program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.2.1"}, {"coremlc-version", "3520.2.1"}, {"mldb_token", "mldb-ndg4j6nglq"}})]
{
    func main<ios16>(tensor<fp16, [198, 40]> input_wav) {
            tensor<int32, [1]> feats_axes_0 = const()[name = tensor<string, []>("feats_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 198, 40]> feats_cast_fp16 = expand_dims(axes = feats_axes_0, x = input_wav)[name = tensor<string, []>("feats_cast_fp16")];
            tensor<int32, []> var_5 = const()[name = tensor<string, []>("op_5"), val = tensor<int32, []>(-1)];
            tensor<int32, []> var_13 = const()[name = tensor<string, []>("op_13"), val = tensor<int32, []>(144)];
            tensor<int32, []> var_15 = const()[name = tensor<string, []>("op_15"), val = tensor<int32, []>(1)];
            tensor<int32, [1]> input_1_axes_0 = const()[name = tensor<string, []>("input_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 198, 40]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = feats_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<int32, [2]> var_41 = const()[name = tensor<string, []>("op_41"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_43 = const()[name = tensor<string, []>("op_43"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_3_pad_type_0 = const()[name = tensor<string, []>("input_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [144, 1, 3, 3]> encoder_conv_subsample_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_conv_subsample_sequential_0_weight_to_fp16"), val = tensor<fp16, [144, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [144]> encoder_conv_subsample_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_conv_subsample_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2752)))];
            tensor<fp16, [1, 144, 98, 19]> input_3_cast_fp16 = conv(bias = encoder_conv_subsample_sequential_0_bias_to_fp16, dilations = var_43, groups = var_15, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = var_41, weight = encoder_conv_subsample_sequential_0_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<fp16, [1, 144, 98, 19]> input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<int32, [2]> var_49 = const()[name = tensor<string, []>("op_49"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_51 = const()[name = tensor<string, []>("op_51"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [144, 144, 3, 3]> encoder_conv_subsample_sequential_2_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_conv_subsample_sequential_2_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3136))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190080))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189824)))];
            tensor<fp16, [144]> encoder_conv_subsample_sequential_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_conv_subsample_sequential_2_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190464)))];
            tensor<fp16, [1, 144, 48, 9]> input_7_cast_fp16 = conv(bias = encoder_conv_subsample_sequential_2_bias_to_fp16, dilations = var_51, groups = var_15, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_49, weight = encoder_conv_subsample_sequential_2_weight_to_fp16_affine_quantized, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<fp16, [1, 144, 48, 9]> outputs_1_cast_fp16 = relu(x = input_7_cast_fp16)[name = tensor<string, []>("outputs_1_cast_fp16")];
            tensor<int32, [4]> var_61 = const()[name = tensor<string, []>("op_61"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_66 = const()[name = tensor<string, []>("op_66"), val = tensor<int32, [3]>([1, 48, 1296])];
            tensor<fp16, [1, 48, 144, 9]> transpose_78 = transpose(perm = var_61, x = outputs_1_cast_fp16)[name = tensor<string, []>("transpose_78")];
            tensor<fp16, [1, 48, 1296]> input_9_cast_fp16 = reshape(shape = var_66, x = transpose_78)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<fp16, [144, 1296]> encoder_input_projection_0_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_input_projection_0_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 1296]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190848))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377792))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377536)))];
            tensor<fp16, [144]> encoder_input_projection_0_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_input_projection_0_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378176)))];
            tensor<fp16, [1, 48, 144]> linear_0_cast_fp16 = linear(bias = encoder_input_projection_0_linear_bias_to_fp16, weight = encoder_input_projection_0_linear_weight_to_fp16_affine_quantized, x = input_9_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<int32, [1]> input_15_axes_0 = const()[name = tensor<string, []>("input_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_0_sequential_0_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_0_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378560)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_0_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_0_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378944)))];
            tensor<fp16, []> var_12_to_fp16 = const()[name = tensor<string, []>("op_12_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 48, 144]> input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = encoder_layers_0_sequential_0_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_0_sequential_0_module_sequential_0_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_0_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379328))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462976))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462336)))];
            tensor<fp16, [576]> encoder_layers_0_sequential_0_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_0_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464192)))];
            tensor<fp16, [1, 48, 576]> linear_1_cast_fp16 = linear(bias = encoder_layers_0_sequential_0_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_0_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_15_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_17_cast_fp16 = silu(x = linear_1_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_0_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465408))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548672))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548416)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_0_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_0_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549056)))];
            tensor<fp16, [1, 48, 144]> linear_2_cast_fp16 = linear(bias = encoder_layers_0_sequential_0_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_0_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_17_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<fp16, []> var_101_to_fp16 = const()[name = tensor<string, []>("op_101_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_102_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_101_to_fp16)[name = tensor<string, []>("op_102_cast_fp16")];
            tensor<fp16, [1, 48, 144]> inputs_3_cast_fp16 = add(x = var_102_cast_fp16, y = linear_0_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
            tensor<int32, [1]> query_1_axes_0 = const()[name = tensor<string, []>("query_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_0_sequential_1_module_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_1_module_layer_norm_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549440)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_1_module_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_1_module_layer_norm_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549824)))];
            tensor<fp16, [1, 48, 144]> query_1_cast_fp16 = layer_norm(axes = query_1_axes_0, beta = encoder_layers_0_sequential_1_module_layer_norm_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_0_sequential_1_module_layer_norm_weight_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_0_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550208))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571264))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571008)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_1_module_attention_query_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_query_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571648)))];
            tensor<fp16, [1, 48, 144]> linear_3_cast_fp16 = linear(bias = encoder_layers_0_sequential_1_module_attention_query_proj_linear_bias_to_fp16, weight = encoder_layers_0_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized, x = query_1_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<int32, [4]> var_133 = const()[name = tensor<string, []>("op_133"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> query_3_cast_fp16 = reshape(shape = var_133, x = linear_3_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_0_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(572032))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(593088))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(592832)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_1_module_attention_key_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_key_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(593472)))];
            tensor<fp16, [1, 48, 144]> linear_4_cast_fp16 = linear(bias = encoder_layers_0_sequential_1_module_attention_key_proj_linear_bias_to_fp16, weight = encoder_layers_0_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized, x = query_1_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<int32, [4]> var_139 = const()[name = tensor<string, []>("op_139"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_140_cast_fp16 = reshape(shape = var_139, x = linear_4_cast_fp16)[name = tensor<string, []>("op_140_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_0_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(593856))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(614912))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(614656)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_1_module_attention_value_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_value_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(615296)))];
            tensor<fp16, [1, 48, 144]> linear_5_cast_fp16 = linear(bias = encoder_layers_0_sequential_1_module_attention_value_proj_linear_bias_to_fp16, weight = encoder_layers_0_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized, x = query_1_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<int32, [4]> var_147 = const()[name = tensor<string, []>("op_147"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_148_cast_fp16 = reshape(shape = var_147, x = linear_5_cast_fp16)[name = tensor<string, []>("op_148_cast_fp16")];
            tensor<int32, [4]> var_149 = const()[name = tensor<string, []>("op_149"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [4, 36]> encoder_layers_0_sequential_1_module_attention_u_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_u_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(615680)))];
            tensor<fp16, [1, 48, 4, 36]> var_156_cast_fp16 = add(x = query_3_cast_fp16, y = encoder_layers_0_sequential_1_module_attention_u_bias_to_fp16)[name = tensor<string, []>("op_156_cast_fp16")];
            tensor<bool, []> content_score_1_transpose_x_0 = const()[name = tensor<string, []>("content_score_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> content_score_1_transpose_y_0 = const()[name = tensor<string, []>("content_score_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = tensor<string, []>("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_25_perm_0 = const()[name = tensor<string, []>("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 36, 48]> transpose_76 = transpose(perm = transpose_25_perm_0, x = var_140_cast_fp16)[name = tensor<string, []>("transpose_76")];
            tensor<fp16, [1, 4, 48, 36]> transpose_77 = transpose(perm = transpose_24_perm_0, x = var_156_cast_fp16)[name = tensor<string, []>("transpose_77")];
            tensor<fp16, [1, 4, 48, 48]> content_score_1_cast_fp16 = matmul(transpose_x = content_score_1_transpose_x_0, transpose_y = content_score_1_transpose_y_0, x = transpose_77, y = transpose_76)[name = tensor<string, []>("content_score_1_cast_fp16")];
            tensor<fp16, [4, 36]> encoder_layers_0_sequential_1_module_attention_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_v_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(616064)))];
            tensor<fp16, [1, 48, 4, 36]> var_160_cast_fp16 = add(x = query_3_cast_fp16, y = encoder_layers_0_sequential_1_module_attention_v_bias_to_fp16)[name = tensor<string, []>("op_160_cast_fp16")];
            tensor<int32, [4]> var_161_perm_0 = const()[name = tensor<string, []>("op_161_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> pos_score_1_transpose_x_0 = const()[name = tensor<string, []>("pos_score_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_score_1_transpose_y_0 = const()[name = tensor<string, []>("pos_score_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 36, 48]> op_163_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(-1), name = tensor<string, []>("op_163_to_fp16_affine_quantized"), quantized_data = tensor<int8, [1, 4, 36, 48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(616448))), scale = tensor<fp16, []>(0x1.698p-6), zero_point = tensor<int8, []>(12)];
            tensor<fp16, [1, 4, 48, 36]> transpose_75 = transpose(perm = var_161_perm_0, x = var_160_cast_fp16)[name = tensor<string, []>("transpose_75")];
            tensor<fp16, [1, 4, 48, 48]> pos_score_1_cast_fp16 = matmul(transpose_x = pos_score_1_transpose_x_0, transpose_y = pos_score_1_transpose_y_0, x = transpose_75, y = op_163_to_fp16_affine_quantized)[name = tensor<string, []>("pos_score_1_cast_fp16")];
            tensor<bool, []> padded_pos_score_1_interleave_0 = const()[name = tensor<string, []>("padded_pos_score_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 1]> zeros_1_to_fp16 = const()[name = tensor<string, []>("zeros_1_to_fp16"), val = tensor<fp16, [1, 4, 48, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623424)))];
            tensor<fp16, [1, 4, 48, 49]> padded_pos_score_1_cast_fp16 = concat(axis = var_5, interleave = padded_pos_score_1_interleave_0, values = (zeros_1_to_fp16, pos_score_1_cast_fp16))[name = tensor<string, []>("padded_pos_score_1_cast_fp16")];
            tensor<int32, [4]> var_177 = const()[name = tensor<string, []>("op_177"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<fp16, [1, 4, 49, 48]> padded_pos_score_3_cast_fp16 = reshape(shape = var_177, x = padded_pos_score_1_cast_fp16)[name = tensor<string, []>("padded_pos_score_3_cast_fp16")];
            tensor<int32, [4]> var_181_begin_0 = const()[name = tensor<string, []>("op_181_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_181_end_0 = const()[name = tensor<string, []>("op_181_end_0"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<bool, [4]> var_181_end_mask_0 = const()[name = tensor<string, []>("op_181_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 48]> var_181_cast_fp16 = slice_by_index(begin = var_181_begin_0, end = var_181_end_0, end_mask = var_181_end_mask_0, x = padded_pos_score_3_cast_fp16)[name = tensor<string, []>("op_181_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_188_cast_fp16 = add(x = content_score_1_cast_fp16, y = var_181_cast_fp16)[name = tensor<string, []>("op_188_cast_fp16")];
            tensor<fp16, []> _inversed_input_25_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_25_y_0_to_fp16"), val = tensor<fp16, []>(0x1.554p-4)];
            tensor<fp16, [1, 4, 48, 48]> _inversed_input_25_cast_fp16 = mul(x = var_188_cast_fp16, y = _inversed_input_25_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_25_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> input_27_cast_fp16 = softmax(axis = var_5, x = _inversed_input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<bool, []> var_193_transpose_x_0 = const()[name = tensor<string, []>("op_193_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_193_transpose_y_0 = const()[name = tensor<string, []>("op_193_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 36]> transpose_74 = transpose(perm = var_149, x = var_148_cast_fp16)[name = tensor<string, []>("transpose_74")];
            tensor<fp16, [1, 4, 48, 36]> var_193_cast_fp16 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = input_27_cast_fp16, y = transpose_74)[name = tensor<string, []>("op_193_cast_fp16")];
            tensor<int32, [4]> context_1_perm_0 = const()[name = tensor<string, []>("context_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_196 = const()[name = tensor<string, []>("op_196"), val = tensor<int32, [3]>([1, -1, 144])];
            tensor<fp16, [1, 48, 4, 36]> transpose_73 = transpose(perm = context_1_perm_0, x = var_193_cast_fp16)[name = tensor<string, []>("transpose_73")];
            tensor<fp16, [1, 48, 144]> input_29_cast_fp16 = reshape(shape = var_196, x = transpose_73)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_0_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623872))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644928))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644672)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_1_module_attention_out_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_1_module_attention_out_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(645312)))];
            tensor<fp16, [1, 48, 144]> linear_7_cast_fp16 = linear(bias = encoder_layers_0_sequential_1_module_attention_out_proj_linear_bias_to_fp16, weight = encoder_layers_0_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized, x = input_29_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_33_cast_fp16 = add(x = linear_7_cast_fp16, y = inputs_3_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<int32, [1]> x_1_axes_0 = const()[name = tensor<string, []>("x_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_0_sequential_2_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_2_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(645696)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_2_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_2_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(646080)))];
            tensor<fp16, [1, 48, 144]> x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = encoder_layers_0_sequential_2_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_0_sequential_2_module_sequential_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
            tensor<int32, [3]> input_35_perm_0 = const()[name = tensor<string, []>("input_35_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_223 = const()[name = tensor<string, []>("op_223"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_225 = const()[name = tensor<string, []>("op_225"), val = tensor<int32, [1]>([1])];
            tensor<string, []> inputs_5_pad_type_0 = const()[name = tensor<string, []>("inputs_5_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> inputs_5_pad_0 = const()[name = tensor<string, []>("inputs_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [288, 144, 1]> encoder_layers_0_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [288, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(646464))), scale = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(688384))), zero_point = tensor<int8, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(688000)))];
            tensor<fp16, [288]> encoder_layers_0_sequential_2_module_sequential_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_2_module_sequential_2_conv_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(689024)))];
            tensor<fp16, [1, 144, 48]> transpose_72 = transpose(perm = input_35_perm_0, x = x_1_cast_fp16)[name = tensor<string, []>("transpose_72")];
            tensor<fp16, [1, 288, 48]> inputs_5_cast_fp16 = conv(bias = encoder_layers_0_sequential_2_module_sequential_2_conv_bias_to_fp16, dilations = var_225, groups = var_15, pad = inputs_5_pad_0, pad_type = inputs_5_pad_type_0, strides = var_223, weight = encoder_layers_0_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized, x = transpose_72)[name = tensor<string, []>("inputs_5_cast_fp16")];
            tensor<int32, [2]> var_228_split_sizes_0 = const()[name = tensor<string, []>("op_228_split_sizes_0"), val = tensor<int32, [2]>([144, 144])];
            tensor<int32, []> var_228_axis_0 = const()[name = tensor<string, []>("op_228_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 144, 48]> var_228_cast_fp16_0, tensor<fp16, [1, 144, 48]> var_228_cast_fp16_1 = split(axis = var_228_axis_0, split_sizes = var_228_split_sizes_0, x = inputs_5_cast_fp16)[name = tensor<string, []>("op_228_cast_fp16")];
            tensor<fp16, [1, 144, 48]> var_230_cast_fp16 = sigmoid(x = var_228_cast_fp16_1)[name = tensor<string, []>("op_230_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_37_cast_fp16 = mul(x = var_228_cast_fp16_0, y = var_230_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<int32, [1]> var_234 = const()[name = tensor<string, []>("op_234"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_236 = const()[name = tensor<string, []>("op_236"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_39_pad_type_0 = const()[name = tensor<string, []>("input_39_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_39_pad_0 = const()[name = tensor<string, []>("input_39_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [144, 1, 3]> inputs_7_weight_0_to_fp16 = const()[name = tensor<string, []>("inputs_7_weight_0_to_fp16"), val = tensor<fp16, [144, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(689664)))];
            tensor<fp16, [144]> inputs_7_bias_0_to_fp16 = const()[name = tensor<string, []>("inputs_7_bias_0_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(690624)))];
            tensor<fp16, [1, 144, 48]> inputs_7_cast_fp16 = conv(bias = inputs_7_bias_0_to_fp16, dilations = var_236, groups = var_13, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = var_234, weight = inputs_7_weight_0_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_41_cast_fp16 = silu(x = inputs_7_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<int32, [1]> var_249 = const()[name = tensor<string, []>("op_249"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_251 = const()[name = tensor<string, []>("op_251"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_43_pad_type_0 = const()[name = tensor<string, []>("input_43_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_43_pad_0 = const()[name = tensor<string, []>("input_43_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [144, 144, 1]> encoder_layers_0_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(691008))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712064))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(711808)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_2_module_sequential_7_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_2_module_sequential_7_conv_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712448)))];
            tensor<fp16, [1, 144, 48]> input_43_cast_fp16 = conv(bias = encoder_layers_0_sequential_2_module_sequential_7_conv_bias_to_fp16, dilations = var_251, groups = var_15, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = var_249, weight = encoder_layers_0_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized, x = input_41_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<int32, [3]> var_255_perm_0 = const()[name = tensor<string, []>("op_255_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 48, 144]> transpose_71 = transpose(perm = var_255_perm_0, x = input_43_cast_fp16)[name = tensor<string, []>("transpose_71")];
            tensor<fp16, [1, 48, 144]> input_45_cast_fp16 = add(x = transpose_71, y = input_33_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<int32, [1]> input_47_axes_0 = const()[name = tensor<string, []>("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_0_sequential_3_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_3_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(712832)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_3_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_3_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713216)))];
            tensor<fp16, [1, 48, 144]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = encoder_layers_0_sequential_3_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_0_sequential_3_module_sequential_0_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_0_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713600))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(797248))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(796608)))];
            tensor<fp16, [576]> encoder_layers_0_sequential_3_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_3_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798464)))];
            tensor<fp16, [1, 48, 576]> linear_8_cast_fp16 = linear(bias = encoder_layers_0_sequential_3_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_0_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_47_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_49_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_0_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(799680))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(882944))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(882688)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_3_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_3_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883328)))];
            tensor<fp16, [1, 48, 144]> linear_9_cast_fp16 = linear(bias = encoder_layers_0_sequential_3_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_0_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_49_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<fp16, []> var_282_to_fp16 = const()[name = tensor<string, []>("op_282_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_283_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_282_to_fp16)[name = tensor<string, []>("op_283_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_55_cast_fp16 = add(x = var_283_cast_fp16, y = input_45_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<int32, [1]> input_57_axes_0 = const()[name = tensor<string, []>("input_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_0_sequential_4_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_4_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883712)))];
            tensor<fp16, [144]> encoder_layers_0_sequential_4_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_sequential_4_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(884096)))];
            tensor<fp16, [1, 48, 144]> input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = encoder_layers_0_sequential_4_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_0_sequential_4_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<int32, [1]> input_59_axes_0 = const()[name = tensor<string, []>("input_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_1_sequential_0_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_0_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(884480)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_0_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_0_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(884864)))];
            tensor<fp16, [1, 48, 144]> input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = encoder_layers_1_sequential_0_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_1_sequential_0_module_sequential_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_1_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(885248))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(968896))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(968256)))];
            tensor<fp16, [576]> encoder_layers_1_sequential_0_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_0_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(970112)))];
            tensor<fp16, [1, 48, 576]> linear_10_cast_fp16 = linear(bias = encoder_layers_1_sequential_0_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_1_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_59_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_61_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_1_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(971328))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1054592))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1054336)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_0_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_0_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1054976)))];
            tensor<fp16, [1, 48, 144]> linear_11_cast_fp16 = linear(bias = encoder_layers_1_sequential_0_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_1_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_61_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, []> var_318_to_fp16 = const()[name = tensor<string, []>("op_318_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_319_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_318_to_fp16)[name = tensor<string, []>("op_319_cast_fp16")];
            tensor<fp16, [1, 48, 144]> inputs_13_cast_fp16 = add(x = var_319_cast_fp16, y = input_57_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
            tensor<int32, [1]> query_5_axes_0 = const()[name = tensor<string, []>("query_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_1_sequential_1_module_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_1_module_layer_norm_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1055360)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_1_module_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_1_module_layer_norm_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1055744)))];
            tensor<fp16, [1, 48, 144]> query_5_cast_fp16 = layer_norm(axes = query_5_axes_0, beta = encoder_layers_1_sequential_1_module_layer_norm_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_1_sequential_1_module_layer_norm_weight_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_1_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1056128))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1077184))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1076928)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_1_module_attention_query_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_query_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1077568)))];
            tensor<fp16, [1, 48, 144]> linear_12_cast_fp16 = linear(bias = encoder_layers_1_sequential_1_module_attention_query_proj_linear_bias_to_fp16, weight = encoder_layers_1_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized, x = query_5_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<int32, [4]> var_350 = const()[name = tensor<string, []>("op_350"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> query_7_cast_fp16 = reshape(shape = var_350, x = linear_12_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_1_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1077952))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1099008))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1098752)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_1_module_attention_key_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_key_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1099392)))];
            tensor<fp16, [1, 48, 144]> linear_13_cast_fp16 = linear(bias = encoder_layers_1_sequential_1_module_attention_key_proj_linear_bias_to_fp16, weight = encoder_layers_1_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized, x = query_5_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<int32, [4]> var_356 = const()[name = tensor<string, []>("op_356"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_357_cast_fp16 = reshape(shape = var_356, x = linear_13_cast_fp16)[name = tensor<string, []>("op_357_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_1_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1099776))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1120832))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1120576)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_1_module_attention_value_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_value_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1121216)))];
            tensor<fp16, [1, 48, 144]> linear_14_cast_fp16 = linear(bias = encoder_layers_1_sequential_1_module_attention_value_proj_linear_bias_to_fp16, weight = encoder_layers_1_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized, x = query_5_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<int32, [4]> var_364 = const()[name = tensor<string, []>("op_364"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_365_cast_fp16 = reshape(shape = var_364, x = linear_14_cast_fp16)[name = tensor<string, []>("op_365_cast_fp16")];
            tensor<int32, [4]> var_366 = const()[name = tensor<string, []>("op_366"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [4, 36]> encoder_layers_1_sequential_1_module_attention_u_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_u_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1121600)))];
            tensor<fp16, [1, 48, 4, 36]> var_373_cast_fp16 = add(x = query_7_cast_fp16, y = encoder_layers_1_sequential_1_module_attention_u_bias_to_fp16)[name = tensor<string, []>("op_373_cast_fp16")];
            tensor<bool, []> content_score_3_transpose_x_0 = const()[name = tensor<string, []>("content_score_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> content_score_3_transpose_y_0 = const()[name = tensor<string, []>("content_score_3_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = tensor<string, []>("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_27_perm_0 = const()[name = tensor<string, []>("transpose_27_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 36, 48]> transpose_69 = transpose(perm = transpose_27_perm_0, x = var_357_cast_fp16)[name = tensor<string, []>("transpose_69")];
            tensor<fp16, [1, 4, 48, 36]> transpose_70 = transpose(perm = transpose_26_perm_0, x = var_373_cast_fp16)[name = tensor<string, []>("transpose_70")];
            tensor<fp16, [1, 4, 48, 48]> content_score_3_cast_fp16 = matmul(transpose_x = content_score_3_transpose_x_0, transpose_y = content_score_3_transpose_y_0, x = transpose_70, y = transpose_69)[name = tensor<string, []>("content_score_3_cast_fp16")];
            tensor<fp16, [4, 36]> encoder_layers_1_sequential_1_module_attention_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_v_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1121984)))];
            tensor<fp16, [1, 48, 4, 36]> var_377_cast_fp16 = add(x = query_7_cast_fp16, y = encoder_layers_1_sequential_1_module_attention_v_bias_to_fp16)[name = tensor<string, []>("op_377_cast_fp16")];
            tensor<int32, [4]> var_378_perm_0 = const()[name = tensor<string, []>("op_378_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> pos_score_5_transpose_x_0 = const()[name = tensor<string, []>("pos_score_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_score_5_transpose_y_0 = const()[name = tensor<string, []>("pos_score_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 36, 48]> op_380_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(-1), name = tensor<string, []>("op_380_to_fp16_affine_quantized"), quantized_data = tensor<int8, [1, 4, 36, 48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1122368))), scale = tensor<fp16, []>(0x1.014p-4), zero_point = tensor<int8, []>(-9)];
            tensor<fp16, [1, 4, 48, 36]> transpose_68 = transpose(perm = var_378_perm_0, x = var_377_cast_fp16)[name = tensor<string, []>("transpose_68")];
            tensor<fp16, [1, 4, 48, 48]> pos_score_5_cast_fp16 = matmul(transpose_x = pos_score_5_transpose_x_0, transpose_y = pos_score_5_transpose_y_0, x = transpose_68, y = op_380_to_fp16_affine_quantized)[name = tensor<string, []>("pos_score_5_cast_fp16")];
            tensor<bool, []> padded_pos_score_5_interleave_0 = const()[name = tensor<string, []>("padded_pos_score_5_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 49]> padded_pos_score_5_cast_fp16 = concat(axis = var_5, interleave = padded_pos_score_5_interleave_0, values = (zeros_1_to_fp16, pos_score_5_cast_fp16))[name = tensor<string, []>("padded_pos_score_5_cast_fp16")];
            tensor<int32, [4]> var_394 = const()[name = tensor<string, []>("op_394"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<fp16, [1, 4, 49, 48]> padded_pos_score_7_cast_fp16 = reshape(shape = var_394, x = padded_pos_score_5_cast_fp16)[name = tensor<string, []>("padded_pos_score_7_cast_fp16")];
            tensor<int32, [4]> var_398_begin_0 = const()[name = tensor<string, []>("op_398_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_398_end_0 = const()[name = tensor<string, []>("op_398_end_0"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<bool, [4]> var_398_end_mask_0 = const()[name = tensor<string, []>("op_398_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 48]> var_398_cast_fp16 = slice_by_index(begin = var_398_begin_0, end = var_398_end_0, end_mask = var_398_end_mask_0, x = padded_pos_score_7_cast_fp16)[name = tensor<string, []>("op_398_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_405_cast_fp16 = add(x = content_score_3_cast_fp16, y = var_398_cast_fp16)[name = tensor<string, []>("op_405_cast_fp16")];
            tensor<fp16, []> _inversed_input_69_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_69_y_0_to_fp16"), val = tensor<fp16, []>(0x1.554p-4)];
            tensor<fp16, [1, 4, 48, 48]> _inversed_input_69_cast_fp16 = mul(x = var_405_cast_fp16, y = _inversed_input_69_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_69_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> input_71_cast_fp16 = softmax(axis = var_5, x = _inversed_input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<bool, []> var_410_transpose_x_0 = const()[name = tensor<string, []>("op_410_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_410_transpose_y_0 = const()[name = tensor<string, []>("op_410_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 36]> transpose_67 = transpose(perm = var_366, x = var_365_cast_fp16)[name = tensor<string, []>("transpose_67")];
            tensor<fp16, [1, 4, 48, 36]> var_410_cast_fp16 = matmul(transpose_x = var_410_transpose_x_0, transpose_y = var_410_transpose_y_0, x = input_71_cast_fp16, y = transpose_67)[name = tensor<string, []>("op_410_cast_fp16")];
            tensor<int32, [4]> context_3_perm_0 = const()[name = tensor<string, []>("context_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_413 = const()[name = tensor<string, []>("op_413"), val = tensor<int32, [3]>([1, -1, 144])];
            tensor<fp16, [1, 48, 4, 36]> transpose_66 = transpose(perm = context_3_perm_0, x = var_410_cast_fp16)[name = tensor<string, []>("transpose_66")];
            tensor<fp16, [1, 48, 144]> input_73_cast_fp16 = reshape(shape = var_413, x = transpose_66)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_1_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1129344))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1150400))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1150144)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_1_module_attention_out_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_1_module_attention_out_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1150784)))];
            tensor<fp16, [1, 48, 144]> linear_16_cast_fp16 = linear(bias = encoder_layers_1_sequential_1_module_attention_out_proj_linear_bias_to_fp16, weight = encoder_layers_1_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized, x = input_73_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_77_cast_fp16 = add(x = linear_16_cast_fp16, y = inputs_13_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<int32, [1]> x_3_axes_0 = const()[name = tensor<string, []>("x_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_1_sequential_2_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_2_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1151168)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_2_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_2_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1151552)))];
            tensor<fp16, [1, 48, 144]> x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = encoder_layers_1_sequential_2_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_1_sequential_2_module_sequential_0_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
            tensor<int32, [3]> input_79_perm_0 = const()[name = tensor<string, []>("input_79_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_440 = const()[name = tensor<string, []>("op_440"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_442 = const()[name = tensor<string, []>("op_442"), val = tensor<int32, [1]>([1])];
            tensor<string, []> inputs_15_pad_type_0 = const()[name = tensor<string, []>("inputs_15_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> inputs_15_pad_0 = const()[name = tensor<string, []>("inputs_15_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [288, 144, 1]> encoder_layers_1_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [288, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1151936))), scale = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1193856))), zero_point = tensor<int8, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1193472)))];
            tensor<fp16, [288]> encoder_layers_1_sequential_2_module_sequential_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_2_module_sequential_2_conv_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1194496)))];
            tensor<fp16, [1, 144, 48]> transpose_65 = transpose(perm = input_79_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_65")];
            tensor<fp16, [1, 288, 48]> inputs_15_cast_fp16 = conv(bias = encoder_layers_1_sequential_2_module_sequential_2_conv_bias_to_fp16, dilations = var_442, groups = var_15, pad = inputs_15_pad_0, pad_type = inputs_15_pad_type_0, strides = var_440, weight = encoder_layers_1_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized, x = transpose_65)[name = tensor<string, []>("inputs_15_cast_fp16")];
            tensor<int32, [2]> var_445_split_sizes_0 = const()[name = tensor<string, []>("op_445_split_sizes_0"), val = tensor<int32, [2]>([144, 144])];
            tensor<int32, []> var_445_axis_0 = const()[name = tensor<string, []>("op_445_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 144, 48]> var_445_cast_fp16_0, tensor<fp16, [1, 144, 48]> var_445_cast_fp16_1 = split(axis = var_445_axis_0, split_sizes = var_445_split_sizes_0, x = inputs_15_cast_fp16)[name = tensor<string, []>("op_445_cast_fp16")];
            tensor<fp16, [1, 144, 48]> var_447_cast_fp16 = sigmoid(x = var_445_cast_fp16_1)[name = tensor<string, []>("op_447_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_81_cast_fp16 = mul(x = var_445_cast_fp16_0, y = var_447_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<int32, [1]> var_451 = const()[name = tensor<string, []>("op_451"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_453 = const()[name = tensor<string, []>("op_453"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_83_pad_type_0 = const()[name = tensor<string, []>("input_83_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_83_pad_0 = const()[name = tensor<string, []>("input_83_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [144, 1, 3]> inputs_17_weight_0_to_fp16 = const()[name = tensor<string, []>("inputs_17_weight_0_to_fp16"), val = tensor<fp16, [144, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195136)))];
            tensor<fp16, [144]> inputs_17_bias_0_to_fp16 = const()[name = tensor<string, []>("inputs_17_bias_0_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1196096)))];
            tensor<fp16, [1, 144, 48]> inputs_17_cast_fp16 = conv(bias = inputs_17_bias_0_to_fp16, dilations = var_453, groups = var_13, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = var_451, weight = inputs_17_weight_0_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_85_cast_fp16 = silu(x = inputs_17_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<int32, [1]> var_466 = const()[name = tensor<string, []>("op_466"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_468 = const()[name = tensor<string, []>("op_468"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_87_pad_type_0 = const()[name = tensor<string, []>("input_87_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_87_pad_0 = const()[name = tensor<string, []>("input_87_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [144, 144, 1]> encoder_layers_1_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1196480))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1217536))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1217280)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_2_module_sequential_7_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_2_module_sequential_7_conv_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1217920)))];
            tensor<fp16, [1, 144, 48]> input_87_cast_fp16 = conv(bias = encoder_layers_1_sequential_2_module_sequential_7_conv_bias_to_fp16, dilations = var_468, groups = var_15, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = var_466, weight = encoder_layers_1_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<int32, [3]> var_472_perm_0 = const()[name = tensor<string, []>("op_472_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 48, 144]> transpose_64 = transpose(perm = var_472_perm_0, x = input_87_cast_fp16)[name = tensor<string, []>("transpose_64")];
            tensor<fp16, [1, 48, 144]> input_89_cast_fp16 = add(x = transpose_64, y = input_77_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<int32, [1]> input_91_axes_0 = const()[name = tensor<string, []>("input_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_1_sequential_3_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_3_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1218304)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_3_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_3_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1218688)))];
            tensor<fp16, [1, 48, 144]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = encoder_layers_1_sequential_3_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_1_sequential_3_module_sequential_0_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_1_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1219072))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1302720))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1302080)))];
            tensor<fp16, [576]> encoder_layers_1_sequential_3_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_3_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1303936)))];
            tensor<fp16, [1, 48, 576]> linear_17_cast_fp16 = linear(bias = encoder_layers_1_sequential_3_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_1_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_91_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_93_cast_fp16 = silu(x = linear_17_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_1_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1305152))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1388416))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1388160)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_3_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_3_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1388800)))];
            tensor<fp16, [1, 48, 144]> linear_18_cast_fp16 = linear(bias = encoder_layers_1_sequential_3_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_1_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_93_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<fp16, []> var_499_to_fp16 = const()[name = tensor<string, []>("op_499_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_500_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_499_to_fp16)[name = tensor<string, []>("op_500_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_99_cast_fp16 = add(x = var_500_cast_fp16, y = input_89_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<int32, [1]> input_101_axes_0 = const()[name = tensor<string, []>("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_1_sequential_4_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_4_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1389184)))];
            tensor<fp16, [144]> encoder_layers_1_sequential_4_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_sequential_4_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1389568)))];
            tensor<fp16, [1, 48, 144]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = encoder_layers_1_sequential_4_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_1_sequential_4_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<int32, [1]> input_103_axes_0 = const()[name = tensor<string, []>("input_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_2_sequential_0_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_0_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1389952)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_0_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_0_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1390336)))];
            tensor<fp16, [1, 48, 144]> input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = encoder_layers_2_sequential_0_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_2_sequential_0_module_sequential_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_2_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1390720))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1474368))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1473728)))];
            tensor<fp16, [576]> encoder_layers_2_sequential_0_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_0_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1475584)))];
            tensor<fp16, [1, 48, 576]> linear_19_cast_fp16 = linear(bias = encoder_layers_2_sequential_0_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_2_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_103_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_105_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_2_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1476800))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1560064))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1559808)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_0_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_0_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1560448)))];
            tensor<fp16, [1, 48, 144]> linear_20_cast_fp16 = linear(bias = encoder_layers_2_sequential_0_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_2_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_105_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<fp16, []> var_535_to_fp16 = const()[name = tensor<string, []>("op_535_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_536_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_535_to_fp16)[name = tensor<string, []>("op_536_cast_fp16")];
            tensor<fp16, [1, 48, 144]> inputs_23_cast_fp16 = add(x = var_536_cast_fp16, y = input_101_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
            tensor<int32, [1]> query_9_axes_0 = const()[name = tensor<string, []>("query_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_2_sequential_1_module_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_1_module_layer_norm_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1560832)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_1_module_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_1_module_layer_norm_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1561216)))];
            tensor<fp16, [1, 48, 144]> query_9_cast_fp16 = layer_norm(axes = query_9_axes_0, beta = encoder_layers_2_sequential_1_module_layer_norm_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_2_sequential_1_module_layer_norm_weight_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_2_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1561600))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1582656))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1582400)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_1_module_attention_query_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_query_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1583040)))];
            tensor<fp16, [1, 48, 144]> linear_21_cast_fp16 = linear(bias = encoder_layers_2_sequential_1_module_attention_query_proj_linear_bias_to_fp16, weight = encoder_layers_2_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized, x = query_9_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<int32, [4]> var_567 = const()[name = tensor<string, []>("op_567"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> query_11_cast_fp16 = reshape(shape = var_567, x = linear_21_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_2_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1583424))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1604480))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1604224)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_1_module_attention_key_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_key_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1604864)))];
            tensor<fp16, [1, 48, 144]> linear_22_cast_fp16 = linear(bias = encoder_layers_2_sequential_1_module_attention_key_proj_linear_bias_to_fp16, weight = encoder_layers_2_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized, x = query_9_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<int32, [4]> var_573 = const()[name = tensor<string, []>("op_573"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_574_cast_fp16 = reshape(shape = var_573, x = linear_22_cast_fp16)[name = tensor<string, []>("op_574_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_2_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1605248))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1626304))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1626048)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_1_module_attention_value_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_value_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1626688)))];
            tensor<fp16, [1, 48, 144]> linear_23_cast_fp16 = linear(bias = encoder_layers_2_sequential_1_module_attention_value_proj_linear_bias_to_fp16, weight = encoder_layers_2_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized, x = query_9_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<int32, [4]> var_581 = const()[name = tensor<string, []>("op_581"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_582_cast_fp16 = reshape(shape = var_581, x = linear_23_cast_fp16)[name = tensor<string, []>("op_582_cast_fp16")];
            tensor<int32, [4]> var_583 = const()[name = tensor<string, []>("op_583"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [4, 36]> encoder_layers_2_sequential_1_module_attention_u_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_u_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1627072)))];
            tensor<fp16, [1, 48, 4, 36]> var_590_cast_fp16 = add(x = query_11_cast_fp16, y = encoder_layers_2_sequential_1_module_attention_u_bias_to_fp16)[name = tensor<string, []>("op_590_cast_fp16")];
            tensor<bool, []> content_score_5_transpose_x_0 = const()[name = tensor<string, []>("content_score_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> content_score_5_transpose_y_0 = const()[name = tensor<string, []>("content_score_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = tensor<string, []>("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_29_perm_0 = const()[name = tensor<string, []>("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 36, 48]> transpose_62 = transpose(perm = transpose_29_perm_0, x = var_574_cast_fp16)[name = tensor<string, []>("transpose_62")];
            tensor<fp16, [1, 4, 48, 36]> transpose_63 = transpose(perm = transpose_28_perm_0, x = var_590_cast_fp16)[name = tensor<string, []>("transpose_63")];
            tensor<fp16, [1, 4, 48, 48]> content_score_5_cast_fp16 = matmul(transpose_x = content_score_5_transpose_x_0, transpose_y = content_score_5_transpose_y_0, x = transpose_63, y = transpose_62)[name = tensor<string, []>("content_score_5_cast_fp16")];
            tensor<fp16, [4, 36]> encoder_layers_2_sequential_1_module_attention_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_v_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1627456)))];
            tensor<fp16, [1, 48, 4, 36]> var_594_cast_fp16 = add(x = query_11_cast_fp16, y = encoder_layers_2_sequential_1_module_attention_v_bias_to_fp16)[name = tensor<string, []>("op_594_cast_fp16")];
            tensor<int32, [4]> var_595_perm_0 = const()[name = tensor<string, []>("op_595_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> pos_score_9_transpose_x_0 = const()[name = tensor<string, []>("pos_score_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_score_9_transpose_y_0 = const()[name = tensor<string, []>("pos_score_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 36, 48]> op_597_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(-1), name = tensor<string, []>("op_597_to_fp16_affine_quantized"), quantized_data = tensor<int8, [1, 4, 36, 48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1627840))), scale = tensor<fp16, []>(0x1.b34p-5), zero_point = tensor<int8, []>(-15)];
            tensor<fp16, [1, 4, 48, 36]> transpose_61 = transpose(perm = var_595_perm_0, x = var_594_cast_fp16)[name = tensor<string, []>("transpose_61")];
            tensor<fp16, [1, 4, 48, 48]> pos_score_9_cast_fp16 = matmul(transpose_x = pos_score_9_transpose_x_0, transpose_y = pos_score_9_transpose_y_0, x = transpose_61, y = op_597_to_fp16_affine_quantized)[name = tensor<string, []>("pos_score_9_cast_fp16")];
            tensor<bool, []> padded_pos_score_9_interleave_0 = const()[name = tensor<string, []>("padded_pos_score_9_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 49]> padded_pos_score_9_cast_fp16 = concat(axis = var_5, interleave = padded_pos_score_9_interleave_0, values = (zeros_1_to_fp16, pos_score_9_cast_fp16))[name = tensor<string, []>("padded_pos_score_9_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = tensor<string, []>("op_611"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<fp16, [1, 4, 49, 48]> padded_pos_score_11_cast_fp16 = reshape(shape = var_611, x = padded_pos_score_9_cast_fp16)[name = tensor<string, []>("padded_pos_score_11_cast_fp16")];
            tensor<int32, [4]> var_615_begin_0 = const()[name = tensor<string, []>("op_615_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_615_end_0 = const()[name = tensor<string, []>("op_615_end_0"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<bool, [4]> var_615_end_mask_0 = const()[name = tensor<string, []>("op_615_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 48]> var_615_cast_fp16 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = padded_pos_score_11_cast_fp16)[name = tensor<string, []>("op_615_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_622_cast_fp16 = add(x = content_score_5_cast_fp16, y = var_615_cast_fp16)[name = tensor<string, []>("op_622_cast_fp16")];
            tensor<fp16, []> _inversed_input_113_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_113_y_0_to_fp16"), val = tensor<fp16, []>(0x1.554p-4)];
            tensor<fp16, [1, 4, 48, 48]> _inversed_input_113_cast_fp16 = mul(x = var_622_cast_fp16, y = _inversed_input_113_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_113_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> input_115_cast_fp16 = softmax(axis = var_5, x = _inversed_input_113_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<bool, []> var_627_transpose_x_0 = const()[name = tensor<string, []>("op_627_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_627_transpose_y_0 = const()[name = tensor<string, []>("op_627_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 36]> transpose_60 = transpose(perm = var_583, x = var_582_cast_fp16)[name = tensor<string, []>("transpose_60")];
            tensor<fp16, [1, 4, 48, 36]> var_627_cast_fp16 = matmul(transpose_x = var_627_transpose_x_0, transpose_y = var_627_transpose_y_0, x = input_115_cast_fp16, y = transpose_60)[name = tensor<string, []>("op_627_cast_fp16")];
            tensor<int32, [4]> context_5_perm_0 = const()[name = tensor<string, []>("context_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_630 = const()[name = tensor<string, []>("op_630"), val = tensor<int32, [3]>([1, -1, 144])];
            tensor<fp16, [1, 48, 4, 36]> transpose_59 = transpose(perm = context_5_perm_0, x = var_627_cast_fp16)[name = tensor<string, []>("transpose_59")];
            tensor<fp16, [1, 48, 144]> input_117_cast_fp16 = reshape(shape = var_630, x = transpose_59)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_2_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1634816))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1655872))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1655616)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_1_module_attention_out_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_1_module_attention_out_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1656256)))];
            tensor<fp16, [1, 48, 144]> linear_25_cast_fp16 = linear(bias = encoder_layers_2_sequential_1_module_attention_out_proj_linear_bias_to_fp16, weight = encoder_layers_2_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized, x = input_117_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_121_cast_fp16 = add(x = linear_25_cast_fp16, y = inputs_23_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<int32, [1]> x_5_axes_0 = const()[name = tensor<string, []>("x_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_2_sequential_2_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_2_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1656640)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_2_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_2_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1657024)))];
            tensor<fp16, [1, 48, 144]> x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = encoder_layers_2_sequential_2_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_2_sequential_2_module_sequential_0_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
            tensor<int32, [3]> input_123_perm_0 = const()[name = tensor<string, []>("input_123_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_657 = const()[name = tensor<string, []>("op_657"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_659 = const()[name = tensor<string, []>("op_659"), val = tensor<int32, [1]>([1])];
            tensor<string, []> inputs_25_pad_type_0 = const()[name = tensor<string, []>("inputs_25_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> inputs_25_pad_0 = const()[name = tensor<string, []>("inputs_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [288, 144, 1]> encoder_layers_2_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [288, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1657408))), scale = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1699328))), zero_point = tensor<int8, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1698944)))];
            tensor<fp16, [288]> encoder_layers_2_sequential_2_module_sequential_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_2_module_sequential_2_conv_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1699968)))];
            tensor<fp16, [1, 144, 48]> transpose_58 = transpose(perm = input_123_perm_0, x = x_5_cast_fp16)[name = tensor<string, []>("transpose_58")];
            tensor<fp16, [1, 288, 48]> inputs_25_cast_fp16 = conv(bias = encoder_layers_2_sequential_2_module_sequential_2_conv_bias_to_fp16, dilations = var_659, groups = var_15, pad = inputs_25_pad_0, pad_type = inputs_25_pad_type_0, strides = var_657, weight = encoder_layers_2_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized, x = transpose_58)[name = tensor<string, []>("inputs_25_cast_fp16")];
            tensor<int32, [2]> var_662_split_sizes_0 = const()[name = tensor<string, []>("op_662_split_sizes_0"), val = tensor<int32, [2]>([144, 144])];
            tensor<int32, []> var_662_axis_0 = const()[name = tensor<string, []>("op_662_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 144, 48]> var_662_cast_fp16_0, tensor<fp16, [1, 144, 48]> var_662_cast_fp16_1 = split(axis = var_662_axis_0, split_sizes = var_662_split_sizes_0, x = inputs_25_cast_fp16)[name = tensor<string, []>("op_662_cast_fp16")];
            tensor<fp16, [1, 144, 48]> var_664_cast_fp16 = sigmoid(x = var_662_cast_fp16_1)[name = tensor<string, []>("op_664_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_125_cast_fp16 = mul(x = var_662_cast_fp16_0, y = var_664_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
            tensor<int32, [1]> var_668 = const()[name = tensor<string, []>("op_668"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_670 = const()[name = tensor<string, []>("op_670"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_127_pad_type_0 = const()[name = tensor<string, []>("input_127_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_127_pad_0 = const()[name = tensor<string, []>("input_127_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [144, 1, 3]> inputs_27_weight_0_to_fp16 = const()[name = tensor<string, []>("inputs_27_weight_0_to_fp16"), val = tensor<fp16, [144, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1700608)))];
            tensor<fp16, [144]> inputs_27_bias_0_to_fp16 = const()[name = tensor<string, []>("inputs_27_bias_0_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1701568)))];
            tensor<fp16, [1, 144, 48]> inputs_27_cast_fp16 = conv(bias = inputs_27_bias_0_to_fp16, dilations = var_670, groups = var_13, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = var_668, weight = inputs_27_weight_0_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("inputs_27_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_129_cast_fp16 = silu(x = inputs_27_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<int32, [1]> var_683 = const()[name = tensor<string, []>("op_683"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_685 = const()[name = tensor<string, []>("op_685"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_131_pad_type_0 = const()[name = tensor<string, []>("input_131_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_131_pad_0 = const()[name = tensor<string, []>("input_131_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [144, 144, 1]> encoder_layers_2_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1701952))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1723008))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1722752)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_2_module_sequential_7_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_2_module_sequential_7_conv_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1723392)))];
            tensor<fp16, [1, 144, 48]> input_131_cast_fp16 = conv(bias = encoder_layers_2_sequential_2_module_sequential_7_conv_bias_to_fp16, dilations = var_685, groups = var_15, pad = input_131_pad_0, pad_type = input_131_pad_type_0, strides = var_683, weight = encoder_layers_2_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<int32, [3]> var_689_perm_0 = const()[name = tensor<string, []>("op_689_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 48, 144]> transpose_57 = transpose(perm = var_689_perm_0, x = input_131_cast_fp16)[name = tensor<string, []>("transpose_57")];
            tensor<fp16, [1, 48, 144]> input_133_cast_fp16 = add(x = transpose_57, y = input_121_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<int32, [1]> input_135_axes_0 = const()[name = tensor<string, []>("input_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_2_sequential_3_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_3_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1723776)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_3_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_3_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1724160)))];
            tensor<fp16, [1, 48, 144]> input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = encoder_layers_2_sequential_3_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_2_sequential_3_module_sequential_0_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_2_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1724544))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1808192))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1807552)))];
            tensor<fp16, [576]> encoder_layers_2_sequential_3_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_3_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1809408)))];
            tensor<fp16, [1, 48, 576]> linear_26_cast_fp16 = linear(bias = encoder_layers_2_sequential_3_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_2_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_135_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_137_cast_fp16 = silu(x = linear_26_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_2_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1810624))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1893888))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1893632)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_3_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_3_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1894272)))];
            tensor<fp16, [1, 48, 144]> linear_27_cast_fp16 = linear(bias = encoder_layers_2_sequential_3_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_2_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_137_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<fp16, []> var_716_to_fp16 = const()[name = tensor<string, []>("op_716_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_717_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_716_to_fp16)[name = tensor<string, []>("op_717_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_143_cast_fp16 = add(x = var_717_cast_fp16, y = input_133_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<int32, [1]> input_145_axes_0 = const()[name = tensor<string, []>("input_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_2_sequential_4_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_4_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1894656)))];
            tensor<fp16, [144]> encoder_layers_2_sequential_4_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_sequential_4_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1895040)))];
            tensor<fp16, [1, 48, 144]> input_145_cast_fp16 = layer_norm(axes = input_145_axes_0, beta = encoder_layers_2_sequential_4_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_2_sequential_4_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
            tensor<int32, [1]> input_147_axes_0 = const()[name = tensor<string, []>("input_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_3_sequential_0_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_0_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1895424)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_0_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_0_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1895808)))];
            tensor<fp16, [1, 48, 144]> input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = encoder_layers_3_sequential_0_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_3_sequential_0_module_sequential_0_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_3_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1896192))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1979840))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1979200)))];
            tensor<fp16, [576]> encoder_layers_3_sequential_0_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_0_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1981056)))];
            tensor<fp16, [1, 48, 576]> linear_28_cast_fp16 = linear(bias = encoder_layers_3_sequential_0_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_3_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_147_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_149_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_3_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1982272))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2065536))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2065280)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_0_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_0_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2065920)))];
            tensor<fp16, [1, 48, 144]> linear_29_cast_fp16 = linear(bias = encoder_layers_3_sequential_0_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_3_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_149_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<fp16, []> var_752_to_fp16 = const()[name = tensor<string, []>("op_752_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_753_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_752_to_fp16)[name = tensor<string, []>("op_753_cast_fp16")];
            tensor<fp16, [1, 48, 144]> inputs_33_cast_fp16 = add(x = var_753_cast_fp16, y = input_145_cast_fp16)[name = tensor<string, []>("inputs_33_cast_fp16")];
            tensor<int32, [1]> query_13_axes_0 = const()[name = tensor<string, []>("query_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_3_sequential_1_module_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_1_module_layer_norm_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2066304)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_1_module_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_1_module_layer_norm_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2066688)))];
            tensor<fp16, [1, 48, 144]> query_13_cast_fp16 = layer_norm(axes = query_13_axes_0, beta = encoder_layers_3_sequential_1_module_layer_norm_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_3_sequential_1_module_layer_norm_weight_to_fp16, x = inputs_33_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_3_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2067072))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2088128))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2087872)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_1_module_attention_query_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_query_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2088512)))];
            tensor<fp16, [1, 48, 144]> linear_30_cast_fp16 = linear(bias = encoder_layers_3_sequential_1_module_attention_query_proj_linear_bias_to_fp16, weight = encoder_layers_3_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized, x = query_13_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<int32, [4]> var_784 = const()[name = tensor<string, []>("op_784"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> query_15_cast_fp16 = reshape(shape = var_784, x = linear_30_cast_fp16)[name = tensor<string, []>("query_15_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_3_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2088896))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2109952))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2109696)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_1_module_attention_key_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_key_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2110336)))];
            tensor<fp16, [1, 48, 144]> linear_31_cast_fp16 = linear(bias = encoder_layers_3_sequential_1_module_attention_key_proj_linear_bias_to_fp16, weight = encoder_layers_3_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized, x = query_13_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<int32, [4]> var_790 = const()[name = tensor<string, []>("op_790"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_791_cast_fp16 = reshape(shape = var_790, x = linear_31_cast_fp16)[name = tensor<string, []>("op_791_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_3_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2110720))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2131776))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2131520)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_1_module_attention_value_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_value_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2132160)))];
            tensor<fp16, [1, 48, 144]> linear_32_cast_fp16 = linear(bias = encoder_layers_3_sequential_1_module_attention_value_proj_linear_bias_to_fp16, weight = encoder_layers_3_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized, x = query_13_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<int32, [4]> var_798 = const()[name = tensor<string, []>("op_798"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_799_cast_fp16 = reshape(shape = var_798, x = linear_32_cast_fp16)[name = tensor<string, []>("op_799_cast_fp16")];
            tensor<int32, [4]> var_800 = const()[name = tensor<string, []>("op_800"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [4, 36]> encoder_layers_3_sequential_1_module_attention_u_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_u_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2132544)))];
            tensor<fp16, [1, 48, 4, 36]> var_807_cast_fp16 = add(x = query_15_cast_fp16, y = encoder_layers_3_sequential_1_module_attention_u_bias_to_fp16)[name = tensor<string, []>("op_807_cast_fp16")];
            tensor<bool, []> content_score_7_transpose_x_0 = const()[name = tensor<string, []>("content_score_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> content_score_7_transpose_y_0 = const()[name = tensor<string, []>("content_score_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = tensor<string, []>("transpose_30_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_31_perm_0 = const()[name = tensor<string, []>("transpose_31_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 36, 48]> transpose_55 = transpose(perm = transpose_31_perm_0, x = var_791_cast_fp16)[name = tensor<string, []>("transpose_55")];
            tensor<fp16, [1, 4, 48, 36]> transpose_56 = transpose(perm = transpose_30_perm_0, x = var_807_cast_fp16)[name = tensor<string, []>("transpose_56")];
            tensor<fp16, [1, 4, 48, 48]> content_score_7_cast_fp16 = matmul(transpose_x = content_score_7_transpose_x_0, transpose_y = content_score_7_transpose_y_0, x = transpose_56, y = transpose_55)[name = tensor<string, []>("content_score_7_cast_fp16")];
            tensor<fp16, [4, 36]> encoder_layers_3_sequential_1_module_attention_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_v_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2132928)))];
            tensor<fp16, [1, 48, 4, 36]> var_811_cast_fp16 = add(x = query_15_cast_fp16, y = encoder_layers_3_sequential_1_module_attention_v_bias_to_fp16)[name = tensor<string, []>("op_811_cast_fp16")];
            tensor<int32, [4]> var_812_perm_0 = const()[name = tensor<string, []>("op_812_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> pos_score_13_transpose_x_0 = const()[name = tensor<string, []>("pos_score_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_score_13_transpose_y_0 = const()[name = tensor<string, []>("pos_score_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 36, 48]> op_814_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(-1), name = tensor<string, []>("op_814_to_fp16_affine_quantized"), quantized_data = tensor<int8, [1, 4, 36, 48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2133312))), scale = tensor<fp16, []>(0x1.d1cp-5), zero_point = tensor<int8, []>(-22)];
            tensor<fp16, [1, 4, 48, 36]> transpose_54 = transpose(perm = var_812_perm_0, x = var_811_cast_fp16)[name = tensor<string, []>("transpose_54")];
            tensor<fp16, [1, 4, 48, 48]> pos_score_13_cast_fp16 = matmul(transpose_x = pos_score_13_transpose_x_0, transpose_y = pos_score_13_transpose_y_0, x = transpose_54, y = op_814_to_fp16_affine_quantized)[name = tensor<string, []>("pos_score_13_cast_fp16")];
            tensor<bool, []> padded_pos_score_13_interleave_0 = const()[name = tensor<string, []>("padded_pos_score_13_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 49]> padded_pos_score_13_cast_fp16 = concat(axis = var_5, interleave = padded_pos_score_13_interleave_0, values = (zeros_1_to_fp16, pos_score_13_cast_fp16))[name = tensor<string, []>("padded_pos_score_13_cast_fp16")];
            tensor<int32, [4]> var_828 = const()[name = tensor<string, []>("op_828"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<fp16, [1, 4, 49, 48]> padded_pos_score_15_cast_fp16 = reshape(shape = var_828, x = padded_pos_score_13_cast_fp16)[name = tensor<string, []>("padded_pos_score_15_cast_fp16")];
            tensor<int32, [4]> var_832_begin_0 = const()[name = tensor<string, []>("op_832_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_832_end_0 = const()[name = tensor<string, []>("op_832_end_0"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<bool, [4]> var_832_end_mask_0 = const()[name = tensor<string, []>("op_832_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 48]> var_832_cast_fp16 = slice_by_index(begin = var_832_begin_0, end = var_832_end_0, end_mask = var_832_end_mask_0, x = padded_pos_score_15_cast_fp16)[name = tensor<string, []>("op_832_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_839_cast_fp16 = add(x = content_score_7_cast_fp16, y = var_832_cast_fp16)[name = tensor<string, []>("op_839_cast_fp16")];
            tensor<fp16, []> _inversed_input_157_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_157_y_0_to_fp16"), val = tensor<fp16, []>(0x1.554p-4)];
            tensor<fp16, [1, 4, 48, 48]> _inversed_input_157_cast_fp16 = mul(x = var_839_cast_fp16, y = _inversed_input_157_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_157_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> input_159_cast_fp16 = softmax(axis = var_5, x = _inversed_input_157_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<bool, []> var_844_transpose_x_0 = const()[name = tensor<string, []>("op_844_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_844_transpose_y_0 = const()[name = tensor<string, []>("op_844_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 36]> transpose_53 = transpose(perm = var_800, x = var_799_cast_fp16)[name = tensor<string, []>("transpose_53")];
            tensor<fp16, [1, 4, 48, 36]> var_844_cast_fp16 = matmul(transpose_x = var_844_transpose_x_0, transpose_y = var_844_transpose_y_0, x = input_159_cast_fp16, y = transpose_53)[name = tensor<string, []>("op_844_cast_fp16")];
            tensor<int32, [4]> context_7_perm_0 = const()[name = tensor<string, []>("context_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_847 = const()[name = tensor<string, []>("op_847"), val = tensor<int32, [3]>([1, -1, 144])];
            tensor<fp16, [1, 48, 4, 36]> transpose_52 = transpose(perm = context_7_perm_0, x = var_844_cast_fp16)[name = tensor<string, []>("transpose_52")];
            tensor<fp16, [1, 48, 144]> input_161_cast_fp16 = reshape(shape = var_847, x = transpose_52)[name = tensor<string, []>("input_161_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_3_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2140288))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2161344))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2161088)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_1_module_attention_out_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_1_module_attention_out_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2161728)))];
            tensor<fp16, [1, 48, 144]> linear_34_cast_fp16 = linear(bias = encoder_layers_3_sequential_1_module_attention_out_proj_linear_bias_to_fp16, weight = encoder_layers_3_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized, x = input_161_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_165_cast_fp16 = add(x = linear_34_cast_fp16, y = inputs_33_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
            tensor<int32, [1]> x_7_axes_0 = const()[name = tensor<string, []>("x_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_3_sequential_2_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_2_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2162112)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_2_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_2_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2162496)))];
            tensor<fp16, [1, 48, 144]> x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = encoder_layers_3_sequential_2_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_3_sequential_2_module_sequential_0_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
            tensor<int32, [3]> input_167_perm_0 = const()[name = tensor<string, []>("input_167_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_874 = const()[name = tensor<string, []>("op_874"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_876 = const()[name = tensor<string, []>("op_876"), val = tensor<int32, [1]>([1])];
            tensor<string, []> inputs_35_pad_type_0 = const()[name = tensor<string, []>("inputs_35_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> inputs_35_pad_0 = const()[name = tensor<string, []>("inputs_35_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [288, 144, 1]> encoder_layers_3_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [288, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2162880))), scale = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2204800))), zero_point = tensor<int8, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2204416)))];
            tensor<fp16, [288]> encoder_layers_3_sequential_2_module_sequential_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_2_module_sequential_2_conv_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2205440)))];
            tensor<fp16, [1, 144, 48]> transpose_51 = transpose(perm = input_167_perm_0, x = x_7_cast_fp16)[name = tensor<string, []>("transpose_51")];
            tensor<fp16, [1, 288, 48]> inputs_35_cast_fp16 = conv(bias = encoder_layers_3_sequential_2_module_sequential_2_conv_bias_to_fp16, dilations = var_876, groups = var_15, pad = inputs_35_pad_0, pad_type = inputs_35_pad_type_0, strides = var_874, weight = encoder_layers_3_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized, x = transpose_51)[name = tensor<string, []>("inputs_35_cast_fp16")];
            tensor<int32, [2]> var_879_split_sizes_0 = const()[name = tensor<string, []>("op_879_split_sizes_0"), val = tensor<int32, [2]>([144, 144])];
            tensor<int32, []> var_879_axis_0 = const()[name = tensor<string, []>("op_879_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 144, 48]> var_879_cast_fp16_0, tensor<fp16, [1, 144, 48]> var_879_cast_fp16_1 = split(axis = var_879_axis_0, split_sizes = var_879_split_sizes_0, x = inputs_35_cast_fp16)[name = tensor<string, []>("op_879_cast_fp16")];
            tensor<fp16, [1, 144, 48]> var_881_cast_fp16 = sigmoid(x = var_879_cast_fp16_1)[name = tensor<string, []>("op_881_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_169_cast_fp16 = mul(x = var_879_cast_fp16_0, y = var_881_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
            tensor<int32, [1]> var_885 = const()[name = tensor<string, []>("op_885"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_887 = const()[name = tensor<string, []>("op_887"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_171_pad_type_0 = const()[name = tensor<string, []>("input_171_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_171_pad_0 = const()[name = tensor<string, []>("input_171_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [144, 1, 3]> inputs_37_weight_0_to_fp16 = const()[name = tensor<string, []>("inputs_37_weight_0_to_fp16"), val = tensor<fp16, [144, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2206080)))];
            tensor<fp16, [144]> inputs_37_bias_0_to_fp16 = const()[name = tensor<string, []>("inputs_37_bias_0_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2207040)))];
            tensor<fp16, [1, 144, 48]> inputs_37_cast_fp16 = conv(bias = inputs_37_bias_0_to_fp16, dilations = var_887, groups = var_13, pad = input_171_pad_0, pad_type = input_171_pad_type_0, strides = var_885, weight = inputs_37_weight_0_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("inputs_37_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_173_cast_fp16 = silu(x = inputs_37_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
            tensor<int32, [1]> var_900 = const()[name = tensor<string, []>("op_900"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_902 = const()[name = tensor<string, []>("op_902"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_175_pad_type_0 = const()[name = tensor<string, []>("input_175_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_175_pad_0 = const()[name = tensor<string, []>("input_175_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [144, 144, 1]> encoder_layers_3_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2207424))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2228480))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2228224)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_2_module_sequential_7_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_2_module_sequential_7_conv_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2228864)))];
            tensor<fp16, [1, 144, 48]> input_175_cast_fp16 = conv(bias = encoder_layers_3_sequential_2_module_sequential_7_conv_bias_to_fp16, dilations = var_902, groups = var_15, pad = input_175_pad_0, pad_type = input_175_pad_type_0, strides = var_900, weight = encoder_layers_3_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized, x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
            tensor<int32, [3]> var_906_perm_0 = const()[name = tensor<string, []>("op_906_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 48, 144]> transpose_50 = transpose(perm = var_906_perm_0, x = input_175_cast_fp16)[name = tensor<string, []>("transpose_50")];
            tensor<fp16, [1, 48, 144]> input_177_cast_fp16 = add(x = transpose_50, y = input_165_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
            tensor<int32, [1]> input_179_axes_0 = const()[name = tensor<string, []>("input_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_3_sequential_3_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_3_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2229248)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_3_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_3_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2229632)))];
            tensor<fp16, [1, 48, 144]> input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = encoder_layers_3_sequential_3_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_3_sequential_3_module_sequential_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_3_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2230016))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2313664))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2313024)))];
            tensor<fp16, [576]> encoder_layers_3_sequential_3_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_3_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2314880)))];
            tensor<fp16, [1, 48, 576]> linear_35_cast_fp16 = linear(bias = encoder_layers_3_sequential_3_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_3_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_179_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_181_cast_fp16 = silu(x = linear_35_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_3_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2316096))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2399360))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2399104)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_3_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_3_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2399744)))];
            tensor<fp16, [1, 48, 144]> linear_36_cast_fp16 = linear(bias = encoder_layers_3_sequential_3_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_3_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_181_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<fp16, []> var_933_to_fp16 = const()[name = tensor<string, []>("op_933_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_934_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_933_to_fp16)[name = tensor<string, []>("op_934_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_187_cast_fp16 = add(x = var_934_cast_fp16, y = input_177_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
            tensor<int32, [1]> input_189_axes_0 = const()[name = tensor<string, []>("input_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_3_sequential_4_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_4_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2400128)))];
            tensor<fp16, [144]> encoder_layers_3_sequential_4_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_sequential_4_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2400512)))];
            tensor<fp16, [1, 48, 144]> input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = encoder_layers_3_sequential_4_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_3_sequential_4_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
            tensor<int32, [1]> input_191_axes_0 = const()[name = tensor<string, []>("input_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_4_sequential_0_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_0_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2400896)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_0_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_0_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2401280)))];
            tensor<fp16, [1, 48, 144]> input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = encoder_layers_4_sequential_0_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_4_sequential_0_module_sequential_0_weight_to_fp16, x = input_189_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_4_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2401664))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2485312))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2484672)))];
            tensor<fp16, [576]> encoder_layers_4_sequential_0_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_0_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2486528)))];
            tensor<fp16, [1, 48, 576]> linear_37_cast_fp16 = linear(bias = encoder_layers_4_sequential_0_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_4_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_191_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_193_cast_fp16 = silu(x = linear_37_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_4_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2487744))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2571008))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2570752)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_0_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_0_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2571392)))];
            tensor<fp16, [1, 48, 144]> linear_38_cast_fp16 = linear(bias = encoder_layers_4_sequential_0_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_4_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_193_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<fp16, []> var_969_to_fp16 = const()[name = tensor<string, []>("op_969_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_970_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_969_to_fp16)[name = tensor<string, []>("op_970_cast_fp16")];
            tensor<fp16, [1, 48, 144]> inputs_43_cast_fp16 = add(x = var_970_cast_fp16, y = input_189_cast_fp16)[name = tensor<string, []>("inputs_43_cast_fp16")];
            tensor<int32, [1]> query_17_axes_0 = const()[name = tensor<string, []>("query_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_4_sequential_1_module_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_1_module_layer_norm_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2571776)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_1_module_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_1_module_layer_norm_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2572160)))];
            tensor<fp16, [1, 48, 144]> query_17_cast_fp16 = layer_norm(axes = query_17_axes_0, beta = encoder_layers_4_sequential_1_module_layer_norm_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_4_sequential_1_module_layer_norm_weight_to_fp16, x = inputs_43_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_4_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2572544))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2593600))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2593344)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_1_module_attention_query_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_query_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2593984)))];
            tensor<fp16, [1, 48, 144]> linear_39_cast_fp16 = linear(bias = encoder_layers_4_sequential_1_module_attention_query_proj_linear_bias_to_fp16, weight = encoder_layers_4_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized, x = query_17_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<int32, [4]> var_1001 = const()[name = tensor<string, []>("op_1001"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> query_19_cast_fp16 = reshape(shape = var_1001, x = linear_39_cast_fp16)[name = tensor<string, []>("query_19_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_4_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2594368))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2615424))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2615168)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_1_module_attention_key_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_key_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2615808)))];
            tensor<fp16, [1, 48, 144]> linear_40_cast_fp16 = linear(bias = encoder_layers_4_sequential_1_module_attention_key_proj_linear_bias_to_fp16, weight = encoder_layers_4_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized, x = query_17_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<int32, [4]> var_1007 = const()[name = tensor<string, []>("op_1007"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_1008_cast_fp16 = reshape(shape = var_1007, x = linear_40_cast_fp16)[name = tensor<string, []>("op_1008_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_4_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2616192))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2637248))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2636992)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_1_module_attention_value_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_value_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2637632)))];
            tensor<fp16, [1, 48, 144]> linear_41_cast_fp16 = linear(bias = encoder_layers_4_sequential_1_module_attention_value_proj_linear_bias_to_fp16, weight = encoder_layers_4_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized, x = query_17_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<int32, [4]> var_1015 = const()[name = tensor<string, []>("op_1015"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_1016_cast_fp16 = reshape(shape = var_1015, x = linear_41_cast_fp16)[name = tensor<string, []>("op_1016_cast_fp16")];
            tensor<int32, [4]> var_1017 = const()[name = tensor<string, []>("op_1017"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [4, 36]> encoder_layers_4_sequential_1_module_attention_u_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_u_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2638016)))];
            tensor<fp16, [1, 48, 4, 36]> var_1024_cast_fp16 = add(x = query_19_cast_fp16, y = encoder_layers_4_sequential_1_module_attention_u_bias_to_fp16)[name = tensor<string, []>("op_1024_cast_fp16")];
            tensor<bool, []> content_score_9_transpose_x_0 = const()[name = tensor<string, []>("content_score_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> content_score_9_transpose_y_0 = const()[name = tensor<string, []>("content_score_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = tensor<string, []>("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_33_perm_0 = const()[name = tensor<string, []>("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 36, 48]> transpose_48 = transpose(perm = transpose_33_perm_0, x = var_1008_cast_fp16)[name = tensor<string, []>("transpose_48")];
            tensor<fp16, [1, 4, 48, 36]> transpose_49 = transpose(perm = transpose_32_perm_0, x = var_1024_cast_fp16)[name = tensor<string, []>("transpose_49")];
            tensor<fp16, [1, 4, 48, 48]> content_score_9_cast_fp16 = matmul(transpose_x = content_score_9_transpose_x_0, transpose_y = content_score_9_transpose_y_0, x = transpose_49, y = transpose_48)[name = tensor<string, []>("content_score_9_cast_fp16")];
            tensor<fp16, [4, 36]> encoder_layers_4_sequential_1_module_attention_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_v_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2638400)))];
            tensor<fp16, [1, 48, 4, 36]> var_1028_cast_fp16 = add(x = query_19_cast_fp16, y = encoder_layers_4_sequential_1_module_attention_v_bias_to_fp16)[name = tensor<string, []>("op_1028_cast_fp16")];
            tensor<int32, [4]> var_1029_perm_0 = const()[name = tensor<string, []>("op_1029_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> pos_score_17_transpose_x_0 = const()[name = tensor<string, []>("pos_score_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_score_17_transpose_y_0 = const()[name = tensor<string, []>("pos_score_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 36, 48]> op_1031_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(-1), name = tensor<string, []>("op_1031_to_fp16_affine_quantized"), quantized_data = tensor<int8, [1, 4, 36, 48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2638784))), scale = tensor<fp16, []>(0x1.66p-5), zero_point = tensor<int8, []>(-13)];
            tensor<fp16, [1, 4, 48, 36]> transpose_47 = transpose(perm = var_1029_perm_0, x = var_1028_cast_fp16)[name = tensor<string, []>("transpose_47")];
            tensor<fp16, [1, 4, 48, 48]> pos_score_17_cast_fp16 = matmul(transpose_x = pos_score_17_transpose_x_0, transpose_y = pos_score_17_transpose_y_0, x = transpose_47, y = op_1031_to_fp16_affine_quantized)[name = tensor<string, []>("pos_score_17_cast_fp16")];
            tensor<bool, []> padded_pos_score_17_interleave_0 = const()[name = tensor<string, []>("padded_pos_score_17_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 49]> padded_pos_score_17_cast_fp16 = concat(axis = var_5, interleave = padded_pos_score_17_interleave_0, values = (zeros_1_to_fp16, pos_score_17_cast_fp16))[name = tensor<string, []>("padded_pos_score_17_cast_fp16")];
            tensor<int32, [4]> var_1045 = const()[name = tensor<string, []>("op_1045"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<fp16, [1, 4, 49, 48]> padded_pos_score_19_cast_fp16 = reshape(shape = var_1045, x = padded_pos_score_17_cast_fp16)[name = tensor<string, []>("padded_pos_score_19_cast_fp16")];
            tensor<int32, [4]> var_1049_begin_0 = const()[name = tensor<string, []>("op_1049_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1049_end_0 = const()[name = tensor<string, []>("op_1049_end_0"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<bool, [4]> var_1049_end_mask_0 = const()[name = tensor<string, []>("op_1049_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 48]> var_1049_cast_fp16 = slice_by_index(begin = var_1049_begin_0, end = var_1049_end_0, end_mask = var_1049_end_mask_0, x = padded_pos_score_19_cast_fp16)[name = tensor<string, []>("op_1049_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_1056_cast_fp16 = add(x = content_score_9_cast_fp16, y = var_1049_cast_fp16)[name = tensor<string, []>("op_1056_cast_fp16")];
            tensor<fp16, []> _inversed_input_201_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_201_y_0_to_fp16"), val = tensor<fp16, []>(0x1.554p-4)];
            tensor<fp16, [1, 4, 48, 48]> _inversed_input_201_cast_fp16 = mul(x = var_1056_cast_fp16, y = _inversed_input_201_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_201_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> input_203_cast_fp16 = softmax(axis = var_5, x = _inversed_input_201_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<bool, []> var_1061_transpose_x_0 = const()[name = tensor<string, []>("op_1061_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1061_transpose_y_0 = const()[name = tensor<string, []>("op_1061_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 36]> transpose_46 = transpose(perm = var_1017, x = var_1016_cast_fp16)[name = tensor<string, []>("transpose_46")];
            tensor<fp16, [1, 4, 48, 36]> var_1061_cast_fp16 = matmul(transpose_x = var_1061_transpose_x_0, transpose_y = var_1061_transpose_y_0, x = input_203_cast_fp16, y = transpose_46)[name = tensor<string, []>("op_1061_cast_fp16")];
            tensor<int32, [4]> context_9_perm_0 = const()[name = tensor<string, []>("context_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1064 = const()[name = tensor<string, []>("op_1064"), val = tensor<int32, [3]>([1, -1, 144])];
            tensor<fp16, [1, 48, 4, 36]> transpose_45 = transpose(perm = context_9_perm_0, x = var_1061_cast_fp16)[name = tensor<string, []>("transpose_45")];
            tensor<fp16, [1, 48, 144]> input_205_cast_fp16 = reshape(shape = var_1064, x = transpose_45)[name = tensor<string, []>("input_205_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_4_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2645760))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2666816))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2666560)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_1_module_attention_out_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_1_module_attention_out_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2667200)))];
            tensor<fp16, [1, 48, 144]> linear_43_cast_fp16 = linear(bias = encoder_layers_4_sequential_1_module_attention_out_proj_linear_bias_to_fp16, weight = encoder_layers_4_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized, x = input_205_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_209_cast_fp16 = add(x = linear_43_cast_fp16, y = inputs_43_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
            tensor<int32, [1]> x_9_axes_0 = const()[name = tensor<string, []>("x_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_4_sequential_2_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_2_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2667584)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_2_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_2_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2667968)))];
            tensor<fp16, [1, 48, 144]> x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = encoder_layers_4_sequential_2_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_4_sequential_2_module_sequential_0_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
            tensor<int32, [3]> input_211_perm_0 = const()[name = tensor<string, []>("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1091 = const()[name = tensor<string, []>("op_1091"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_1093 = const()[name = tensor<string, []>("op_1093"), val = tensor<int32, [1]>([1])];
            tensor<string, []> inputs_45_pad_type_0 = const()[name = tensor<string, []>("inputs_45_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> inputs_45_pad_0 = const()[name = tensor<string, []>("inputs_45_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [288, 144, 1]> encoder_layers_4_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [288, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2668352))), scale = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2710272))), zero_point = tensor<int8, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2709888)))];
            tensor<fp16, [288]> encoder_layers_4_sequential_2_module_sequential_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_2_module_sequential_2_conv_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2710912)))];
            tensor<fp16, [1, 144, 48]> transpose_44 = transpose(perm = input_211_perm_0, x = x_9_cast_fp16)[name = tensor<string, []>("transpose_44")];
            tensor<fp16, [1, 288, 48]> inputs_45_cast_fp16 = conv(bias = encoder_layers_4_sequential_2_module_sequential_2_conv_bias_to_fp16, dilations = var_1093, groups = var_15, pad = inputs_45_pad_0, pad_type = inputs_45_pad_type_0, strides = var_1091, weight = encoder_layers_4_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized, x = transpose_44)[name = tensor<string, []>("inputs_45_cast_fp16")];
            tensor<int32, [2]> var_1096_split_sizes_0 = const()[name = tensor<string, []>("op_1096_split_sizes_0"), val = tensor<int32, [2]>([144, 144])];
            tensor<int32, []> var_1096_axis_0 = const()[name = tensor<string, []>("op_1096_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 144, 48]> var_1096_cast_fp16_0, tensor<fp16, [1, 144, 48]> var_1096_cast_fp16_1 = split(axis = var_1096_axis_0, split_sizes = var_1096_split_sizes_0, x = inputs_45_cast_fp16)[name = tensor<string, []>("op_1096_cast_fp16")];
            tensor<fp16, [1, 144, 48]> var_1098_cast_fp16 = sigmoid(x = var_1096_cast_fp16_1)[name = tensor<string, []>("op_1098_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_213_cast_fp16 = mul(x = var_1096_cast_fp16_0, y = var_1098_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
            tensor<int32, [1]> var_1102 = const()[name = tensor<string, []>("op_1102"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_1104 = const()[name = tensor<string, []>("op_1104"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_215_pad_type_0 = const()[name = tensor<string, []>("input_215_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_215_pad_0 = const()[name = tensor<string, []>("input_215_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [144, 1, 3]> inputs_47_weight_0_to_fp16 = const()[name = tensor<string, []>("inputs_47_weight_0_to_fp16"), val = tensor<fp16, [144, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2711552)))];
            tensor<fp16, [144]> inputs_47_bias_0_to_fp16 = const()[name = tensor<string, []>("inputs_47_bias_0_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2712512)))];
            tensor<fp16, [1, 144, 48]> inputs_47_cast_fp16 = conv(bias = inputs_47_bias_0_to_fp16, dilations = var_1104, groups = var_13, pad = input_215_pad_0, pad_type = input_215_pad_type_0, strides = var_1102, weight = inputs_47_weight_0_to_fp16, x = input_213_cast_fp16)[name = tensor<string, []>("inputs_47_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_217_cast_fp16 = silu(x = inputs_47_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
            tensor<int32, [1]> var_1117 = const()[name = tensor<string, []>("op_1117"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_1119 = const()[name = tensor<string, []>("op_1119"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_219_pad_type_0 = const()[name = tensor<string, []>("input_219_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_219_pad_0 = const()[name = tensor<string, []>("input_219_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [144, 144, 1]> encoder_layers_4_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2712896))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2733952))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2733696)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_2_module_sequential_7_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_2_module_sequential_7_conv_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2734336)))];
            tensor<fp16, [1, 144, 48]> input_219_cast_fp16 = conv(bias = encoder_layers_4_sequential_2_module_sequential_7_conv_bias_to_fp16, dilations = var_1119, groups = var_15, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = var_1117, weight = encoder_layers_4_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized, x = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
            tensor<int32, [3]> var_1123_perm_0 = const()[name = tensor<string, []>("op_1123_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 48, 144]> transpose_43 = transpose(perm = var_1123_perm_0, x = input_219_cast_fp16)[name = tensor<string, []>("transpose_43")];
            tensor<fp16, [1, 48, 144]> input_221_cast_fp16 = add(x = transpose_43, y = input_209_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
            tensor<int32, [1]> input_223_axes_0 = const()[name = tensor<string, []>("input_223_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_4_sequential_3_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_3_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2734720)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_3_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_3_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2735104)))];
            tensor<fp16, [1, 48, 144]> input_223_cast_fp16 = layer_norm(axes = input_223_axes_0, beta = encoder_layers_4_sequential_3_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_4_sequential_3_module_sequential_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_4_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2735488))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2819136))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2818496)))];
            tensor<fp16, [576]> encoder_layers_4_sequential_3_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_3_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2820352)))];
            tensor<fp16, [1, 48, 576]> linear_44_cast_fp16 = linear(bias = encoder_layers_4_sequential_3_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_4_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_223_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_225_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_4_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2821568))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2904832))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2904576)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_3_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_3_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2905216)))];
            tensor<fp16, [1, 48, 144]> linear_45_cast_fp16 = linear(bias = encoder_layers_4_sequential_3_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_4_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_225_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<fp16, []> var_1150_to_fp16 = const()[name = tensor<string, []>("op_1150_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_1151_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1150_to_fp16)[name = tensor<string, []>("op_1151_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_231_cast_fp16 = add(x = var_1151_cast_fp16, y = input_221_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
            tensor<int32, [1]> input_233_axes_0 = const()[name = tensor<string, []>("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_4_sequential_4_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_4_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2905600)))];
            tensor<fp16, [144]> encoder_layers_4_sequential_4_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_sequential_4_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2905984)))];
            tensor<fp16, [1, 48, 144]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_4_sequential_4_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_4_sequential_4_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_5_sequential_0_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_0_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2906368)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_0_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_0_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2906752)))];
            tensor<fp16, [1, 48, 144]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_5_sequential_0_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_5_sequential_0_module_sequential_0_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_5_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2907136))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2990784))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2990144)))];
            tensor<fp16, [576]> encoder_layers_5_sequential_0_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_0_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2992000)))];
            tensor<fp16, [1, 48, 576]> linear_46_cast_fp16 = linear(bias = encoder_layers_5_sequential_0_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_5_sequential_0_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_235_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_237_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor<string, []>("input_237_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_5_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2993216))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3076480))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3076224)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_0_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_0_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3076864)))];
            tensor<fp16, [1, 48, 144]> linear_47_cast_fp16 = linear(bias = encoder_layers_5_sequential_0_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_5_sequential_0_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_237_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<fp16, []> var_1186_to_fp16 = const()[name = tensor<string, []>("op_1186_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_1187_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1186_to_fp16)[name = tensor<string, []>("op_1187_cast_fp16")];
            tensor<fp16, [1, 48, 144]> inputs_53_cast_fp16 = add(x = var_1187_cast_fp16, y = input_233_cast_fp16)[name = tensor<string, []>("inputs_53_cast_fp16")];
            tensor<int32, [1]> query_21_axes_0 = const()[name = tensor<string, []>("query_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_5_sequential_1_module_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_1_module_layer_norm_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3077248)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_1_module_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_1_module_layer_norm_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3077632)))];
            tensor<fp16, [1, 48, 144]> query_21_cast_fp16 = layer_norm(axes = query_21_axes_0, beta = encoder_layers_5_sequential_1_module_layer_norm_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_5_sequential_1_module_layer_norm_weight_to_fp16, x = inputs_53_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_5_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3078016))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3099072))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3098816)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_1_module_attention_query_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_query_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3099456)))];
            tensor<fp16, [1, 48, 144]> linear_48_cast_fp16 = linear(bias = encoder_layers_5_sequential_1_module_attention_query_proj_linear_bias_to_fp16, weight = encoder_layers_5_sequential_1_module_attention_query_proj_linear_weight_to_fp16_affine_quantized, x = query_21_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<int32, [4]> var_1218 = const()[name = tensor<string, []>("op_1218"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> query_cast_fp16 = reshape(shape = var_1218, x = linear_48_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_5_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3099840))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3120896))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3120640)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_1_module_attention_key_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_key_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3121280)))];
            tensor<fp16, [1, 48, 144]> linear_49_cast_fp16 = linear(bias = encoder_layers_5_sequential_1_module_attention_key_proj_linear_bias_to_fp16, weight = encoder_layers_5_sequential_1_module_attention_key_proj_linear_weight_to_fp16_affine_quantized, x = query_21_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1224 = const()[name = tensor<string, []>("op_1224"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_1225_cast_fp16 = reshape(shape = var_1224, x = linear_49_cast_fp16)[name = tensor<string, []>("op_1225_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_5_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3121664))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3142720))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3142464)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_1_module_attention_value_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_value_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3143104)))];
            tensor<fp16, [1, 48, 144]> linear_50_cast_fp16 = linear(bias = encoder_layers_5_sequential_1_module_attention_value_proj_linear_bias_to_fp16, weight = encoder_layers_5_sequential_1_module_attention_value_proj_linear_weight_to_fp16_affine_quantized, x = query_21_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1232 = const()[name = tensor<string, []>("op_1232"), val = tensor<int32, [4]>([1, -1, 4, 36])];
            tensor<fp16, [1, 48, 4, 36]> var_1233_cast_fp16 = reshape(shape = var_1232, x = linear_50_cast_fp16)[name = tensor<string, []>("op_1233_cast_fp16")];
            tensor<int32, [4]> var_1234 = const()[name = tensor<string, []>("op_1234"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [4, 36]> encoder_layers_5_sequential_1_module_attention_u_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_u_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3143488)))];
            tensor<fp16, [1, 48, 4, 36]> var_1241_cast_fp16 = add(x = query_cast_fp16, y = encoder_layers_5_sequential_1_module_attention_u_bias_to_fp16)[name = tensor<string, []>("op_1241_cast_fp16")];
            tensor<bool, []> content_score_transpose_x_0 = const()[name = tensor<string, []>("content_score_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> content_score_transpose_y_0 = const()[name = tensor<string, []>("content_score_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = tensor<string, []>("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_35_perm_0 = const()[name = tensor<string, []>("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 4, 36, 48]> transpose_41 = transpose(perm = transpose_35_perm_0, x = var_1225_cast_fp16)[name = tensor<string, []>("transpose_41")];
            tensor<fp16, [1, 4, 48, 36]> transpose_42 = transpose(perm = transpose_34_perm_0, x = var_1241_cast_fp16)[name = tensor<string, []>("transpose_42")];
            tensor<fp16, [1, 4, 48, 48]> content_score_cast_fp16 = matmul(transpose_x = content_score_transpose_x_0, transpose_y = content_score_transpose_y_0, x = transpose_42, y = transpose_41)[name = tensor<string, []>("content_score_cast_fp16")];
            tensor<fp16, [4, 36]> encoder_layers_5_sequential_1_module_attention_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_v_bias_to_fp16"), val = tensor<fp16, [4, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3143872)))];
            tensor<fp16, [1, 48, 4, 36]> var_1245_cast_fp16 = add(x = query_cast_fp16, y = encoder_layers_5_sequential_1_module_attention_v_bias_to_fp16)[name = tensor<string, []>("op_1245_cast_fp16")];
            tensor<int32, [4]> var_1246_perm_0 = const()[name = tensor<string, []>("op_1246_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> pos_score_21_transpose_x_0 = const()[name = tensor<string, []>("pos_score_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_score_21_transpose_y_0 = const()[name = tensor<string, []>("pos_score_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 36, 48]> op_1248_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(-1), name = tensor<string, []>("op_1248_to_fp16_affine_quantized"), quantized_data = tensor<int8, [1, 4, 36, 48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3144256))), scale = tensor<fp16, []>(0x1.76cp-5), zero_point = tensor<int8, []>(9)];
            tensor<fp16, [1, 4, 48, 36]> transpose_40 = transpose(perm = var_1246_perm_0, x = var_1245_cast_fp16)[name = tensor<string, []>("transpose_40")];
            tensor<fp16, [1, 4, 48, 48]> pos_score_21_cast_fp16 = matmul(transpose_x = pos_score_21_transpose_x_0, transpose_y = pos_score_21_transpose_y_0, x = transpose_40, y = op_1248_to_fp16_affine_quantized)[name = tensor<string, []>("pos_score_21_cast_fp16")];
            tensor<bool, []> padded_pos_score_21_interleave_0 = const()[name = tensor<string, []>("padded_pos_score_21_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 49]> padded_pos_score_21_cast_fp16 = concat(axis = var_5, interleave = padded_pos_score_21_interleave_0, values = (zeros_1_to_fp16, pos_score_21_cast_fp16))[name = tensor<string, []>("padded_pos_score_21_cast_fp16")];
            tensor<int32, [4]> var_1262 = const()[name = tensor<string, []>("op_1262"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<fp16, [1, 4, 49, 48]> padded_pos_score_cast_fp16 = reshape(shape = var_1262, x = padded_pos_score_21_cast_fp16)[name = tensor<string, []>("padded_pos_score_cast_fp16")];
            tensor<int32, [4]> var_1266_begin_0 = const()[name = tensor<string, []>("op_1266_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1266_end_0 = const()[name = tensor<string, []>("op_1266_end_0"), val = tensor<int32, [4]>([1, 4, 49, 48])];
            tensor<bool, [4]> var_1266_end_mask_0 = const()[name = tensor<string, []>("op_1266_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 48]> var_1266_cast_fp16 = slice_by_index(begin = var_1266_begin_0, end = var_1266_end_0, end_mask = var_1266_end_mask_0, x = padded_pos_score_cast_fp16)[name = tensor<string, []>("op_1266_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_1273_cast_fp16 = add(x = content_score_cast_fp16, y = var_1266_cast_fp16)[name = tensor<string, []>("op_1273_cast_fp16")];
            tensor<fp16, []> _inversed_input_245_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_245_y_0_to_fp16"), val = tensor<fp16, []>(0x1.554p-4)];
            tensor<fp16, [1, 4, 48, 48]> _inversed_input_245_cast_fp16 = mul(x = var_1273_cast_fp16, y = _inversed_input_245_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_245_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> input_247_cast_fp16 = softmax(axis = var_5, x = _inversed_input_245_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
            tensor<bool, []> var_1278_transpose_x_0 = const()[name = tensor<string, []>("op_1278_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1278_transpose_y_0 = const()[name = tensor<string, []>("op_1278_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 48, 36]> transpose_39 = transpose(perm = var_1234, x = var_1233_cast_fp16)[name = tensor<string, []>("transpose_39")];
            tensor<fp16, [1, 4, 48, 36]> var_1278_cast_fp16 = matmul(transpose_x = var_1278_transpose_x_0, transpose_y = var_1278_transpose_y_0, x = input_247_cast_fp16, y = transpose_39)[name = tensor<string, []>("op_1278_cast_fp16")];
            tensor<int32, [4]> context_perm_0 = const()[name = tensor<string, []>("context_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1281 = const()[name = tensor<string, []>("op_1281"), val = tensor<int32, [3]>([1, -1, 144])];
            tensor<fp16, [1, 48, 4, 36]> transpose_38 = transpose(perm = context_perm_0, x = var_1278_cast_fp16)[name = tensor<string, []>("transpose_38")];
            tensor<fp16, [1, 48, 144]> input_249_cast_fp16 = reshape(shape = var_1281, x = transpose_38)[name = tensor<string, []>("input_249_cast_fp16")];
            tensor<fp16, [144, 144]> encoder_layers_5_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3151232))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3172288))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3172032)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_1_module_attention_out_proj_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_1_module_attention_out_proj_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3172672)))];
            tensor<fp16, [1, 48, 144]> linear_52_cast_fp16 = linear(bias = encoder_layers_5_sequential_1_module_attention_out_proj_linear_bias_to_fp16, weight = encoder_layers_5_sequential_1_module_attention_out_proj_linear_weight_to_fp16_affine_quantized, x = input_249_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_253_cast_fp16 = add(x = linear_52_cast_fp16, y = inputs_53_cast_fp16)[name = tensor<string, []>("input_253_cast_fp16")];
            tensor<int32, [1]> x_axes_0 = const()[name = tensor<string, []>("x_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_5_sequential_2_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_2_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3173056)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_2_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_2_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3173440)))];
            tensor<fp16, [1, 48, 144]> x_cast_fp16 = layer_norm(axes = x_axes_0, beta = encoder_layers_5_sequential_2_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_5_sequential_2_module_sequential_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
            tensor<int32, [3]> input_255_perm_0 = const()[name = tensor<string, []>("input_255_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1308 = const()[name = tensor<string, []>("op_1308"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_1310 = const()[name = tensor<string, []>("op_1310"), val = tensor<int32, [1]>([1])];
            tensor<string, []> inputs_55_pad_type_0 = const()[name = tensor<string, []>("inputs_55_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> inputs_55_pad_0 = const()[name = tensor<string, []>("inputs_55_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [288, 144, 1]> encoder_layers_5_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [288, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3173824))), scale = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3215744))), zero_point = tensor<int8, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3215360)))];
            tensor<fp16, [288]> encoder_layers_5_sequential_2_module_sequential_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_2_module_sequential_2_conv_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3216384)))];
            tensor<fp16, [1, 144, 48]> transpose_37 = transpose(perm = input_255_perm_0, x = x_cast_fp16)[name = tensor<string, []>("transpose_37")];
            tensor<fp16, [1, 288, 48]> inputs_55_cast_fp16 = conv(bias = encoder_layers_5_sequential_2_module_sequential_2_conv_bias_to_fp16, dilations = var_1310, groups = var_15, pad = inputs_55_pad_0, pad_type = inputs_55_pad_type_0, strides = var_1308, weight = encoder_layers_5_sequential_2_module_sequential_2_conv_weight_to_fp16_affine_quantized, x = transpose_37)[name = tensor<string, []>("inputs_55_cast_fp16")];
            tensor<int32, [2]> var_1313_split_sizes_0 = const()[name = tensor<string, []>("op_1313_split_sizes_0"), val = tensor<int32, [2]>([144, 144])];
            tensor<int32, []> var_1313_axis_0 = const()[name = tensor<string, []>("op_1313_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 144, 48]> var_1313_cast_fp16_0, tensor<fp16, [1, 144, 48]> var_1313_cast_fp16_1 = split(axis = var_1313_axis_0, split_sizes = var_1313_split_sizes_0, x = inputs_55_cast_fp16)[name = tensor<string, []>("op_1313_cast_fp16")];
            tensor<fp16, [1, 144, 48]> var_1315_cast_fp16 = sigmoid(x = var_1313_cast_fp16_1)[name = tensor<string, []>("op_1315_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_257_cast_fp16 = mul(x = var_1313_cast_fp16_0, y = var_1315_cast_fp16)[name = tensor<string, []>("input_257_cast_fp16")];
            tensor<int32, [1]> var_1319 = const()[name = tensor<string, []>("op_1319"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_1321 = const()[name = tensor<string, []>("op_1321"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_259_pad_type_0 = const()[name = tensor<string, []>("input_259_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_259_pad_0 = const()[name = tensor<string, []>("input_259_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [144, 1, 3]> inputs_57_weight_0_to_fp16 = const()[name = tensor<string, []>("inputs_57_weight_0_to_fp16"), val = tensor<fp16, [144, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3217024)))];
            tensor<fp16, [144]> inputs_57_bias_0_to_fp16 = const()[name = tensor<string, []>("inputs_57_bias_0_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3217984)))];
            tensor<fp16, [1, 144, 48]> inputs_57_cast_fp16 = conv(bias = inputs_57_bias_0_to_fp16, dilations = var_1321, groups = var_13, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = var_1319, weight = inputs_57_weight_0_to_fp16, x = input_257_cast_fp16)[name = tensor<string, []>("inputs_57_cast_fp16")];
            tensor<fp16, [1, 144, 48]> input_261_cast_fp16 = silu(x = inputs_57_cast_fp16)[name = tensor<string, []>("input_261_cast_fp16")];
            tensor<int32, [1]> var_1334 = const()[name = tensor<string, []>("op_1334"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_1336 = const()[name = tensor<string, []>("op_1336"), val = tensor<int32, [1]>([1])];
            tensor<string, []> input_263_pad_type_0 = const()[name = tensor<string, []>("input_263_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_263_pad_0 = const()[name = tensor<string, []>("input_263_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [144, 144, 1]> encoder_layers_5_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 144, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3218368))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3239424))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3239168)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_2_module_sequential_7_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_2_module_sequential_7_conv_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3239808)))];
            tensor<fp16, [1, 144, 48]> input_263_cast_fp16 = conv(bias = encoder_layers_5_sequential_2_module_sequential_7_conv_bias_to_fp16, dilations = var_1336, groups = var_15, pad = input_263_pad_0, pad_type = input_263_pad_type_0, strides = var_1334, weight = encoder_layers_5_sequential_2_module_sequential_7_conv_weight_to_fp16_affine_quantized, x = input_261_cast_fp16)[name = tensor<string, []>("input_263_cast_fp16")];
            tensor<int32, [3]> var_1340_perm_0 = const()[name = tensor<string, []>("op_1340_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 48, 144]> transpose_36 = transpose(perm = var_1340_perm_0, x = input_263_cast_fp16)[name = tensor<string, []>("transpose_36")];
            tensor<fp16, [1, 48, 144]> input_265_cast_fp16 = add(x = transpose_36, y = input_253_cast_fp16)[name = tensor<string, []>("input_265_cast_fp16")];
            tensor<int32, [1]> input_267_axes_0 = const()[name = tensor<string, []>("input_267_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_5_sequential_3_module_sequential_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_3_module_sequential_0_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3240192)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_3_module_sequential_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_3_module_sequential_0_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3240576)))];
            tensor<fp16, [1, 48, 144]> input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = encoder_layers_5_sequential_3_module_sequential_0_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_5_sequential_3_module_sequential_0_weight_to_fp16, x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
            tensor<fp16, [576, 144]> encoder_layers_5_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [576, 144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3240960))), scale = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3324608))), zero_point = tensor<int8, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3323968)))];
            tensor<fp16, [576]> encoder_layers_5_sequential_3_module_sequential_1_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_3_module_sequential_1_linear_bias_to_fp16"), val = tensor<fp16, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3325824)))];
            tensor<fp16, [1, 48, 576]> linear_53_cast_fp16 = linear(bias = encoder_layers_5_sequential_3_module_sequential_1_linear_bias_to_fp16, weight = encoder_layers_5_sequential_3_module_sequential_1_linear_weight_to_fp16_affine_quantized, x = input_267_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
            tensor<fp16, [1, 48, 576]> input_269_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor<string, []>("input_269_cast_fp16")];
            tensor<fp16, [144, 576]> encoder_layers_5_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [144, 576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3327040))), scale = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3410304))), zero_point = tensor<int8, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3410048)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_3_module_sequential_4_linear_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_3_module_sequential_4_linear_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3410688)))];
            tensor<fp16, [1, 48, 144]> linear_54_cast_fp16 = linear(bias = encoder_layers_5_sequential_3_module_sequential_4_linear_bias_to_fp16, weight = encoder_layers_5_sequential_3_module_sequential_4_linear_weight_to_fp16_affine_quantized, x = input_269_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
            tensor<fp16, []> var_1367_to_fp16 = const()[name = tensor<string, []>("op_1367_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 48, 144]> var_1368_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1367_to_fp16)[name = tensor<string, []>("op_1368_cast_fp16")];
            tensor<fp16, [1, 48, 144]> input_cast_fp16 = add(x = var_1368_cast_fp16, y = input_265_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<int32, [1]> var_1375_axes_0 = const()[name = tensor<string, []>("op_1375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [144]> encoder_layers_5_sequential_4_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_4_weight_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3411072)))];
            tensor<fp16, [144]> encoder_layers_5_sequential_4_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_sequential_4_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3411456)))];
            tensor<fp16, [1, 48, 144]> enc_3d_output = layer_norm(axes = var_1375_axes_0, beta = encoder_layers_5_sequential_4_bias_to_fp16, epsilon = var_12_to_fp16, gamma = encoder_layers_5_sequential_4_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("op_1375_cast_fp16")];
        } -> (enc_3d_output);
}