm_0 = const()[name = string("value_121_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1815_perm_0 = const()[name = string("input_1815_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1817_pad_0 = const()[name = string("input_1817_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1817_mode_0 = const()[name = string("input_1817_mode_0"), val = string("constant")];
            fp16 const_690_to_fp16 = const()[name = string("const_690_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1815_cast_fp16 = transpose(perm = input_1815_perm_0, x = var_6408_cast_fp16_2)[name = string("transpose_408")];
            tensor<fp16, [1, 512, 314]> input_1817_cast_fp16 = pad(constant_val = const_690_to_fp16, mode = input_1817_mode_0, pad = input_1817_pad_0, x = input_1815_cast_fp16)[name = string("input_1817_cast_fp16")];
            string x_605_pad_type_0 = const()[name = string("x_605_pad_type_0"), val = string("valid")];
            int32 x_605_groups_0 = const()[name = string("x_605_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_605_strides_0 = const()[name = string("x_605_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_605_pad_0 = const()[name = string("x_605_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_605_dilations_0 = const()[name = string("x_605_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_10_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_10_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381085312)))];
            tensor<fp16, [1, 512, 304]> x_605_cast_fp16 = conv(dilations = x_605_dilations_0, groups = x_605_groups_0, pad = x_605_pad_0, pad_type = x_605_pad_type_0, strides = x_605_strides_0, weight = m_encoder_tp_encoders_10_self_attn_fsmn_block_weight_to_fp16, x = input_1817_cast_fp16)[name = string("x_605_cast_fp16")];
            tensor<int32, [3]> x_607_perm_0 = const()[name = string("x_607_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_607_cast_fp16 = transpose(perm = x_607_perm_0, x = x_605_cast_fp16)[name = string("transpose_407")];
            tensor<fp16, [1, 304, 512]> input_1819_cast_fp16 = add(x = x_607_cast_fp16, y = var_6408_cast_fp16_2)[name = string("input_1819_cast_fp16")];
            fp16 var_6438_to_fp16 = const()[name = string("op_6438_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_243_cast_fp16 = mul(x = var_6413_cast_fp16, y = var_6438_to_fp16)[name = string("q_h_243_cast_fp16")];
            bool scores_241_transpose_x_0 = const()[name = string("scores_241_transpose_x_0"), val = bool(false)];
            bool scores_241_transpose_y_0 = const()[name = string("scores_241_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_330_perm_0 = const()[name = string("transpose_330_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_331_perm_0 = const()[name = string("transpose_331_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_331 = transpose(perm = transpose_331_perm_0, x = var_6416_cast_fp16)[name = string("transpose_405")];
            tensor<fp16, [1, 4, 304, 128]> transpose_330 = transpose(perm = transpose_330_perm_0, x = q_h_243_cast_fp16)[name = string("transpose_406")];
            tensor<fp16, [1, 4, 304, 304]> scores_241_cast_fp16 = matmul(transpose_x = scores_241_transpose_x_0, transpose_y = scores_241_transpose_y_0, x = transpose_330, y = transpose_331)[name = string("scores_241_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6446_cast_fp16 = softmax(axis = var_29, x = scores_241_cast_fp16)[name = string("op_6446_cast_fp16")];
            bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)];
            bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_121_cast_fp16 = transpose(perm = value_121_perm_0, x = var_6419_cast_fp16)[name = string("transpose_409")];
            tensor<fp16, [1, 4, 304, 128]> x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = var_6446_cast_fp16, y = value_121_cast_fp16)[name = string("x_611_cast_fp16")];
            tensor<int32, [4]> var_6450_perm_0 = const()[name = string("op_6450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6452 = const()[name = string("op_6452"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6450_cast_fp16 = transpose(perm = var_6450_perm_0, x = x_611_cast_fp16)[name = string("transpose_404")];
            tensor<fp16, [1, 304, 512]> input_1823_cast_fp16 = reshape(shape = var_6452, x = var_6450_cast_fp16)[name = string("input_1823_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_10_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_10_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381096640)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_10_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381620992)))];
            tensor<fp16, [1, 304, 512]> linear_241_cast_fp16 = linear(bias = m_encoder_tp_encoders_10_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_10_self_attn_linear_out_weight_to_fp16, x = input_1823_cast_fp16)[name = string("linear_241_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1825_cast_fp16 = add(x = linear_241_cast_fp16, y = input_1819_cast_fp16)[name = string("input_1825_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1827_cast_fp16 = add(x = input_1811_cast_fp16, y = input_1825_cast_fp16)[name = string("input_1827_cast_fp16")];
            tensor<int32, [1]> output_245_axes_0 = const()[name = string("output_245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_692_to_fp16 = const()[name = string("const_692_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381622080)))];
            tensor<fp16, [512]> const_693_to_fp16 = const()[name = string("const_693_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381623168)))];
            tensor<fp16, [1, 304, 512]> output_245_cast_fp16 = layer_norm(axes = output_245_axes_0, beta = const_693_to_fp16, epsilon = var_22_to_fp16, gamma = const_692_to_fp16, x = input_1827_cast_fp16)[name = string("output_245_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_10_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_10_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381624256)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_10_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_10_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383721472)))];
            tensor<fp16, [1, 304, 2048]> linear_242_cast_fp16 = linear(bias = m_encoder_tp_encoders_10_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_10_feed_forward_w_1_weight_to_fp16, x = output_245_cast_fp16)[name = string("linear_242_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1835_cast_fp16 = relu(x = linear_242_cast_fp16)[name = string("input_1835_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_10_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_10_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383725632)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_10_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_10_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385822848)))];
            tensor<fp16, [1, 304, 512]> linear_243_cast_fp16 = linear(bias = m_encoder_tp_encoders_10_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_10_feed_forward_w_2_weight_to_fp16, x = input_1835_cast_fp16)[name = string("linear_243_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1841_cast_fp16 = add(x = input_1827_cast_fp16, y = linear_243_cast_fp16)[name = string("input_1841_cast_fp16")];
            tensor<int32, [1]> output_247_axes_0 = const()[name = string("output_247_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_694_to_fp16 = const()[name = string("const_694_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385823936)))];
            tensor<fp16, [512]> const_695_to_fp16 = const()[name = string("const_695_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385825024)))];
            tensor<fp16, [1, 304, 512]> output_247_cast_fp16 = layer_norm(axes = output_247_axes_0, beta = const_695_to_fp16, epsilon = var_22_to_fp16, gamma = const_694_to_fp16, x = input_1841_cast_fp16)[name = string("output_247_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_11_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_11_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385826112)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_11_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_11_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387399040)))];
            tensor<fp16, [1, 304, 1536]> linear_244_cast_fp16 = linear(bias = m_encoder_tp_encoders_11_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_11_self_attn_linear_q_k_v_weight_to_fp16, x = output_247_cast_fp16)[name = string("linear_244_cast_fp16")];
            tensor<int32, [3]> tile_61 = const()[name = string("tile_61"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6510_axis_0 = const()[name = string("op_6510_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6510_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6510_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6510_cast_fp16_2 = split(axis = var_6510_axis_0, split_sizes = tile_61, x = linear_244_cast_fp16)[name = string("op_6510_cast_fp16")];
            tensor<int32, [4]> var_6514 = const()[name = string("op_6514"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6515_cast_fp16 = reshape(shape = var_6514, x = var_6510_cast_fp16_0)[name = string("op_6515_cast_fp16")];
            tensor<int32, [4]> var_6517 = const()[name = string("op_6517"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6518_cast_fp16 = reshape(shape = var_6517, x = var_6510_cast_fp16_1)[name = string("op_6518_cast_fp16")];
            tensor<int32, [4]> var_6520 = const()[name = string("op_6520"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6521_cast_fp16 = reshape(shape = var_6520, x = var_6510_cast_fp16_2)[name = string("op_6521_cast_fp16")];
            tensor<int32, [4]> value_123_perm_0 = const()[name = string("value_123_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1845_perm_0 = const()[name = string("input_1845_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1847_pad_0 = const()[name = string("input_1847_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1847_mode_0 = const()[name = string("input_1847_mode_0"), val = string("constant")];
            fp16 const_701_to_fp16 = const()[name = string("const_701_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1845_cast_fp16 = transpose(perm = input_1845_perm_0, x = var_6510_cast_fp16_2)[name = string("transpose_402")];
            tensor<fp16, [1, 512, 314]> input_1847_cast_fp16 = pad(constant_val = const_701_to_fp16, mode = input_1847_mode_0, pad = input_1847_pad_0, x = input_1845_cast_fp16)[name = string("input_1847_cast_fp16")];
            string x_615_pad_type_0 = const()[name = string("x_615_pad_type_0"), val = string("valid")];
            int32 x_615_groups_0 = const()[name = string("x_615_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_615_strides_0 = const()[name = string("x_615_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_615_pad_0 = const()[name = string("x_615_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_615_dilations_0 = const()[name = string("x_615_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_11_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_11_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387402176)))];
            tensor<fp16, [1, 512, 304]> x_615_cast_fp16 = conv(dilations = x_615_dilations_0, groups = x_615_groups_0, pad = x_615_pad_0, pad_type = x_615_pad_type_0, strides = x_615_strides_0, weight = m_encoder_tp_encoders_11_self_attn_fsmn_block_weight_to_fp16, x = input_1847_cast_fp16)[name = string("x_615_cast_fp16")];
            tensor<int32, [3]> x_617_perm_0 = const()[name = string("x_617_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_617_cast_fp16 = transpose(perm = x_617_perm_0, x = x_615_cast_fp16)[name = string("transpose_401")];
            tensor<fp16, [1, 304, 512]> input_1849_cast_fp16 = add(x = x_617_cast_fp16, y = var_6510_cast_fp16_2)[name = string("input_1849_cast_fp16")];
            fp16 var_6540_to_fp16 = const()[name = string("op_6540_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_247_cast_fp16 = mul(x = var_6515_cast_fp16, y = var_6540_to_fp16)[name = string("q_h_247_cast_fp16")];
            bool scores_245_transpose_x_0 = const()[name = string("scores_245_transpose_x_0"), val = bool(false)];
            bool scores_245_transpose_y_0 = const()[name = string("scores_245_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_332_perm_0 = const()[name = string("transpose_332_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_333_perm_0 = const()[name = string("transpose_333_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_333 = transpose(perm = transpose_333_perm_0, x = var_6518_cast_fp16)[name = string("transpose_399")];
            tensor<fp16, [1, 4, 304, 128]> transpose_332 = transpose(perm = transpose_332_perm_0, x = q_h_247_cast_fp16)[name = string("transpose_400")];
            tensor<fp16, [1, 4, 304, 304]> scores_245_cast_fp16 = matmul(transpose_x = scores_245_transpose_x_0, transpose_y = scores_245_transpose_y_0, x = transpose_332, y = transpose_333)[name = string("scores_245_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6548_cast_fp16 = softmax(axis = var_29, x = scores_245_cast_fp16)[name = string("op_6548_cast_fp16")];
            bool x_621_transpose_x_0 = const()[name = string("x_621_transpose_x_0"), val = bool(false)];
            bool x_621_transpose_y_0 = const()[name = string("x_621_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_123_cast_fp16 = transpose(perm = value_123_perm_0, x = var_6521_cast_fp16)[name = string("transpose_403")];
            tensor<fp16, [1, 4, 304, 128]> x_621_cast_fp16 = matmul(transpose_x = x_621_transpose_x_0, transpose_y = x_621_transpose_y_0, x = var_6548_cast_fp16, y = value_123_cast_fp16)[name = string("x_621_cast_fp16")];
            tensor<int32, [4]> var_6552_perm_0 = const()[name = string("op_6552_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6554 = const()[name = string("op_6554"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6552_cast_fp16 = transpose(perm = var_6552_perm_0, x = x_621_cast_fp16)[name = string("transpose_398")];
            tensor<fp16, [1, 304, 512]> input_1853_cast_fp16 = reshape(shape = var_6554, x = var_6552_cast_fp16)[name = string("input_1853_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_11_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_11_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387413504)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_11_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387937856)))];
            tensor<fp16, [1, 304, 512]> linear_245_cast_fp16 = linear(bias = m_encoder_tp_encoders_11_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_11_self_attn_linear_out_weight_to_fp16, x = input_1853_cast_fp16)[name = string("linear_245_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1855_cast_fp16 = add(x = linear_245_cast_fp16, y = input_1849_cast_fp16)[name = string("input_1855_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1857_cast_fp16 = add(x = input_1841_cast_fp16, y = input_1855_cast_fp16)[name = string("input_1857_cast_fp16")];
            tensor<int32, [1]> output_249_axes_0 = const()[name = string("output_249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_703_to_fp16 = const()[name = string("const_703_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387938944)))];
            tensor<fp16, [512]> const_704_to_fp16 = const()[name = string("const_704_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387940032)))];
            tensor<fp16, [1, 304, 512]> output_249_cast_fp16 = layer_norm(axes = output_249_axes_0, beta = const_704_to_fp16, epsilon = var_22_to_fp16, gamma = const_703_to_fp16, x = input_1857_cast_fp16)[name = string("output_249_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_11_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_11_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387941120)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_11_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_11_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390038336)))];
            tensor<fp16, [1, 304, 2048]> linear_246_cast_fp16 = linear(bias = m_encoder_tp_encoders_11_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_11_feed_forward_w_1_weight_to_fp16, x = output_249_cast_fp16)[name = string("linear_246_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1865_cast_fp16 = relu(x = linear_246_cast_fp16)[name = string("input_1865_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_11_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_11_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390042496)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_11_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_11_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392139712)))];
            tensor<fp16, [1, 304, 512]> linear_247_cast_fp16 = linear(bias = m_encoder_tp_encoders_11_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_11_feed_forward_w_2_weight_to_fp16, x = input_1865_cast_fp16)[name = string("linear_247_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1871_cast_fp16 = add(x = input_1857_cast_fp16, y = linear_247_cast_fp16)[name = string("input_1871_cast_fp16")];
            tensor<int32, [1]> output_251_axes_0 = const()[name = string("output_251_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_705_to_fp16 = const()[name = string("const_705_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392140800)))];
            tensor<fp16, [512]> const_706_to_fp16 = const()[name = string("const_706_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392141888)))];
            tensor<fp16, [1, 304, 512]> output_251_cast_fp16 = layer_norm(axes = output_251_axes_0, beta = const_706_to_fp16, epsilon = var_22_to_fp16, gamma = const_705_to_fp16, x = input_1871_cast_fp16)[name = string("output_251_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_12_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_12_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392142976)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_12_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_12_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393715904)))];
            tensor<fp16, [1, 304, 1536]> linear_248_cast_fp16 = linear(bias = m_encoder_tp_encoders_12_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_12_self_attn_linear_q_k_v_weight_to_fp16, x = output_251_cast_fp16)[name = string("linear_248_cast_fp16")];
            tensor<int32, [3]> tile_62 = const()[name = string("tile_62"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6612_axis_0 = const()[name = string("op_6612_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6612_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6612_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6612_cast_fp16_2 = split(axis = var_6612_axis_0, split_sizes = tile_62, x = linear_248_cast_fp16)[name = string("op_6612_cast_fp16")];
            tensor<int32, [4]> var_6616 = const()[name = string("op_6616"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6617_cast_fp16 = reshape(shape = var_6616, x = var_6612_cast_fp16_0)[name = string("op_6617_cast_fp16")];
            tensor<int32, [4]> var_6619 = const()[name = string("op_6619"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6620_cast_fp16 = reshape(shape = var_6619, x = var_6612_cast_fp16_1)[name = string("op_6620_cast_fp16")];
            tensor<int32, [4]> var_6622 = const()[name = string("op_6622"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6623_cast_fp16 = reshape(shape = var_6622, x = var_6612_cast_fp16_2)[name = string("op_6623_cast_fp16")];
            tensor<int32, [4]> value_125_perm_0 = const()[name = string("value_125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1875_perm_0 = const()[name = string("input_1875_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1877_pad_0 = const()[name = string("input_1877_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1877_mode_0 = const()[name = string("input_1877_mode_0"), val = string("constant")];
            fp16 const_712_to_fp16 = const()[name = string("const_712_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1875_cast_fp16 = transpose(perm = input_1875_perm_0, x = var_6612_cast_fp16_2)[name = string("transpose_396")];
            tensor<fp16, [1, 512, 314]> input_1877_cast_fp16 = pad(constant_val = const_712_to_fp16, mode = input_1877_mode_0, pad = input_1877_pad_0, x = input_1875_cast_fp16)[name = string("input_1877_cast_fp16")];
            string x_625_pad_type_0 = const()[name = string("x_625_pad_type_0"), val = string("valid")];
            int32 x_625_groups_0 = const()[name = string("x_625_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_625_strides_0 = const()[name = string("x_625_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_625_pad_0 = const()[name = string("x_625_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_625_dilations_0 = const()[name = string("x_625_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_12_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_12_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393719040)))];
            tensor<fp16, [1, 512, 304]> x_625_cast_fp16 = conv(dilations = x_625_dilations_0, groups = x_625_groups_0, pad = x_625_pad_0, pad_type = x_625_pad_type_0, strides = x_625_strides_0, weight = m_encoder_tp_encoders_12_self_attn_fsmn_block_weight_to_fp16, x = input_1877_cast_fp16)[name = string("x_625_cast_fp16")];
            tensor<int32, [3]> x_627_perm_0 = const()[name = string("x_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_627_cast_fp16 = transpose(perm = x_627_perm_0, x = x_625_cast_fp16)[name = string("transpose_395")];
            tensor<fp16, [1, 304, 512]> input_1879_cast_fp16 = add(x = x_627_cast_fp16, y = var_6612_cast_fp16_2)[name = string("input_1879_cast_fp16")];
            fp16 var_6642_to_fp16 = const()[name = string("op_6642_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_251_cast_fp16 = mul(x = var_6617_cast_fp16, y = var_6642_to_fp16)[name = string("q_h_251_cast_fp16")];
            bool scores_249_transpose_x_0 = const()[name = string("scores_249_transpose_x_0"), val = bool(false)];
            bool scores_249_transpose_y_0 = const()[name = string("scores_249_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_334_perm_0 = const()[name = string("transpose_334_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_335_perm_0 = const()[name = string("transpose_335_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_335 = transpose(perm = transpose_335_perm_0, x = var_6620_cast_fp16)[name = string("transpose_393")];
            tensor<fp16, [1, 4, 304, 128]> transpose_334 = transpose(perm = transpose_334_perm_0, x = q_h_251_cast_fp16)[name = string("transpose_394")];
            tensor<fp16, [1, 4, 304, 304]> scores_249_cast_fp16 = matmul(transpose_x = scores_249_transpose_x_0, transpose_y = scores_249_transpose_y_0, x = transpose_334, y = transpose_335)[name = string("scores_249_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6650_cast_fp16 = softmax(axis = var_29, x = scores_249_cast_fp16)[name = string("op_6650_cast_fp16")];
            bool x_631_transpose_x_0 = const()[name = string("x_631_transpose_x_0"), val = bool(false)];
            bool x_631_transpose_y_0 = const()[name = string("x_631_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_125_cast_fp16 = transpose(perm = value_125_perm_0, x = var_6623_cast_fp16)[name = string("transpose_397")];
            tensor<fp16, [1, 4, 304, 128]> x_631_cast_fp16 = matmul(transpose_x = x_631_transpose_x_0, transpose_y = x_631_transpose_y_0, x = var_6650_cast_fp16, y = value_125_cast_fp16)[name = string("x_631_cast_fp16")];
            tensor<int32, [4]> var_6654_perm_0 = const()[name = string("op_6654_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6656 = const()[name = string("op_6656"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6654_cast_fp16 = transpose(perm = var_6654_perm_0, x = x_631_cast_fp16)[name = string("transpose_392")];
            tensor<fp16, [1, 304, 512]> input_1883_cast_fp16 = reshape(shape = var_6656, x = var_6654_cast_fp16)[name = string("input_1883_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_12_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_12_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393730368)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_12_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394254720)))];
            tensor<fp16, [1, 304, 512]> linear_249_cast_fp16 = linear(bias = m_encoder_tp_encoders_12_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_12_self_attn_linear_out_weight_to_fp16, x = input_1883_cast_fp16)[name = string("linear_249_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1885_cast_fp16 = add(x = linear_249_cast_fp16, y = input_1879_cast_fp16)[name = string("input_1885_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1887_cast_fp16 = add(x = input_1871_cast_fp16, y = input_1885_cast_fp16)[name = string("input_1887_cast_fp16")];
            tensor<int32, [1]> output_253_axes_0 = const()[name = string("output_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_714_to_fp16 = const()[name = string("const_714_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394255808)))];
            tensor<fp16, [512]> const_715_to_fp16 = const()[name = string("const_715_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394256896)))];
            tensor<fp16, [1, 304, 512]> output_253_cast_fp16 = layer_norm(axes = output_253_axes_0, beta = const_715_to_fp16, epsilon = var_22_to_fp16, gamma = const_714_to_fp16, x = input_1887_cast_fp16)[name = string("output_253_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_12_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_12_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394257984)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_12_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_12_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396355200)))];
            tensor<fp16, [1, 304, 2048]> linear_250_cast_fp16 = linear(bias = m_encoder_tp_encoders_12_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_12_feed_forward_w_1_weight_to_fp16, x = output_253_cast_fp16)[name = string("linear_250_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1895_cast_fp16 = relu(x = linear_250_cast_fp16)[name = string("input_1895_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_12_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_12_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396359360)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_12_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_12_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398456576)))];
            tensor<fp16, [1, 304, 512]> linear_251_cast_fp16 = linear(bias = m_encoder_tp_encoders_12_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_12_feed_forward_w_2_weight_to_fp16, x = input_1895_cast_fp16)[name = string("linear_251_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1901_cast_fp16 = add(x = input_1887_cast_fp16, y = linear_251_cast_fp16)[name = string("input_1901_cast_fp16")];
            tensor<int32, [1]> output_255_axes_0 = const()[name = string("output_255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_716_to_fp16 = const()[name = string("const_716_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398457664)))];
            tensor<fp16, [512]> const_717_to_fp16 = const()[name = string("const_717_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398458752)))];
            tensor<fp16, [1, 304, 512]> output_255_cast_fp16 = layer_norm(axes = output_255_axes_0, beta = const_717_to_fp16, epsilon = var_22_to_fp16, gamma = const_716_to_fp16, x = input_1901_cast_fp16)[name = string("output_255_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_13_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_13_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398459840)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_13_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_13_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400032768)))];
            tensor<fp16, [1, 304, 1536]> linear_252_cast_fp16 = linear(bias = m_encoder_tp_encoders_13_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_13_self_attn_linear_q_k_v_weight_to_fp16, x = output_255_cast_fp16)[name = string("linear_252_cast_fp16")];
            tensor<int32, [3]> tile_63 = const()[name = string("tile_63"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6714_axis_0 = const()[name = string("op_6714_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6714_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6714_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6714_cast_fp16_2 = split(axis = var_6714_axis_0, split_sizes = tile_63, x = linear_252_cast_fp16)[name = string("op_6714_cast_fp16")];
            tensor<int32, [4]> var_6718 = const()[name = string("op_6718"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6719_cast_fp16 = reshape(shape = var_6718, x = var_6714_cast_fp16_0)[name = string("op_6719_cast_fp16")];
            tensor<int32, [4]> var_6721 = const()[name = string("op_6721"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6722_cast_fp16 = reshape(shape = var_6721, x = var_6714_cast_fp16_1)[name = string("op_6722_cast_fp16")];
            tensor<int32, [4]> var_6724 = const()[name = string("op_6724"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6725_cast_fp16 = reshape(shape = var_6724, x = var_6714_cast_fp16_2)[name = string("op_6725_cast_fp16")];
            tensor<int32, [4]> value_127_perm_0 = const()[name = string("value_127_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1905_perm_0 = const()[name = string("input_1905_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1907_pad_0 = const()[name = string("input_1907_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1907_mode_0 = const()[name = string("input_1907_mode_0"), val = string("constant")];
            fp16 const_723_to_fp16 = const()[name = string("const_723_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1905_cast_fp16 = transpose(perm = input_1905_perm_0, x = var_6714_cast_fp16_2)[name = string("transpose_390")];
            tensor<fp16, [1, 512, 314]> input_1907_cast_fp16 = pad(constant_val = const_723_to_fp16, mode = input_1907_mode_0, pad = input_1907_pad_0, x = input_1905_cast_fp16)[name = string("input_1907_cast_fp16")];
            string x_635_pad_type_0 = const()[name = string("x_635_pad_type_0"), val = string("valid")];
            int32 x_635_groups_0 = const()[name = string("x_635_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_635_strides_0 = const()[name = string("x_635_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_635_pad_0 = const()[name = string("x_635_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_635_dilations_0 = const()[name = string("x_635_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_13_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_13_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400035904)))];
            tensor<fp16, [1, 512, 304]> x_635_cast_fp16 = conv(dilations = x_635_dilations_0, groups = x_635_groups_0, pad = x_635_pad_0, pad_type = x_635_pad_type_0, strides = x_635_strides_0, weight = m_encoder_tp_encoders_13_self_attn_fsmn_block_weight_to_fp16, x = input_1907_cast_fp16)[name = string("x_635_cast_fp16")];
            tensor<int32, [3]> x_637_perm_0 = const()[name = string("x_637_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_637_cast_fp16 = transpose(perm = x_637_perm_0, x = x_635_cast_fp16)[name = string("transpose_389")];
            tensor<fp16, [1, 304, 512]> input_1909_cast_fp16 = add(x = x_637_cast_fp16, y = var_6714_cast_fp16_2)[name = string("input_1909_cast_fp16")];
            fp16 var_6744_to_fp16 = const()[name = string("op_6744_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_255_cast_fp16 = mul(x = var_6719_cast_fp16, y = var_6744_to_fp16)[name = string("q_h_255_cast_fp16")];
            bool scores_253_transpose_x_0 = const()[name = string("scores_253_transpose_x_0"), val = bool(false)];
            bool scores_253_transpose_y_0 = const()[name = string("scores_253_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_336_perm_0 = const()[name = string("transpose_336_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_337_perm_0 = const()[name = string("transpose_337_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_337 = transpose(perm = transpose_337_perm_0, x = var_6722_cast_fp16)[name = string("transpose_387")];
            tensor<fp16, [1, 4, 304, 128]> transpose_336 = transpose(perm = transpose_336_perm_0, x = q_h_255_cast_fp16)[name = string("transpose_388")];
            tensor<fp16, [1, 4, 304, 304]> scores_253_cast_fp16 = matmul(transpose_x = scores_253_transpose_x_0, transpose_y = scores_253_transpose_y_0, x = transpose_336, y = transpose_337)[name = string("scores_253_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6752_cast_fp16 = softmax(axis = var_29, x = scores_253_cast_fp16)[name = string("op_6752_cast_fp16")];
            bool x_641_transpose_x_0 = const()[name = string("x_641_transpose_x_0"), val = bool(false)];
            bool x_641_transpose_y_0 = const()[name = string("x_641_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_127_cast_fp16 = transpose(perm = value_127_perm_0, x = var_6725_cast_fp16)[name = string("transpose_391")];
            tensor<fp16, [1, 4, 304, 128]> x_641_cast_fp16 = matmul(transpose_x = x_641_transpose_x_0, transpose_y = x_641_transpose_y_0, x = var_6752_cast_fp16, y = value_127_cast_fp16)[name = string("x_641_cast_fp16")];
            tensor<int32, [4]> var_6756_perm_0 = const()[name = string("op_6756_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6758 = const()[name = string("op_6758"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6756_cast_fp16 = transpose(perm = var_6756_perm_0, x = x_641_cast_fp16)[name = string("transpose_386")];
            tensor<fp16, [1, 304, 512]> input_1913_cast_fp16 = reshape(shape = var_6758, x = var_6756_cast_fp16)[name = string("input_1913_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_13_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_13_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400047232)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_13_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400571584)))];
            tensor<fp16, [1, 304, 512]> linear_253_cast_fp16 = linear(bias = m_encoder_tp_encoders_13_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_13_self_attn_linear_out_weight_to_fp16, x = input_1913_cast_fp16)[name = string("linear_253_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1915_cast_fp16 = add(x = linear_253_cast_fp16, y = input_1909_cast_fp16)[name = string("input_1915_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1917_cast_fp16 = add(x = input_1901_cast_fp16, y = input_1915_cast_fp16)[name = string("input_1917_cast_fp16")];
            tensor<int32, [1]> output_257_axes_0 = const()[name = string("output_257_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_725_to_fp16 = const()[name = string("const_725_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400572672)))];
            tensor<fp16, [512]> const_726_to_fp16 = const()[name = string("const_726_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400573760)))];
            tensor<fp16, [1, 304, 512]> output_257_cast_fp16 = layer_norm(axes = output_257_axes_0, beta = const_726_to_fp16, epsilon = var_22_to_fp16, gamma = const_725_to_fp16, x = input_1917_cast_fp16)[name = string("output_257_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_13_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_13_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400574848)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_13_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_13_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402672064)))];
            tensor<fp16, [1, 304, 2048]> linear_254_cast_fp16 = linear(bias = m_encoder_tp_encoders_13_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_13_feed_forward_w_1_weight_to_fp16, x = output_257_cast_fp16)[name = string("linear_254_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1925_cast_fp16 = relu(x = linear_254_cast_fp16)[name = string("input_1925_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_13_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_13_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402676224)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_13_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_13_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404773440)))];
            tensor<fp16, [1, 304, 512]> linear_255_cast_fp16 = linear(bias = m_encoder_tp_encoders_13_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_13_feed_forward_w_2_weight_to_fp16, x = input_1925_cast_fp16)[name = string("linear_255_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1931_cast_fp16 = add(x = input_1917_cast_fp16, y = linear_255_cast_fp16)[name = string("input_1931_cast_fp16")];
            tensor<int32, [1]> output_259_axes_0 = const()[name = string("output_259_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_727_to_fp16 = const()[name = string("const_727_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404774528)))];
            tensor<fp16, [512]> const_728_to_fp16 = const()[name = string("const_728_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404775616)))];
            tensor<fp16, [1, 304, 512]> output_259_cast_fp16 = layer_norm(axes = output_259_axes_0, beta = const_728_to_fp16, epsilon = var_22_to_fp16, gamma = const_727_to_fp16, x = input_1931_cast_fp16)[name = string("output_259_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_14_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_14_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404776704)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_14_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_14_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406349632)))];
            tensor<fp16, [1, 304, 1536]> linear_256_cast_fp16 = linear(bias = m_encoder_tp_encoders_14_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_14_self_attn_linear_q_k_v_weight_to_fp16, x = output_259_cast_fp16)[name = string("linear_256_cast_fp16")];
            tensor<int32, [3]> tile_64 = const()[name = string("tile_64"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6816_axis_0 = const()[name = string("op_6816_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6816_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6816_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6816_cast_fp16_2 = split(axis = var_6816_axis_0, split_sizes = tile_64, x = linear_256_cast_fp16)[name = string("op_6816_cast_fp16")];
            tensor<int32, [4]> var_6820 = const()[name = string("op_6820"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6821_cast_fp16 = reshape(shape = var_6820, x = var_6816_cast_fp16_0)[name = string("op_6821_cast_fp16")];
            tensor<int32, [4]> var_6823 = const()[name = string("op_6823"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6824_cast_fp16 = reshape(shape = var_6823, x = var_6816_cast_fp16_1)[name = string("op_6824_cast_fp16")];
            tensor<int32, [4]> var_6826 = const()[name = string("op_6826"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6827_cast_fp16 = reshape(shape = var_6826, x = var_6816_cast_fp16_2)[name = string("op_6827_cast_fp16")];
            tensor<int32, [4]> value_129_perm_0 = const()[name = string("value_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1935_perm_0 = const()[name = string("input_1935_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1937_pad_0 = const()[name = string("input_1937_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1937_mode_0 = const()[name = string("input_1937_mode_0"), val = string("constant")];
            fp16 const_734_to_fp16 = const()[name = string("const_734_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1935_cast_fp16 = transpose(perm = input_1935_perm_0, x = var_6816_cast_fp16_2)[name = string("transpose_384")];
            tensor<fp16, [1, 512, 314]> input_1937_cast_fp16 = pad(constant_val = const_734_to_fp16, mode = input_1937_mode_0, pad = input_1937_pad_0, x = input_1935_cast_fp16)[name = string("input_1937_cast_fp16")];
            string x_645_pad_type_0 = const()[name = string("x_645_pad_type_0"), val = string("valid")];
            int32 x_645_groups_0 = const()[name = string("x_645_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_645_strides_0 = const()[name = string("x_645_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_645_pad_0 = const()[name = string("x_645_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_645_dilations_0 = const()[name = string("x_645_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_14_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_14_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406352768)))];
            tensor<fp16, [1, 512, 304]> x_645_cast_fp16 = conv(dilations = x_645_dilations_0, groups = x_645_groups_0, pad = x_645_pad_0, pad_type = x_645_pad_type_0, strides = x_645_strides_0, weight = m_encoder_tp_encoders_14_self_attn_fsmn_block_weight_to_fp16, x = input_1937_cast_fp16)[name = string("x_645_cast_fp16")];
            tensor<int32, [3]> x_647_perm_0 = const()[name = string("x_647_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_647_cast_fp16 = transpose(perm = x_647_perm_0, x = x_645_cast_fp16)[name = string("transpose_383")];
            tensor<fp16, [1, 304, 512]> input_1939_cast_fp16 = add(x = x_647_cast_fp16, y = var_6816_cast_fp16_2)[name = string("input_1939_cast_fp16")];
            fp16 var_6846_to_fp16 = const()[name = string("op_6846_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_259_cast_fp16 = mul(x = var_6821_cast_fp16, y = var_6846_to_fp16)[name = string("q_h_259_cast_fp16")];
            bool scores_257_transpose_x_0 = const()[name = string("scores_257_transpose_x_0"), val = bool(false)];
            bool scores_257_transpose_y_0 = const()[name = string("scores_257_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_338_perm_0 = const()[name = string("transpose_338_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_339_perm_0 = const()[name = string("transpose_339_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_339 = transpose(perm = transpose_339_perm_0, x = var_6824_cast_fp16)[name = string("transpose_381")];
            tensor<fp16, [1, 4, 304, 128]> transpose_338 = transpose(perm = transpose_338_perm_0, x = q_h_259_cast_fp16)[name = string("transpose_382")];
            tensor<fp16, [1, 4, 304, 304]> scores_257_cast_fp16 = matmul(transpose_x = scores_257_transpose_x_0, transpose_y = scores_257_transpose_y_0, x = transpose_338, y = transpose_339)[name = string("scores_257_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6854_cast_fp16 = softmax(axis = var_29, x = scores_257_cast_fp16)[name = string("op_6854_cast_fp16")];
            bool x_651_transpose_x_0 = const()[name = string("x_651_transpose_x_0"), val = bool(false)];
            bool x_651_transpose_y_0 = const()[name = string("x_651_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_129_cast_fp16 = transpose(perm = value_129_perm_0, x = var_6827_cast_fp16)[name = string("transpose_385")];
            tensor<fp16, [1, 4, 304, 128]> x_651_cast_fp16 = matmul(transpose_x = x_651_transpose_x_0, transpose_y = x_651_transpose_y_0, x = var_6854_cast_fp16, y = value_129_cast_fp16)[name = string("x_651_cast_fp16")];
            tensor<int32, [4]> var_6858_perm_0 = const()[name = string("op_6858_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6860 = const()[name = string("op_6860"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6858_cast_fp16 = transpose(perm = var_6858_perm_0, x = x_651_cast_fp16)[name = string("transpose_380")];
            tensor<fp16, [1, 304, 512]> input_1943_cast_fp16 = reshape(shape = var_6860, x = var_6858_cast_fp16)[name = string("input_1943_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_14_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_14_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406364096)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_14_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406888448)))];
            tensor<fp16, [1, 304, 512]> linear_257_cast_fp16 = linear(bias = m_encoder_tp_encoders_14_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_14_self_attn_linear_out_weight_to_fp16, x = input_1943_cast_fp16)[name = string("linear_257_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1945_cast_fp16 = add(x = linear_257_cast_fp16, y = input_1939_cast_fp16)[name = string("input_1945_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1947_cast_fp16 = add(x = input_1931_cast_fp16, y = input_1945_cast_fp16)[name = string("input_1947_cast_fp16")];
            tensor<int32, [1]> output_261_axes_0 = const()[name = string("output_261_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_736_to_fp16 = const()[name = string("const_736_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406889536)))];
            tensor<fp16, [512]> const_737_to_fp16 = const()[name = string("const_737_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406890624)))];
            tensor<fp16, [1, 304, 512]> output_261_cast_fp16 = layer_norm(axes = output_261_axes_0, beta = const_737_to_fp16, epsilon = var_22_to_fp16, gamma = const_736_to_fp16, x = input_1947_cast_fp16)[name = string("output_261_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_14_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_14_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406891712)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_14_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_14_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408988928)))];
            tensor<fp16, [1, 304, 2048]> linear_258_cast_fp16 = linear(bias = m_encoder_tp_encoders_14_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_14_feed_forward_w_1_weight_to_fp16, x = output_261_cast_fp16)[name = string("linear_258_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1955_cast_fp16 = relu(x = linear_258_cast_fp16)[name = string("input_1955_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_14_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_14_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408993088)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_14_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_14_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411090304)))];
            tensor<fp16, [1, 304, 512]> linear_259_cast_fp16 = linear(bias = m_encoder_tp_encoders_14_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_14_feed_forward_w_2_weight_to_fp16, x = input_1955_cast_fp16)[name = string("linear_259_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1961_cast_fp16 = add(x = input_1947_cast_fp16, y = linear_259_cast_fp16)[name = string("input_1961_cast_fp16")];
            tensor<int32, [1]> output_263_axes_0 = const()[name = string("output_263_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_738_to_fp16 = const()[name = string("const_738_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411091392)))];
            tensor<fp16, [512]> const_739_to_fp16 = const()[name = string("const_739_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411092480)))];
            tensor<fp16, [1, 304, 512]> output_263_cast_fp16 = layer_norm(axes = output_263_axes_0, beta = const_739_to_fp16, epsilon = var_22_to_fp16, gamma = const_738_to_fp16, x = input_1961_cast_fp16)[name = string("output_263_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_15_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_15_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411093568)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_15_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_15_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412666496)))];
            tensor<fp16, [1, 304, 1536]> linear_260_cast_fp16 = linear(bias = m_encoder_tp_encoders_15_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_15_self_attn_linear_q_k_v_weight_to_fp16, x = output_263_cast_fp16)[name = string("linear_260_cast_fp16")];
            tensor<int32, [3]> tile_65 = const()[name = string("tile_65"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_6918_axis_0 = const()[name = string("op_6918_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_6918_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_6918_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_6918_cast_fp16_2 = split(axis = var_6918_axis_0, split_sizes = tile_65, x = linear_260_cast_fp16)[name = string("op_6918_cast_fp16")];
            tensor<int32, [4]> var_6922 = const()[name = string("op_6922"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6923_cast_fp16 = reshape(shape = var_6922, x = var_6918_cast_fp16_0)[name = string("op_6923_cast_fp16")];
            tensor<int32, [4]> var_6925 = const()[name = string("op_6925"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6926_cast_fp16 = reshape(shape = var_6925, x = var_6918_cast_fp16_1)[name = string("op_6926_cast_fp16")];
            tensor<int32, [4]> var_6928 = const()[name = string("op_6928"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_6929_cast_fp16 = reshape(shape = var_6928, x = var_6918_cast_fp16_2)[name = string("op_6929_cast_fp16")];
            tensor<int32, [4]> value_131_perm_0 = const()[name = string("value_131_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1965_perm_0 = const()[name = string("input_1965_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1967_pad_0 = const()[name = string("input_1967_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1967_mode_0 = const()[name = string("input_1967_mode_0"), val = string("constant")];
            fp16 const_745_to_fp16 = const()[name = string("const_745_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1965_cast_fp16 = transpose(perm = input_1965_perm_0, x = var_6918_cast_fp16_2)[name = string("transpose_378")];
            tensor<fp16, [1, 512, 314]> input_1967_cast_fp16 = pad(constant_val = const_745_to_fp16, mode = input_1967_mode_0, pad = input_1967_pad_0, x = input_1965_cast_fp16)[name = string("input_1967_cast_fp16")];
            string x_655_pad_type_0 = const()[name = string("x_655_pad_type_0"), val = string("valid")];
            int32 x_655_groups_0 = const()[name = string("x_655_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_655_strides_0 = const()[name = string("x_655_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_655_pad_0 = const()[name = string("x_655_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_655_dilations_0 = const()[name = string("x_655_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_15_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_15_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412669632)))];
            tensor<fp16, [1, 512, 304]> x_655_cast_fp16 = conv(dilations = x_655_dilations_0, groups = x_655_groups_0, pad = x_655_pad_0, pad_type = x_655_pad_type_0, strides = x_655_strides_0, weight = m_encoder_tp_encoders_15_self_attn_fsmn_block_weight_to_fp16, x = input_1967_cast_fp16)[name = string("x_655_cast_fp16")];
            tensor<int32, [3]> x_657_perm_0 = const()[name = string("x_657_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_657_cast_fp16 = transpose(perm = x_657_perm_0, x = x_655_cast_fp16)[name = string("transpose_377")];
            tensor<fp16, [1, 304, 512]> input_1969_cast_fp16 = add(x = x_657_cast_fp16, y = var_6918_cast_fp16_2)[name = string("input_1969_cast_fp16")];
            fp16 var_6948_to_fp16 = const()[name = string("op_6948_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_263_cast_fp16 = mul(x = var_6923_cast_fp16, y = var_6948_to_fp16)[name = string("q_h_263_cast_fp16")];
            bool scores_261_transpose_x_0 = const()[name = string("scores_261_transpose_x_0"), val = bool(false)];
            bool scores_261_transpose_y_0 = const()[name = string("scores_261_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_340_perm_0 = const()[name = string("transpose_340_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_341_perm_0 = const()[name = string("transpose_341_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_341 = transpose(perm = transpose_341_perm_0, x = var_6926_cast_fp16)[name = string("transpose_375")];
            tensor<fp16, [1, 4, 304, 128]> transpose_340 = transpose(perm = transpose_340_perm_0, x = q_h_263_cast_fp16)[name = string("transpose_376")];
            tensor<fp16, [1, 4, 304, 304]> scores_261_cast_fp16 = matmul(transpose_x = scores_261_transpose_x_0, transpose_y = scores_261_transpose_y_0, x = transpose_340, y = transpose_341)[name = string("scores_261_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_6956_cast_fp16 = softmax(axis = var_29, x = scores_261_cast_fp16)[name = string("op_6956_cast_fp16")];
            bool x_661_transpose_x_0 = const()[name = string("x_661_transpose_x_0"), val = bool(false)];
            bool x_661_transpose_y_0 = const()[name = string("x_661_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_131_cast_fp16 = transpose(perm = value_131_perm_0, x = var_6929_cast_fp16)[name = string("transpose_379")];
            tensor<fp16, [1, 4, 304, 128]> x_661_cast_fp16 = matmul(transpose_x = x_661_transpose_x_0, transpose_y = x_661_transpose_y_0, x = var_6956_cast_fp16, y = value_131_cast_fp16)[name = string("x_661_cast_fp16")];
            tensor<int32, [4]> var_6960_perm_0 = const()[name = string("op_6960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6962 = const()[name = string("op_6962"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_6960_cast_fp16 = transpose(perm = var_6960_perm_0, x = x_661_cast_fp16)[name = string("transpose_374")];
            tensor<fp16, [1, 304, 512]> input_1973_cast_fp16 = reshape(shape = var_6962, x = var_6960_cast_fp16)[name = string("input_1973_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_15_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_15_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412680960)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_15_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413205312)))];
            tensor<fp16, [1, 304, 512]> linear_261_cast_fp16 = linear(bias = m_encoder_tp_encoders_15_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_15_self_attn_linear_out_weight_to_fp16, x = input_1973_cast_fp16)[name = string("linear_261_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1975_cast_fp16 = add(x = linear_261_cast_fp16, y = input_1969_cast_fp16)[name = string("input_1975_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1977_cast_fp16 = add(x = input_1961_cast_fp16, y = input_1975_cast_fp16)[name = string("input_1977_cast_fp16")];
            tensor<int32, [1]> output_265_axes_0 = const()[name = string("output_265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_747_to_fp16 = const()[name = string("const_747_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413206400)))];
            tensor<fp16, [512]> const_748_to_fp16 = const()[name = string("const_748_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413207488)))];
            tensor<fp16, [1, 304, 512]> output_265_cast_fp16 = layer_norm(axes = output_265_axes_0, beta = const_748_to_fp16, epsilon = var_22_to_fp16, gamma = const_747_to_fp16, x = input_1977_cast_fp16)[name = string("output_265_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_15_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_15_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413208576)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_15_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_15_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415305792)))];
            tensor<fp16, [1, 304, 2048]> linear_262_cast_fp16 = linear(bias = m_encoder_tp_encoders_15_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_15_feed_forward_w_1_weight_to_fp16, x = output_265_cast_fp16)[name = string("linear_262_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_1985_cast_fp16 = relu(x = linear_262_cast_fp16)[name = string("input_1985_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_15_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_15_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415309952)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_15_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_15_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417407168)))];
            tensor<fp16, [1, 304, 512]> linear_263_cast_fp16 = linear(bias = m_encoder_tp_encoders_15_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_15_feed_forward_w_2_weight_to_fp16, x = input_1985_cast_fp16)[name = string("linear_263_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_1991_cast_fp16 = add(x = input_1977_cast_fp16, y = linear_263_cast_fp16)[name = string("input_1991_cast_fp16")];
            tensor<int32, [1]> output_267_axes_0 = const()[name = string("output_267_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_749_to_fp16 = const()[name = string("const_749_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417408256)))];
            tensor<fp16, [512]> const_750_to_fp16 = const()[name = string("const_750_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417409344)))];
            tensor<fp16, [1, 304, 512]> output_267_cast_fp16 = layer_norm(axes = output_267_axes_0, beta = const_750_to_fp16, epsilon = var_22_to_fp16, gamma = const_749_to_fp16, x = input_1991_cast_fp16)[name = string("output_267_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_16_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_16_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417410432)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_16_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_16_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418983360)))];
            tensor<fp16, [1, 304, 1536]> linear_264_cast_fp16 = linear(bias = m_encoder_tp_encoders_16_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_16_self_attn_linear_q_k_v_weight_to_fp16, x = output_267_cast_fp16)[name = string("linear_264_cast_fp16")];
            tensor<int32, [3]> tile_66 = const()[name = string("tile_66"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_7020_axis_0 = const()[name = string("op_7020_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_7020_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_7020_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_7020_cast_fp16_2 = split(axis = var_7020_axis_0, split_sizes = tile_66, x = linear_264_cast_fp16)[name = string("op_7020_cast_fp16")];
            tensor<int32, [4]> var_7024 = const()[name = string("op_7024"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7025_cast_fp16 = reshape(shape = var_7024, x = var_7020_cast_fp16_0)[name = string("op_7025_cast_fp16")];
            tensor<int32, [4]> var_7027 = const()[name = string("op_7027"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7028_cast_fp16 = reshape(shape = var_7027, x = var_7020_cast_fp16_1)[name = string("op_7028_cast_fp16")];
            tensor<int32, [4]> var_7030 = const()[name = string("op_7030"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7031_cast_fp16 = reshape(shape = var_7030, x = var_7020_cast_fp16_2)[name = string("op_7031_cast_fp16")];
            tensor<int32, [4]> value_133_perm_0 = const()[name = string("value_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_1995_perm_0 = const()[name = string("input_1995_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_1997_pad_0 = const()[name = string("input_1997_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_1997_mode_0 = const()[name = string("input_1997_mode_0"), val = string("constant")];
            fp16 const_756_to_fp16 = const()[name = string("const_756_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_1995_cast_fp16 = transpose(perm = input_1995_perm_0, x = var_7020_cast_fp16_2)[name = string("transpose_372")];
            tensor<fp16, [1, 512, 314]> input_1997_cast_fp16 = pad(constant_val = const_756_to_fp16, mode = input_1997_mode_0, pad = input_1997_pad_0, x = input_1995_cast_fp16)[name = string("input_1997_cast_fp16")];
            string x_665_pad_type_0 = const()[name = string("x_665_pad_type_0"), val = string("valid")];
            int32 x_665_groups_0 = const()[name = string("x_665_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_665_strides_0 = const()[name = string("x_665_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_665_pad_0 = const()[name = string("x_665_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_665_dilations_0 = const()[name = string("x_665_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_16_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_16_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418986496)))];
            tensor<fp16, [1, 512, 304]> x_665_cast_fp16 = conv(dilations = x_665_dilations_0, groups = x_665_groups_0, pad = x_665_pad_0, pad_type = x_665_pad_type_0, strides = x_665_strides_0, weight = m_encoder_tp_encoders_16_self_attn_fsmn_block_weight_to_fp16, x = input_1997_cast_fp16)[name = string("x_665_cast_fp16")];
            tensor<int32, [3]> x_667_perm_0 = const()[name = string("x_667_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_667_cast_fp16 = transpose(perm = x_667_perm_0, x = x_665_cast_fp16)[name = string("transpose_371")];
            tensor<fp16, [1, 304, 512]> input_1999_cast_fp16 = add(x = x_667_cast_fp16, y = var_7020_cast_fp16_2)[name = string("input_1999_cast_fp16")];
            fp16 var_7050_to_fp16 = const()[name = string("op_7050_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_267_cast_fp16 = mul(x = var_7025_cast_fp16, y = var_7050_to_fp16)[name = string("q_h_267_cast_fp16")];
            bool scores_265_transpose_x_0 = const()[name = string("scores_265_transpose_x_0"), val = bool(false)];
            bool scores_265_transpose_y_0 = const()[name = string("scores_265_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_342_perm_0 = const()[name = string("transpose_342_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_343_perm_0 = const()[name = string("transpose_343_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_343 = transpose(perm = transpose_343_perm_0, x = var_7028_cast_fp16)[name = string("transpose_369")];
            tensor<fp16, [1, 4, 304, 128]> transpose_342 = transpose(perm = transpose_342_perm_0, x = q_h_267_cast_fp16)[name = string("transpose_370")];
            tensor<fp16, [1, 4, 304, 304]> scores_265_cast_fp16 = matmul(transpose_x = scores_265_transpose_x_0, transpose_y = scores_265_transpose_y_0, x = transpose_342, y = transpose_343)[name = string("scores_265_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_7058_cast_fp16 = softmax(axis = var_29, x = scores_265_cast_fp16)[name = string("op_7058_cast_fp16")];
            bool x_671_transpose_x_0 = const()[name = string("x_671_transpose_x_0"), val = bool(false)];
            bool x_671_transpose_y_0 = const()[name = string("x_671_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_133_cast_fp16 = transpose(perm = value_133_perm_0, x = var_7031_cast_fp16)[name = string("transpose_373")];
            tensor<fp16, [1, 4, 304, 128]> x_671_cast_fp16 = matmul(transpose_x = x_671_transpose_x_0, transpose_y = x_671_transpose_y_0, x = var_7058_cast_fp16, y = value_133_cast_fp16)[name = string("x_671_cast_fp16")];
            tensor<int32, [4]> var_7062_perm_0 = const()[name = string("op_7062_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7064 = const()[name = string("op_7064"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_7062_cast_fp16 = transpose(perm = var_7062_perm_0, x = x_671_cast_fp16)[name = string("transpose_368")];
            tensor<fp16, [1, 304, 512]> input_2003_cast_fp16 = reshape(shape = var_7064, x = var_7062_cast_fp16)[name = string("input_2003_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_16_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_16_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418997824)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_16_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419522176)))];
            tensor<fp16, [1, 304, 512]> linear_265_cast_fp16 = linear(bias = m_encoder_tp_encoders_16_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_16_self_attn_linear_out_weight_to_fp16, x = input_2003_cast_fp16)[name = string("linear_265_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2005_cast_fp16 = add(x = linear_265_cast_fp16, y = input_1999_cast_fp16)[name = string("input_2005_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2007_cast_fp16 = add(x = input_1991_cast_fp16, y = input_2005_cast_fp16)[name = string("input_2007_cast_fp16")];
            tensor<int32, [1]> output_269_axes_0 = const()[name = string("output_269_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_758_to_fp16 = const()[name = string("const_758_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419523264)))];
            tensor<fp16, [512]> const_759_to_fp16 = const()[name = string("const_759_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419524352)))];
            tensor<fp16, [1, 304, 512]> output_269_cast_fp16 = layer_norm(axes = output_269_axes_0, beta = const_759_to_fp16, epsilon = var_22_to_fp16, gamma = const_758_to_fp16, x = input_2007_cast_fp16)[name = string("output_269_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_16_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_16_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419525440)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_16_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_16_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421622656)))];
            tensor<fp16, [1, 304, 2048]> linear_266_cast_fp16 = linear(bias = m_encoder_tp_encoders_16_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_16_feed_forward_w_1_weight_to_fp16, x = output_269_cast_fp16)[name = string("linear_266_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_2015_cast_fp16 = relu(x = linear_266_cast_fp16)[name = string("input_2015_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_16_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_16_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421626816)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_16_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_16_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423724032)))];
            tensor<fp16, [1, 304, 512]> linear_267_cast_fp16 = linear(bias = m_encoder_tp_encoders_16_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_16_feed_forward_w_2_weight_to_fp16, x = input_2015_cast_fp16)[name = string("linear_267_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2021_cast_fp16 = add(x = input_2007_cast_fp16, y = linear_267_cast_fp16)[name = string("input_2021_cast_fp16")];
            tensor<int32, [1]> output_271_axes_0 = const()[name = string("output_271_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_760_to_fp16 = const()[name = string("const_760_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423725120)))];
            tensor<fp16, [512]> const_761_to_fp16 = const()[name = string("const_761_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423726208)))];
            tensor<fp16, [1, 304, 512]> output_271_cast_fp16 = layer_norm(axes = output_271_axes_0, beta = const_761_to_fp16, epsilon = var_22_to_fp16, gamma = const_760_to_fp16, x = input_2021_cast_fp16)[name = string("output_271_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_17_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_17_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423727296)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_17_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_17_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425300224)))];
            tensor<fp16, [1, 304, 1536]> linear_268_cast_fp16 = linear(bias = m_encoder_tp_encoders_17_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_17_self_attn_linear_q_k_v_weight_to_fp16, x = output_271_cast_fp16)[name = string("linear_268_cast_fp16")];
            tensor<int32, [3]> tile_67 = const()[name = string("tile_67"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_7122_axis_0 = const()[name = string("op_7122_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_7122_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_7122_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_7122_cast_fp16_2 = split(axis = var_7122_axis_0, split_sizes = tile_67, x = linear_268_cast_fp16)[name = string("op_7122_cast_fp16")];
            tensor<int32, [4]> var_7126 = const()[name = string("op_7126"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7127_cast_fp16 = reshape(shape = var_7126, x = var_7122_cast_fp16_0)[name = string("op_7127_cast_fp16")];
            tensor<int32, [4]> var_7129 = const()[name = string("op_7129"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7130_cast_fp16 = reshape(shape = var_7129, x = var_7122_cast_fp16_1)[name = string("op_7130_cast_fp16")];
            tensor<int32, [4]> var_7132 = const()[name = string("op_7132"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7133_cast_fp16 = reshape(shape = var_7132, x = var_7122_cast_fp16_2)[name = string("op_7133_cast_fp16")];
            tensor<int32, [4]> value_135_perm_0 = const()[name = string("value_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_2025_perm_0 = const()[name = string("input_2025_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_2027_pad_0 = const()[name = string("input_2027_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_2027_mode_0 = const()[name = string("input_2027_mode_0"), val = string("constant")];
            fp16 const_767_to_fp16 = const()[name = string("const_767_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_2025_cast_fp16 = transpose(perm = input_2025_perm_0, x = var_7122_cast_fp16_2)[name = string("transpose_366")];
            tensor<fp16, [1, 512, 314]> input_2027_cast_fp16 = pad(constant_val = const_767_to_fp16, mode = input_2027_mode_0, pad = input_2027_pad_0, x = input_2025_cast_fp16)[name = string("input_2027_cast_fp16")];
            string x_675_pad_type_0 = const()[name = string("x_675_pad_type_0"), val = string("valid")];
            int32 x_675_groups_0 = const()[name = string("x_675_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_675_strides_0 = const()[name = string("x_675_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_675_pad_0 = const()[name = string("x_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_675_dilations_0 = const()[name = string("x_675_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_17_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_17_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425303360)))];
            tensor<fp16, [1, 512, 304]> x_675_cast_fp16 = conv(dilations = x_675_dilations_0, groups = x_675_groups_0, pad = x_675_pad_0, pad_type = x_675_pad_type_0, strides = x_675_strides_0, weight = m_encoder_tp_encoders_17_self_attn_fsmn_block_weight_to_fp16, x = input_2027_cast_fp16)[name = string("x_675_cast_fp16")];
            tensor<int32, [3]> x_677_perm_0 = const()[name = string("x_677_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_677_cast_fp16 = transpose(perm = x_677_perm_0, x = x_675_cast_fp16)[name = string("transpose_365")];
            tensor<fp16, [1, 304, 512]> input_2029_cast_fp16 = add(x = x_677_cast_fp16, y = var_7122_cast_fp16_2)[name = string("input_2029_cast_fp16")];
            fp16 var_7152_to_fp16 = const()[name = string("op_7152_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_271_cast_fp16 = mul(x = var_7127_cast_fp16, y = var_7152_to_fp16)[name = string("q_h_271_cast_fp16")];
            bool scores_269_transpose_x_0 = const()[name = string("scores_269_transpose_x_0"), val = bool(false)];
            bool scores_269_transpose_y_0 = const()[name = string("scores_269_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_344_perm_0 = const()[name = string("transpose_344_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_345_perm_0 = const()[name = string("transpose_345_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_345 = transpose(perm = transpose_345_perm_0, x = var_7130_cast_fp16)[name = string("transpose_363")];
            tensor<fp16, [1, 4, 304, 128]> transpose_344 = transpose(perm = transpose_344_perm_0, x = q_h_271_cast_fp16)[name = string("transpose_364")];
            tensor<fp16, [1, 4, 304, 304]> scores_269_cast_fp16 = matmul(transpose_x = scores_269_transpose_x_0, transpose_y = scores_269_transpose_y_0, x = transpose_344, y = transpose_345)[name = string("scores_269_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_7160_cast_fp16 = softmax(axis = var_29, x = scores_269_cast_fp16)[name = string("op_7160_cast_fp16")];
            bool x_681_transpose_x_0 = const()[name = string("x_681_transpose_x_0"), val = bool(false)];
            bool x_681_transpose_y_0 = const()[name = string("x_681_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_135_cast_fp16 = transpose(perm = value_135_perm_0, x = var_7133_cast_fp16)[name = string("transpose_367")];
            tensor<fp16, [1, 4, 304, 128]> x_681_cast_fp16 = matmul(transpose_x = x_681_transpose_x_0, transpose_y = x_681_transpose_y_0, x = var_7160_cast_fp16, y = value_135_cast_fp16)[name = string("x_681_cast_fp16")];
            tensor<int32, [4]> var_7164_perm_0 = const()[name = string("op_7164_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7166 = const()[name = string("op_7166"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_7164_cast_fp16 = transpose(perm = var_7164_perm_0, x = x_681_cast_fp16)[name = string("transpose_362")];
            tensor<fp16, [1, 304, 512]> input_2033_cast_fp16 = reshape(shape = var_7166, x = var_7164_cast_fp16)[name = string("input_2033_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_17_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_17_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425314688)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_17_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425839040)))];
            tensor<fp16, [1, 304, 512]> linear_269_cast_fp16 = linear(bias = m_encoder_tp_encoders_17_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_17_self_attn_linear_out_weight_to_fp16, x = input_2033_cast_fp16)[name = string("linear_269_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2035_cast_fp16 = add(x = linear_269_cast_fp16, y = input_2029_cast_fp16)[name = string("input_2035_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2037_cast_fp16 = add(x = input_2021_cast_fp16, y = input_2035_cast_fp16)[name = string("input_2037_cast_fp16")];
            tensor<int32, [1]> output_273_axes_0 = const()[name = string("output_273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_769_to_fp16 = const()[name = string("const_769_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425840128)))];
            tensor<fp16, [512]> const_770_to_fp16 = const()[name = string("const_770_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425841216)))];
            tensor<fp16, [1, 304, 512]> output_273_cast_fp16 = layer_norm(axes = output_273_axes_0, beta = const_770_to_fp16, epsilon = var_22_to_fp16, gamma = const_769_to_fp16, x = input_2037_cast_fp16)[name = string("output_273_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_17_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_17_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425842304)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_17_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_17_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427939520)))];
            tensor<fp16, [1, 304, 2048]> linear_270_cast_fp16 = linear(bias = m_encoder_tp_encoders_17_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_17_feed_forward_w_1_weight_to_fp16, x = output_273_cast_fp16)[name = string("linear_270_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_2045_cast_fp16 = relu(x = linear_270_cast_fp16)[name = string("input_2045_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_17_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_17_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427943680)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_17_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_17_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430040896)))];
            tensor<fp16, [1, 304, 512]> linear_271_cast_fp16 = linear(bias = m_encoder_tp_encoders_17_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_17_feed_forward_w_2_weight_to_fp16, x = input_2045_cast_fp16)[name = string("linear_271_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2051_cast_fp16 = add(x = input_2037_cast_fp16, y = linear_271_cast_fp16)[name = string("input_2051_cast_fp16")];
            tensor<int32, [1]> output_275_axes_0 = const()[name = string("output_275_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_771_to_fp16 = const()[name = string("const_771_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430041984)))];
            tensor<fp16, [512]> const_772_to_fp16 = const()[name = string("const_772_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430043072)))];
            tensor<fp16, [1, 304, 512]> output_275_cast_fp16 = layer_norm(axes = output_275_axes_0, beta = const_772_to_fp16, epsilon = var_22_to_fp16, gamma = const_771_to_fp16, x = input_2051_cast_fp16)[name = string("output_275_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_18_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_18_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430044160)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_18_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_18_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431617088)))];
            tensor<fp16, [1, 304, 1536]> linear_272_cast_fp16 = linear(bias = m_encoder_tp_encoders_18_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_18_self_attn_linear_q_k_v_weight_to_fp16, x = output_275_cast_fp16)[name = string("linear_272_cast_fp16")];
            tensor<int32, [3]> tile_68 = const()[name = string("tile_68"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_7224_axis_0 = const()[name = string("op_7224_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_7224_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_7224_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_7224_cast_fp16_2 = split(axis = var_7224_axis_0, split_sizes = tile_68, x = linear_272_cast_fp16)[name = string("op_7224_cast_fp16")];
            tensor<int32, [4]> var_7228 = const()[name = string("op_7228"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7229_cast_fp16 = reshape(shape = var_7228, x = var_7224_cast_fp16_0)[name = string("op_7229_cast_fp16")];
            tensor<int32, [4]> var_7231 = const()[name = string("op_7231"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7232_cast_fp16 = reshape(shape = var_7231, x = var_7224_cast_fp16_1)[name = string("op_7232_cast_fp16")];
            tensor<int32, [4]> var_7234 = const()[name = string("op_7234"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7235_cast_fp16 = reshape(shape = var_7234, x = var_7224_cast_fp16_2)[name = string("op_7235_cast_fp16")];
            tensor<int32, [4]> value_137_perm_0 = const()[name = string("value_137_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_2055_perm_0 = const()[name = string("input_2055_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_2057_pad_0 = const()[name = string("input_2057_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_2057_mode_0 = const()[name = string("input_2057_mode_0"), val = string("constant")];
            fp16 const_778_to_fp16 = const()[name = string("const_778_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_2055_cast_fp16 = transpose(perm = input_2055_perm_0, x = var_7224_cast_fp16_2)[name = string("transpose_360")];
            tensor<fp16, [1, 512, 314]> input_2057_cast_fp16 = pad(constant_val = const_778_to_fp16, mode = input_2057_mode_0, pad = input_2057_pad_0, x = input_2055_cast_fp16)[name = string("input_2057_cast_fp16")];
            string x_685_pad_type_0 = const()[name = string("x_685_pad_type_0"), val = string("valid")];
            int32 x_685_groups_0 = const()[name = string("x_685_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_685_strides_0 = const()[name = string("x_685_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_685_pad_0 = const()[name = string("x_685_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_685_dilations_0 = const()[name = string("x_685_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_18_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_18_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431620224)))];
            tensor<fp16, [1, 512, 304]> x_685_cast_fp16 = conv(dilations = x_685_dilations_0, groups = x_685_groups_0, pad = x_685_pad_0, pad_type = x_685_pad_type_0, strides = x_685_strides_0, weight = m_encoder_tp_encoders_18_self_attn_fsmn_block_weight_to_fp16, x = input_2057_cast_fp16)[name = string("x_685_cast_fp16")];
            tensor<int32, [3]> x_687_perm_0 = const()[name = string("x_687_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_687_cast_fp16 = transpose(perm = x_687_perm_0, x = x_685_cast_fp16)[name = string("transpose_359")];
            tensor<fp16, [1, 304, 512]> input_2059_cast_fp16 = add(x = x_687_cast_fp16, y = var_7224_cast_fp16_2)[name = string("input_2059_cast_fp16")];
            fp16 var_7254_to_fp16 = const()[name = string("op_7254_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_275_cast_fp16 = mul(x = var_7229_cast_fp16, y = var_7254_to_fp16)[name = string("q_h_275_cast_fp16")];
            bool scores_273_transpose_x_0 = const()[name = string("scores_273_transpose_x_0"), val = bool(false)];
            bool scores_273_transpose_y_0 = const()[name = string("scores_273_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_346_perm_0 = const()[name = string("transpose_346_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_347_perm_0 = const()[name = string("transpose_347_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_347 = transpose(perm = transpose_347_perm_0, x = var_7232_cast_fp16)[name = string("transpose_357")];
            tensor<fp16, [1, 4, 304, 128]> transpose_346 = transpose(perm = transpose_346_perm_0, x = q_h_275_cast_fp16)[name = string("transpose_358")];
            tensor<fp16, [1, 4, 304, 304]> scores_273_cast_fp16 = matmul(transpose_x = scores_273_transpose_x_0, transpose_y = scores_273_transpose_y_0, x = transpose_346, y = transpose_347)[name = string("scores_273_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_7262_cast_fp16 = softmax(axis = var_29, x = scores_273_cast_fp16)[name = string("op_7262_cast_fp16")];
            bool x_691_transpose_x_0 = const()[name = string("x_691_transpose_x_0"), val = bool(false)];
            bool x_691_transpose_y_0 = const()[name = string("x_691_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_137_cast_fp16 = transpose(perm = value_137_perm_0, x = var_7235_cast_fp16)[name = string("transpose_361")];
            tensor<fp16, [1, 4, 304, 128]> x_691_cast_fp16 = matmul(transpose_x = x_691_transpose_x_0, transpose_y = x_691_transpose_y_0, x = var_7262_cast_fp16, y = value_137_cast_fp16)[name = string("x_691_cast_fp16")];
            tensor<int32, [4]> var_7266_perm_0 = const()[name = string("op_7266_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7268 = const()[name = string("op_7268"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_7266_cast_fp16 = transpose(perm = var_7266_perm_0, x = x_691_cast_fp16)[name = string("transpose_356")];
            tensor<fp16, [1, 304, 512]> input_2063_cast_fp16 = reshape(shape = var_7268, x = var_7266_cast_fp16)[name = string("input_2063_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_18_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_18_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431631552)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_18_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432155904)))];
            tensor<fp16, [1, 304, 512]> linear_273_cast_fp16 = linear(bias = m_encoder_tp_encoders_18_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_18_self_attn_linear_out_weight_to_fp16, x = input_2063_cast_fp16)[name = string("linear_273_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2065_cast_fp16 = add(x = linear_273_cast_fp16, y = input_2059_cast_fp16)[name = string("input_2065_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2067_cast_fp16 = add(x = input_2051_cast_fp16, y = input_2065_cast_fp16)[name = string("input_2067_cast_fp16")];
            tensor<int32, [1]> output_277_axes_0 = const()[name = string("output_277_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_780_to_fp16 = const()[name = string("const_780_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432156992)))];
            tensor<fp16, [512]> const_781_to_fp16 = const()[name = string("const_781_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432158080)))];
            tensor<fp16, [1, 304, 512]> output_277_cast_fp16 = layer_norm(axes = output_277_axes_0, beta = const_781_to_fp16, epsilon = var_22_to_fp16, gamma = const_780_to_fp16, x = input_2067_cast_fp16)[name = string("output_277_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_18_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_18_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432159168)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_18_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_18_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434256384)))];
            tensor<fp16, [1, 304, 2048]> linear_274_cast_fp16 = linear(bias = m_encoder_tp_encoders_18_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_18_feed_forward_w_1_weight_to_fp16, x = output_277_cast_fp16)[name = string("linear_274_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_2075_cast_fp16 = relu(x = linear_274_cast_fp16)[name = string("input_2075_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_18_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_18_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434260544)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_18_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_18_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436357760)))];
            tensor<fp16, [1, 304, 512]> linear_275_cast_fp16 = linear(bias = m_encoder_tp_encoders_18_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_18_feed_forward_w_2_weight_to_fp16, x = input_2075_cast_fp16)[name = string("linear_275_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2081_cast_fp16 = add(x = input_2067_cast_fp16, y = linear_275_cast_fp16)[name = string("input_2081_cast_fp16")];
            tensor<int32, [1]> output_279_axes_0 = const()[name = string("output_279_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_782_to_fp16 = const()[name = string("const_782_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436358848)))];
            tensor<fp16, [512]> const_783_to_fp16 = const()[name = string("const_783_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436359936)))];
            tensor<fp16, [1, 304, 512]> output_279_cast_fp16 = layer_norm(axes = output_279_axes_0, beta = const_783_to_fp16, epsilon = var_22_to_fp16, gamma = const_782_to_fp16, x = input_2081_cast_fp16)[name = string("output_279_cast_fp16")];
            tensor<fp16, [1536, 512]> m_encoder_tp_encoders_19_self_attn_linear_q_k_v_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_19_self_attn_linear_q_k_v_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436361024)))];
            tensor<fp16, [1536]> m_encoder_tp_encoders_19_self_attn_linear_q_k_v_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_19_self_attn_linear_q_k_v_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437933952)))];
            tensor<fp16, [1, 304, 1536]> linear_276_cast_fp16 = linear(bias = m_encoder_tp_encoders_19_self_attn_linear_q_k_v_bias_to_fp16, weight = m_encoder_tp_encoders_19_self_attn_linear_q_k_v_weight_to_fp16, x = output_279_cast_fp16)[name = string("linear_276_cast_fp16")];
            tensor<int32, [3]> tile_69 = const()[name = string("tile_69"), val = tensor<int32, [3]>([512, 512, 512])];
            int32 var_7326_axis_0 = const()[name = string("op_7326_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 304, 512]> var_7326_cast_fp16_0, tensor<fp16, [1, 304, 512]> var_7326_cast_fp16_1, tensor<fp16, [1, 304, 512]> var_7326_cast_fp16_2 = split(axis = var_7326_axis_0, split_sizes = tile_69, x = linear_276_cast_fp16)[name = string("op_7326_cast_fp16")];
            tensor<int32, [4]> var_7330 = const()[name = string("op_7330"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7331_cast_fp16 = reshape(shape = var_7330, x = var_7326_cast_fp16_0)[name = string("op_7331_cast_fp16")];
            tensor<int32, [4]> var_7333 = const()[name = string("op_7333"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7334_cast_fp16 = reshape(shape = var_7333, x = var_7326_cast_fp16_1)[name = string("op_7334_cast_fp16")];
            tensor<int32, [4]> var_7336 = const()[name = string("op_7336"), val = tensor<int32, [4]>([1, 304, 4, 128])];
            tensor<fp16, [1, 304, 4, 128]> var_7337_cast_fp16 = reshape(shape = var_7336, x = var_7326_cast_fp16_2)[name = string("op_7337_cast_fp16")];
            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> input_2085_perm_0 = const()[name = string("input_2085_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> input_2087_pad_0 = const()[name = string("input_2087_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 5])];
            string input_2087_mode_0 = const()[name = string("input_2087_mode_0"), val = string("constant")];
            fp16 const_789_to_fp16 = const()[name = string("const_789_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 512, 304]> input_2085_cast_fp16 = transpose(perm = input_2085_perm_0, x = var_7326_cast_fp16_2)[name = string("transpose_354")];
            tensor<fp16, [1, 512, 314]> input_2087_cast_fp16 = pad(constant_val = const_789_to_fp16, mode = input_2087_mode_0, pad = input_2087_pad_0, x = input_2085_cast_fp16)[name = string("input_2087_cast_fp16")];
            string x_695_pad_type_0 = const()[name = string("x_695_pad_type_0"), val = string("valid")];
            int32 x_695_groups_0 = const()[name = string("x_695_groups_0"), val = int32(512)];
            tensor<int32, [1]> x_695_strides_0 = const()[name = string("x_695_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_695_pad_0 = const()[name = string("x_695_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_695_dilations_0 = const()[name = string("x_695_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 11]> m_encoder_tp_encoders_19_self_attn_fsmn_block_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_19_self_attn_fsmn_block_weight_to_fp16"), val = tensor<fp16, [512, 1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437937088)))];
            tensor<fp16, [1, 512, 304]> x_695_cast_fp16 = conv(dilations = x_695_dilations_0, groups = x_695_groups_0, pad = x_695_pad_0, pad_type = x_695_pad_type_0, strides = x_695_strides_0, weight = m_encoder_tp_encoders_19_self_attn_fsmn_block_weight_to_fp16, x = input_2087_cast_fp16)[name = string("x_695_cast_fp16")];
            tensor<int32, [3]> x_697_perm_0 = const()[name = string("x_697_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 304, 512]> x_697_cast_fp16 = transpose(perm = x_697_perm_0, x = x_695_cast_fp16)[name = string("transpose_353")];
            tensor<fp16, [1, 304, 512]> input_2089_cast_fp16 = add(x = x_697_cast_fp16, y = var_7326_cast_fp16_2)[name = string("input_2089_cast_fp16")];
            fp16 var_7356_to_fp16 = const()[name = string("op_7356_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 304, 4, 128]> q_h_cast_fp16 = mul(x = var_7331_cast_fp16, y = var_7356_to_fp16)[name = string("q_h_cast_fp16")];
            bool scores_277_transpose_x_0 = const()[name = string("scores_277_transpose_x_0"), val = bool(false)];
            bool scores_277_transpose_y_0 = const()[name = string("scores_277_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_348_perm_0 = const()[name = string("transpose_348_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_349_perm_0 = const()[name = string("transpose_349_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 128, 304]> transpose_349 = transpose(perm = transpose_349_perm_0, x = var_7334_cast_fp16)[name = string("transpose_351")];
            tensor<fp16, [1, 4, 304, 128]> transpose_348 = transpose(perm = transpose_348_perm_0, x = q_h_cast_fp16)[name = string("transpose_352")];
            tensor<fp16, [1, 4, 304, 304]> scores_277_cast_fp16 = matmul(transpose_x = scores_277_transpose_x_0, transpose_y = scores_277_transpose_y_0, x = transpose_348, y = transpose_349)[name = string("scores_277_cast_fp16")];
            tensor<fp16, [1, 4, 304, 304]> var_7364_cast_fp16 = softmax(axis = var_29, x = scores_277_cast_fp16)[name = string("op_7364_cast_fp16")];
            bool x_transpose_x_0 = const()[name = string("x_transpose_x_0"), val = bool(false)];
            bool x_transpose_y_0 = const()[name = string("x_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 304, 128]> value_cast_fp16 = transpose(perm = value_perm_0, x = var_7337_cast_fp16)[name = string("transpose_355")];
            tensor<fp16, [1, 4, 304, 128]> x_cast_fp16 = matmul(transpose_x = x_transpose_x_0, transpose_y = x_transpose_y_0, x = var_7364_cast_fp16, y = value_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [4]> var_7368_perm_0 = const()[name = string("op_7368_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7370 = const()[name = string("op_7370"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 304, 4, 128]> var_7368_cast_fp16 = transpose(perm = var_7368_perm_0, x = x_cast_fp16)[name = string("transpose_350")];
            tensor<fp16, [1, 304, 512]> input_2093_cast_fp16 = reshape(shape = var_7370, x = var_7368_cast_fp16)[name = string("input_2093_cast_fp16")];
            tensor<fp16, [512, 512]> m_encoder_tp_encoders_19_self_attn_linear_out_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_19_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437948416)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_19_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(438472768)))];
            tensor<fp16, [1, 304, 512]> linear_277_cast_fp16 = linear(bias = m_encoder_tp_encoders_19_self_attn_linear_out_bias_to_fp16, weight = m_encoder_tp_encoders_19_self_attn_linear_out_weight_to_fp16, x = input_2093_cast_fp16)[name = string("linear_277_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2095_cast_fp16 = add(x = linear_277_cast_fp16, y = input_2089_cast_fp16)[name = string("input_2095_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2097_cast_fp16 = add(x = input_2081_cast_fp16, y = input_2095_cast_fp16)[name = string("input_2097_cast_fp16")];
            tensor<int32, [1]> output_281_axes_0 = const()[name = string("output_281_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_791_to_fp16 = const()[name = string("const_791_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(438473856)))];
            tensor<fp16, [512]> const_792_to_fp16 = const()[name = string("const_792_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(438474944)))];
            tensor<fp16, [1, 304, 512]> output_281_cast_fp16 = layer_norm(axes = output_281_axes_0, beta = const_792_to_fp16, epsilon = var_22_to_fp16, gamma = const_791_to_fp16, x = input_2097_cast_fp16)[name = string("output_281_cast_fp16")];
            tensor<fp16, [2048, 512]> m_encoder_tp_encoders_19_feed_forward_w_1_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_19_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(438476032)))];
            tensor<fp16, [2048]> m_encoder_tp_encoders_19_feed_forward_w_1_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_19_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440573248)))];
            tensor<fp16, [1, 304, 2048]> linear_278_cast_fp16 = linear(bias = m_encoder_tp_encoders_19_feed_forward_w_1_bias_to_fp16, weight = m_encoder_tp_encoders_19_feed_forward_w_1_weight_to_fp16, x = output_281_cast_fp16)[name = string("linear_278_cast_fp16")];
            tensor<fp16, [1, 304, 2048]> input_2105_cast_fp16 = relu(x = linear_278_cast_fp16)[name = string("input_2105_cast_fp16")];
            tensor<fp16, [512, 2048]> m_encoder_tp_encoders_19_feed_forward_w_2_weight_to_fp16 = const()[name = string("m_encoder_tp_encoders_19_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440577408)))];
            tensor<fp16, [512]> m_encoder_tp_encoders_19_feed_forward_w_2_bias_to_fp16 = const()[name = string("m_encoder_tp_encoders_19_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442674624)))];
            tensor<fp16, [1, 304, 512]> linear_279_cast_fp16 = linear(bias = m_encoder_tp_encoders_19_feed_forward_w_2_bias_to_fp16, weight = m_encoder_tp_encoders_19_feed_forward_w_2_weight_to_fp16, x = input_2105_cast_fp16)[name = string("linear_279_cast_fp16")];
            tensor<fp16, [1, 304, 512]> input_2111_cast_fp16 = add(x = input_2097_cast_fp16, y = linear_279_cast_fp16)[name = string("input_2111_cast_fp16")];
            tensor<int32, [1]> output_axes_0 = const()[name = string("output_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> const_793_to_fp16 = const()[name = string("const_793_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442675712)))];
            tensor<fp16, [512]> const_794_to_fp16 = const()[name = string("const_794_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442676800)))];
            tensor<fp16, [1, 304, 512]> output_cast_fp16 = layer_norm(axes = output_axes_0, beta = const_794_to_fp16, epsilon = var_22_to_fp16, gamma = const_793_to_fp16, x = input_2111_cast_fp16)[name = string("output_cast_fp16")];
            tensor<fp16, [25055, 512]> m_ctc_ctc_lo_weight_to_fp16 = const()[name = string("m_ctc_ctc_lo_weight_to_fp16"), val = tensor<fp16, [25055, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442677888)))];
            tensor<fp16, [25055]> m_ctc_ctc_lo_bias_to_fp16 = const()[name = string("m_ctc_ctc_lo_bias_to_fp16"), val = tensor<fp16, [25055]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468334272)))];
            tensor<fp16, [1, 304, 25055]> ctc_logits = linear(bias = m_ctc_ctc_lo_bias_to_fp16, weight = m_ctc_ctc_lo_weight_to_fp16, x = output_cast_fp16)[name = string("linear_280_cast_fp16")];
        } -> (ctc_logits, encoder_out_lens);
}