program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.7.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios17>(tensor<int32, [1, 512]> tokens, tensor<fp32, [1, 1, 512]> x_mask) {
            tensor<int32, []> var_13_batch_dims_0 = const()[name = tensor<string, []>("op_13_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_13_validate_indices_0 = const()[name = tensor<string, []>("op_13_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [178, 72]> enc_p_emb_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_emb_weight_to_fp16"), val = tensor<fp16, [178, 72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<string, []> tokens_to_int16_dtype_0 = const()[name = tensor<string, []>("tokens_to_int16_dtype_0"), val = tensor<string, []>("int16")];
            tensor<string, []> cast_40_dtype_0 = const()[name = tensor<string, []>("cast_40_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
            tensor<int16, [1, 512]> tokens_to_int16 = cast(dtype = tokens_to_int16_dtype_0, x = tokens)[name = tensor<string, []>("cast_44")];
            tensor<int32, [1, 512]> cast_40 = cast(dtype = cast_40_dtype_0, x = tokens_to_int16)[name = tensor<string, []>("cast_43")];
            tensor<bool, [1, 512]> greater_equal_0 = greater_equal(x = cast_40, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(178)];
            tensor<int32, [1, 512]> add_0 = add(x = cast_40, y = slice_by_index_0)[name = tensor<string, []>("add_0")];
            tensor<int32, [1, 512]> select_0 = select(a = cast_40, b = add_0, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
            tensor<int32, []> var_13_cast_fp16_cast_uint16_axis_0 = const()[name = tensor<string, []>("op_13_cast_fp16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
            tensor<string, []> select_0_to_int16_dtype_0 = const()[name = tensor<string, []>("select_0_to_int16_dtype_0"), val = tensor<string, []>("int16")];
            tensor<int16, [1, 512]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor<string, []>("cast_42")];
            tensor<fp16, [1, 512, 72]> var_13_cast_fp16_cast_uint16_cast_uint16 = gather(axis = var_13_cast_fp16_cast_uint16_axis_0, batch_dims = var_13_batch_dims_0, indices = select_0_to_int16, validate_indices = var_13_validate_indices_0, x = enc_p_emb_weight_to_fp16)[name = tensor<string, []>("op_13_cast_fp16_cast_uint16_cast_uint16")];
            tensor<fp16, []> var_14_to_fp16 = const()[name = tensor<string, []>("op_14_to_fp16"), val = tensor<fp16, []>(0x1.0f8p+3)];
            tensor<fp16, [1, 512, 72]> x_1_cast_fp16 = mul(x = var_13_cast_fp16_cast_uint16_cast_uint16, y = var_14_to_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
            tensor<int32, [3]> x_3_perm_0 = const()[name = tensor<string, []>("x_3_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<string, []> x_mask_to_fp16_dtype_0 = const()[name = tensor<string, []>("x_mask_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 512]> x_mask_to_fp16 = cast(dtype = x_mask_to_fp16_dtype_0, x = x_mask)[name = tensor<string, []>("cast_41")];
            tensor<fp16, [1, 72, 512]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = x_1_cast_fp16)[name = tensor<string, []>("transpose_22")];
            tensor<fp16, [1, 72, 512]> x_5_cast_fp16 = mul(x = x_3_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
            tensor<int32, []> var_42 = const()[name = tensor<string, []>("op_42"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> var_68_axes_0 = const()[name = tensor<string, []>("op_68_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512]> var_68_cast_fp16 = expand_dims(axes = var_68_axes_0, x = x_mask_to_fp16)[name = tensor<string, []>("op_68_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = tensor<string, []>("op_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 512, 1]> var_69_cast_fp16 = expand_dims(axes = var_69_axes_0, x = x_mask_to_fp16)[name = tensor<string, []>("op_69_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> mask_cast_fp16 = mul(x = var_68_cast_fp16, y = var_69_cast_fp16)[name = tensor<string, []>("mask_cast_fp16")];
            tensor<fp16, [1, 72, 512]> input_1_cast_fp16 = mul(x = x_5_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> query_1_strides_0 = const()[name = tensor<string, []>("query_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> query_1_dilations_0 = const()[name = tensor<string, []>("query_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> query_1_groups_0 = const()[name = tensor<string, []>("query_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_0_conv_q_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_0_conv_q_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25792)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_0_conv_q_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_0_conv_q_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36224)))];
            tensor<fp16, [1, 72, 512]> query_1_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_0_conv_q_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = enc_p_encoder_attn_layers_0_conv_q_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
            tensor<string, []> key_1_pad_type_0 = const()[name = tensor<string, []>("key_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> key_1_strides_0 = const()[name = tensor<string, []>("key_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> key_1_pad_0 = const()[name = tensor<string, []>("key_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> key_1_dilations_0 = const()[name = tensor<string, []>("key_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> key_1_groups_0 = const()[name = tensor<string, []>("key_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_0_conv_k_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_0_conv_k_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36480)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_0_conv_k_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_0_conv_k_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46912)))];
            tensor<fp16, [1, 72, 512]> key_1_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_0_conv_k_bias_to_fp16, dilations = key_1_dilations_0, groups = key_1_groups_0, pad = key_1_pad_0, pad_type = key_1_pad_type_0, strides = key_1_strides_0, weight = enc_p_encoder_attn_layers_0_conv_k_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
            tensor<string, []> value_1_pad_type_0 = const()[name = tensor<string, []>("value_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> value_1_strides_0 = const()[name = tensor<string, []>("value_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> value_1_pad_0 = const()[name = tensor<string, []>("value_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> value_1_dilations_0 = const()[name = tensor<string, []>("value_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> value_1_groups_0 = const()[name = tensor<string, []>("value_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_0_conv_v_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_0_conv_v_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47168)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_0_conv_v_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_0_conv_v_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57600)))];
            tensor<fp16, [1, 72, 512]> value_1_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_0_conv_v_bias_to_fp16, dilations = value_1_dilations_0, groups = value_1_groups_0, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = value_1_strides_0, weight = enc_p_encoder_attn_layers_0_conv_v_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
            tensor<int32, [4]> var_99 = const()[name = tensor<string, []>("op_99"), val = tensor<int32, [4]>([1, 2, 36, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_100_cast_fp16 = reshape(shape = var_99, x = query_1_cast_fp16)[name = tensor<string, []>("op_100_cast_fp16")];
            tensor<int32, [4]> query_3_perm_0 = const()[name = tensor<string, []>("query_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_102 = const()[name = tensor<string, []>("op_102"), val = tensor<int32, [4]>([1, 2, 36, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_103_cast_fp16 = reshape(shape = var_102, x = key_1_cast_fp16)[name = tensor<string, []>("op_103_cast_fp16")];
            tensor<int32, [4]> var_105 = const()[name = tensor<string, []>("op_105"), val = tensor<int32, [4]>([1, 2, 36, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_106_cast_fp16 = reshape(shape = var_105, x = value_1_cast_fp16)[name = tensor<string, []>("op_106_cast_fp16")];
            tensor<fp16, []> _inversed_109_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_109_y_0_to_fp16"), val = tensor<fp16, []>(0x1.554p-3)];
            tensor<fp16, [1, 2, 512, 36]> query_3_cast_fp16 = transpose(perm = query_3_perm_0, x = var_100_cast_fp16)[name = tensor<string, []>("transpose_21")];
            tensor<fp16, [1, 2, 512, 36]> _inversed_109_cast_fp16 = mul(x = query_3_cast_fp16, y = _inversed_109_y_0_to_fp16)[name = tensor<string, []>("_inversed_109_cast_fp16")];
            tensor<bool, []> scores_1_transpose_x_0 = const()[name = tensor<string, []>("scores_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> scores_1_transpose_y_0 = const()[name = tensor<string, []>("scores_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 2, 512, 512]> scores_1_cast_fp16 = matmul(transpose_x = scores_1_transpose_x_0, transpose_y = scores_1_transpose_y_0, x = _inversed_109_cast_fp16, y = var_103_cast_fp16)[name = tensor<string, []>("scores_1_cast_fp16")];
            tensor<bool, []> x_9_transpose_x_0 = const()[name = tensor<string, []>("x_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_9_transpose_y_0 = const()[name = tensor<string, []>("x_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 36, 1023]> var_119_to_fp16 = const()[name = tensor<string, []>("op_119_to_fp16"), val = tensor<fp16, [1, 1, 36, 1023]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57856)))];
            tensor<fp16, [1, 2, 512, 1023]> x_9_cast_fp16 = matmul(transpose_x = x_9_transpose_x_0, transpose_y = x_9_transpose_y_0, x = _inversed_109_cast_fp16, y = var_119_to_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
            tensor<int32, [8]> x_11_pad_0 = const()[name = tensor<string, []>("x_11_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 1])];
            tensor<string, []> x_11_mode_0 = const()[name = tensor<string, []>("x_11_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 512, 1024]> x_11_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = x_11_mode_0, pad = x_11_pad_0, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<int32, [3]> var_123 = const()[name = tensor<string, []>("op_123"), val = tensor<int32, [3]>([1, 2, 524288])];
            tensor<fp16, [1, 2, 524288]> input_5_cast_fp16 = reshape(shape = var_123, x = x_11_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<int32, [6]> x_flat_1_pad_0 = const()[name = tensor<string, []>("x_flat_1_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 511])];
            tensor<string, []> x_flat_1_mode_0 = const()[name = tensor<string, []>("x_flat_1_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 524799]> x_flat_1_cast_fp16 = pad(constant_val = const_2_to_fp16, mode = x_flat_1_mode_0, pad = x_flat_1_pad_0, x = input_5_cast_fp16)[name = tensor<string, []>("x_flat_1_cast_fp16")];
            tensor<int32, [4]> var_127 = const()[name = tensor<string, []>("op_127"), val = tensor<int32, [4]>([1, 2, 513, 1023])];
            tensor<fp16, [1, 2, 513, 1023]> var_128_cast_fp16 = reshape(shape = var_127, x = x_flat_1_cast_fp16)[name = tensor<string, []>("op_128_cast_fp16")];
            tensor<int32, [4]> var_131_begin_0 = const()[name = tensor<string, []>("op_131_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_131_end_0 = const()[name = tensor<string, []>("op_131_end_0"), val = tensor<int32, [4]>([1, 2, 512, 1023])];
            tensor<bool, [4]> var_131_end_mask_0 = const()[name = tensor<string, []>("op_131_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, 512, 1023]> var_131_cast_fp16 = slice_by_index(begin = var_131_begin_0, end = var_131_end_0, end_mask = var_131_end_mask_0, x = var_128_cast_fp16)[name = tensor<string, []>("op_131_cast_fp16")];
            tensor<int32, [4]> var_132_begin_0 = const()[name = tensor<string, []>("op_132_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 511])];
            tensor<int32, [4]> var_132_end_0 = const()[name = tensor<string, []>("op_132_end_0"), val = tensor<int32, [4]>([1, 2, 512, 1023])];
            tensor<bool, [4]> var_132_end_mask_0 = const()[name = tensor<string, []>("op_132_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 512]> var_132_cast_fp16 = slice_by_index(begin = var_132_begin_0, end = var_132_end_0, end_mask = var_132_end_mask_0, x = var_131_cast_fp16)[name = tensor<string, []>("op_132_cast_fp16")];
            tensor<fp16, [1, 2, 512, 512]> scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_132_cast_fp16)[name = tensor<string, []>("scores_3_cast_fp16")];
            tensor<fp16, []> var_24_promoted_to_fp16 = const()[name = tensor<string, []>("op_24_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<bool, [1, 1, 512, 512]> var_134_cast_fp16 = equal(x = mask_cast_fp16, y = var_24_promoted_to_fp16)[name = tensor<string, []>("op_134_cast_fp16")];
            tensor<fp16, []> var_38_to_fp16 = const()[name = tensor<string, []>("op_38_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
            tensor<fp16, [1, 2, 512, 512]> scores_5_cast_fp16 = select(a = var_38_to_fp16, b = scores_3_cast_fp16, cond = var_134_cast_fp16)[name = tensor<string, []>("scores_5_cast_fp16")];
            tensor<fp16, [1, 2, 512, 512]> p_attn_1_cast_fp16 = softmax(axis = var_42, x = scores_5_cast_fp16)[name = tensor<string, []>("p_attn_1_cast_fp16")];
            tensor<bool, []> output_1_transpose_x_1 = const()[name = tensor<string, []>("output_1_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> output_1_transpose_y_1 = const()[name = tensor<string, []>("output_1_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 2, 512, 36]> output_1_cast_fp16 = matmul(transpose_x = output_1_transpose_x_1, transpose_y = output_1_transpose_y_1, x = p_attn_1_cast_fp16, y = var_106_cast_fp16)[name = tensor<string, []>("output_1_cast_fp16")];
            tensor<int32, [8]> x_13_pad_0 = const()[name = tensor<string, []>("x_13_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 511])];
            tensor<string, []> x_13_mode_0 = const()[name = tensor<string, []>("x_13_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_3_to_fp16 = const()[name = tensor<string, []>("const_3_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 512, 1023]> x_13_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = x_13_mode_0, pad = x_13_pad_0, x = p_attn_1_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
            tensor<int32, [3]> var_140 = const()[name = tensor<string, []>("op_140"), val = tensor<int32, [3]>([1, 2, 523776])];
            tensor<fp16, [1, 2, 523776]> input_7_cast_fp16 = reshape(shape = var_140, x = x_13_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<int32, [6]> x_flat_3_pad_0 = const()[name = tensor<string, []>("x_flat_3_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 512, 0])];
            tensor<string, []> x_flat_3_mode_0 = const()[name = tensor<string, []>("x_flat_3_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_4_to_fp16 = const()[name = tensor<string, []>("const_4_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 524288]> x_flat_3_cast_fp16 = pad(constant_val = const_4_to_fp16, mode = x_flat_3_mode_0, pad = x_flat_3_pad_0, x = input_7_cast_fp16)[name = tensor<string, []>("x_flat_3_cast_fp16")];
            tensor<int32, [4]> var_144 = const()[name = tensor<string, []>("op_144"), val = tensor<int32, [4]>([1, 2, 512, 1024])];
            tensor<fp16, [1, 2, 512, 1024]> var_145_cast_fp16 = reshape(shape = var_144, x = x_flat_3_cast_fp16)[name = tensor<string, []>("op_145_cast_fp16")];
            tensor<int32, [4]> x_15_begin_0 = const()[name = tensor<string, []>("x_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x_15_end_0 = const()[name = tensor<string, []>("x_15_end_0"), val = tensor<int32, [4]>([1, 2, 512, 1024])];
            tensor<bool, [4]> x_15_end_mask_0 = const()[name = tensor<string, []>("x_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 1023]> x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, x = var_145_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
            tensor<bool, []> var_155_transpose_x_0 = const()[name = tensor<string, []>("op_155_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_155_transpose_y_0 = const()[name = tensor<string, []>("op_155_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 1023, 36]> var_154_to_fp16 = const()[name = tensor<string, []>("op_154_to_fp16"), val = tensor<fp16, [1, 1, 1023, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131584)))];
            tensor<fp16, [1, 2, 512, 36]> var_155_cast_fp16 = matmul(transpose_x = var_155_transpose_x_0, transpose_y = var_155_transpose_y_0, x = x_15_cast_fp16, y = var_154_to_fp16)[name = tensor<string, []>("op_155_cast_fp16")];
            tensor<fp16, [1, 2, 512, 36]> output_3_cast_fp16 = add(x = output_1_cast_fp16, y = var_155_cast_fp16)[name = tensor<string, []>("output_3_cast_fp16")];
            tensor<int32, [4]> var_157_perm_0 = const()[name = tensor<string, []>("op_157_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> var_159 = const()[name = tensor<string, []>("op_159"), val = tensor<int32, [3]>([1, 72, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_157_cast_fp16 = transpose(perm = var_157_perm_0, x = output_3_cast_fp16)[name = tensor<string, []>("transpose_20")];
            tensor<fp16, [1, 72, 512]> input_11_cast_fp16 = reshape(shape = var_159, x = var_157_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_0_conv_o_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_0_conv_o_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205312)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_0_conv_o_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_0_conv_o_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215744)))];
            tensor<fp16, [1, 72, 512]> input_13_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_0_conv_o_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = enc_p_encoder_attn_layers_0_conv_o_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<fp16, [1, 72, 512]> x_17_cast_fp16 = add(x = input_1_cast_fp16, y = input_13_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
            tensor<int32, [3]> input_15_perm_0 = const()[name = tensor<string, []>("input_15_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<int32, [1]> x_19_axes_0 = const()[name = tensor<string, []>("x_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_1_0_gamma_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_1_0_gamma_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216000)))];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_1_0_beta_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_1_0_beta_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216256)))];
            tensor<fp16, []> var_20_to_fp16 = const()[name = tensor<string, []>("op_20_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 512, 72]> input_15_cast_fp16 = transpose(perm = input_15_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_19")];
            tensor<fp16, [1, 512, 72]> x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = enc_p_encoder_norm_layers_1_0_beta_to_fp16, epsilon = var_20_to_fp16, gamma = enc_p_encoder_norm_layers_1_0_gamma_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
            tensor<int32, [3]> x_21_perm_0 = const()[name = tensor<string, []>("x_21_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<fp16, [1, 72, 512]> x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = x_19_cast_fp16)[name = tensor<string, []>("transpose_18")];
            tensor<fp16, [1, 72, 512]> input_17_cast_fp16 = mul(x = x_21_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<int32, [6]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 1])];
            tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_6_to_fp16 = const()[name = tensor<string, []>("const_6_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 72, 514]> input_19_cast_fp16 = pad(constant_val = const_6_to_fp16, mode = input_19_mode_0, pad = input_19_pad_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<string, []> x_23_pad_type_0 = const()[name = tensor<string, []>("x_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_23_strides_0 = const()[name = tensor<string, []>("x_23_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_23_pad_0 = const()[name = tensor<string, []>("x_23_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_23_dilations_0 = const()[name = tensor<string, []>("x_23_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_23_groups_0 = const()[name = tensor<string, []>("x_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 72, 3]> enc_p_encoder_ffn_layers_0_conv_1_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_0_conv_1_weight_to_fp16"), val = tensor<fp16, [384, 72, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216512)))];
            tensor<fp16, [384]> enc_p_encoder_ffn_layers_0_conv_1_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_0_conv_1_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382464)))];
            tensor<fp16, [1, 384, 512]> x_23_cast_fp16 = conv(bias = enc_p_encoder_ffn_layers_0_conv_1_bias_to_fp16, dilations = x_23_dilations_0, groups = x_23_groups_0, pad = x_23_pad_0, pad_type = x_23_pad_type_0, strides = x_23_strides_0, weight = enc_p_encoder_ffn_layers_0_conv_1_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
            tensor<fp16, [1, 384, 512]> input_21_cast_fp16 = relu(x = x_23_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<fp16, [1, 384, 512]> input_23_cast_fp16 = mul(x = input_21_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<int32, [6]> input_25_pad_0 = const()[name = tensor<string, []>("input_25_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 1])];
            tensor<string, []> input_25_mode_0 = const()[name = tensor<string, []>("input_25_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_7_to_fp16 = const()[name = tensor<string, []>("const_7_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 384, 514]> input_25_cast_fp16 = pad(constant_val = const_7_to_fp16, mode = input_25_mode_0, pad = input_25_pad_0, x = input_23_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<string, []> x_27_pad_type_0 = const()[name = tensor<string, []>("x_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_27_strides_0 = const()[name = tensor<string, []>("x_27_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_27_pad_0 = const()[name = tensor<string, []>("x_27_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_27_dilations_0 = const()[name = tensor<string, []>("x_27_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_27_groups_0 = const()[name = tensor<string, []>("x_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 384, 3]> enc_p_encoder_ffn_layers_0_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_0_conv_2_weight_to_fp16"), val = tensor<fp16, [72, 384, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383296)))];
            tensor<fp16, [72]> enc_p_encoder_ffn_layers_0_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_0_conv_2_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549248)))];
            tensor<fp16, [1, 72, 512]> x_27_cast_fp16 = conv(bias = enc_p_encoder_ffn_layers_0_conv_2_bias_to_fp16, dilations = x_27_dilations_0, groups = x_27_groups_0, pad = x_27_pad_0, pad_type = x_27_pad_type_0, strides = x_27_strides_0, weight = enc_p_encoder_ffn_layers_0_conv_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
            tensor<fp16, [1, 72, 512]> input_27_cast_fp16 = mul(x = x_27_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<fp16, [1, 72, 512]> x_29_cast_fp16 = add(x = x_21_cast_fp16, y = input_27_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
            tensor<int32, [3]> input_29_perm_0 = const()[name = tensor<string, []>("input_29_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<int32, [1]> x_31_axes_0 = const()[name = tensor<string, []>("x_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_2_0_gamma_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_2_0_gamma_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549504)))];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_2_0_beta_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_2_0_beta_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549760)))];
            tensor<fp16, [1, 512, 72]> input_29_cast_fp16 = transpose(perm = input_29_perm_0, x = x_29_cast_fp16)[name = tensor<string, []>("transpose_17")];
            tensor<fp16, [1, 512, 72]> x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = enc_p_encoder_norm_layers_2_0_beta_to_fp16, epsilon = var_20_to_fp16, gamma = enc_p_encoder_norm_layers_2_0_gamma_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
            tensor<int32, [3]> input_31_perm_0 = const()[name = tensor<string, []>("input_31_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> query_5_strides_0 = const()[name = tensor<string, []>("query_5_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> query_5_dilations_0 = const()[name = tensor<string, []>("query_5_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> query_5_groups_0 = const()[name = tensor<string, []>("query_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_1_conv_q_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_1_conv_q_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550016)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_1_conv_q_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_1_conv_q_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(560448)))];
            tensor<fp16, [1, 72, 512]> input_31_cast_fp16 = transpose(perm = input_31_perm_0, x = x_31_cast_fp16)[name = tensor<string, []>("transpose_16")];
            tensor<fp16, [1, 72, 512]> query_5_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_1_conv_q_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = enc_p_encoder_attn_layers_1_conv_q_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
            tensor<string, []> key_5_pad_type_0 = const()[name = tensor<string, []>("key_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> key_5_strides_0 = const()[name = tensor<string, []>("key_5_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> key_5_pad_0 = const()[name = tensor<string, []>("key_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> key_5_dilations_0 = const()[name = tensor<string, []>("key_5_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> key_5_groups_0 = const()[name = tensor<string, []>("key_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_1_conv_k_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_1_conv_k_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(560704)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_1_conv_k_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_1_conv_k_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571136)))];
            tensor<fp16, [1, 72, 512]> key_5_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_1_conv_k_bias_to_fp16, dilations = key_5_dilations_0, groups = key_5_groups_0, pad = key_5_pad_0, pad_type = key_5_pad_type_0, strides = key_5_strides_0, weight = enc_p_encoder_attn_layers_1_conv_k_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
            tensor<string, []> value_5_pad_type_0 = const()[name = tensor<string, []>("value_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> value_5_strides_0 = const()[name = tensor<string, []>("value_5_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> value_5_pad_0 = const()[name = tensor<string, []>("value_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> value_5_dilations_0 = const()[name = tensor<string, []>("value_5_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> value_5_groups_0 = const()[name = tensor<string, []>("value_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_1_conv_v_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_1_conv_v_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571392)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_1_conv_v_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_1_conv_v_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581824)))];
            tensor<fp16, [1, 72, 512]> value_5_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_1_conv_v_bias_to_fp16, dilations = value_5_dilations_0, groups = value_5_groups_0, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = value_5_strides_0, weight = enc_p_encoder_attn_layers_1_conv_v_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
            tensor<int32, [4]> var_236 = const()[name = tensor<string, []>("op_236"), val = tensor<int32, [4]>([1, 2, 36, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_237_cast_fp16 = reshape(shape = var_236, x = query_5_cast_fp16)[name = tensor<string, []>("op_237_cast_fp16")];
            tensor<int32, [4]> query_7_perm_0 = const()[name = tensor<string, []>("query_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_239 = const()[name = tensor<string, []>("op_239"), val = tensor<int32, [4]>([1, 2, 36, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_240_cast_fp16 = reshape(shape = var_239, x = key_5_cast_fp16)[name = tensor<string, []>("op_240_cast_fp16")];
            tensor<int32, [4]> var_242 = const()[name = tensor<string, []>("op_242"), val = tensor<int32, [4]>([1, 2, 36, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_243_cast_fp16 = reshape(shape = var_242, x = value_5_cast_fp16)[name = tensor<string, []>("op_243_cast_fp16")];
            tensor<fp16, []> _inversed_246_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_246_y_0_to_fp16"), val = tensor<fp16, []>(0x1.554p-3)];
            tensor<fp16, [1, 2, 512, 36]> query_7_cast_fp16 = transpose(perm = query_7_perm_0, x = var_237_cast_fp16)[name = tensor<string, []>("transpose_15")];
            tensor<fp16, [1, 2, 512, 36]> _inversed_246_cast_fp16 = mul(x = query_7_cast_fp16, y = _inversed_246_y_0_to_fp16)[name = tensor<string, []>("_inversed_246_cast_fp16")];
            tensor<bool, []> scores_7_transpose_x_0 = const()[name = tensor<string, []>("scores_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> scores_7_transpose_y_0 = const()[name = tensor<string, []>("scores_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 2, 512, 512]> scores_7_cast_fp16 = matmul(transpose_x = scores_7_transpose_x_0, transpose_y = scores_7_transpose_y_0, x = _inversed_246_cast_fp16, y = var_240_cast_fp16)[name = tensor<string, []>("scores_7_cast_fp16")];
            tensor<bool, []> x_35_transpose_x_0 = const()[name = tensor<string, []>("x_35_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_35_transpose_y_0 = const()[name = tensor<string, []>("x_35_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 36, 1023]> var_256_to_fp16 = const()[name = tensor<string, []>("op_256_to_fp16"), val = tensor<fp16, [1, 1, 36, 1023]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(582080)))];
            tensor<fp16, [1, 2, 512, 1023]> x_35_cast_fp16 = matmul(transpose_x = x_35_transpose_x_0, transpose_y = x_35_transpose_y_0, x = _inversed_246_cast_fp16, y = var_256_to_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<int32, [8]> x_37_pad_0 = const()[name = tensor<string, []>("x_37_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 1])];
            tensor<string, []> x_37_mode_0 = const()[name = tensor<string, []>("x_37_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_9_to_fp16 = const()[name = tensor<string, []>("const_9_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 512, 1024]> x_37_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = x_37_mode_0, pad = x_37_pad_0, x = x_35_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<int32, [3]> var_260 = const()[name = tensor<string, []>("op_260"), val = tensor<int32, [3]>([1, 2, 524288])];
            tensor<fp16, [1, 2, 524288]> input_35_cast_fp16 = reshape(shape = var_260, x = x_37_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<int32, [6]> x_flat_5_pad_0 = const()[name = tensor<string, []>("x_flat_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 511])];
            tensor<string, []> x_flat_5_mode_0 = const()[name = tensor<string, []>("x_flat_5_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_10_to_fp16 = const()[name = tensor<string, []>("const_10_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 524799]> x_flat_5_cast_fp16 = pad(constant_val = const_10_to_fp16, mode = x_flat_5_mode_0, pad = x_flat_5_pad_0, x = input_35_cast_fp16)[name = tensor<string, []>("x_flat_5_cast_fp16")];
            tensor<int32, [4]> var_264 = const()[name = tensor<string, []>("op_264"), val = tensor<int32, [4]>([1, 2, 513, 1023])];
            tensor<fp16, [1, 2, 513, 1023]> var_265_cast_fp16 = reshape(shape = var_264, x = x_flat_5_cast_fp16)[name = tensor<string, []>("op_265_cast_fp16")];
            tensor<int32, [4]> var_268_begin_0 = const()[name = tensor<string, []>("op_268_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_268_end_0 = const()[name = tensor<string, []>("op_268_end_0"), val = tensor<int32, [4]>([1, 2, 512, 1023])];
            tensor<bool, [4]> var_268_end_mask_0 = const()[name = tensor<string, []>("op_268_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, 512, 1023]> var_268_cast_fp16 = slice_by_index(begin = var_268_begin_0, end = var_268_end_0, end_mask = var_268_end_mask_0, x = var_265_cast_fp16)[name = tensor<string, []>("op_268_cast_fp16")];
            tensor<int32, [4]> var_269_begin_0 = const()[name = tensor<string, []>("op_269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 511])];
            tensor<int32, [4]> var_269_end_0 = const()[name = tensor<string, []>("op_269_end_0"), val = tensor<int32, [4]>([1, 2, 512, 1023])];
            tensor<bool, [4]> var_269_end_mask_0 = const()[name = tensor<string, []>("op_269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 512]> var_269_cast_fp16 = slice_by_index(begin = var_269_begin_0, end = var_269_end_0, end_mask = var_269_end_mask_0, x = var_268_cast_fp16)[name = tensor<string, []>("op_269_cast_fp16")];
            tensor<fp16, [1, 2, 512, 512]> scores_9_cast_fp16 = add(x = scores_7_cast_fp16, y = var_269_cast_fp16)[name = tensor<string, []>("scores_9_cast_fp16")];
            tensor<fp16, [1, 2, 512, 512]> scores_11_cast_fp16 = select(a = var_38_to_fp16, b = scores_9_cast_fp16, cond = var_134_cast_fp16)[name = tensor<string, []>("scores_11_cast_fp16")];
            tensor<fp16, [1, 2, 512, 512]> p_attn_3_cast_fp16 = softmax(axis = var_42, x = scores_11_cast_fp16)[name = tensor<string, []>("p_attn_3_cast_fp16")];
            tensor<bool, []> output_5_transpose_x_1 = const()[name = tensor<string, []>("output_5_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> output_5_transpose_y_1 = const()[name = tensor<string, []>("output_5_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 2, 512, 36]> output_5_cast_fp16 = matmul(transpose_x = output_5_transpose_x_1, transpose_y = output_5_transpose_y_1, x = p_attn_3_cast_fp16, y = var_243_cast_fp16)[name = tensor<string, []>("output_5_cast_fp16")];
            tensor<int32, [8]> x_39_pad_0 = const()[name = tensor<string, []>("x_39_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 511])];
            tensor<string, []> x_39_mode_0 = const()[name = tensor<string, []>("x_39_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_11_to_fp16 = const()[name = tensor<string, []>("const_11_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 512, 1023]> x_39_cast_fp16 = pad(constant_val = const_11_to_fp16, mode = x_39_mode_0, pad = x_39_pad_0, x = p_attn_3_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
            tensor<int32, [3]> var_277 = const()[name = tensor<string, []>("op_277"), val = tensor<int32, [3]>([1, 2, 523776])];
            tensor<fp16, [1, 2, 523776]> input_37_cast_fp16 = reshape(shape = var_277, x = x_39_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<int32, [6]> x_flat_7_pad_0 = const()[name = tensor<string, []>("x_flat_7_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 512, 0])];
            tensor<string, []> x_flat_7_mode_0 = const()[name = tensor<string, []>("x_flat_7_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_12_to_fp16 = const()[name = tensor<string, []>("const_12_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 524288]> x_flat_7_cast_fp16 = pad(constant_val = const_12_to_fp16, mode = x_flat_7_mode_0, pad = x_flat_7_pad_0, x = input_37_cast_fp16)[name = tensor<string, []>("x_flat_7_cast_fp16")];
            tensor<int32, [4]> var_281 = const()[name = tensor<string, []>("op_281"), val = tensor<int32, [4]>([1, 2, 512, 1024])];
            tensor<fp16, [1, 2, 512, 1024]> var_282_cast_fp16 = reshape(shape = var_281, x = x_flat_7_cast_fp16)[name = tensor<string, []>("op_282_cast_fp16")];
            tensor<int32, [4]> x_41_begin_0 = const()[name = tensor<string, []>("x_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x_41_end_0 = const()[name = tensor<string, []>("x_41_end_0"), val = tensor<int32, [4]>([1, 2, 512, 1024])];
            tensor<bool, [4]> x_41_end_mask_0 = const()[name = tensor<string, []>("x_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 1023]> x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, x = var_282_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")];
            tensor<bool, []> var_292_transpose_x_0 = const()[name = tensor<string, []>("op_292_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_292_transpose_y_0 = const()[name = tensor<string, []>("op_292_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 1023, 36]> var_291_to_fp16 = const()[name = tensor<string, []>("op_291_to_fp16"), val = tensor<fp16, [1, 1, 1023, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655808)))];
            tensor<fp16, [1, 2, 512, 36]> var_292_cast_fp16 = matmul(transpose_x = var_292_transpose_x_0, transpose_y = var_292_transpose_y_0, x = x_41_cast_fp16, y = var_291_to_fp16)[name = tensor<string, []>("op_292_cast_fp16")];
            tensor<fp16, [1, 2, 512, 36]> output_7_cast_fp16 = add(x = output_5_cast_fp16, y = var_292_cast_fp16)[name = tensor<string, []>("output_7_cast_fp16")];
            tensor<int32, [4]> var_294_perm_0 = const()[name = tensor<string, []>("op_294_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> var_296 = const()[name = tensor<string, []>("op_296"), val = tensor<int32, [3]>([1, 72, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_294_cast_fp16 = transpose(perm = var_294_perm_0, x = output_7_cast_fp16)[name = tensor<string, []>("transpose_14")];
            tensor<fp16, [1, 72, 512]> input_41_cast_fp16 = reshape(shape = var_296, x = var_294_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<string, []> input_43_pad_type_0 = const()[name = tensor<string, []>("input_43_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_43_strides_0 = const()[name = tensor<string, []>("input_43_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_43_pad_0 = const()[name = tensor<string, []>("input_43_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_43_dilations_0 = const()[name = tensor<string, []>("input_43_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_43_groups_0 = const()[name = tensor<string, []>("input_43_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_1_conv_o_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_1_conv_o_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(729536)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_1_conv_o_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_1_conv_o_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(739968)))];
            tensor<fp16, [1, 72, 512]> input_43_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_1_conv_o_bias_to_fp16, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = enc_p_encoder_attn_layers_1_conv_o_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<fp16, [1, 72, 512]> x_43_cast_fp16 = add(x = input_31_cast_fp16, y = input_43_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
            tensor<int32, [3]> input_45_perm_0 = const()[name = tensor<string, []>("input_45_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<int32, [1]> x_45_axes_0 = const()[name = tensor<string, []>("x_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_1_1_gamma_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_1_1_gamma_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(740224)))];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_1_1_beta_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_1_1_beta_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(740480)))];
            tensor<fp16, [1, 512, 72]> input_45_cast_fp16 = transpose(perm = input_45_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_13")];
            tensor<fp16, [1, 512, 72]> x_45_cast_fp16 = layer_norm(axes = x_45_axes_0, beta = enc_p_encoder_norm_layers_1_1_beta_to_fp16, epsilon = var_20_to_fp16, gamma = enc_p_encoder_norm_layers_1_1_gamma_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
            tensor<int32, [3]> x_47_perm_0 = const()[name = tensor<string, []>("x_47_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<fp16, [1, 72, 512]> x_47_cast_fp16 = transpose(perm = x_47_perm_0, x = x_45_cast_fp16)[name = tensor<string, []>("transpose_12")];
            tensor<fp16, [1, 72, 512]> input_47_cast_fp16 = mul(x = x_47_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<int32, [6]> input_49_pad_0 = const()[name = tensor<string, []>("input_49_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 1])];
            tensor<string, []> input_49_mode_0 = const()[name = tensor<string, []>("input_49_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_14_to_fp16 = const()[name = tensor<string, []>("const_14_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 72, 514]> input_49_cast_fp16 = pad(constant_val = const_14_to_fp16, mode = input_49_mode_0, pad = input_49_pad_0, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<string, []> x_49_pad_type_0 = const()[name = tensor<string, []>("x_49_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_49_strides_0 = const()[name = tensor<string, []>("x_49_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_49_pad_0 = const()[name = tensor<string, []>("x_49_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_49_dilations_0 = const()[name = tensor<string, []>("x_49_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_49_groups_0 = const()[name = tensor<string, []>("x_49_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 72, 3]> enc_p_encoder_ffn_layers_1_conv_1_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_1_conv_1_weight_to_fp16"), val = tensor<fp16, [384, 72, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(740736)))];
            tensor<fp16, [384]> enc_p_encoder_ffn_layers_1_conv_1_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_1_conv_1_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(906688)))];
            tensor<fp16, [1, 384, 512]> x_49_cast_fp16 = conv(bias = enc_p_encoder_ffn_layers_1_conv_1_bias_to_fp16, dilations = x_49_dilations_0, groups = x_49_groups_0, pad = x_49_pad_0, pad_type = x_49_pad_type_0, strides = x_49_strides_0, weight = enc_p_encoder_ffn_layers_1_conv_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
            tensor<fp16, [1, 384, 512]> input_51_cast_fp16 = relu(x = x_49_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<fp16, [1, 384, 512]> input_53_cast_fp16 = mul(x = input_51_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<int32, [6]> input_55_pad_0 = const()[name = tensor<string, []>("input_55_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 1])];
            tensor<string, []> input_55_mode_0 = const()[name = tensor<string, []>("input_55_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_15_to_fp16 = const()[name = tensor<string, []>("const_15_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 384, 514]> input_55_cast_fp16 = pad(constant_val = const_15_to_fp16, mode = input_55_mode_0, pad = input_55_pad_0, x = input_53_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<string, []> x_53_pad_type_0 = const()[name = tensor<string, []>("x_53_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_53_strides_0 = const()[name = tensor<string, []>("x_53_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_53_pad_0 = const()[name = tensor<string, []>("x_53_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_53_dilations_0 = const()[name = tensor<string, []>("x_53_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_53_groups_0 = const()[name = tensor<string, []>("x_53_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 384, 3]> enc_p_encoder_ffn_layers_1_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_1_conv_2_weight_to_fp16"), val = tensor<fp16, [72, 384, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(907520)))];
            tensor<fp16, [72]> enc_p_encoder_ffn_layers_1_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_1_conv_2_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1073472)))];
            tensor<fp16, [1, 72, 512]> x_53_cast_fp16 = conv(bias = enc_p_encoder_ffn_layers_1_conv_2_bias_to_fp16, dilations = x_53_dilations_0, groups = x_53_groups_0, pad = x_53_pad_0, pad_type = x_53_pad_type_0, strides = x_53_strides_0, weight = enc_p_encoder_ffn_layers_1_conv_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")];
            tensor<fp16, [1, 72, 512]> input_57_cast_fp16 = mul(x = x_53_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<fp16, [1, 72, 512]> x_55_cast_fp16 = add(x = x_47_cast_fp16, y = input_57_cast_fp16)[name = tensor<string, []>("x_55_cast_fp16")];
            tensor<int32, [3]> input_59_perm_0 = const()[name = tensor<string, []>("input_59_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<int32, [1]> x_57_axes_0 = const()[name = tensor<string, []>("x_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_2_1_gamma_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_2_1_gamma_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1073728)))];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_2_1_beta_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_2_1_beta_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1073984)))];
            tensor<fp16, [1, 512, 72]> input_59_cast_fp16 = transpose(perm = input_59_perm_0, x = x_55_cast_fp16)[name = tensor<string, []>("transpose_11")];
            tensor<fp16, [1, 512, 72]> x_57_cast_fp16 = layer_norm(axes = x_57_axes_0, beta = enc_p_encoder_norm_layers_2_1_beta_to_fp16, epsilon = var_20_to_fp16, gamma = enc_p_encoder_norm_layers_2_1_gamma_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("x_57_cast_fp16")];
            tensor<int32, [3]> input_61_perm_0 = const()[name = tensor<string, []>("input_61_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> query_9_strides_0 = const()[name = tensor<string, []>("query_9_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> query_9_dilations_0 = const()[name = tensor<string, []>("query_9_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> query_9_groups_0 = const()[name = tensor<string, []>("query_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_2_conv_q_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_2_conv_q_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1074240)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_2_conv_q_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_2_conv_q_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1084672)))];
            tensor<fp16, [1, 72, 512]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_57_cast_fp16)[name = tensor<string, []>("transpose_10")];
            tensor<fp16, [1, 72, 512]> query_9_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_2_conv_q_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = enc_p_encoder_attn_layers_2_conv_q_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
            tensor<string, []> key_9_pad_type_0 = const()[name = tensor<string, []>("key_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> key_9_strides_0 = const()[name = tensor<string, []>("key_9_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> key_9_pad_0 = const()[name = tensor<string, []>("key_9_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> key_9_dilations_0 = const()[name = tensor<string, []>("key_9_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> key_9_groups_0 = const()[name = tensor<string, []>("key_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_2_conv_k_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_2_conv_k_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1084928)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_2_conv_k_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_2_conv_k_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1095360)))];
            tensor<fp16, [1, 72, 512]> key_9_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_2_conv_k_bias_to_fp16, dilations = key_9_dilations_0, groups = key_9_groups_0, pad = key_9_pad_0, pad_type = key_9_pad_type_0, strides = key_9_strides_0, weight = enc_p_encoder_attn_layers_2_conv_k_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
            tensor<string, []> value_9_pad_type_0 = const()[name = tensor<string, []>("value_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> value_9_strides_0 = const()[name = tensor<string, []>("value_9_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> value_9_pad_0 = const()[name = tensor<string, []>("value_9_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> value_9_dilations_0 = const()[name = tensor<string, []>("value_9_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> value_9_groups_0 = const()[name = tensor<string, []>("value_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_2_conv_v_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_2_conv_v_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1095616)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_2_conv_v_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_2_conv_v_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1106048)))];
            tensor<fp16, [1, 72, 512]> value_9_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_2_conv_v_bias_to_fp16, dilations = value_9_dilations_0, groups = value_9_groups_0, pad = value_9_pad_0, pad_type = value_9_pad_type_0, strides = value_9_strides_0, weight = enc_p_encoder_attn_layers_2_conv_v_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = tensor<string, []>("op_373"), val = tensor<int32, [4]>([1, 2, 36, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_374_cast_fp16 = reshape(shape = var_373, x = query_9_cast_fp16)[name = tensor<string, []>("op_374_cast_fp16")];
            tensor<int32, [4]> query_perm_0 = const()[name = tensor<string, []>("query_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_376 = const()[name = tensor<string, []>("op_376"), val = tensor<int32, [4]>([1, 2, 36, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_377_cast_fp16 = reshape(shape = var_376, x = key_9_cast_fp16)[name = tensor<string, []>("op_377_cast_fp16")];
            tensor<int32, [4]> var_379 = const()[name = tensor<string, []>("op_379"), val = tensor<int32, [4]>([1, 2, 36, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_380_cast_fp16 = reshape(shape = var_379, x = value_9_cast_fp16)[name = tensor<string, []>("op_380_cast_fp16")];
            tensor<fp16, []> _inversed_383_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_383_y_0_to_fp16"), val = tensor<fp16, []>(0x1.554p-3)];
            tensor<fp16, [1, 2, 512, 36]> query_cast_fp16 = transpose(perm = query_perm_0, x = var_374_cast_fp16)[name = tensor<string, []>("transpose_9")];
            tensor<fp16, [1, 2, 512, 36]> _inversed_383_cast_fp16 = mul(x = query_cast_fp16, y = _inversed_383_y_0_to_fp16)[name = tensor<string, []>("_inversed_383_cast_fp16")];
            tensor<bool, []> scores_13_transpose_x_0 = const()[name = tensor<string, []>("scores_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> scores_13_transpose_y_0 = const()[name = tensor<string, []>("scores_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 2, 512, 512]> scores_13_cast_fp16 = matmul(transpose_x = scores_13_transpose_x_0, transpose_y = scores_13_transpose_y_0, x = _inversed_383_cast_fp16, y = var_377_cast_fp16)[name = tensor<string, []>("scores_13_cast_fp16")];
            tensor<bool, []> x_61_transpose_x_0 = const()[name = tensor<string, []>("x_61_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_61_transpose_y_0 = const()[name = tensor<string, []>("x_61_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 36, 1023]> var_393_to_fp16 = const()[name = tensor<string, []>("op_393_to_fp16"), val = tensor<fp16, [1, 1, 36, 1023]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1106304)))];
            tensor<fp16, [1, 2, 512, 1023]> x_61_cast_fp16 = matmul(transpose_x = x_61_transpose_x_0, transpose_y = x_61_transpose_y_0, x = _inversed_383_cast_fp16, y = var_393_to_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
            tensor<int32, [8]> x_63_pad_0 = const()[name = tensor<string, []>("x_63_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 1])];
            tensor<string, []> x_63_mode_0 = const()[name = tensor<string, []>("x_63_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_17_to_fp16 = const()[name = tensor<string, []>("const_17_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 512, 1024]> x_63_cast_fp16 = pad(constant_val = const_17_to_fp16, mode = x_63_mode_0, pad = x_63_pad_0, x = x_61_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
            tensor<int32, [3]> var_397 = const()[name = tensor<string, []>("op_397"), val = tensor<int32, [3]>([1, 2, 524288])];
            tensor<fp16, [1, 2, 524288]> input_65_cast_fp16 = reshape(shape = var_397, x = x_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<int32, [6]> x_flat_9_pad_0 = const()[name = tensor<string, []>("x_flat_9_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 511])];
            tensor<string, []> x_flat_9_mode_0 = const()[name = tensor<string, []>("x_flat_9_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_18_to_fp16 = const()[name = tensor<string, []>("const_18_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 524799]> x_flat_9_cast_fp16 = pad(constant_val = const_18_to_fp16, mode = x_flat_9_mode_0, pad = x_flat_9_pad_0, x = input_65_cast_fp16)[name = tensor<string, []>("x_flat_9_cast_fp16")];
            tensor<int32, [4]> var_401 = const()[name = tensor<string, []>("op_401"), val = tensor<int32, [4]>([1, 2, 513, 1023])];
            tensor<fp16, [1, 2, 513, 1023]> var_402_cast_fp16 = reshape(shape = var_401, x = x_flat_9_cast_fp16)[name = tensor<string, []>("op_402_cast_fp16")];
            tensor<int32, [4]> var_405_begin_0 = const()[name = tensor<string, []>("op_405_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_405_end_0 = const()[name = tensor<string, []>("op_405_end_0"), val = tensor<int32, [4]>([1, 2, 512, 1023])];
            tensor<bool, [4]> var_405_end_mask_0 = const()[name = tensor<string, []>("op_405_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, 512, 1023]> var_405_cast_fp16 = slice_by_index(begin = var_405_begin_0, end = var_405_end_0, end_mask = var_405_end_mask_0, x = var_402_cast_fp16)[name = tensor<string, []>("op_405_cast_fp16")];
            tensor<int32, [4]> var_406_begin_0 = const()[name = tensor<string, []>("op_406_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 511])];
            tensor<int32, [4]> var_406_end_0 = const()[name = tensor<string, []>("op_406_end_0"), val = tensor<int32, [4]>([1, 2, 512, 1023])];
            tensor<bool, [4]> var_406_end_mask_0 = const()[name = tensor<string, []>("op_406_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 512]> var_406_cast_fp16 = slice_by_index(begin = var_406_begin_0, end = var_406_end_0, end_mask = var_406_end_mask_0, x = var_405_cast_fp16)[name = tensor<string, []>("op_406_cast_fp16")];
            tensor<fp16, [1, 2, 512, 512]> scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_406_cast_fp16)[name = tensor<string, []>("scores_15_cast_fp16")];
            tensor<fp16, [1, 2, 512, 512]> scores_cast_fp16 = select(a = var_38_to_fp16, b = scores_15_cast_fp16, cond = var_134_cast_fp16)[name = tensor<string, []>("scores_cast_fp16")];
            tensor<fp16, [1, 2, 512, 512]> p_attn_cast_fp16 = softmax(axis = var_42, x = scores_cast_fp16)[name = tensor<string, []>("p_attn_cast_fp16")];
            tensor<bool, []> output_9_transpose_x_1 = const()[name = tensor<string, []>("output_9_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> output_9_transpose_y_1 = const()[name = tensor<string, []>("output_9_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 2, 512, 36]> output_9_cast_fp16 = matmul(transpose_x = output_9_transpose_x_1, transpose_y = output_9_transpose_y_1, x = p_attn_cast_fp16, y = var_380_cast_fp16)[name = tensor<string, []>("output_9_cast_fp16")];
            tensor<int32, [8]> x_65_pad_0 = const()[name = tensor<string, []>("x_65_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 511])];
            tensor<string, []> x_65_mode_0 = const()[name = tensor<string, []>("x_65_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_19_to_fp16 = const()[name = tensor<string, []>("const_19_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 512, 1023]> x_65_cast_fp16 = pad(constant_val = const_19_to_fp16, mode = x_65_mode_0, pad = x_65_pad_0, x = p_attn_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
            tensor<int32, [3]> var_414 = const()[name = tensor<string, []>("op_414"), val = tensor<int32, [3]>([1, 2, 523776])];
            tensor<fp16, [1, 2, 523776]> input_67_cast_fp16 = reshape(shape = var_414, x = x_65_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<int32, [6]> x_flat_pad_0 = const()[name = tensor<string, []>("x_flat_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 512, 0])];
            tensor<string, []> x_flat_mode_0 = const()[name = tensor<string, []>("x_flat_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_20_to_fp16 = const()[name = tensor<string, []>("const_20_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 2, 524288]> x_flat_cast_fp16 = pad(constant_val = const_20_to_fp16, mode = x_flat_mode_0, pad = x_flat_pad_0, x = input_67_cast_fp16)[name = tensor<string, []>("x_flat_cast_fp16")];
            tensor<int32, [4]> var_418 = const()[name = tensor<string, []>("op_418"), val = tensor<int32, [4]>([1, 2, 512, 1024])];
            tensor<fp16, [1, 2, 512, 1024]> var_419_cast_fp16 = reshape(shape = var_418, x = x_flat_cast_fp16)[name = tensor<string, []>("op_419_cast_fp16")];
            tensor<int32, [4]> x_67_begin_0 = const()[name = tensor<string, []>("x_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x_67_end_0 = const()[name = tensor<string, []>("x_67_end_0"), val = tensor<int32, [4]>([1, 2, 512, 1024])];
            tensor<bool, [4]> x_67_end_mask_0 = const()[name = tensor<string, []>("x_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 1023]> x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, x = var_419_cast_fp16)[name = tensor<string, []>("x_67_cast_fp16")];
            tensor<bool, []> var_429_transpose_x_0 = const()[name = tensor<string, []>("op_429_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_429_transpose_y_0 = const()[name = tensor<string, []>("op_429_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 1023, 36]> var_428_to_fp16 = const()[name = tensor<string, []>("op_428_to_fp16"), val = tensor<fp16, [1, 1, 1023, 36]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1180032)))];
            tensor<fp16, [1, 2, 512, 36]> var_429_cast_fp16 = matmul(transpose_x = var_429_transpose_x_0, transpose_y = var_429_transpose_y_0, x = x_67_cast_fp16, y = var_428_to_fp16)[name = tensor<string, []>("op_429_cast_fp16")];
            tensor<fp16, [1, 2, 512, 36]> output_cast_fp16 = add(x = output_9_cast_fp16, y = var_429_cast_fp16)[name = tensor<string, []>("output_cast_fp16")];
            tensor<int32, [4]> var_431_perm_0 = const()[name = tensor<string, []>("op_431_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> var_433 = const()[name = tensor<string, []>("op_433"), val = tensor<int32, [3]>([1, 72, 512])];
            tensor<fp16, [1, 2, 36, 512]> var_431_cast_fp16 = transpose(perm = var_431_perm_0, x = output_cast_fp16)[name = tensor<string, []>("transpose_8")];
            tensor<fp16, [1, 72, 512]> input_71_cast_fp16 = reshape(shape = var_433, x = var_431_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<string, []> input_73_pad_type_0 = const()[name = tensor<string, []>("input_73_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_73_strides_0 = const()[name = tensor<string, []>("input_73_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_73_pad_0 = const()[name = tensor<string, []>("input_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_73_dilations_0 = const()[name = tensor<string, []>("input_73_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_73_groups_0 = const()[name = tensor<string, []>("input_73_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 72, 1]> enc_p_encoder_attn_layers_2_conv_o_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_2_conv_o_weight_to_fp16"), val = tensor<fp16, [72, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1253760)))];
            tensor<fp16, [72]> enc_p_encoder_attn_layers_2_conv_o_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_attn_layers_2_conv_o_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1264192)))];
            tensor<fp16, [1, 72, 512]> input_73_cast_fp16 = conv(bias = enc_p_encoder_attn_layers_2_conv_o_bias_to_fp16, dilations = input_73_dilations_0, groups = input_73_groups_0, pad = input_73_pad_0, pad_type = input_73_pad_type_0, strides = input_73_strides_0, weight = enc_p_encoder_attn_layers_2_conv_o_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<fp16, [1, 72, 512]> x_69_cast_fp16 = add(x = input_61_cast_fp16, y = input_73_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
            tensor<int32, [3]> input_75_perm_0 = const()[name = tensor<string, []>("input_75_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<int32, [1]> x_71_axes_0 = const()[name = tensor<string, []>("x_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_1_2_gamma_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_1_2_gamma_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1264448)))];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_1_2_beta_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_1_2_beta_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1264704)))];
            tensor<fp16, [1, 512, 72]> input_75_cast_fp16 = transpose(perm = input_75_perm_0, x = x_69_cast_fp16)[name = tensor<string, []>("transpose_7")];
            tensor<fp16, [1, 512, 72]> x_71_cast_fp16 = layer_norm(axes = x_71_axes_0, beta = enc_p_encoder_norm_layers_1_2_beta_to_fp16, epsilon = var_20_to_fp16, gamma = enc_p_encoder_norm_layers_1_2_gamma_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
            tensor<int32, [3]> x_73_perm_0 = const()[name = tensor<string, []>("x_73_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<fp16, [1, 72, 512]> x_73_cast_fp16 = transpose(perm = x_73_perm_0, x = x_71_cast_fp16)[name = tensor<string, []>("transpose_6")];
            tensor<fp16, [1, 72, 512]> input_77_cast_fp16 = mul(x = x_73_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<int32, [6]> input_79_pad_0 = const()[name = tensor<string, []>("input_79_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 1])];
            tensor<string, []> input_79_mode_0 = const()[name = tensor<string, []>("input_79_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_22_to_fp16 = const()[name = tensor<string, []>("const_22_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 72, 514]> input_79_cast_fp16 = pad(constant_val = const_22_to_fp16, mode = input_79_mode_0, pad = input_79_pad_0, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<string, []> x_75_pad_type_0 = const()[name = tensor<string, []>("x_75_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_75_strides_0 = const()[name = tensor<string, []>("x_75_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_75_pad_0 = const()[name = tensor<string, []>("x_75_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_75_dilations_0 = const()[name = tensor<string, []>("x_75_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_75_groups_0 = const()[name = tensor<string, []>("x_75_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 72, 3]> enc_p_encoder_ffn_layers_2_conv_1_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_2_conv_1_weight_to_fp16"), val = tensor<fp16, [384, 72, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1264960)))];
            tensor<fp16, [384]> enc_p_encoder_ffn_layers_2_conv_1_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_2_conv_1_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1430912)))];
            tensor<fp16, [1, 384, 512]> x_75_cast_fp16 = conv(bias = enc_p_encoder_ffn_layers_2_conv_1_bias_to_fp16, dilations = x_75_dilations_0, groups = x_75_groups_0, pad = x_75_pad_0, pad_type = x_75_pad_type_0, strides = x_75_strides_0, weight = enc_p_encoder_ffn_layers_2_conv_1_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
            tensor<fp16, [1, 384, 512]> input_81_cast_fp16 = relu(x = x_75_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<fp16, [1, 384, 512]> input_83_cast_fp16 = mul(x = input_81_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<int32, [6]> input_85_pad_0 = const()[name = tensor<string, []>("input_85_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 1])];
            tensor<string, []> input_85_mode_0 = const()[name = tensor<string, []>("input_85_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_23_to_fp16 = const()[name = tensor<string, []>("const_23_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 384, 514]> input_85_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_85_mode_0, pad = input_85_pad_0, x = input_83_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<string, []> x_79_pad_type_0 = const()[name = tensor<string, []>("x_79_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_79_strides_0 = const()[name = tensor<string, []>("x_79_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_79_pad_0 = const()[name = tensor<string, []>("x_79_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_79_dilations_0 = const()[name = tensor<string, []>("x_79_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_79_groups_0 = const()[name = tensor<string, []>("x_79_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [72, 384, 3]> enc_p_encoder_ffn_layers_2_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_2_conv_2_weight_to_fp16"), val = tensor<fp16, [72, 384, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1431744)))];
            tensor<fp16, [72]> enc_p_encoder_ffn_layers_2_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_ffn_layers_2_conv_2_bias_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1597696)))];
            tensor<fp16, [1, 72, 512]> x_79_cast_fp16 = conv(bias = enc_p_encoder_ffn_layers_2_conv_2_bias_to_fp16, dilations = x_79_dilations_0, groups = x_79_groups_0, pad = x_79_pad_0, pad_type = x_79_pad_type_0, strides = x_79_strides_0, weight = enc_p_encoder_ffn_layers_2_conv_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("x_79_cast_fp16")];
            tensor<fp16, [1, 72, 512]> input_87_cast_fp16 = mul(x = x_79_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<fp16, [1, 72, 512]> x_81_cast_fp16 = add(x = x_73_cast_fp16, y = input_87_cast_fp16)[name = tensor<string, []>("x_81_cast_fp16")];
            tensor<int32, [3]> input_89_perm_0 = const()[name = tensor<string, []>("input_89_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<int32, [1]> x_83_axes_0 = const()[name = tensor<string, []>("x_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_2_2_gamma_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_2_2_gamma_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1597952)))];
            tensor<fp16, [72]> enc_p_encoder_norm_layers_2_2_beta_to_fp16 = const()[name = tensor<string, []>("enc_p_encoder_norm_layers_2_2_beta_to_fp16"), val = tensor<fp16, [72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1598208)))];
            tensor<fp16, [1, 512, 72]> input_89_cast_fp16 = transpose(perm = input_89_perm_0, x = x_81_cast_fp16)[name = tensor<string, []>("transpose_5")];
            tensor<fp16, [1, 512, 72]> x_83_cast_fp16 = layer_norm(axes = x_83_axes_0, beta = enc_p_encoder_norm_layers_2_2_beta_to_fp16, epsilon = var_20_to_fp16, gamma = enc_p_encoder_norm_layers_2_2_gamma_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
            tensor<int32, [3]> x_85_perm_0 = const()[name = tensor<string, []>("x_85_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<fp16, [1, 72, 512]> x_85_cast_fp16 = transpose(perm = x_85_perm_0, x = x_83_cast_fp16)[name = tensor<string, []>("transpose_4")];
            tensor<fp16, [1, 72, 512]> input_91_cast_fp16 = mul(x = x_85_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<string, []> var_494_pad_type_0 = const()[name = tensor<string, []>("op_494_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> var_494_strides_0 = const()[name = tensor<string, []>("op_494_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_494_pad_0 = const()[name = tensor<string, []>("op_494_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_494_dilations_0 = const()[name = tensor<string, []>("op_494_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> var_494_groups_0 = const()[name = tensor<string, []>("op_494_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [256, 72, 1]> enc_p_proj_weight_to_fp16 = const()[name = tensor<string, []>("enc_p_proj_weight_to_fp16"), val = tensor<fp16, [256, 72, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1598464)))];
            tensor<fp16, [256]> enc_p_proj_bias_to_fp16 = const()[name = tensor<string, []>("enc_p_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1635392)))];
            tensor<fp16, [1, 256, 512]> var_494_cast_fp16 = conv(bias = enc_p_proj_bias_to_fp16, dilations = var_494_dilations_0, groups = var_494_groups_0, pad = var_494_pad_0, pad_type = var_494_pad_type_0, strides = var_494_strides_0, weight = enc_p_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("op_494_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_495_cast_fp16 = mul(x = var_494_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("op_495_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_498_axis_0 = const()[name = tensor<string, []>("op_498_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 128, 512]> m_p, tensor<fp16, [1, 128, 512]> logs_p = split(axis = var_498_axis_0, split_sizes = tile_0, x = var_495_cast_fp16)[name = tensor<string, []>("op_498_cast_fp16")];
            tensor<fp16, [1, 72, 512]> input_93_cast_fp16 = mul(x = input_91_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<string, []> x_89_pad_type_0 = const()[name = tensor<string, []>("x_89_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_89_pad_0 = const()[name = tensor<string, []>("x_89_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> x_89_strides_0 = const()[name = tensor<string, []>("x_89_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_89_dilations_0 = const()[name = tensor<string, []>("x_89_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_89_groups_0 = const()[name = tensor<string, []>("x_89_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [256, 72, 3]> dp_conv_1_weight_to_fp16 = const()[name = tensor<string, []>("dp_conv_1_weight_to_fp16"), val = tensor<fp16, [256, 72, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1635968)))];
            tensor<fp16, [256]> dp_conv_1_bias_to_fp16 = const()[name = tensor<string, []>("dp_conv_1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1746624)))];
            tensor<fp16, [1, 256, 512]> x_89_cast_fp16 = conv(bias = dp_conv_1_bias_to_fp16, dilations = x_89_dilations_0, groups = x_89_groups_0, pad = x_89_pad_0, pad_type = x_89_pad_type_0, strides = x_89_strides_0, weight = dp_conv_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_91_cast_fp16 = relu(x = x_89_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
            tensor<int32, [3]> input_95_perm_0 = const()[name = tensor<string, []>("input_95_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<int32, [1]> x_93_axes_0 = const()[name = tensor<string, []>("x_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> dp_norm_1_gamma_to_fp16 = const()[name = tensor<string, []>("dp_norm_1_gamma_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1747200)))];
            tensor<fp16, [256]> dp_norm_1_beta_to_fp16 = const()[name = tensor<string, []>("dp_norm_1_beta_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1747776)))];
            tensor<fp16, []> var_504_to_fp16 = const()[name = tensor<string, []>("op_504_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 512, 256]> input_95_cast_fp16 = transpose(perm = input_95_perm_0, x = x_91_cast_fp16)[name = tensor<string, []>("transpose_3")];
            tensor<fp16, [1, 512, 256]> x_93_cast_fp16 = layer_norm(axes = x_93_axes_0, beta = dp_norm_1_beta_to_fp16, epsilon = var_504_to_fp16, gamma = dp_norm_1_gamma_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("x_93_cast_fp16")];
            tensor<int32, [3]> input_97_perm_0 = const()[name = tensor<string, []>("input_97_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<fp16, [1, 256, 512]> input_97_cast_fp16 = transpose(perm = input_97_perm_0, x = x_93_cast_fp16)[name = tensor<string, []>("transpose_2")];
            tensor<fp16, [1, 256, 512]> input_99_cast_fp16 = mul(x = input_97_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<string, []> x_97_pad_type_0 = const()[name = tensor<string, []>("x_97_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_97_pad_0 = const()[name = tensor<string, []>("x_97_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> x_97_strides_0 = const()[name = tensor<string, []>("x_97_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_97_dilations_0 = const()[name = tensor<string, []>("x_97_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_97_groups_0 = const()[name = tensor<string, []>("x_97_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [256, 256, 3]> dp_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("dp_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 256, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1748352)))];
            tensor<fp16, [256]> dp_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("dp_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2141632)))];
            tensor<fp16, [1, 256, 512]> x_97_cast_fp16 = conv(bias = dp_conv_2_bias_to_fp16, dilations = x_97_dilations_0, groups = x_97_groups_0, pad = x_97_pad_0, pad_type = x_97_pad_type_0, strides = x_97_strides_0, weight = dp_conv_2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_99_cast_fp16 = relu(x = x_97_cast_fp16)[name = tensor<string, []>("x_99_cast_fp16")];
            tensor<int32, [3]> input_101_perm_0 = const()[name = tensor<string, []>("input_101_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<int32, [1]> x_101_axes_0 = const()[name = tensor<string, []>("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> dp_norm_2_gamma_to_fp16 = const()[name = tensor<string, []>("dp_norm_2_gamma_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2142208)))];
            tensor<fp16, [256]> dp_norm_2_beta_to_fp16 = const()[name = tensor<string, []>("dp_norm_2_beta_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2142784)))];
            tensor<fp16, [1, 512, 256]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_1")];
            tensor<fp16, [1, 512, 256]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = dp_norm_2_beta_to_fp16, epsilon = var_504_to_fp16, gamma = dp_norm_2_gamma_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
            tensor<int32, [3]> input_103_perm_0 = const()[name = tensor<string, []>("input_103_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
            tensor<fp16, [1, 256, 512]> input_103_cast_fp16 = transpose(perm = input_103_perm_0, x = x_101_cast_fp16)[name = tensor<string, []>("transpose_0")];
            tensor<fp16, [1, 256, 512]> input_cast_fp16 = mul(x = input_103_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<string, []> x_pad_type_0 = const()[name = tensor<string, []>("x_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_strides_0 = const()[name = tensor<string, []>("x_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_pad_0 = const()[name = tensor<string, []>("x_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_dilations_0 = const()[name = tensor<string, []>("x_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_groups_0 = const()[name = tensor<string, []>("x_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 256, 1]> dp_proj_weight_to_fp16 = const()[name = tensor<string, []>("dp_proj_weight_to_fp16"), val = tensor<fp16, [1, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2143360)))];
            tensor<fp16, [1]> dp_proj_bias_to_fp16 = const()[name = tensor<string, []>("dp_proj_bias_to_fp16"), val = tensor<fp16, [1]>([0x1.3e4p-1])];
            tensor<fp16, [1, 1, 512]> x_cast_fp16 = conv(bias = dp_proj_bias_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = dp_proj_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
            tensor<fp16, [1, 1, 512]> logw = mul(x = x_cast_fp16, y = x_mask_to_fp16)[name = tensor<string, []>("op_555_cast_fp16")];
        } -> (m_p, logs_p, logw);
}