program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
{
    func main<ios17>(tensor<int32, [1]> audio_length, tensor<fp32, [1, 128, 17]> audio_signal, tensor<fp32, [17, 1, 70, 512]> cache_last_channel, tensor<int32, [1]> cache_last_channel_len, tensor<fp32, [17, 1, 512, 8]> cache_last_time, tensor<fp32, [1, 128, 16]> pre_cache) {
            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(2)];
            tensor<bool, []> full_input_interleave_0 = const()[name = tensor<string, []>("full_input_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> pre_cache_to_fp16_dtype_0 = const()[name = tensor<string, []>("pre_cache_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<string, []> audio_signal_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_signal_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 17]> audio_signal_to_fp16 = cast(dtype = audio_signal_to_fp16_dtype_0, x = audio_signal)[name = tensor<string, []>("cast_195")];
            tensor<fp16, [1, 128, 16]> pre_cache_to_fp16 = cast(dtype = pre_cache_to_fp16_dtype_0, x = pre_cache)[name = tensor<string, []>("cast_196")];
            tensor<fp16, [1, 128, 33]> full_input_cast_fp16 = concat(axis = var_9, interleave = full_input_interleave_0, values = (pre_cache_to_fp16, audio_signal_to_fp16))[name = tensor<string, []>("full_input_cast_fp16")];
            tensor<int32, []> var_12 = const()[name = tensor<string, []>("op_12"), val = tensor<int32, []>(16)];
            tensor<int32, [1]> value_1 = add(x = audio_length, y = var_12)[name = tensor<string, []>("value_1")];
            tensor<int32, [3]> var_28_begin_0 = const()[name = tensor<string, []>("op_28_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
            tensor<int32, [3]> var_28_end_0 = const()[name = tensor<string, []>("op_28_end_0"), val = tensor<int32, [3]>([1, 128, 33])];
            tensor<bool, [3]> var_28_end_mask_0 = const()[name = tensor<string, []>("op_28_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 16]> var_28_cast_fp16 = slice_by_index(begin = var_28_begin_0, end = var_28_end_0, end_mask = var_28_end_mask_0, x = full_input_cast_fp16)[name = tensor<string, []>("op_28_cast_fp16")];
            tensor<string, []> var_28_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_28_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> var_62 = const()[name = tensor<string, []>("op_62"), val = tensor<int32, []>(-1)];
            tensor<int32, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<int32, []>(1)];
            tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> cast_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> _inversed_108_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_108_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1]> value_1_to_fp16 = cast(dtype = cast_0_to_fp16_dtype_0, x = value_1)[name = tensor<string, []>("cast_193")];
            tensor<fp16, [1]> _inversed_108_cast_fp16 = mul(x = value_1_to_fp16, y = _inversed_108_y_0_to_fp16)[name = tensor<string, []>("_inversed_108_cast_fp16")];
            tensor<fp16, []> var_109_to_fp16 = const()[name = tensor<string, []>("op_109_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1]> lengths_1_cast_fp16 = add(x = _inversed_108_cast_fp16, y = var_109_to_fp16)[name = tensor<string, []>("lengths_1_cast_fp16")];
            tensor<fp16, [1]> lengths_3_cast_fp16 = floor(x = lengths_1_cast_fp16)[name = tensor<string, []>("lengths_3_cast_fp16")];
            tensor<fp16, []> _inversed_116_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_116_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1]> _inversed_116_cast_fp16 = mul(x = lengths_3_cast_fp16, y = _inversed_116_y_0_to_fp16)[name = tensor<string, []>("_inversed_116_cast_fp16")];
            tensor<fp16, []> var_117_to_fp16 = const()[name = tensor<string, []>("op_117_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1]> lengths_7_cast_fp16 = add(x = _inversed_116_cast_fp16, y = var_117_to_fp16)[name = tensor<string, []>("lengths_7_cast_fp16")];
            tensor<fp16, [1]> lengths_9_cast_fp16 = floor(x = lengths_7_cast_fp16)[name = tensor<string, []>("lengths_9_cast_fp16")];
            tensor<fp16, []> _inversed_124_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_124_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1]> _inversed_124_cast_fp16 = mul(x = lengths_9_cast_fp16, y = _inversed_124_y_0_to_fp16)[name = tensor<string, []>("_inversed_124_cast_fp16")];
            tensor<fp16, []> var_125_to_fp16 = const()[name = tensor<string, []>("op_125_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1]> lengths_13_cast_fp16 = add(x = _inversed_124_cast_fp16, y = var_125_to_fp16)[name = tensor<string, []>("lengths_13_cast_fp16")];
            tensor<fp16, [1]> lengths_cast_fp16 = floor(x = lengths_13_cast_fp16)[name = tensor<string, []>("lengths_cast_fp16")];
            tensor<string, []> cast_9_dtype_0 = const()[name = tensor<string, []>("cast_9_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [1]> input_1_axes_0 = const()[name = tensor<string, []>("input_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 33, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = full_input_cast_fp16)[name = tensor<string, []>("transpose_241")];
            tensor<fp16, [1, 1, 33, 128]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = x_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<int32, [8]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
            tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 1, 36, 131]> input_3_cast_fp16 = pad(constant_val = const_0_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_5_strides_0 = const()[name = tensor<string, []>("input_5_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [4]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = tensor<string, []>("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_5_groups_0 = const()[name = tensor<string, []>("input_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736)))];
            tensor<fp16, [1, 256, 17, 65]> input_5_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<fp16, [1, 256, 17, 65]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<int32, [8]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_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, 256, 20, 68]> input_9_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(256)];
            tensor<int32, [4]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5312)))];
            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9984)))];
            tensor<fp16, [1, 256, 9, 33]> input_11_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16, x = input_9_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, [2]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10560)))];
            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141696)))];
            tensor<fp16, [1, 256, 9, 33]> input_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_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 = encoder_pre_encode_conv_3_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<fp16, [1, 256, 9, 33]> input_15_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<int32, [8]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
            tensor<string, []> input_17_mode_0 = const()[name = tensor<string, []>("input_17_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, 256, 12, 36]> input_17_cast_fp16 = pad(constant_val = const_2_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(256)];
            tensor<int32, [4]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142272)))];
            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146944)))];
            tensor<fp16, [1, 256, 5, 17]> input_19_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<string, []> input_21_pad_type_0 = const()[name = tensor<string, []>("input_21_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = tensor<string, []>("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_21_pad_0 = const()[name = tensor<string, []>("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = tensor<string, []>("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_21_groups_0 = const()[name = tensor<string, []>("input_21_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147520)))];
            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278656)))];
            tensor<fp16, [1, 256, 5, 17]> input_21_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<fp16, [1, 256, 5, 17]> x_3_cast_fp16 = relu(x = input_21_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
            tensor<int32, [4]> var_181_perm_0 = const()[name = tensor<string, []>("op_181_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_182 = const()[name = tensor<string, []>("op_182"), val = tensor<int32, [3]>([1, 5, -1])];
            tensor<fp16, [1, 5, 256, 17]> var_181_cast_fp16 = transpose(perm = var_181_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_240")];
            tensor<fp16, [1, 5, 4352]> input_23_cast_fp16 = reshape(shape = var_182, x = var_181_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<fp16, [512, 4352]> encoder_pre_encode_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_weight_to_fp16"), val = tensor<fp16, [512, 4352]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279232)))];
            tensor<fp16, [512]> encoder_pre_encode_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4735744)))];
            tensor<fp16, [1, 5, 512]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<int32, [3]> var_192_begin_0 = const()[name = tensor<string, []>("op_192_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_192_end_0 = const()[name = tensor<string, []>("op_192_end_0"), val = tensor<int32, [3]>([1, 5, 512])];
            tensor<bool, [3]> var_192_end_mask_0 = const()[name = tensor<string, []>("op_192_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 3, 512]> var_192_cast_fp16 = slice_by_index(begin = var_192_begin_0, end = var_192_end_0, end_mask = var_192_end_mask_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_192_cast_fp16")];
            tensor<int32, []> var_194 = const()[name = tensor<string, []>("op_194"), val = tensor<int32, []>(2)];
            tensor<int32, [1]> lengths_cast_fp16_to_int32 = cast(dtype = cast_9_dtype_0, x = lengths_cast_fp16)[name = tensor<string, []>("cast_192")];
            tensor<int32, [1]> var_195 = sub(x = lengths_cast_fp16_to_int32, y = var_194)[name = tensor<string, []>("op_195")];
            tensor<string, []> var_195_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_195_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_60_promoted_to_fp16 = const()[name = tensor<string, []>("op_60_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, []> const_5_to_fp16 = const()[name = tensor<string, []>("const_5_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<fp16, [1]> var_195_to_fp16 = cast(dtype = var_195_promoted_to_fp16_dtype_0, x = var_195)[name = tensor<string, []>("cast_191")];
            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_60_promoted_to_fp16, beta = const_5_to_fp16, x = var_195_to_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
            tensor<int32, [1]> cache_keep_size = const()[name = tensor<string, []>("cache_keep_size"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_211_promoted_to_fp16 = const()[name = tensor<string, []>("op_211_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_211_promoted_to_fp16)[name = tensor<string, []>("padding_length_cast_fp16")];
            tensor<int32, []> const_7 = const()[name = tensor<string, []>("const_7"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> var_213 = mul(x = cache_last_channel_len, y = const_7)[name = tensor<string, []>("op_213")];
            tensor<int32, []> var_214 = const()[name = tensor<string, []>("op_214"), val = tensor<int32, []>(70)];
            tensor<int32, [1]> offset = add(x = var_213, y = var_214)[name = tensor<string, []>("offset")];
            tensor<int32, [1]> var_254_axes_0 = const()[name = tensor<string, []>("op_254_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> var_254_cast_fp16 = expand_dims(axes = var_254_axes_0, x = padding_length_cast_fp16)[name = tensor<string, []>("op_254_cast_fp16")];
            tensor<fp16, [1, 73]> var_253_promoted_to_fp16 = const()[name = tensor<string, []>("op_253_promoted_to_fp16"), val = tensor<fp16, [1, 73]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736832)))];
            tensor<bool, [1, 73]> pad_mask_1_cast_fp16 = less(x = var_253_promoted_to_fp16, y = var_254_cast_fp16)[name = tensor<string, []>("pad_mask_1_cast_fp16")];
            tensor<int32, [1, 73]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<int32, [1, 73]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72]])];
            tensor<int32, [1]> var_260_axes_0 = const()[name = tensor<string, []>("op_260_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 1]> var_260 = expand_dims(axes = var_260_axes_0, x = offset)[name = tensor<string, []>("op_260")];
            tensor<bool, [1, 73]> pad_mask_off = greater_equal(x = expand_dims_1, y = var_260)[name = tensor<string, []>("pad_mask_off")];
            tensor<bool, [1, 73]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = tensor<string, []>("pad_mask_3")];
            tensor<int32, [1]> var_263_axes_0 = const()[name = tensor<string, []>("op_263_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 73]> var_263 = expand_dims(axes = var_263_axes_0, x = pad_mask_3)[name = tensor<string, []>("op_263")];
            tensor<int32, [3]> var_264 = const()[name = tensor<string, []>("op_264"), val = tensor<int32, [3]>([1, 73, 1])];
            tensor<bool, [1, 73, 73]> pad_mask_for_att_mask_1 = tile(reps = var_264, x = var_263)[name = tensor<string, []>("pad_mask_for_att_mask_1")];
            tensor<int32, [3]> var_266_perm_0 = const()[name = tensor<string, []>("op_266_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<bool, [1, 73, 73]> var_266 = transpose(perm = var_266_perm_0, x = pad_mask_for_att_mask_1)[name = tensor<string, []>("transpose_239")];
            tensor<bool, [1, 73, 73]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_266)[name = tensor<string, []>("pad_mask_for_att_mask")];
            tensor<bool, [1, 73, 73]> const_15 = const()[name = tensor<string, []>("const_15"), val = tensor<bool, [1, 73, 73]>([[[true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false], [false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
            tensor<bool, [1, 73, 73]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_15)[name = tensor<string, []>("att_mask_9")];
            tensor<bool, [1, 73, 73]> att_mask = logical_not(x = att_mask_9)[name = tensor<string, []>("att_mask")];
            tensor<bool, [1, 73]> pad_mask_5 = logical_not(x = pad_mask_3)[name = tensor<string, []>("pad_mask_5")];
            tensor<int32, [2]> pad_mask_begin_0 = const()[name = tensor<string, []>("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 70])];
            tensor<int32, [2]> pad_mask_end_0 = const()[name = tensor<string, []>("pad_mask_end_0"), val = tensor<int32, [2]>([1, 73])];
            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = tensor<string, []>("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<bool, [1, 3]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = tensor<string, []>("pad_mask")];
            tensor<int32, [3]> mask_1_begin_0 = const()[name = tensor<string, []>("mask_1_begin_0"), val = tensor<int32, [3]>([0, 70, 0])];
            tensor<int32, [3]> mask_1_end_0 = const()[name = tensor<string, []>("mask_1_end_0"), val = tensor<int32, [3]>([1, 73, 73])];
            tensor<bool, [3]> mask_1_end_mask_0 = const()[name = tensor<string, []>("mask_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<bool, [1, 3, 73]> mask_1 = slice_by_index(begin = mask_1_begin_0, end = mask_1_end_0, end_mask = mask_1_end_mask_0, x = att_mask)[name = tensor<string, []>("mask_1")];
            tensor<int32, [4]> cache_1_begin_0 = const()[name = tensor<string, []>("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> cache_1_end_0 = const()[name = tensor<string, []>("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 70, 512])];
            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = tensor<string, []>("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = tensor<string, []>("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<string, []> cache_last_channel_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_last_channel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [17, 1, 70, 512]> cache_last_channel_to_fp16 = cast(dtype = cache_last_channel_to_fp16_dtype_0, x = cache_last_channel)[name = tensor<string, []>("cast_190")];
            tensor<fp16, [1, 70, 512]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_1_cast_fp16")];
            tensor<int32, [4]> cache_3_begin_0 = const()[name = tensor<string, []>("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> cache_3_end_0 = const()[name = tensor<string, []>("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 512, 8])];
            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = tensor<string, []>("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = tensor<string, []>("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<string, []> cache_last_time_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_last_time_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [17, 1, 512, 8]> cache_last_time_to_fp16 = cast(dtype = cache_last_time_to_fp16_dtype_0, x = cache_last_time)[name = tensor<string, []>("cast_189")];
            tensor<fp16, [1, 512, 8]> cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_3_cast_fp16")];
            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4737088)))];
            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4738176)))];
            tensor<fp16, []> var_38_to_fp16 = const()[name = tensor<string, []>("op_38_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 3, 512]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_192_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4739264)))];
            tensor<fp16, [2048]> linear_1_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_1_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6836480)))];
            tensor<fp16, [1, 3, 2048]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6840640)))];
            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8937856)))];
            tensor<fp16, [1, 3, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<fp16, []> var_303_to_fp16 = const()[name = tensor<string, []>("op_303_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_304_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_303_to_fp16)[name = tensor<string, []>("op_304_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_37_cast_fp16 = add(x = var_192_cast_fp16, y = var_304_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<int32, [1]> key_1_axes_0 = const()[name = tensor<string, []>("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8938944)))];
            tensor<fp16, [512]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8940032)))];
            tensor<fp16, [1, 3, 512]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
            tensor<bool, []> input_39_interleave_0 = const()[name = tensor<string, []>("input_39_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_39_cast_fp16 = concat(axis = var_64, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<int32, [3]> var_326_begin_0 = const()[name = tensor<string, []>("op_326_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_326_end_0 = const()[name = tensor<string, []>("op_326_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_326_end_mask_0 = const()[name = tensor<string, []>("op_326_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_326_cast_fp16 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = cache_1_cast_fp16)[name = tensor<string, []>("op_326_cast_fp16")];
            tensor<int32, [3]> var_329_begin_0 = const()[name = tensor<string, []>("op_329_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_329_end_0 = const()[name = tensor<string, []>("op_329_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_329_end_mask_0 = const()[name = tensor<string, []>("op_329_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_329_cast_fp16 = slice_by_index(begin = var_329_begin_0, end = var_329_end_0, end_mask = var_329_end_mask_0, x = key_1_cast_fp16)[name = tensor<string, []>("op_329_cast_fp16")];
            tensor<bool, []> var_332_interleave_0 = const()[name = tensor<string, []>("op_332_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_332_cast_fp16 = concat(axis = var_64, interleave = var_332_interleave_0, values = (var_326_cast_fp16, var_329_cast_fp16))[name = tensor<string, []>("op_332_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8941120)))];
            tensor<fp16, [1, 3, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16, x = key_1_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<int32, [4]> var_336 = const()[name = tensor<string, []>("op_336"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_1_cast_fp16 = reshape(shape = var_336, x = linear_3_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9465472)))];
            tensor<fp16, [1, 73, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<int32, [4]> var_340 = const()[name = tensor<string, []>("op_340"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_1_cast_fp16 = reshape(shape = var_340, x = linear_4_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9989824)))];
            tensor<fp16, [1, 73, 512]> linear_5_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<int32, [4]> var_344 = const()[name = tensor<string, []>("op_344"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_1_cast_fp16 = reshape(shape = var_344, x = linear_5_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
            tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10514176)))];
            tensor<fp16, [1, 3, 8, 64]> var_356_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_356_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10515264)))];
            tensor<fp16, [1, 3, 8, 64]> var_358_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_358_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_7_transpose_x_0 = const()[name = tensor<string, []>("x_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_7_transpose_y_0 = const()[name = tensor<string, []>("x_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_360_to_fp16 = const()[name = tensor<string, []>("op_360_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10516352)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_358_cast_fp16)[name = tensor<string, []>("transpose_237")];
            tensor<fp16, [1, 8, 3, 145]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = var_360_to_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
            tensor<int32, [8]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_9_mode_0 = const()[name = tensor<string, []>("x_9_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, 8, 3, 146]> x_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
            tensor<int32, [4]> var_368 = const()[name = tensor<string, []>("op_368"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_11_cast_fp16 = reshape(shape = var_368, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<int32, [4]> var_372_begin_0 = const()[name = tensor<string, []>("op_372_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_372_end_0 = const()[name = tensor<string, []>("op_372_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_372_end_mask_0 = const()[name = tensor<string, []>("op_372_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_372_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = tensor<string, []>("op_373"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_1_cast_fp16 = reshape(shape = var_373, x = var_372_cast_fp16)[name = tensor<string, []>("matrix_bd_1_cast_fp16")];
            tensor<bool, []> matrix_ac_1_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_1_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_51_perm_0 = const()[name = tensor<string, []>("transpose_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = tensor<string, []>("transpose_52_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_52 = transpose(perm = transpose_52_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_235")];
            tensor<fp16, [1, 8, 3, 64]> transpose_51 = transpose(perm = transpose_51_perm_0, x = var_356_cast_fp16)[name = tensor<string, []>("transpose_236")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_51, y = transpose_52)[name = tensor<string, []>("matrix_ac_1_cast_fp16")];
            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = tensor<string, []>("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = tensor<string, []>("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = tensor<string, []>("matrix_bd_3_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_382_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = tensor<string, []>("op_382_cast_fp16")];
            tensor<fp16, []> _inversed_scores_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_1_cast_fp16 = mul(x = var_382_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_1_cast_fp16")];
            tensor<int32, [1]> mask_3_axes_0 = const()[name = tensor<string, []>("mask_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 3, 73]> mask_3 = expand_dims(axes = mask_3_axes_0, x = mask_1)[name = tensor<string, []>("mask_3")];
            tensor<fp16, []> var_41_to_fp16 = const()[name = tensor<string, []>("op_41_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
            tensor<fp16, [1, 8, 3, 73]> scores_3_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_3_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_388_cast_fp16 = softmax(axis = var_62, x = scores_3_cast_fp16)[name = tensor<string, []>("op_388_cast_fp16")];
            tensor<fp16, []> var_40_to_fp16 = const()[name = tensor<string, []>("op_40_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 73]> input_41_cast_fp16 = select(a = var_40_to_fp16, b = var_388_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<bool, []> x_13_transpose_x_0 = const()[name = tensor<string, []>("x_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_13_transpose_y_0 = const()[name = tensor<string, []>("x_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = v_1_cast_fp16)[name = tensor<string, []>("transpose_238")];
            tensor<fp16, [1, 8, 3, 64]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_3_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
            tensor<int32, [4]> var_392_perm_0 = const()[name = tensor<string, []>("op_392_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_393 = const()[name = tensor<string, []>("op_393"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_392_cast_fp16 = transpose(perm = var_392_perm_0, x = x_13_cast_fp16)[name = tensor<string, []>("transpose_234")];
            tensor<fp16, [1, 3, 512]> input_43_cast_fp16 = reshape(shape = var_393, x = var_392_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10664896)))];
            tensor<fp16, [1, 3, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11189248)))];
            tensor<fp16, [512]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11190336)))];
            tensor<fp16, [1, 3, 512]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_51_pad_type_0 = const()[name = tensor<string, []>("input_51_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_51_strides_0 = const()[name = tensor<string, []>("input_51_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_51_pad_0 = const()[name = tensor<string, []>("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_51_dilations_0 = const()[name = tensor<string, []>("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_51_groups_0 = const()[name = tensor<string, []>("input_51_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11191424)))];
            tensor<fp16, [1, 512, 3]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_233")];
            tensor<fp16, [1, 1024, 3]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<int32, []> x_19_split_num_splits_0 = const()[name = tensor<string, []>("x_19_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_19_split_axis_0 = const()[name = tensor<string, []>("x_19_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_19_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = tensor<string, []>("x_19_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = tensor<string, []>("x_19_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
            tensor<int32, [1]> var_418_axes_0 = const()[name = tensor<string, []>("op_418_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 3]> var_418 = expand_dims(axes = var_418_axes_0, x = pad_mask)[name = tensor<string, []>("op_418")];
            tensor<fp16, [1, 512, 3]> input_53_cast_fp16 = select(a = var_40_to_fp16, b = x_19_cast_fp16, cond = var_418)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<bool, []> new_x_3_interleave_0 = const()[name = tensor<string, []>("new_x_3_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_3_cast_fp16 = concat(axis = var_62, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = tensor<string, []>("new_x_3_cast_fp16")];
            tensor<int32, [3]> next_cache_1_begin_0 = const()[name = tensor<string, []>("next_cache_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_1_end_0 = const()[name = tensor<string, []>("next_cache_1_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_1_end_mask_0 = const()[name = tensor<string, []>("next_cache_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_1_cast_fp16 = slice_by_index(begin = next_cache_1_begin_0, end = next_cache_1_end_0, end_mask = next_cache_1_end_mask_0, x = new_x_3_cast_fp16)[name = tensor<string, []>("next_cache_1_cast_fp16")];
            tensor<int32, [3]> var_434_begin_0 = const()[name = tensor<string, []>("op_434_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_434_end_0 = const()[name = tensor<string, []>("op_434_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_434_end_mask_0 = const()[name = tensor<string, []>("op_434_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_434_cast_fp16 = slice_by_index(begin = var_434_begin_0, end = var_434_end_0, end_mask = var_434_end_mask_0, x = next_cache_1_cast_fp16)[name = tensor<string, []>("op_434_cast_fp16")];
            tensor<string, []> x_21_pad_type_0 = const()[name = tensor<string, []>("x_21_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_21_groups_0 = const()[name = tensor<string, []>("x_21_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_21_strides_0 = const()[name = tensor<string, []>("x_21_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_21_pad_0 = const()[name = tensor<string, []>("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_21_dilations_0 = const()[name = tensor<string, []>("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12240064)))];
            tensor<fp16, [1, 512, 3]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16, x = new_x_3_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
            tensor<int32, [3]> input_55_perm_0 = const()[name = tensor<string, []>("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12249344)))];
            tensor<fp16, [512]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12250432)))];
            tensor<fp16, [1, 3, 512]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = tensor<string, []>("transpose_232")];
            tensor<fp16, [1, 3, 512]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
            tensor<int32, [3]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_231")];
            tensor<fp16, [1, 512, 3]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<string, []> x_25_pad_type_0 = const()[name = tensor<string, []>("x_25_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_25_strides_0 = const()[name = tensor<string, []>("x_25_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_25_pad_0 = const()[name = tensor<string, []>("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_25_dilations_0 = const()[name = tensor<string, []>("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_25_groups_0 = const()[name = tensor<string, []>("x_25_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12251520)))];
            tensor<fp16, [1, 512, 3]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
            tensor<int32, [3]> input_61_perm_0 = const()[name = tensor<string, []>("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = tensor<string, []>("transpose_230")];
            tensor<fp16, [1, 3, 512]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<int32, [1]> input_65_axes_0 = const()[name = tensor<string, []>("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12775872)))];
            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12776960)))];
            tensor<fp16, [1, 3, 512]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12778048)))];
            tensor<fp16, [1, 3, 2048]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14875264)))];
            tensor<fp16, [1, 3, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<fp16, []> var_475_to_fp16 = const()[name = tensor<string, []>("op_475_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_476_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_475_to_fp16)[name = tensor<string, []>("op_476_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_476_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16972480)))];
            tensor<fp16, [512]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16973568)))];
            tensor<fp16, [1, 3, 512]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<int32, [4]> cache_5_begin_0 = const()[name = tensor<string, []>("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> cache_5_end_0 = const()[name = tensor<string, []>("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 70, 512])];
            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = tensor<string, []>("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = tensor<string, []>("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_5_cast_fp16")];
            tensor<int32, [4]> cache_7_begin_0 = const()[name = tensor<string, []>("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> cache_7_end_0 = const()[name = tensor<string, []>("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 512, 8])];
            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = tensor<string, []>("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = tensor<string, []>("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_7_cast_fp16")];
            tensor<int32, [1]> input_79_axes_0 = const()[name = tensor<string, []>("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16974656)))];
            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16975744)))];
            tensor<fp16, [1, 3, 512]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16976832)))];
            tensor<fp16, [1, 3, 2048]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19074048)))];
            tensor<fp16, [1, 3, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, []> var_510_to_fp16 = const()[name = tensor<string, []>("op_510_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_511_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_510_to_fp16)[name = tensor<string, []>("op_511_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_511_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<int32, [1]> key_3_axes_0 = const()[name = tensor<string, []>("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21171264)))];
            tensor<fp16, [512]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21172352)))];
            tensor<fp16, [1, 3, 512]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
            tensor<bool, []> input_91_interleave_0 = const()[name = tensor<string, []>("input_91_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_91_cast_fp16 = concat(axis = var_64, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<int32, [3]> var_533_begin_0 = const()[name = tensor<string, []>("op_533_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_533_end_0 = const()[name = tensor<string, []>("op_533_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_533_end_mask_0 = const()[name = tensor<string, []>("op_533_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_533_cast_fp16 = slice_by_index(begin = var_533_begin_0, end = var_533_end_0, end_mask = var_533_end_mask_0, x = cache_5_cast_fp16)[name = tensor<string, []>("op_533_cast_fp16")];
            tensor<int32, [3]> var_536_begin_0 = const()[name = tensor<string, []>("op_536_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_536_end_0 = const()[name = tensor<string, []>("op_536_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_536_end_mask_0 = const()[name = tensor<string, []>("op_536_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_536_cast_fp16 = slice_by_index(begin = var_536_begin_0, end = var_536_end_0, end_mask = var_536_end_mask_0, x = key_3_cast_fp16)[name = tensor<string, []>("op_536_cast_fp16")];
            tensor<bool, []> var_539_interleave_0 = const()[name = tensor<string, []>("op_539_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_539_cast_fp16 = concat(axis = var_64, interleave = var_539_interleave_0, values = (var_533_cast_fp16, var_536_cast_fp16))[name = tensor<string, []>("op_539_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21173440)))];
            tensor<fp16, [1, 3, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16, x = key_3_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<int32, [4]> var_543 = const()[name = tensor<string, []>("op_543"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_7_cast_fp16 = reshape(shape = var_543, x = linear_12_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21697792)))];
            tensor<fp16, [1, 73, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<int32, [4]> var_547 = const()[name = tensor<string, []>("op_547"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_5_cast_fp16 = reshape(shape = var_547, x = linear_13_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22222144)))];
            tensor<fp16, [1, 73, 512]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<int32, [4]> var_551 = const()[name = tensor<string, []>("op_551"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_3_cast_fp16 = reshape(shape = var_551, x = linear_14_cast_fp16)[name = tensor<string, []>("v_3_cast_fp16")];
            tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22746496)))];
            tensor<fp16, [1, 3, 8, 64]> var_563_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_563_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22747584)))];
            tensor<fp16, [1, 3, 8, 64]> var_565_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_565_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_33_transpose_x_0 = const()[name = tensor<string, []>("x_33_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_33_transpose_y_0 = const()[name = tensor<string, []>("x_33_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_567_to_fp16 = const()[name = tensor<string, []>("op_567_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22748672)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_565_cast_fp16)[name = tensor<string, []>("transpose_228")];
            tensor<fp16, [1, 8, 3, 145]> x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = var_567_to_fp16)[name = tensor<string, []>("x_33_cast_fp16")];
            tensor<int32, [8]> x_35_pad_0 = const()[name = tensor<string, []>("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_35_mode_0 = const()[name = tensor<string, []>("x_35_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_36_to_fp16 = const()[name = tensor<string, []>("const_36_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_35_cast_fp16 = pad(constant_val = const_36_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<int32, [4]> var_575 = const()[name = tensor<string, []>("op_575"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_37_cast_fp16 = reshape(shape = var_575, x = x_35_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<int32, [4]> var_579_begin_0 = const()[name = tensor<string, []>("op_579_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_579_end_0 = const()[name = tensor<string, []>("op_579_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_579_end_mask_0 = const()[name = tensor<string, []>("op_579_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_579_cast_fp16 = slice_by_index(begin = var_579_begin_0, end = var_579_end_0, end_mask = var_579_end_mask_0, x = x_37_cast_fp16)[name = tensor<string, []>("op_579_cast_fp16")];
            tensor<int32, [4]> var_580 = const()[name = tensor<string, []>("op_580"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_5_cast_fp16 = reshape(shape = var_580, x = var_579_cast_fp16)[name = tensor<string, []>("matrix_bd_5_cast_fp16")];
            tensor<bool, []> matrix_ac_3_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_3_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_53_perm_0 = const()[name = tensor<string, []>("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = tensor<string, []>("transpose_54_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_54 = transpose(perm = transpose_54_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_226")];
            tensor<fp16, [1, 8, 3, 64]> transpose_53 = transpose(perm = transpose_53_perm_0, x = var_563_cast_fp16)[name = tensor<string, []>("transpose_227")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_53, y = transpose_54)[name = tensor<string, []>("matrix_ac_3_cast_fp16")];
            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = tensor<string, []>("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = tensor<string, []>("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = tensor<string, []>("matrix_bd_7_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_589_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = tensor<string, []>("op_589_cast_fp16")];
            tensor<fp16, []> _inversed_scores_5_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_5_cast_fp16 = mul(x = var_589_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_5_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_7_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_7_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_595_cast_fp16 = softmax(axis = var_62, x = scores_7_cast_fp16)[name = tensor<string, []>("op_595_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_93_cast_fp16 = select(a = var_40_to_fp16, b = var_595_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<bool, []> x_39_transpose_x_0 = const()[name = tensor<string, []>("x_39_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_39_transpose_y_0 = const()[name = tensor<string, []>("x_39_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = v_3_cast_fp16)[name = tensor<string, []>("transpose_229")];
            tensor<fp16, [1, 8, 3, 64]> x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_5_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
            tensor<int32, [4]> var_599_perm_0 = const()[name = tensor<string, []>("op_599_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_600 = const()[name = tensor<string, []>("op_600"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_599_cast_fp16 = transpose(perm = var_599_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_225")];
            tensor<fp16, [1, 3, 512]> input_95_cast_fp16 = reshape(shape = var_600, x = var_599_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22897216)))];
            tensor<fp16, [1, 3, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23421568)))];
            tensor<fp16, [512]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23422656)))];
            tensor<fp16, [1, 3, 512]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
            tensor<int32, [3]> input_101_perm_0 = const()[name = tensor<string, []>("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23423744)))];
            tensor<fp16, [1, 512, 3]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_224")];
            tensor<fp16, [1, 1024, 3]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<int32, []> x_45_split_num_splits_0 = const()[name = tensor<string, []>("x_45_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_45_split_axis_0 = const()[name = tensor<string, []>("x_45_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_45_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = tensor<string, []>("x_45_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = tensor<string, []>("x_45_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_105_cast_fp16 = select(a = var_40_to_fp16, b = x_45_cast_fp16, cond = var_418)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<bool, []> new_x_7_interleave_0 = const()[name = tensor<string, []>("new_x_7_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_7_cast_fp16 = concat(axis = var_62, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = tensor<string, []>("new_x_7_cast_fp16")];
            tensor<int32, [3]> next_cache_3_begin_0 = const()[name = tensor<string, []>("next_cache_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_3_end_0 = const()[name = tensor<string, []>("next_cache_3_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_3_end_mask_0 = const()[name = tensor<string, []>("next_cache_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_3_cast_fp16 = slice_by_index(begin = next_cache_3_begin_0, end = next_cache_3_end_0, end_mask = next_cache_3_end_mask_0, x = new_x_7_cast_fp16)[name = tensor<string, []>("next_cache_3_cast_fp16")];
            tensor<int32, [3]> var_641_begin_0 = const()[name = tensor<string, []>("op_641_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_641_end_0 = const()[name = tensor<string, []>("op_641_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_641_end_mask_0 = const()[name = tensor<string, []>("op_641_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_641_cast_fp16 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = next_cache_3_cast_fp16)[name = tensor<string, []>("op_641_cast_fp16")];
            tensor<string, []> x_47_pad_type_0 = const()[name = tensor<string, []>("x_47_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_47_groups_0 = const()[name = tensor<string, []>("x_47_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_47_strides_0 = const()[name = tensor<string, []>("x_47_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_47_pad_0 = const()[name = tensor<string, []>("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_47_dilations_0 = const()[name = tensor<string, []>("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24472384)))];
            tensor<fp16, [1, 512, 3]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16, x = new_x_7_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
            tensor<int32, [3]> input_107_perm_0 = const()[name = tensor<string, []>("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_49_axes_0 = const()[name = tensor<string, []>("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24481664)))];
            tensor<fp16, [512]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24482752)))];
            tensor<fp16, [1, 3, 512]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_223")];
            tensor<fp16, [1, 3, 512]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
            tensor<int32, [3]> input_109_perm_0 = const()[name = tensor<string, []>("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = tensor<string, []>("transpose_222")];
            tensor<fp16, [1, 512, 3]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
            tensor<string, []> x_51_pad_type_0 = const()[name = tensor<string, []>("x_51_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_51_strides_0 = const()[name = tensor<string, []>("x_51_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_51_pad_0 = const()[name = tensor<string, []>("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_51_dilations_0 = const()[name = tensor<string, []>("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_51_groups_0 = const()[name = tensor<string, []>("x_51_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24483840)))];
            tensor<fp16, [1, 512, 3]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
            tensor<int32, [3]> input_113_perm_0 = const()[name = tensor<string, []>("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_221")];
            tensor<fp16, [1, 3, 512]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<int32, [1]> input_117_axes_0 = const()[name = tensor<string, []>("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25008192)))];
            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25009280)))];
            tensor<fp16, [1, 3, 512]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25010368)))];
            tensor<fp16, [1, 3, 2048]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27107584)))];
            tensor<fp16, [1, 3, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<fp16, []> var_682_to_fp16 = const()[name = tensor<string, []>("op_682_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_683_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_682_to_fp16)[name = tensor<string, []>("op_683_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_683_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<int32, [1]> input_129_axes_0 = const()[name = tensor<string, []>("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29204800)))];
            tensor<fp16, [512]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29205888)))];
            tensor<fp16, [1, 3, 512]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<int32, [4]> cache_9_begin_0 = const()[name = tensor<string, []>("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> cache_9_end_0 = const()[name = tensor<string, []>("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 70, 512])];
            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = tensor<string, []>("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = tensor<string, []>("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_9_cast_fp16")];
            tensor<int32, [4]> cache_11_begin_0 = const()[name = tensor<string, []>("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> cache_11_end_0 = const()[name = tensor<string, []>("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 512, 8])];
            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = tensor<string, []>("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = tensor<string, []>("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_11_cast_fp16")];
            tensor<int32, [1]> input_131_axes_0 = const()[name = tensor<string, []>("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29206976)))];
            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29208064)))];
            tensor<fp16, [1, 3, 512]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29209152)))];
            tensor<fp16, [1, 3, 2048]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31306368)))];
            tensor<fp16, [1, 3, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<fp16, []> var_717_to_fp16 = const()[name = tensor<string, []>("op_717_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_718_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_717_to_fp16)[name = tensor<string, []>("op_718_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_718_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<int32, [1]> key_5_axes_0 = const()[name = tensor<string, []>("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33403584)))];
            tensor<fp16, [512]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33404672)))];
            tensor<fp16, [1, 3, 512]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_143_cast_fp16 = concat(axis = var_64, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<int32, [3]> var_740_begin_0 = const()[name = tensor<string, []>("op_740_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_740_end_0 = const()[name = tensor<string, []>("op_740_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_740_end_mask_0 = const()[name = tensor<string, []>("op_740_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_740_cast_fp16 = slice_by_index(begin = var_740_begin_0, end = var_740_end_0, end_mask = var_740_end_mask_0, x = cache_9_cast_fp16)[name = tensor<string, []>("op_740_cast_fp16")];
            tensor<int32, [3]> var_743_begin_0 = const()[name = tensor<string, []>("op_743_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_743_end_0 = const()[name = tensor<string, []>("op_743_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_743_end_mask_0 = const()[name = tensor<string, []>("op_743_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_743_cast_fp16 = slice_by_index(begin = var_743_begin_0, end = var_743_end_0, end_mask = var_743_end_mask_0, x = key_5_cast_fp16)[name = tensor<string, []>("op_743_cast_fp16")];
            tensor<bool, []> var_746_interleave_0 = const()[name = tensor<string, []>("op_746_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_746_cast_fp16 = concat(axis = var_64, interleave = var_746_interleave_0, values = (var_740_cast_fp16, var_743_cast_fp16))[name = tensor<string, []>("op_746_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33405760)))];
            tensor<fp16, [1, 3, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16, x = key_5_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<int32, [4]> var_750 = const()[name = tensor<string, []>("op_750"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_13_cast_fp16 = reshape(shape = var_750, x = linear_21_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33930112)))];
            tensor<fp16, [1, 73, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<int32, [4]> var_754 = const()[name = tensor<string, []>("op_754"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_9_cast_fp16 = reshape(shape = var_754, x = linear_22_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34454464)))];
            tensor<fp16, [1, 73, 512]> linear_23_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<int32, [4]> var_758 = const()[name = tensor<string, []>("op_758"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_5_cast_fp16 = reshape(shape = var_758, x = linear_23_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
            tensor<int32, [4]> value_7_perm_0 = const()[name = tensor<string, []>("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34978816)))];
            tensor<fp16, [1, 3, 8, 64]> var_770_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_770_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34979904)))];
            tensor<fp16, [1, 3, 8, 64]> var_772_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_772_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_59_transpose_x_0 = const()[name = tensor<string, []>("x_59_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_59_transpose_y_0 = const()[name = tensor<string, []>("x_59_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_774_to_fp16 = const()[name = tensor<string, []>("op_774_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34980992)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_772_cast_fp16)[name = tensor<string, []>("transpose_219")];
            tensor<fp16, [1, 8, 3, 145]> x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = var_774_to_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
            tensor<int32, [8]> x_61_pad_0 = const()[name = tensor<string, []>("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_61_mode_0 = const()[name = tensor<string, []>("x_61_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_49_to_fp16 = const()[name = tensor<string, []>("const_49_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_61_cast_fp16 = pad(constant_val = const_49_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
            tensor<int32, [4]> var_782 = const()[name = tensor<string, []>("op_782"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_63_cast_fp16 = reshape(shape = var_782, x = x_61_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
            tensor<int32, [4]> var_786_begin_0 = const()[name = tensor<string, []>("op_786_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_786_end_0 = const()[name = tensor<string, []>("op_786_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_786_end_mask_0 = const()[name = tensor<string, []>("op_786_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_786_cast_fp16 = slice_by_index(begin = var_786_begin_0, end = var_786_end_0, end_mask = var_786_end_mask_0, x = x_63_cast_fp16)[name = tensor<string, []>("op_786_cast_fp16")];
            tensor<int32, [4]> var_787 = const()[name = tensor<string, []>("op_787"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_9_cast_fp16 = reshape(shape = var_787, x = var_786_cast_fp16)[name = tensor<string, []>("matrix_bd_9_cast_fp16")];
            tensor<bool, []> matrix_ac_5_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_5_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_55_perm_0 = const()[name = tensor<string, []>("transpose_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = tensor<string, []>("transpose_56_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_56 = transpose(perm = transpose_56_perm_0, x = k_9_cast_fp16)[name = tensor<string, []>("transpose_217")];
            tensor<fp16, [1, 8, 3, 64]> transpose_55 = transpose(perm = transpose_55_perm_0, x = var_770_cast_fp16)[name = tensor<string, []>("transpose_218")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_55, y = transpose_56)[name = tensor<string, []>("matrix_ac_5_cast_fp16")];
            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = tensor<string, []>("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = tensor<string, []>("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = tensor<string, []>("matrix_bd_11_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_796_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = tensor<string, []>("op_796_cast_fp16")];
            tensor<fp16, []> _inversed_scores_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_9_cast_fp16 = mul(x = var_796_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_9_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_11_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_11_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_802_cast_fp16 = softmax(axis = var_62, x = scores_11_cast_fp16)[name = tensor<string, []>("op_802_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_145_cast_fp16 = select(a = var_40_to_fp16, b = var_802_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_145_cast_fp16")];
            tensor<bool, []> x_65_transpose_x_0 = const()[name = tensor<string, []>("x_65_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_65_transpose_y_0 = const()[name = tensor<string, []>("x_65_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = v_5_cast_fp16)[name = tensor<string, []>("transpose_220")];
            tensor<fp16, [1, 8, 3, 64]> x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_7_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
            tensor<int32, [4]> var_806_perm_0 = const()[name = tensor<string, []>("op_806_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_807 = const()[name = tensor<string, []>("op_807"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_806_cast_fp16 = transpose(perm = var_806_perm_0, x = x_65_cast_fp16)[name = tensor<string, []>("transpose_216")];
            tensor<fp16, [1, 3, 512]> input_147_cast_fp16 = reshape(shape = var_807, x = var_806_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35129536)))];
            tensor<fp16, [1, 3, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
            tensor<int32, [1]> x_69_axes_0 = const()[name = tensor<string, []>("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35653888)))];
            tensor<fp16, [512]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35654976)))];
            tensor<fp16, [1, 3, 512]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
            tensor<int32, [3]> input_153_perm_0 = const()[name = tensor<string, []>("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_155_pad_type_0 = const()[name = tensor<string, []>("input_155_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_155_strides_0 = const()[name = tensor<string, []>("input_155_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_155_pad_0 = const()[name = tensor<string, []>("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_155_dilations_0 = const()[name = tensor<string, []>("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_155_groups_0 = const()[name = tensor<string, []>("input_155_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35656064)))];
            tensor<fp16, [1, 512, 3]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = tensor<string, []>("transpose_215")];
            tensor<fp16, [1, 1024, 3]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<int32, []> x_71_split_num_splits_0 = const()[name = tensor<string, []>("x_71_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_71_split_axis_0 = const()[name = tensor<string, []>("x_71_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_71_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = tensor<string, []>("x_71_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = tensor<string, []>("x_71_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_157_cast_fp16 = select(a = var_40_to_fp16, b = x_71_cast_fp16, cond = var_418)[name = tensor<string, []>("input_157_cast_fp16")];
            tensor<bool, []> new_x_11_interleave_0 = const()[name = tensor<string, []>("new_x_11_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_11_cast_fp16 = concat(axis = var_62, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = tensor<string, []>("new_x_11_cast_fp16")];
            tensor<int32, [3]> next_cache_5_begin_0 = const()[name = tensor<string, []>("next_cache_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_5_end_0 = const()[name = tensor<string, []>("next_cache_5_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_5_end_mask_0 = const()[name = tensor<string, []>("next_cache_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_5_cast_fp16 = slice_by_index(begin = next_cache_5_begin_0, end = next_cache_5_end_0, end_mask = next_cache_5_end_mask_0, x = new_x_11_cast_fp16)[name = tensor<string, []>("next_cache_5_cast_fp16")];
            tensor<int32, [3]> var_848_begin_0 = const()[name = tensor<string, []>("op_848_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_848_end_0 = const()[name = tensor<string, []>("op_848_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_848_end_mask_0 = const()[name = tensor<string, []>("op_848_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_848_cast_fp16 = slice_by_index(begin = var_848_begin_0, end = var_848_end_0, end_mask = var_848_end_mask_0, x = next_cache_5_cast_fp16)[name = tensor<string, []>("op_848_cast_fp16")];
            tensor<string, []> x_73_pad_type_0 = const()[name = tensor<string, []>("x_73_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_73_groups_0 = const()[name = tensor<string, []>("x_73_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_73_strides_0 = const()[name = tensor<string, []>("x_73_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_73_pad_0 = const()[name = tensor<string, []>("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_73_dilations_0 = const()[name = tensor<string, []>("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36704704)))];
            tensor<fp16, [1, 512, 3]> x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16, x = new_x_11_cast_fp16)[name = tensor<string, []>("x_73_cast_fp16")];
            tensor<int32, [3]> input_159_perm_0 = const()[name = tensor<string, []>("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_75_axes_0 = const()[name = tensor<string, []>("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36713984)))];
            tensor<fp16, [512]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36715072)))];
            tensor<fp16, [1, 3, 512]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = tensor<string, []>("transpose_214")];
            tensor<fp16, [1, 3, 512]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
            tensor<int32, [3]> input_161_perm_0 = const()[name = tensor<string, []>("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = tensor<string, []>("transpose_213")];
            tensor<fp16, [1, 512, 3]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
            tensor<string, []> x_77_pad_type_0 = const()[name = tensor<string, []>("x_77_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_77_strides_0 = const()[name = tensor<string, []>("x_77_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_77_pad_0 = const()[name = tensor<string, []>("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_77_dilations_0 = const()[name = tensor<string, []>("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_77_groups_0 = const()[name = tensor<string, []>("x_77_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36716160)))];
            tensor<fp16, [1, 512, 3]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
            tensor<int32, [3]> input_165_perm_0 = const()[name = tensor<string, []>("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = tensor<string, []>("transpose_212")];
            tensor<fp16, [1, 3, 512]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
            tensor<int32, [1]> input_169_axes_0 = const()[name = tensor<string, []>("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37240512)))];
            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37241600)))];
            tensor<fp16, [1, 3, 512]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37242688)))];
            tensor<fp16, [1, 3, 2048]> linear_26_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39339904)))];
            tensor<fp16, [1, 3, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<fp16, []> var_889_to_fp16 = const()[name = tensor<string, []>("op_889_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_890_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_889_to_fp16)[name = tensor<string, []>("op_890_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_890_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<int32, [1]> input_181_axes_0 = const()[name = tensor<string, []>("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41437120)))];
            tensor<fp16, [512]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41438208)))];
            tensor<fp16, [1, 3, 512]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
            tensor<int32, [4]> cache_13_begin_0 = const()[name = tensor<string, []>("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> cache_13_end_0 = const()[name = tensor<string, []>("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 70, 512])];
            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = tensor<string, []>("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = tensor<string, []>("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_13_cast_fp16")];
            tensor<int32, [4]> cache_15_begin_0 = const()[name = tensor<string, []>("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> cache_15_end_0 = const()[name = tensor<string, []>("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 512, 8])];
            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = tensor<string, []>("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = tensor<string, []>("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_15_cast_fp16")];
            tensor<int32, [1]> input_183_axes_0 = const()[name = tensor<string, []>("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41439296)))];
            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41440384)))];
            tensor<fp16, [1, 3, 512]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41441472)))];
            tensor<fp16, [1, 3, 2048]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43538688)))];
            tensor<fp16, [1, 3, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<fp16, []> var_924_to_fp16 = const()[name = tensor<string, []>("op_924_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_925_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_924_to_fp16)[name = tensor<string, []>("op_925_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_925_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<int32, [1]> key_7_axes_0 = const()[name = tensor<string, []>("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45635904)))];
            tensor<fp16, [512]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45636992)))];
            tensor<fp16, [1, 3, 512]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
            tensor<bool, []> input_195_interleave_0 = const()[name = tensor<string, []>("input_195_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_195_cast_fp16 = concat(axis = var_64, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = tensor<string, []>("input_195_cast_fp16")];
            tensor<int32, [3]> var_947_begin_0 = const()[name = tensor<string, []>("op_947_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_947_end_0 = const()[name = tensor<string, []>("op_947_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_947_end_mask_0 = const()[name = tensor<string, []>("op_947_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_947_cast_fp16 = slice_by_index(begin = var_947_begin_0, end = var_947_end_0, end_mask = var_947_end_mask_0, x = cache_13_cast_fp16)[name = tensor<string, []>("op_947_cast_fp16")];
            tensor<int32, [3]> var_950_begin_0 = const()[name = tensor<string, []>("op_950_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_950_end_0 = const()[name = tensor<string, []>("op_950_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_950_end_mask_0 = const()[name = tensor<string, []>("op_950_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_950_cast_fp16 = slice_by_index(begin = var_950_begin_0, end = var_950_end_0, end_mask = var_950_end_mask_0, x = key_7_cast_fp16)[name = tensor<string, []>("op_950_cast_fp16")];
            tensor<bool, []> var_953_interleave_0 = const()[name = tensor<string, []>("op_953_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_953_cast_fp16 = concat(axis = var_64, interleave = var_953_interleave_0, values = (var_947_cast_fp16, var_950_cast_fp16))[name = tensor<string, []>("op_953_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45638080)))];
            tensor<fp16, [1, 3, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16, x = key_7_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<int32, [4]> var_957 = const()[name = tensor<string, []>("op_957"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_19_cast_fp16 = reshape(shape = var_957, x = linear_30_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46162432)))];
            tensor<fp16, [1, 73, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<int32, [4]> var_961 = const()[name = tensor<string, []>("op_961"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_13_cast_fp16 = reshape(shape = var_961, x = linear_31_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46686784)))];
            tensor<fp16, [1, 73, 512]> linear_32_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<int32, [4]> var_965 = const()[name = tensor<string, []>("op_965"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_7_cast_fp16 = reshape(shape = var_965, x = linear_32_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
            tensor<int32, [4]> value_9_perm_0 = const()[name = tensor<string, []>("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47211136)))];
            tensor<fp16, [1, 3, 8, 64]> var_977_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_977_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47212224)))];
            tensor<fp16, [1, 3, 8, 64]> var_979_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_979_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_85_transpose_x_0 = const()[name = tensor<string, []>("x_85_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_85_transpose_y_0 = const()[name = tensor<string, []>("x_85_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_981_to_fp16 = const()[name = tensor<string, []>("op_981_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47213312)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_979_cast_fp16)[name = tensor<string, []>("transpose_210")];
            tensor<fp16, [1, 8, 3, 145]> x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = var_981_to_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
            tensor<int32, [8]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_87_mode_0 = const()[name = tensor<string, []>("x_87_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_62_to_fp16 = const()[name = tensor<string, []>("const_62_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_87_cast_fp16 = pad(constant_val = const_62_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
            tensor<int32, [4]> var_989 = const()[name = tensor<string, []>("op_989"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_89_cast_fp16 = reshape(shape = var_989, x = x_87_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
            tensor<int32, [4]> var_993_begin_0 = const()[name = tensor<string, []>("op_993_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_993_end_0 = const()[name = tensor<string, []>("op_993_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_993_end_mask_0 = const()[name = tensor<string, []>("op_993_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_993_cast_fp16 = slice_by_index(begin = var_993_begin_0, end = var_993_end_0, end_mask = var_993_end_mask_0, x = x_89_cast_fp16)[name = tensor<string, []>("op_993_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = tensor<string, []>("op_994"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_13_cast_fp16 = reshape(shape = var_994, x = var_993_cast_fp16)[name = tensor<string, []>("matrix_bd_13_cast_fp16")];
            tensor<bool, []> matrix_ac_7_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_7_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_57_perm_0 = const()[name = tensor<string, []>("transpose_57_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = tensor<string, []>("transpose_58_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_58 = transpose(perm = transpose_58_perm_0, x = k_13_cast_fp16)[name = tensor<string, []>("transpose_208")];
            tensor<fp16, [1, 8, 3, 64]> transpose_57 = transpose(perm = transpose_57_perm_0, x = var_977_cast_fp16)[name = tensor<string, []>("transpose_209")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_57, y = transpose_58)[name = tensor<string, []>("matrix_ac_7_cast_fp16")];
            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = tensor<string, []>("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = tensor<string, []>("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = tensor<string, []>("matrix_bd_15_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1003_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = tensor<string, []>("op_1003_cast_fp16")];
            tensor<fp16, []> _inversed_scores_13_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_13_cast_fp16 = mul(x = var_1003_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_13_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_15_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_15_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1009_cast_fp16 = softmax(axis = var_62, x = scores_15_cast_fp16)[name = tensor<string, []>("op_1009_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_197_cast_fp16 = select(a = var_40_to_fp16, b = var_1009_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_197_cast_fp16")];
            tensor<bool, []> x_91_transpose_x_0 = const()[name = tensor<string, []>("x_91_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_91_transpose_y_0 = const()[name = tensor<string, []>("x_91_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_7_cast_fp16)[name = tensor<string, []>("transpose_211")];
            tensor<fp16, [1, 8, 3, 64]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_9_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
            tensor<int32, [4]> var_1013_perm_0 = const()[name = tensor<string, []>("op_1013_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1014 = const()[name = tensor<string, []>("op_1014"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_1013_cast_fp16 = transpose(perm = var_1013_perm_0, x = x_91_cast_fp16)[name = tensor<string, []>("transpose_207")];
            tensor<fp16, [1, 3, 512]> input_199_cast_fp16 = reshape(shape = var_1014, x = var_1013_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47361856)))];
            tensor<fp16, [1, 3, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<int32, [1]> x_95_axes_0 = const()[name = tensor<string, []>("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47886208)))];
            tensor<fp16, [512]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47887296)))];
            tensor<fp16, [1, 3, 512]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
            tensor<int32, [3]> input_205_perm_0 = const()[name = tensor<string, []>("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_207_strides_0 = const()[name = tensor<string, []>("input_207_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_207_dilations_0 = const()[name = tensor<string, []>("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_207_groups_0 = const()[name = tensor<string, []>("input_207_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47888384)))];
            tensor<fp16, [1, 512, 3]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = tensor<string, []>("transpose_206")];
            tensor<fp16, [1, 1024, 3]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<int32, []> x_97_split_num_splits_0 = const()[name = tensor<string, []>("x_97_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_97_split_axis_0 = const()[name = tensor<string, []>("x_97_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_97_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = tensor<string, []>("x_97_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = tensor<string, []>("x_97_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_209_cast_fp16 = select(a = var_40_to_fp16, b = x_97_cast_fp16, cond = var_418)[name = tensor<string, []>("input_209_cast_fp16")];
            tensor<bool, []> new_x_15_interleave_0 = const()[name = tensor<string, []>("new_x_15_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_15_cast_fp16 = concat(axis = var_62, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = tensor<string, []>("new_x_15_cast_fp16")];
            tensor<int32, [3]> next_cache_7_begin_0 = const()[name = tensor<string, []>("next_cache_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_7_end_0 = const()[name = tensor<string, []>("next_cache_7_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_7_end_mask_0 = const()[name = tensor<string, []>("next_cache_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_7_cast_fp16 = slice_by_index(begin = next_cache_7_begin_0, end = next_cache_7_end_0, end_mask = next_cache_7_end_mask_0, x = new_x_15_cast_fp16)[name = tensor<string, []>("next_cache_7_cast_fp16")];
            tensor<int32, [3]> var_1055_begin_0 = const()[name = tensor<string, []>("op_1055_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_1055_end_0 = const()[name = tensor<string, []>("op_1055_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_1055_end_mask_0 = const()[name = tensor<string, []>("op_1055_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_1055_cast_fp16 = slice_by_index(begin = var_1055_begin_0, end = var_1055_end_0, end_mask = var_1055_end_mask_0, x = next_cache_7_cast_fp16)[name = tensor<string, []>("op_1055_cast_fp16")];
            tensor<string, []> x_99_pad_type_0 = const()[name = tensor<string, []>("x_99_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_99_groups_0 = const()[name = tensor<string, []>("x_99_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_99_strides_0 = const()[name = tensor<string, []>("x_99_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_99_pad_0 = const()[name = tensor<string, []>("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_99_dilations_0 = const()[name = tensor<string, []>("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48937024)))];
            tensor<fp16, [1, 512, 3]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16, x = new_x_15_cast_fp16)[name = tensor<string, []>("x_99_cast_fp16")];
            tensor<int32, [3]> input_211_perm_0 = const()[name = tensor<string, []>("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_101_axes_0 = const()[name = tensor<string, []>("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48946304)))];
            tensor<fp16, [512]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48947392)))];
            tensor<fp16, [1, 3, 512]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_205")];
            tensor<fp16, [1, 3, 512]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
            tensor<int32, [3]> input_213_perm_0 = const()[name = tensor<string, []>("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = tensor<string, []>("transpose_204")];
            tensor<fp16, [1, 512, 3]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
            tensor<string, []> x_103_pad_type_0 = const()[name = tensor<string, []>("x_103_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_103_strides_0 = const()[name = tensor<string, []>("x_103_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_103_pad_0 = const()[name = tensor<string, []>("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_103_dilations_0 = const()[name = tensor<string, []>("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_103_groups_0 = const()[name = tensor<string, []>("x_103_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48948480)))];
            tensor<fp16, [1, 512, 3]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("x_103_cast_fp16")];
            tensor<int32, [3]> input_217_perm_0 = const()[name = tensor<string, []>("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = tensor<string, []>("transpose_203")];
            tensor<fp16, [1, 3, 512]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
            tensor<int32, [1]> input_221_axes_0 = const()[name = tensor<string, []>("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49472832)))];
            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49473920)))];
            tensor<fp16, [1, 3, 512]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49475008)))];
            tensor<fp16, [1, 3, 2048]> linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51572224)))];
            tensor<fp16, [1, 3, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<fp16, []> var_1096_to_fp16 = const()[name = tensor<string, []>("op_1096_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1097_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1096_to_fp16)[name = tensor<string, []>("op_1097_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1097_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
            tensor<int32, [1]> input_233_axes_0 = const()[name = tensor<string, []>("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53669440)))];
            tensor<fp16, [512]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53670528)))];
            tensor<fp16, [1, 3, 512]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
            tensor<int32, [4]> cache_17_begin_0 = const()[name = tensor<string, []>("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> cache_17_end_0 = const()[name = tensor<string, []>("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 70, 512])];
            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = tensor<string, []>("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = tensor<string, []>("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_17_cast_fp16")];
            tensor<int32, [4]> cache_19_begin_0 = const()[name = tensor<string, []>("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> cache_19_end_0 = const()[name = tensor<string, []>("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 512, 8])];
            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = tensor<string, []>("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = tensor<string, []>("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_19_cast_fp16")];
            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53671616)))];
            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53672704)))];
            tensor<fp16, [1, 3, 512]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53673792)))];
            tensor<fp16, [1, 3, 2048]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55771008)))];
            tensor<fp16, [1, 3, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<fp16, []> var_1131_to_fp16 = const()[name = tensor<string, []>("op_1131_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1132_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1131_to_fp16)[name = tensor<string, []>("op_1132_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1132_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
            tensor<int32, [1]> key_9_axes_0 = const()[name = tensor<string, []>("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57868224)))];
            tensor<fp16, [512]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57869312)))];
            tensor<fp16, [1, 3, 512]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
            tensor<bool, []> input_247_interleave_0 = const()[name = tensor<string, []>("input_247_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_247_cast_fp16 = concat(axis = var_64, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = tensor<string, []>("input_247_cast_fp16")];
            tensor<int32, [3]> var_1154_begin_0 = const()[name = tensor<string, []>("op_1154_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1154_end_0 = const()[name = tensor<string, []>("op_1154_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1154_end_mask_0 = const()[name = tensor<string, []>("op_1154_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_1154_cast_fp16 = slice_by_index(begin = var_1154_begin_0, end = var_1154_end_0, end_mask = var_1154_end_mask_0, x = cache_17_cast_fp16)[name = tensor<string, []>("op_1154_cast_fp16")];
            tensor<int32, [3]> var_1157_begin_0 = const()[name = tensor<string, []>("op_1157_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1157_end_0 = const()[name = tensor<string, []>("op_1157_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_1157_end_mask_0 = const()[name = tensor<string, []>("op_1157_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_1157_cast_fp16 = slice_by_index(begin = var_1157_begin_0, end = var_1157_end_0, end_mask = var_1157_end_mask_0, x = key_9_cast_fp16)[name = tensor<string, []>("op_1157_cast_fp16")];
            tensor<bool, []> var_1160_interleave_0 = const()[name = tensor<string, []>("op_1160_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1160_cast_fp16 = concat(axis = var_64, interleave = var_1160_interleave_0, values = (var_1154_cast_fp16, var_1157_cast_fp16))[name = tensor<string, []>("op_1160_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57870400)))];
            tensor<fp16, [1, 3, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16, x = key_9_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<int32, [4]> var_1164 = const()[name = tensor<string, []>("op_1164"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_25_cast_fp16 = reshape(shape = var_1164, x = linear_39_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58394752)))];
            tensor<fp16, [1, 73, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<int32, [4]> var_1168 = const()[name = tensor<string, []>("op_1168"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_17_cast_fp16 = reshape(shape = var_1168, x = linear_40_cast_fp16)[name = tensor<string, []>("k_17_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58919104)))];
            tensor<fp16, [1, 73, 512]> linear_41_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<int32, [4]> var_1172 = const()[name = tensor<string, []>("op_1172"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_9_cast_fp16 = reshape(shape = var_1172, x = linear_41_cast_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
            tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<string, []>("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59443456)))];
            tensor<fp16, [1, 3, 8, 64]> var_1184_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1184_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59444544)))];
            tensor<fp16, [1, 3, 8, 64]> var_1186_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1186_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_111_transpose_x_0 = const()[name = tensor<string, []>("x_111_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_111_transpose_y_0 = const()[name = tensor<string, []>("x_111_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_1188_to_fp16 = const()[name = tensor<string, []>("op_1188_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59445632)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1186_cast_fp16)[name = tensor<string, []>("transpose_201")];
            tensor<fp16, [1, 8, 3, 145]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = var_1188_to_fp16)[name = tensor<string, []>("x_111_cast_fp16")];
            tensor<int32, [8]> x_113_pad_0 = const()[name = tensor<string, []>("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_113_mode_0 = const()[name = tensor<string, []>("x_113_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_75_to_fp16 = const()[name = tensor<string, []>("const_75_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_113_cast_fp16 = pad(constant_val = const_75_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = tensor<string, []>("x_113_cast_fp16")];
            tensor<int32, [4]> var_1196 = const()[name = tensor<string, []>("op_1196"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_115_cast_fp16 = reshape(shape = var_1196, x = x_113_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")];
            tensor<int32, [4]> var_1200_begin_0 = const()[name = tensor<string, []>("op_1200_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1200_end_0 = const()[name = tensor<string, []>("op_1200_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_1200_end_mask_0 = const()[name = tensor<string, []>("op_1200_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_1200_cast_fp16 = slice_by_index(begin = var_1200_begin_0, end = var_1200_end_0, end_mask = var_1200_end_mask_0, x = x_115_cast_fp16)[name = tensor<string, []>("op_1200_cast_fp16")];
            tensor<int32, [4]> var_1201 = const()[name = tensor<string, []>("op_1201"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_17_cast_fp16 = reshape(shape = var_1201, x = var_1200_cast_fp16)[name = tensor<string, []>("matrix_bd_17_cast_fp16")];
            tensor<bool, []> matrix_ac_9_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_9_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_59_perm_0 = const()[name = tensor<string, []>("transpose_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = tensor<string, []>("transpose_60_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_60 = transpose(perm = transpose_60_perm_0, x = k_17_cast_fp16)[name = tensor<string, []>("transpose_199")];
            tensor<fp16, [1, 8, 3, 64]> transpose_59 = transpose(perm = transpose_59_perm_0, x = var_1184_cast_fp16)[name = tensor<string, []>("transpose_200")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_59, y = transpose_60)[name = tensor<string, []>("matrix_ac_9_cast_fp16")];
            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = tensor<string, []>("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = tensor<string, []>("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = tensor<string, []>("matrix_bd_19_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1210_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = tensor<string, []>("op_1210_cast_fp16")];
            tensor<fp16, []> _inversed_scores_17_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_17_cast_fp16 = mul(x = var_1210_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_17_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_19_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_19_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1216_cast_fp16 = softmax(axis = var_62, x = scores_19_cast_fp16)[name = tensor<string, []>("op_1216_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_249_cast_fp16 = select(a = var_40_to_fp16, b = var_1216_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_249_cast_fp16")];
            tensor<bool, []> x_117_transpose_x_0 = const()[name = tensor<string, []>("x_117_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_117_transpose_y_0 = const()[name = tensor<string, []>("x_117_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_9_cast_fp16)[name = tensor<string, []>("transpose_202")];
            tensor<fp16, [1, 8, 3, 64]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_11_cast_fp16)[name = tensor<string, []>("x_117_cast_fp16")];
            tensor<int32, [4]> var_1220_perm_0 = const()[name = tensor<string, []>("op_1220_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1221 = const()[name = tensor<string, []>("op_1221"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_1220_cast_fp16 = transpose(perm = var_1220_perm_0, x = x_117_cast_fp16)[name = tensor<string, []>("transpose_198")];
            tensor<fp16, [1, 3, 512]> input_251_cast_fp16 = reshape(shape = var_1221, x = var_1220_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59594176)))];
            tensor<fp16, [1, 3, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16, x = input_251_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
            tensor<int32, [1]> x_121_axes_0 = const()[name = tensor<string, []>("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60118528)))];
            tensor<fp16, [512]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60119616)))];
            tensor<fp16, [1, 3, 512]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("x_121_cast_fp16")];
            tensor<int32, [3]> input_257_perm_0 = const()[name = tensor<string, []>("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_259_pad_type_0 = const()[name = tensor<string, []>("input_259_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_259_strides_0 = const()[name = tensor<string, []>("input_259_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_259_pad_0 = const()[name = tensor<string, []>("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_259_dilations_0 = const()[name = tensor<string, []>("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_259_groups_0 = const()[name = tensor<string, []>("input_259_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60120704)))];
            tensor<fp16, [1, 512, 3]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = tensor<string, []>("transpose_197")];
            tensor<fp16, [1, 1024, 3]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16, x = input_257_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
            tensor<int32, []> x_123_split_num_splits_0 = const()[name = tensor<string, []>("x_123_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_123_split_axis_0 = const()[name = tensor<string, []>("x_123_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_123_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = tensor<string, []>("x_123_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = tensor<string, []>("x_123_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_123_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_261_cast_fp16 = select(a = var_40_to_fp16, b = x_123_cast_fp16, cond = var_418)[name = tensor<string, []>("input_261_cast_fp16")];
            tensor<bool, []> new_x_19_interleave_0 = const()[name = tensor<string, []>("new_x_19_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_19_cast_fp16 = concat(axis = var_62, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = tensor<string, []>("new_x_19_cast_fp16")];
            tensor<int32, [3]> next_cache_9_begin_0 = const()[name = tensor<string, []>("next_cache_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_9_end_0 = const()[name = tensor<string, []>("next_cache_9_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_9_end_mask_0 = const()[name = tensor<string, []>("next_cache_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_9_cast_fp16 = slice_by_index(begin = next_cache_9_begin_0, end = next_cache_9_end_0, end_mask = next_cache_9_end_mask_0, x = new_x_19_cast_fp16)[name = tensor<string, []>("next_cache_9_cast_fp16")];
            tensor<int32, [3]> var_1262_begin_0 = const()[name = tensor<string, []>("op_1262_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_1262_end_0 = const()[name = tensor<string, []>("op_1262_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_1262_end_mask_0 = const()[name = tensor<string, []>("op_1262_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = var_1262_end_0, end_mask = var_1262_end_mask_0, x = next_cache_9_cast_fp16)[name = tensor<string, []>("op_1262_cast_fp16")];
            tensor<string, []> x_125_pad_type_0 = const()[name = tensor<string, []>("x_125_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_125_groups_0 = const()[name = tensor<string, []>("x_125_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_125_strides_0 = const()[name = tensor<string, []>("x_125_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_125_pad_0 = const()[name = tensor<string, []>("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_125_dilations_0 = const()[name = tensor<string, []>("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61169344)))];
            tensor<fp16, [1, 512, 3]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16, x = new_x_19_cast_fp16)[name = tensor<string, []>("x_125_cast_fp16")];
            tensor<int32, [3]> input_263_perm_0 = const()[name = tensor<string, []>("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_127_axes_0 = const()[name = tensor<string, []>("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61178624)))];
            tensor<fp16, [512]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61179712)))];
            tensor<fp16, [1, 3, 512]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = tensor<string, []>("transpose_196")];
            tensor<fp16, [1, 3, 512]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")];
            tensor<int32, [3]> input_265_perm_0 = const()[name = tensor<string, []>("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = tensor<string, []>("transpose_195")];
            tensor<fp16, [1, 512, 3]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
            tensor<string, []> x_129_pad_type_0 = const()[name = tensor<string, []>("x_129_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_129_strides_0 = const()[name = tensor<string, []>("x_129_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_129_pad_0 = const()[name = tensor<string, []>("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_129_dilations_0 = const()[name = tensor<string, []>("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_129_groups_0 = const()[name = tensor<string, []>("x_129_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61180800)))];
            tensor<fp16, [1, 512, 3]> x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16, x = input_267_cast_fp16)[name = tensor<string, []>("x_129_cast_fp16")];
            tensor<int32, [3]> input_269_perm_0 = const()[name = tensor<string, []>("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = tensor<string, []>("transpose_194")];
            tensor<fp16, [1, 3, 512]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
            tensor<int32, [1]> input_273_axes_0 = const()[name = tensor<string, []>("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61705152)))];
            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61706240)))];
            tensor<fp16, [1, 3, 512]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor<string, []>("input_273_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61707328)))];
            tensor<fp16, [1, 3, 2048]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63804544)))];
            tensor<fp16, [1, 3, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<fp16, []> var_1303_to_fp16 = const()[name = tensor<string, []>("op_1303_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1304_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1303_to_fp16)[name = tensor<string, []>("op_1304_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1304_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
            tensor<int32, [1]> input_285_axes_0 = const()[name = tensor<string, []>("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65901760)))];
            tensor<fp16, [512]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65902848)))];
            tensor<fp16, [1, 3, 512]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
            tensor<int32, [4]> cache_21_begin_0 = const()[name = tensor<string, []>("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> cache_21_end_0 = const()[name = tensor<string, []>("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 70, 512])];
            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = tensor<string, []>("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = tensor<string, []>("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_21_cast_fp16")];
            tensor<int32, [4]> cache_23_begin_0 = const()[name = tensor<string, []>("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> cache_23_end_0 = const()[name = tensor<string, []>("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 512, 8])];
            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = tensor<string, []>("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = tensor<string, []>("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_23_cast_fp16")];
            tensor<int32, [1]> input_287_axes_0 = const()[name = tensor<string, []>("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65903936)))];
            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65905024)))];
            tensor<fp16, [1, 3, 512]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65906112)))];
            tensor<fp16, [1, 3, 2048]> linear_46_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor<string, []>("input_291_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68003328)))];
            tensor<fp16, [1, 3, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16, x = input_291_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<fp16, []> var_1338_to_fp16 = const()[name = tensor<string, []>("op_1338_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1339_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1338_to_fp16)[name = tensor<string, []>("op_1339_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1339_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
            tensor<int32, [1]> key_11_axes_0 = const()[name = tensor<string, []>("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70100544)))];
            tensor<fp16, [512]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70101632)))];
            tensor<fp16, [1, 3, 512]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
            tensor<bool, []> input_299_interleave_0 = const()[name = tensor<string, []>("input_299_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_299_cast_fp16 = concat(axis = var_64, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = tensor<string, []>("input_299_cast_fp16")];
            tensor<int32, [3]> var_1361_begin_0 = const()[name = tensor<string, []>("op_1361_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1361_end_0 = const()[name = tensor<string, []>("op_1361_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1361_end_mask_0 = const()[name = tensor<string, []>("op_1361_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_1361_cast_fp16 = slice_by_index(begin = var_1361_begin_0, end = var_1361_end_0, end_mask = var_1361_end_mask_0, x = cache_21_cast_fp16)[name = tensor<string, []>("op_1361_cast_fp16")];
            tensor<int32, [3]> var_1364_begin_0 = const()[name = tensor<string, []>("op_1364_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1364_end_0 = const()[name = tensor<string, []>("op_1364_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_1364_end_mask_0 = const()[name = tensor<string, []>("op_1364_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_1364_cast_fp16 = slice_by_index(begin = var_1364_begin_0, end = var_1364_end_0, end_mask = var_1364_end_mask_0, x = key_11_cast_fp16)[name = tensor<string, []>("op_1364_cast_fp16")];
            tensor<bool, []> var_1367_interleave_0 = const()[name = tensor<string, []>("op_1367_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1367_cast_fp16 = concat(axis = var_64, interleave = var_1367_interleave_0, values = (var_1361_cast_fp16, var_1364_cast_fp16))[name = tensor<string, []>("op_1367_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70102720)))];
            tensor<fp16, [1, 3, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16, x = key_11_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<int32, [4]> var_1371 = const()[name = tensor<string, []>("op_1371"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_31_cast_fp16 = reshape(shape = var_1371, x = linear_48_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70627072)))];
            tensor<fp16, [1, 73, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1375 = const()[name = tensor<string, []>("op_1375"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_21_cast_fp16 = reshape(shape = var_1375, x = linear_49_cast_fp16)[name = tensor<string, []>("k_21_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71151424)))];
            tensor<fp16, [1, 73, 512]> linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1379 = const()[name = tensor<string, []>("op_1379"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_11_cast_fp16 = reshape(shape = var_1379, x = linear_50_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
            tensor<int32, [4]> value_13_perm_0 = const()[name = tensor<string, []>("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71675776)))];
            tensor<fp16, [1, 3, 8, 64]> var_1391_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1391_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71676864)))];
            tensor<fp16, [1, 3, 8, 64]> var_1393_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1393_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_137_transpose_x_0 = const()[name = tensor<string, []>("x_137_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_137_transpose_y_0 = const()[name = tensor<string, []>("x_137_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_1395_to_fp16 = const()[name = tensor<string, []>("op_1395_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71677952)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1393_cast_fp16)[name = tensor<string, []>("transpose_192")];
            tensor<fp16, [1, 8, 3, 145]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = var_1395_to_fp16)[name = tensor<string, []>("x_137_cast_fp16")];
            tensor<int32, [8]> x_139_pad_0 = const()[name = tensor<string, []>("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_139_mode_0 = const()[name = tensor<string, []>("x_139_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_88_to_fp16 = const()[name = tensor<string, []>("const_88_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_139_cast_fp16 = pad(constant_val = const_88_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = tensor<string, []>("x_139_cast_fp16")];
            tensor<int32, [4]> var_1403 = const()[name = tensor<string, []>("op_1403"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_141_cast_fp16 = reshape(shape = var_1403, x = x_139_cast_fp16)[name = tensor<string, []>("x_141_cast_fp16")];
            tensor<int32, [4]> var_1407_begin_0 = const()[name = tensor<string, []>("op_1407_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1407_end_0 = const()[name = tensor<string, []>("op_1407_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_1407_end_mask_0 = const()[name = tensor<string, []>("op_1407_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_1407_cast_fp16 = slice_by_index(begin = var_1407_begin_0, end = var_1407_end_0, end_mask = var_1407_end_mask_0, x = x_141_cast_fp16)[name = tensor<string, []>("op_1407_cast_fp16")];
            tensor<int32, [4]> var_1408 = const()[name = tensor<string, []>("op_1408"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_21_cast_fp16 = reshape(shape = var_1408, x = var_1407_cast_fp16)[name = tensor<string, []>("matrix_bd_21_cast_fp16")];
            tensor<bool, []> matrix_ac_11_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_11_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_61_perm_0 = const()[name = tensor<string, []>("transpose_61_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = tensor<string, []>("transpose_62_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_62 = transpose(perm = transpose_62_perm_0, x = k_21_cast_fp16)[name = tensor<string, []>("transpose_190")];
            tensor<fp16, [1, 8, 3, 64]> transpose_61 = transpose(perm = transpose_61_perm_0, x = var_1391_cast_fp16)[name = tensor<string, []>("transpose_191")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_61, y = transpose_62)[name = tensor<string, []>("matrix_ac_11_cast_fp16")];
            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = tensor<string, []>("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = tensor<string, []>("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = tensor<string, []>("matrix_bd_23_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1417_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = tensor<string, []>("op_1417_cast_fp16")];
            tensor<fp16, []> _inversed_scores_21_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_21_cast_fp16 = mul(x = var_1417_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_21_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_23_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_23_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1423_cast_fp16 = softmax(axis = var_62, x = scores_23_cast_fp16)[name = tensor<string, []>("op_1423_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_301_cast_fp16 = select(a = var_40_to_fp16, b = var_1423_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_301_cast_fp16")];
            tensor<bool, []> x_143_transpose_x_0 = const()[name = tensor<string, []>("x_143_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_143_transpose_y_0 = const()[name = tensor<string, []>("x_143_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_11_cast_fp16)[name = tensor<string, []>("transpose_193")];
            tensor<fp16, [1, 8, 3, 64]> x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_13_cast_fp16)[name = tensor<string, []>("x_143_cast_fp16")];
            tensor<int32, [4]> var_1427_perm_0 = const()[name = tensor<string, []>("op_1427_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1428 = const()[name = tensor<string, []>("op_1428"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_1427_cast_fp16 = transpose(perm = var_1427_perm_0, x = x_143_cast_fp16)[name = tensor<string, []>("transpose_189")];
            tensor<fp16, [1, 3, 512]> input_303_cast_fp16 = reshape(shape = var_1428, x = var_1427_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71826496)))];
            tensor<fp16, [1, 3, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16, x = input_303_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
            tensor<int32, [1]> x_147_axes_0 = const()[name = tensor<string, []>("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72350848)))];
            tensor<fp16, [512]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72351936)))];
            tensor<fp16, [1, 3, 512]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = tensor<string, []>("x_147_cast_fp16")];
            tensor<int32, [3]> input_309_perm_0 = const()[name = tensor<string, []>("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_311_pad_type_0 = const()[name = tensor<string, []>("input_311_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_311_strides_0 = const()[name = tensor<string, []>("input_311_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_311_pad_0 = const()[name = tensor<string, []>("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_311_dilations_0 = const()[name = tensor<string, []>("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_311_groups_0 = const()[name = tensor<string, []>("input_311_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72353024)))];
            tensor<fp16, [1, 512, 3]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = tensor<string, []>("transpose_188")];
            tensor<fp16, [1, 1024, 3]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16, x = input_309_cast_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
            tensor<int32, []> x_149_split_num_splits_0 = const()[name = tensor<string, []>("x_149_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_149_split_axis_0 = const()[name = tensor<string, []>("x_149_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_149_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = tensor<string, []>("x_149_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = tensor<string, []>("x_149_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_313_cast_fp16 = select(a = var_40_to_fp16, b = x_149_cast_fp16, cond = var_418)[name = tensor<string, []>("input_313_cast_fp16")];
            tensor<bool, []> new_x_23_interleave_0 = const()[name = tensor<string, []>("new_x_23_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_23_cast_fp16 = concat(axis = var_62, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = tensor<string, []>("new_x_23_cast_fp16")];
            tensor<int32, [3]> next_cache_11_begin_0 = const()[name = tensor<string, []>("next_cache_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_11_end_0 = const()[name = tensor<string, []>("next_cache_11_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_11_end_mask_0 = const()[name = tensor<string, []>("next_cache_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_11_cast_fp16 = slice_by_index(begin = next_cache_11_begin_0, end = next_cache_11_end_0, end_mask = next_cache_11_end_mask_0, x = new_x_23_cast_fp16)[name = tensor<string, []>("next_cache_11_cast_fp16")];
            tensor<int32, [3]> var_1469_begin_0 = const()[name = tensor<string, []>("op_1469_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_1469_end_0 = const()[name = tensor<string, []>("op_1469_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_1469_end_mask_0 = const()[name = tensor<string, []>("op_1469_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_1469_cast_fp16 = slice_by_index(begin = var_1469_begin_0, end = var_1469_end_0, end_mask = var_1469_end_mask_0, x = next_cache_11_cast_fp16)[name = tensor<string, []>("op_1469_cast_fp16")];
            tensor<string, []> x_151_pad_type_0 = const()[name = tensor<string, []>("x_151_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_151_groups_0 = const()[name = tensor<string, []>("x_151_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_151_strides_0 = const()[name = tensor<string, []>("x_151_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_151_pad_0 = const()[name = tensor<string, []>("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_151_dilations_0 = const()[name = tensor<string, []>("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73401664)))];
            tensor<fp16, [1, 512, 3]> x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16, x = new_x_23_cast_fp16)[name = tensor<string, []>("x_151_cast_fp16")];
            tensor<int32, [3]> input_315_perm_0 = const()[name = tensor<string, []>("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_153_axes_0 = const()[name = tensor<string, []>("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73410944)))];
            tensor<fp16, [512]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73412032)))];
            tensor<fp16, [1, 3, 512]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = tensor<string, []>("transpose_187")];
            tensor<fp16, [1, 3, 512]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")];
            tensor<int32, [3]> input_317_perm_0 = const()[name = tensor<string, []>("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = tensor<string, []>("transpose_186")];
            tensor<fp16, [1, 512, 3]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
            tensor<string, []> x_155_pad_type_0 = const()[name = tensor<string, []>("x_155_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_155_strides_0 = const()[name = tensor<string, []>("x_155_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_155_pad_0 = const()[name = tensor<string, []>("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_155_dilations_0 = const()[name = tensor<string, []>("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_155_groups_0 = const()[name = tensor<string, []>("x_155_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73413120)))];
            tensor<fp16, [1, 512, 3]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16, x = input_319_cast_fp16)[name = tensor<string, []>("x_155_cast_fp16")];
            tensor<int32, [3]> input_321_perm_0 = const()[name = tensor<string, []>("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = tensor<string, []>("transpose_185")];
            tensor<fp16, [1, 3, 512]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
            tensor<int32, [1]> input_325_axes_0 = const()[name = tensor<string, []>("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73937472)))];
            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73938560)))];
            tensor<fp16, [1, 3, 512]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = tensor<string, []>("input_325_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73939648)))];
            tensor<fp16, [1, 3, 2048]> linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16, x = input_325_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76036864)))];
            tensor<fp16, [1, 3, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16, x = input_329_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
            tensor<fp16, []> var_1510_to_fp16 = const()[name = tensor<string, []>("op_1510_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1511_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1510_to_fp16)[name = tensor<string, []>("op_1511_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1511_cast_fp16)[name = tensor<string, []>("input_335_cast_fp16")];
            tensor<int32, [1]> input_337_axes_0 = const()[name = tensor<string, []>("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78134080)))];
            tensor<fp16, [512]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78135168)))];
            tensor<fp16, [1, 3, 512]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
            tensor<int32, [4]> cache_25_begin_0 = const()[name = tensor<string, []>("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> cache_25_end_0 = const()[name = tensor<string, []>("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 70, 512])];
            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = tensor<string, []>("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = tensor<string, []>("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_25_cast_fp16")];
            tensor<int32, [4]> cache_27_begin_0 = const()[name = tensor<string, []>("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> cache_27_end_0 = const()[name = tensor<string, []>("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 512, 8])];
            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = tensor<string, []>("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = tensor<string, []>("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_27_cast_fp16")];
            tensor<int32, [1]> input_339_axes_0 = const()[name = tensor<string, []>("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78136256)))];
            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78137344)))];
            tensor<fp16, [1, 3, 512]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<string, []>("input_339_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78138432)))];
            tensor<fp16, [1, 3, 2048]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16, x = input_339_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = tensor<string, []>("input_343_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80235648)))];
            tensor<fp16, [1, 3, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16, x = input_343_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")];
            tensor<fp16, []> var_1545_to_fp16 = const()[name = tensor<string, []>("op_1545_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1546_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1545_to_fp16)[name = tensor<string, []>("op_1546_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1546_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
            tensor<int32, [1]> key_13_axes_0 = const()[name = tensor<string, []>("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82332864)))];
            tensor<fp16, [512]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82333952)))];
            tensor<fp16, [1, 3, 512]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
            tensor<bool, []> input_351_interleave_0 = const()[name = tensor<string, []>("input_351_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_351_cast_fp16 = concat(axis = var_64, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = tensor<string, []>("input_351_cast_fp16")];
            tensor<int32, [3]> var_1568_begin_0 = const()[name = tensor<string, []>("op_1568_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1568_end_0 = const()[name = tensor<string, []>("op_1568_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1568_end_mask_0 = const()[name = tensor<string, []>("op_1568_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_1568_cast_fp16 = slice_by_index(begin = var_1568_begin_0, end = var_1568_end_0, end_mask = var_1568_end_mask_0, x = cache_25_cast_fp16)[name = tensor<string, []>("op_1568_cast_fp16")];
            tensor<int32, [3]> var_1571_begin_0 = const()[name = tensor<string, []>("op_1571_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1571_end_0 = const()[name = tensor<string, []>("op_1571_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_1571_end_mask_0 = const()[name = tensor<string, []>("op_1571_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_1571_cast_fp16 = slice_by_index(begin = var_1571_begin_0, end = var_1571_end_0, end_mask = var_1571_end_mask_0, x = key_13_cast_fp16)[name = tensor<string, []>("op_1571_cast_fp16")];
            tensor<bool, []> var_1574_interleave_0 = const()[name = tensor<string, []>("op_1574_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1574_cast_fp16 = concat(axis = var_64, interleave = var_1574_interleave_0, values = (var_1568_cast_fp16, var_1571_cast_fp16))[name = tensor<string, []>("op_1574_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82335040)))];
            tensor<fp16, [1, 3, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16, x = key_13_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1578 = const()[name = tensor<string, []>("op_1578"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_37_cast_fp16 = reshape(shape = var_1578, x = linear_57_cast_fp16)[name = tensor<string, []>("q_37_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82859392)))];
            tensor<fp16, [1, 73, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1582 = const()[name = tensor<string, []>("op_1582"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_25_cast_fp16 = reshape(shape = var_1582, x = linear_58_cast_fp16)[name = tensor<string, []>("k_25_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83383744)))];
            tensor<fp16, [1, 73, 512]> linear_59_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")];
            tensor<int32, [4]> var_1586 = const()[name = tensor<string, []>("op_1586"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_13_cast_fp16 = reshape(shape = var_1586, x = linear_59_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
            tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<string, []>("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83908096)))];
            tensor<fp16, [1, 3, 8, 64]> var_1598_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1598_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83909184)))];
            tensor<fp16, [1, 3, 8, 64]> var_1600_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1600_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_163_transpose_x_0 = const()[name = tensor<string, []>("x_163_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_163_transpose_y_0 = const()[name = tensor<string, []>("x_163_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_1602_to_fp16 = const()[name = tensor<string, []>("op_1602_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83910272)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1600_cast_fp16)[name = tensor<string, []>("transpose_183")];
            tensor<fp16, [1, 8, 3, 145]> x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = var_1602_to_fp16)[name = tensor<string, []>("x_163_cast_fp16")];
            tensor<int32, [8]> x_165_pad_0 = const()[name = tensor<string, []>("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_165_mode_0 = const()[name = tensor<string, []>("x_165_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_101_to_fp16 = const()[name = tensor<string, []>("const_101_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_165_cast_fp16 = pad(constant_val = const_101_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = tensor<string, []>("x_165_cast_fp16")];
            tensor<int32, [4]> var_1610 = const()[name = tensor<string, []>("op_1610"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_167_cast_fp16 = reshape(shape = var_1610, x = x_165_cast_fp16)[name = tensor<string, []>("x_167_cast_fp16")];
            tensor<int32, [4]> var_1614_begin_0 = const()[name = tensor<string, []>("op_1614_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1614_end_0 = const()[name = tensor<string, []>("op_1614_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_1614_end_mask_0 = const()[name = tensor<string, []>("op_1614_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_1614_cast_fp16 = slice_by_index(begin = var_1614_begin_0, end = var_1614_end_0, end_mask = var_1614_end_mask_0, x = x_167_cast_fp16)[name = tensor<string, []>("op_1614_cast_fp16")];
            tensor<int32, [4]> var_1615 = const()[name = tensor<string, []>("op_1615"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_25_cast_fp16 = reshape(shape = var_1615, x = var_1614_cast_fp16)[name = tensor<string, []>("matrix_bd_25_cast_fp16")];
            tensor<bool, []> matrix_ac_13_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_13_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_63_perm_0 = const()[name = tensor<string, []>("transpose_63_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = tensor<string, []>("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_64 = transpose(perm = transpose_64_perm_0, x = k_25_cast_fp16)[name = tensor<string, []>("transpose_181")];
            tensor<fp16, [1, 8, 3, 64]> transpose_63 = transpose(perm = transpose_63_perm_0, x = var_1598_cast_fp16)[name = tensor<string, []>("transpose_182")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_63, y = transpose_64)[name = tensor<string, []>("matrix_ac_13_cast_fp16")];
            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = tensor<string, []>("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = tensor<string, []>("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = tensor<string, []>("matrix_bd_27_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1624_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = tensor<string, []>("op_1624_cast_fp16")];
            tensor<fp16, []> _inversed_scores_25_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_25_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_25_cast_fp16 = mul(x = var_1624_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_25_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_27_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_27_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1630_cast_fp16 = softmax(axis = var_62, x = scores_27_cast_fp16)[name = tensor<string, []>("op_1630_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_353_cast_fp16 = select(a = var_40_to_fp16, b = var_1630_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_353_cast_fp16")];
            tensor<bool, []> x_169_transpose_x_0 = const()[name = tensor<string, []>("x_169_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_169_transpose_y_0 = const()[name = tensor<string, []>("x_169_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_13_cast_fp16)[name = tensor<string, []>("transpose_184")];
            tensor<fp16, [1, 8, 3, 64]> x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_15_cast_fp16)[name = tensor<string, []>("x_169_cast_fp16")];
            tensor<int32, [4]> var_1634_perm_0 = const()[name = tensor<string, []>("op_1634_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1635 = const()[name = tensor<string, []>("op_1635"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_1634_cast_fp16 = transpose(perm = var_1634_perm_0, x = x_169_cast_fp16)[name = tensor<string, []>("transpose_180")];
            tensor<fp16, [1, 3, 512]> input_355_cast_fp16 = reshape(shape = var_1635, x = var_1634_cast_fp16)[name = tensor<string, []>("input_355_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84058816)))];
            tensor<fp16, [1, 3, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16, x = input_355_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
            tensor<int32, [1]> x_173_axes_0 = const()[name = tensor<string, []>("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84583168)))];
            tensor<fp16, [512]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84584256)))];
            tensor<fp16, [1, 3, 512]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("x_173_cast_fp16")];
            tensor<int32, [3]> input_361_perm_0 = const()[name = tensor<string, []>("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_363_pad_type_0 = const()[name = tensor<string, []>("input_363_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_363_strides_0 = const()[name = tensor<string, []>("input_363_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_363_pad_0 = const()[name = tensor<string, []>("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_363_dilations_0 = const()[name = tensor<string, []>("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_363_groups_0 = const()[name = tensor<string, []>("input_363_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84585344)))];
            tensor<fp16, [1, 512, 3]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = tensor<string, []>("transpose_179")];
            tensor<fp16, [1, 1024, 3]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16, x = input_361_cast_fp16)[name = tensor<string, []>("input_363_cast_fp16")];
            tensor<int32, []> x_175_split_num_splits_0 = const()[name = tensor<string, []>("x_175_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_175_split_axis_0 = const()[name = tensor<string, []>("x_175_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_175_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = tensor<string, []>("x_175_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = tensor<string, []>("x_175_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_365_cast_fp16 = select(a = var_40_to_fp16, b = x_175_cast_fp16, cond = var_418)[name = tensor<string, []>("input_365_cast_fp16")];
            tensor<bool, []> new_x_27_interleave_0 = const()[name = tensor<string, []>("new_x_27_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_27_cast_fp16 = concat(axis = var_62, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = tensor<string, []>("new_x_27_cast_fp16")];
            tensor<int32, [3]> next_cache_13_begin_0 = const()[name = tensor<string, []>("next_cache_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_13_end_0 = const()[name = tensor<string, []>("next_cache_13_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_13_end_mask_0 = const()[name = tensor<string, []>("next_cache_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_13_cast_fp16 = slice_by_index(begin = next_cache_13_begin_0, end = next_cache_13_end_0, end_mask = next_cache_13_end_mask_0, x = new_x_27_cast_fp16)[name = tensor<string, []>("next_cache_13_cast_fp16")];
            tensor<int32, [3]> var_1676_begin_0 = const()[name = tensor<string, []>("op_1676_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_1676_end_0 = const()[name = tensor<string, []>("op_1676_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_1676_end_mask_0 = const()[name = tensor<string, []>("op_1676_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_1676_cast_fp16 = slice_by_index(begin = var_1676_begin_0, end = var_1676_end_0, end_mask = var_1676_end_mask_0, x = next_cache_13_cast_fp16)[name = tensor<string, []>("op_1676_cast_fp16")];
            tensor<string, []> x_177_pad_type_0 = const()[name = tensor<string, []>("x_177_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_177_groups_0 = const()[name = tensor<string, []>("x_177_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_177_strides_0 = const()[name = tensor<string, []>("x_177_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_177_pad_0 = const()[name = tensor<string, []>("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_177_dilations_0 = const()[name = tensor<string, []>("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85633984)))];
            tensor<fp16, [1, 512, 3]> x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16, x = new_x_27_cast_fp16)[name = tensor<string, []>("x_177_cast_fp16")];
            tensor<int32, [3]> input_367_perm_0 = const()[name = tensor<string, []>("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_179_axes_0 = const()[name = tensor<string, []>("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85643264)))];
            tensor<fp16, [512]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85644352)))];
            tensor<fp16, [1, 3, 512]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = tensor<string, []>("transpose_178")];
            tensor<fp16, [1, 3, 512]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = tensor<string, []>("x_179_cast_fp16")];
            tensor<int32, [3]> input_369_perm_0 = const()[name = tensor<string, []>("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = tensor<string, []>("transpose_177")];
            tensor<fp16, [1, 512, 3]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
            tensor<string, []> x_181_pad_type_0 = const()[name = tensor<string, []>("x_181_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_181_strides_0 = const()[name = tensor<string, []>("x_181_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_181_pad_0 = const()[name = tensor<string, []>("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_181_dilations_0 = const()[name = tensor<string, []>("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_181_groups_0 = const()[name = tensor<string, []>("x_181_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85645440)))];
            tensor<fp16, [1, 512, 3]> x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16, x = input_371_cast_fp16)[name = tensor<string, []>("x_181_cast_fp16")];
            tensor<int32, [3]> input_373_perm_0 = const()[name = tensor<string, []>("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = tensor<string, []>("transpose_176")];
            tensor<fp16, [1, 3, 512]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
            tensor<int32, [1]> input_377_axes_0 = const()[name = tensor<string, []>("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86169792)))];
            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86170880)))];
            tensor<fp16, [1, 3, 512]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = tensor<string, []>("input_377_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86171968)))];
            tensor<fp16, [1, 3, 2048]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16, x = input_377_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor<string, []>("input_381_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88269184)))];
            tensor<fp16, [1, 3, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16, x = input_381_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
            tensor<fp16, []> var_1717_to_fp16 = const()[name = tensor<string, []>("op_1717_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1718_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1717_to_fp16)[name = tensor<string, []>("op_1718_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1718_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
            tensor<int32, [1]> input_389_axes_0 = const()[name = tensor<string, []>("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90366400)))];
            tensor<fp16, [512]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90367488)))];
            tensor<fp16, [1, 3, 512]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
            tensor<int32, [4]> cache_29_begin_0 = const()[name = tensor<string, []>("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> cache_29_end_0 = const()[name = tensor<string, []>("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 70, 512])];
            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = tensor<string, []>("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = tensor<string, []>("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_29_cast_fp16")];
            tensor<int32, [4]> cache_31_begin_0 = const()[name = tensor<string, []>("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> cache_31_end_0 = const()[name = tensor<string, []>("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 512, 8])];
            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = tensor<string, []>("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = tensor<string, []>("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_31_cast_fp16")];
            tensor<int32, [1]> input_391_axes_0 = const()[name = tensor<string, []>("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90368576)))];
            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90369664)))];
            tensor<fp16, [1, 3, 512]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = tensor<string, []>("input_391_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90370752)))];
            tensor<fp16, [1, 3, 2048]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16, x = input_391_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = tensor<string, []>("input_395_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92467968)))];
            tensor<fp16, [1, 3, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16, x = input_395_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
            tensor<fp16, []> var_1752_to_fp16 = const()[name = tensor<string, []>("op_1752_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1753_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1752_to_fp16)[name = tensor<string, []>("op_1753_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1753_cast_fp16)[name = tensor<string, []>("input_401_cast_fp16")];
            tensor<int32, [1]> key_15_axes_0 = const()[name = tensor<string, []>("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94565184)))];
            tensor<fp16, [512]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94566272)))];
            tensor<fp16, [1, 3, 512]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = tensor<string, []>("key_15_cast_fp16")];
            tensor<bool, []> input_403_interleave_0 = const()[name = tensor<string, []>("input_403_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_403_cast_fp16 = concat(axis = var_64, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = tensor<string, []>("input_403_cast_fp16")];
            tensor<int32, [3]> var_1775_begin_0 = const()[name = tensor<string, []>("op_1775_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1775_end_0 = const()[name = tensor<string, []>("op_1775_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1775_end_mask_0 = const()[name = tensor<string, []>("op_1775_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_1775_cast_fp16 = slice_by_index(begin = var_1775_begin_0, end = var_1775_end_0, end_mask = var_1775_end_mask_0, x = cache_29_cast_fp16)[name = tensor<string, []>("op_1775_cast_fp16")];
            tensor<int32, [3]> var_1778_begin_0 = const()[name = tensor<string, []>("op_1778_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1778_end_0 = const()[name = tensor<string, []>("op_1778_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_1778_end_mask_0 = const()[name = tensor<string, []>("op_1778_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_1778_cast_fp16 = slice_by_index(begin = var_1778_begin_0, end = var_1778_end_0, end_mask = var_1778_end_mask_0, x = key_15_cast_fp16)[name = tensor<string, []>("op_1778_cast_fp16")];
            tensor<bool, []> var_1781_interleave_0 = const()[name = tensor<string, []>("op_1781_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1781_cast_fp16 = concat(axis = var_64, interleave = var_1781_interleave_0, values = (var_1775_cast_fp16, var_1778_cast_fp16))[name = tensor<string, []>("op_1781_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94567360)))];
            tensor<fp16, [1, 3, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16, x = key_15_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
            tensor<int32, [4]> var_1785 = const()[name = tensor<string, []>("op_1785"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_43_cast_fp16 = reshape(shape = var_1785, x = linear_66_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95091712)))];
            tensor<fp16, [1, 73, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
            tensor<int32, [4]> var_1789 = const()[name = tensor<string, []>("op_1789"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_29_cast_fp16 = reshape(shape = var_1789, x = linear_67_cast_fp16)[name = tensor<string, []>("k_29_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95616064)))];
            tensor<fp16, [1, 73, 512]> linear_68_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
            tensor<int32, [4]> var_1793 = const()[name = tensor<string, []>("op_1793"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_15_cast_fp16 = reshape(shape = var_1793, x = linear_68_cast_fp16)[name = tensor<string, []>("v_15_cast_fp16")];
            tensor<int32, [4]> value_17_perm_0 = const()[name = tensor<string, []>("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96140416)))];
            tensor<fp16, [1, 3, 8, 64]> var_1805_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1805_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96141504)))];
            tensor<fp16, [1, 3, 8, 64]> var_1807_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1807_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_189_transpose_x_0 = const()[name = tensor<string, []>("x_189_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_189_transpose_y_0 = const()[name = tensor<string, []>("x_189_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_1809_to_fp16 = const()[name = tensor<string, []>("op_1809_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96142592)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_1807_cast_fp16)[name = tensor<string, []>("transpose_174")];
            tensor<fp16, [1, 8, 3, 145]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = var_1809_to_fp16)[name = tensor<string, []>("x_189_cast_fp16")];
            tensor<int32, [8]> x_191_pad_0 = const()[name = tensor<string, []>("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_191_mode_0 = const()[name = tensor<string, []>("x_191_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_114_to_fp16 = const()[name = tensor<string, []>("const_114_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_191_cast_fp16 = pad(constant_val = const_114_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = tensor<string, []>("x_191_cast_fp16")];
            tensor<int32, [4]> var_1817 = const()[name = tensor<string, []>("op_1817"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_193_cast_fp16 = reshape(shape = var_1817, x = x_191_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
            tensor<int32, [4]> var_1821_begin_0 = const()[name = tensor<string, []>("op_1821_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1821_end_0 = const()[name = tensor<string, []>("op_1821_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_1821_end_mask_0 = const()[name = tensor<string, []>("op_1821_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, x = x_193_cast_fp16)[name = tensor<string, []>("op_1821_cast_fp16")];
            tensor<int32, [4]> var_1822 = const()[name = tensor<string, []>("op_1822"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_29_cast_fp16 = reshape(shape = var_1822, x = var_1821_cast_fp16)[name = tensor<string, []>("matrix_bd_29_cast_fp16")];
            tensor<bool, []> matrix_ac_15_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_15_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_65_perm_0 = const()[name = tensor<string, []>("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = tensor<string, []>("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_66 = transpose(perm = transpose_66_perm_0, x = k_29_cast_fp16)[name = tensor<string, []>("transpose_172")];
            tensor<fp16, [1, 8, 3, 64]> transpose_65 = transpose(perm = transpose_65_perm_0, x = var_1805_cast_fp16)[name = tensor<string, []>("transpose_173")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_65, y = transpose_66)[name = tensor<string, []>("matrix_ac_15_cast_fp16")];
            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = tensor<string, []>("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = tensor<string, []>("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = tensor<string, []>("matrix_bd_31_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1831_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = tensor<string, []>("op_1831_cast_fp16")];
            tensor<fp16, []> _inversed_scores_29_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_29_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_29_cast_fp16 = mul(x = var_1831_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_29_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_31_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_31_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_1837_cast_fp16 = softmax(axis = var_62, x = scores_31_cast_fp16)[name = tensor<string, []>("op_1837_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_405_cast_fp16 = select(a = var_40_to_fp16, b = var_1837_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_405_cast_fp16")];
            tensor<bool, []> x_195_transpose_x_0 = const()[name = tensor<string, []>("x_195_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_195_transpose_y_0 = const()[name = tensor<string, []>("x_195_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_15_cast_fp16)[name = tensor<string, []>("transpose_175")];
            tensor<fp16, [1, 8, 3, 64]> x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_17_cast_fp16)[name = tensor<string, []>("x_195_cast_fp16")];
            tensor<int32, [4]> var_1841_perm_0 = const()[name = tensor<string, []>("op_1841_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1842 = const()[name = tensor<string, []>("op_1842"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_1841_cast_fp16 = transpose(perm = var_1841_perm_0, x = x_195_cast_fp16)[name = tensor<string, []>("transpose_171")];
            tensor<fp16, [1, 3, 512]> input_407_cast_fp16 = reshape(shape = var_1842, x = var_1841_cast_fp16)[name = tensor<string, []>("input_407_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96291136)))];
            tensor<fp16, [1, 3, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16, x = input_407_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = tensor<string, []>("input_411_cast_fp16")];
            tensor<int32, [1]> x_199_axes_0 = const()[name = tensor<string, []>("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96815488)))];
            tensor<fp16, [512]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96816576)))];
            tensor<fp16, [1, 3, 512]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")];
            tensor<int32, [3]> input_413_perm_0 = const()[name = tensor<string, []>("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_415_pad_type_0 = const()[name = tensor<string, []>("input_415_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_415_strides_0 = const()[name = tensor<string, []>("input_415_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_415_pad_0 = const()[name = tensor<string, []>("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_415_dilations_0 = const()[name = tensor<string, []>("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_415_groups_0 = const()[name = tensor<string, []>("input_415_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96817664)))];
            tensor<fp16, [1, 512, 3]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = tensor<string, []>("transpose_170")];
            tensor<fp16, [1, 1024, 3]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16, x = input_413_cast_fp16)[name = tensor<string, []>("input_415_cast_fp16")];
            tensor<int32, []> x_201_split_num_splits_0 = const()[name = tensor<string, []>("x_201_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_201_split_axis_0 = const()[name = tensor<string, []>("x_201_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_201_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = tensor<string, []>("x_201_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = tensor<string, []>("x_201_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_201_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_417_cast_fp16 = select(a = var_40_to_fp16, b = x_201_cast_fp16, cond = var_418)[name = tensor<string, []>("input_417_cast_fp16")];
            tensor<bool, []> new_x_31_interleave_0 = const()[name = tensor<string, []>("new_x_31_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_31_cast_fp16 = concat(axis = var_62, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = tensor<string, []>("new_x_31_cast_fp16")];
            tensor<int32, [3]> next_cache_15_begin_0 = const()[name = tensor<string, []>("next_cache_15_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_15_end_0 = const()[name = tensor<string, []>("next_cache_15_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_15_end_mask_0 = const()[name = tensor<string, []>("next_cache_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_15_cast_fp16 = slice_by_index(begin = next_cache_15_begin_0, end = next_cache_15_end_0, end_mask = next_cache_15_end_mask_0, x = new_x_31_cast_fp16)[name = tensor<string, []>("next_cache_15_cast_fp16")];
            tensor<int32, [3]> var_1883_begin_0 = const()[name = tensor<string, []>("op_1883_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_1883_end_0 = const()[name = tensor<string, []>("op_1883_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_1883_end_mask_0 = const()[name = tensor<string, []>("op_1883_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_1883_cast_fp16 = slice_by_index(begin = var_1883_begin_0, end = var_1883_end_0, end_mask = var_1883_end_mask_0, x = next_cache_15_cast_fp16)[name = tensor<string, []>("op_1883_cast_fp16")];
            tensor<string, []> x_203_pad_type_0 = const()[name = tensor<string, []>("x_203_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_203_groups_0 = const()[name = tensor<string, []>("x_203_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_203_strides_0 = const()[name = tensor<string, []>("x_203_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_203_pad_0 = const()[name = tensor<string, []>("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_203_dilations_0 = const()[name = tensor<string, []>("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97866304)))];
            tensor<fp16, [1, 512, 3]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16, x = new_x_31_cast_fp16)[name = tensor<string, []>("x_203_cast_fp16")];
            tensor<int32, [3]> input_419_perm_0 = const()[name = tensor<string, []>("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_205_axes_0 = const()[name = tensor<string, []>("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97875584)))];
            tensor<fp16, [512]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97876672)))];
            tensor<fp16, [1, 3, 512]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = tensor<string, []>("transpose_169")];
            tensor<fp16, [1, 3, 512]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = tensor<string, []>("x_205_cast_fp16")];
            tensor<int32, [3]> input_421_perm_0 = const()[name = tensor<string, []>("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = tensor<string, []>("transpose_168")];
            tensor<fp16, [1, 512, 3]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
            tensor<string, []> x_207_pad_type_0 = const()[name = tensor<string, []>("x_207_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_207_strides_0 = const()[name = tensor<string, []>("x_207_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_207_pad_0 = const()[name = tensor<string, []>("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_207_dilations_0 = const()[name = tensor<string, []>("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_207_groups_0 = const()[name = tensor<string, []>("x_207_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97877760)))];
            tensor<fp16, [1, 512, 3]> x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16, x = input_423_cast_fp16)[name = tensor<string, []>("x_207_cast_fp16")];
            tensor<int32, [3]> input_425_perm_0 = const()[name = tensor<string, []>("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = tensor<string, []>("transpose_167")];
            tensor<fp16, [1, 3, 512]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
            tensor<int32, [1]> input_429_axes_0 = const()[name = tensor<string, []>("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98402112)))];
            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98403200)))];
            tensor<fp16, [1, 3, 512]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = tensor<string, []>("input_429_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98404288)))];
            tensor<fp16, [1, 3, 2048]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16, x = input_429_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor<string, []>("input_433_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100501504)))];
            tensor<fp16, [1, 3, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16, x = input_433_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
            tensor<fp16, []> var_1924_to_fp16 = const()[name = tensor<string, []>("op_1924_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1925_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_1924_to_fp16)[name = tensor<string, []>("op_1925_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_1925_cast_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
            tensor<int32, [1]> input_441_axes_0 = const()[name = tensor<string, []>("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102598720)))];
            tensor<fp16, [512]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102599808)))];
            tensor<fp16, [1, 3, 512]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
            tensor<int32, [4]> cache_33_begin_0 = const()[name = tensor<string, []>("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> cache_33_end_0 = const()[name = tensor<string, []>("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 70, 512])];
            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = tensor<string, []>("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = tensor<string, []>("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_33_cast_fp16")];
            tensor<int32, [4]> cache_35_begin_0 = const()[name = tensor<string, []>("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> cache_35_end_0 = const()[name = tensor<string, []>("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 512, 8])];
            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = tensor<string, []>("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = tensor<string, []>("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_35_cast_fp16")];
            tensor<int32, [1]> input_443_axes_0 = const()[name = tensor<string, []>("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102600896)))];
            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102601984)))];
            tensor<fp16, [1, 3, 512]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = tensor<string, []>("input_443_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102603072)))];
            tensor<fp16, [1, 3, 2048]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16, x = input_443_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = tensor<string, []>("input_447_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104700288)))];
            tensor<fp16, [1, 3, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16, x = input_447_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")];
            tensor<fp16, []> var_1959_to_fp16 = const()[name = tensor<string, []>("op_1959_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_1960_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_1959_to_fp16)[name = tensor<string, []>("op_1960_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_1960_cast_fp16)[name = tensor<string, []>("input_453_cast_fp16")];
            tensor<int32, [1]> key_17_axes_0 = const()[name = tensor<string, []>("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106797504)))];
            tensor<fp16, [512]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106798592)))];
            tensor<fp16, [1, 3, 512]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
            tensor<bool, []> input_455_interleave_0 = const()[name = tensor<string, []>("input_455_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_455_cast_fp16 = concat(axis = var_64, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = tensor<string, []>("input_455_cast_fp16")];
            tensor<int32, [3]> var_1982_begin_0 = const()[name = tensor<string, []>("op_1982_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1982_end_0 = const()[name = tensor<string, []>("op_1982_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1982_end_mask_0 = const()[name = tensor<string, []>("op_1982_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_1982_cast_fp16 = slice_by_index(begin = var_1982_begin_0, end = var_1982_end_0, end_mask = var_1982_end_mask_0, x = cache_33_cast_fp16)[name = tensor<string, []>("op_1982_cast_fp16")];
            tensor<int32, [3]> var_1985_begin_0 = const()[name = tensor<string, []>("op_1985_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1985_end_0 = const()[name = tensor<string, []>("op_1985_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_1985_end_mask_0 = const()[name = tensor<string, []>("op_1985_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_1985_cast_fp16 = slice_by_index(begin = var_1985_begin_0, end = var_1985_end_0, end_mask = var_1985_end_mask_0, x = key_17_cast_fp16)[name = tensor<string, []>("op_1985_cast_fp16")];
            tensor<bool, []> var_1988_interleave_0 = const()[name = tensor<string, []>("op_1988_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1988_cast_fp16 = concat(axis = var_64, interleave = var_1988_interleave_0, values = (var_1982_cast_fp16, var_1985_cast_fp16))[name = tensor<string, []>("op_1988_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106799680)))];
            tensor<fp16, [1, 3, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16, x = key_17_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
            tensor<int32, [4]> var_1992 = const()[name = tensor<string, []>("op_1992"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_49_cast_fp16 = reshape(shape = var_1992, x = linear_75_cast_fp16)[name = tensor<string, []>("q_49_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107324032)))];
            tensor<fp16, [1, 73, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
            tensor<int32, [4]> var_1996 = const()[name = tensor<string, []>("op_1996"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_33_cast_fp16 = reshape(shape = var_1996, x = linear_76_cast_fp16)[name = tensor<string, []>("k_33_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107848384)))];
            tensor<fp16, [1, 73, 512]> linear_77_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
            tensor<int32, [4]> var_2000 = const()[name = tensor<string, []>("op_2000"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_17_cast_fp16 = reshape(shape = var_2000, x = linear_77_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
            tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<string, []>("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108372736)))];
            tensor<fp16, [1, 3, 8, 64]> var_2012_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2012_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108373824)))];
            tensor<fp16, [1, 3, 8, 64]> var_2014_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2014_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_215_transpose_x_0 = const()[name = tensor<string, []>("x_215_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_215_transpose_y_0 = const()[name = tensor<string, []>("x_215_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_2016_to_fp16 = const()[name = tensor<string, []>("op_2016_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108374912)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2014_cast_fp16)[name = tensor<string, []>("transpose_165")];
            tensor<fp16, [1, 8, 3, 145]> x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = var_2016_to_fp16)[name = tensor<string, []>("x_215_cast_fp16")];
            tensor<int32, [8]> x_217_pad_0 = const()[name = tensor<string, []>("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_217_mode_0 = const()[name = tensor<string, []>("x_217_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_127_to_fp16 = const()[name = tensor<string, []>("const_127_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_217_cast_fp16 = pad(constant_val = const_127_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = tensor<string, []>("x_217_cast_fp16")];
            tensor<int32, [4]> var_2024 = const()[name = tensor<string, []>("op_2024"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_219_cast_fp16 = reshape(shape = var_2024, x = x_217_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
            tensor<int32, [4]> var_2028_begin_0 = const()[name = tensor<string, []>("op_2028_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2028_end_0 = const()[name = tensor<string, []>("op_2028_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_2028_end_mask_0 = const()[name = tensor<string, []>("op_2028_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_2028_cast_fp16 = slice_by_index(begin = var_2028_begin_0, end = var_2028_end_0, end_mask = var_2028_end_mask_0, x = x_219_cast_fp16)[name = tensor<string, []>("op_2028_cast_fp16")];
            tensor<int32, [4]> var_2029 = const()[name = tensor<string, []>("op_2029"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_33_cast_fp16 = reshape(shape = var_2029, x = var_2028_cast_fp16)[name = tensor<string, []>("matrix_bd_33_cast_fp16")];
            tensor<bool, []> matrix_ac_17_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_17_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_67_perm_0 = const()[name = tensor<string, []>("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = tensor<string, []>("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_68 = transpose(perm = transpose_68_perm_0, x = k_33_cast_fp16)[name = tensor<string, []>("transpose_163")];
            tensor<fp16, [1, 8, 3, 64]> transpose_67 = transpose(perm = transpose_67_perm_0, x = var_2012_cast_fp16)[name = tensor<string, []>("transpose_164")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_67, y = transpose_68)[name = tensor<string, []>("matrix_ac_17_cast_fp16")];
            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = tensor<string, []>("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = tensor<string, []>("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = tensor<string, []>("matrix_bd_35_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2038_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = tensor<string, []>("op_2038_cast_fp16")];
            tensor<fp16, []> _inversed_scores_33_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_33_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_33_cast_fp16 = mul(x = var_2038_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_33_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_35_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_35_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2044_cast_fp16 = softmax(axis = var_62, x = scores_35_cast_fp16)[name = tensor<string, []>("op_2044_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_457_cast_fp16 = select(a = var_40_to_fp16, b = var_2044_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_457_cast_fp16")];
            tensor<bool, []> x_221_transpose_x_0 = const()[name = tensor<string, []>("x_221_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_221_transpose_y_0 = const()[name = tensor<string, []>("x_221_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_17_cast_fp16)[name = tensor<string, []>("transpose_166")];
            tensor<fp16, [1, 8, 3, 64]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_19_cast_fp16)[name = tensor<string, []>("x_221_cast_fp16")];
            tensor<int32, [4]> var_2048_perm_0 = const()[name = tensor<string, []>("op_2048_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2049 = const()[name = tensor<string, []>("op_2049"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_2048_cast_fp16 = transpose(perm = var_2048_perm_0, x = x_221_cast_fp16)[name = tensor<string, []>("transpose_162")];
            tensor<fp16, [1, 3, 512]> input_459_cast_fp16 = reshape(shape = var_2049, x = var_2048_cast_fp16)[name = tensor<string, []>("input_459_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108523456)))];
            tensor<fp16, [1, 3, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16, x = input_459_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = tensor<string, []>("input_463_cast_fp16")];
            tensor<int32, [1]> x_225_axes_0 = const()[name = tensor<string, []>("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109047808)))];
            tensor<fp16, [512]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109048896)))];
            tensor<fp16, [1, 3, 512]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("x_225_cast_fp16")];
            tensor<int32, [3]> input_465_perm_0 = const()[name = tensor<string, []>("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_467_pad_type_0 = const()[name = tensor<string, []>("input_467_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_467_strides_0 = const()[name = tensor<string, []>("input_467_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_467_pad_0 = const()[name = tensor<string, []>("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_467_dilations_0 = const()[name = tensor<string, []>("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_467_groups_0 = const()[name = tensor<string, []>("input_467_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109049984)))];
            tensor<fp16, [1, 512, 3]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = tensor<string, []>("transpose_161")];
            tensor<fp16, [1, 1024, 3]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16, x = input_465_cast_fp16)[name = tensor<string, []>("input_467_cast_fp16")];
            tensor<int32, []> x_227_split_num_splits_0 = const()[name = tensor<string, []>("x_227_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_227_split_axis_0 = const()[name = tensor<string, []>("x_227_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_227_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = tensor<string, []>("x_227_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = tensor<string, []>("x_227_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_227_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_469_cast_fp16 = select(a = var_40_to_fp16, b = x_227_cast_fp16, cond = var_418)[name = tensor<string, []>("input_469_cast_fp16")];
            tensor<bool, []> new_x_35_interleave_0 = const()[name = tensor<string, []>("new_x_35_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_35_cast_fp16 = concat(axis = var_62, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = tensor<string, []>("new_x_35_cast_fp16")];
            tensor<int32, [3]> next_cache_17_begin_0 = const()[name = tensor<string, []>("next_cache_17_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_17_end_0 = const()[name = tensor<string, []>("next_cache_17_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_17_end_mask_0 = const()[name = tensor<string, []>("next_cache_17_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_17_cast_fp16 = slice_by_index(begin = next_cache_17_begin_0, end = next_cache_17_end_0, end_mask = next_cache_17_end_mask_0, x = new_x_35_cast_fp16)[name = tensor<string, []>("next_cache_17_cast_fp16")];
            tensor<int32, [3]> var_2090_begin_0 = const()[name = tensor<string, []>("op_2090_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_2090_end_0 = const()[name = tensor<string, []>("op_2090_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_2090_end_mask_0 = const()[name = tensor<string, []>("op_2090_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2090_cast_fp16 = slice_by_index(begin = var_2090_begin_0, end = var_2090_end_0, end_mask = var_2090_end_mask_0, x = next_cache_17_cast_fp16)[name = tensor<string, []>("op_2090_cast_fp16")];
            tensor<string, []> x_229_pad_type_0 = const()[name = tensor<string, []>("x_229_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_229_groups_0 = const()[name = tensor<string, []>("x_229_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_229_strides_0 = const()[name = tensor<string, []>("x_229_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_229_pad_0 = const()[name = tensor<string, []>("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_229_dilations_0 = const()[name = tensor<string, []>("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110098624)))];
            tensor<fp16, [1, 512, 3]> x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16, x = new_x_35_cast_fp16)[name = tensor<string, []>("x_229_cast_fp16")];
            tensor<int32, [3]> input_471_perm_0 = const()[name = tensor<string, []>("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_231_axes_0 = const()[name = tensor<string, []>("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110107904)))];
            tensor<fp16, [512]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110108992)))];
            tensor<fp16, [1, 3, 512]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = tensor<string, []>("transpose_160")];
            tensor<fp16, [1, 3, 512]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = tensor<string, []>("x_231_cast_fp16")];
            tensor<int32, [3]> input_473_perm_0 = const()[name = tensor<string, []>("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = tensor<string, []>("transpose_159")];
            tensor<fp16, [1, 512, 3]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
            tensor<string, []> x_233_pad_type_0 = const()[name = tensor<string, []>("x_233_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_233_strides_0 = const()[name = tensor<string, []>("x_233_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_233_pad_0 = const()[name = tensor<string, []>("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_233_dilations_0 = const()[name = tensor<string, []>("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_233_groups_0 = const()[name = tensor<string, []>("x_233_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110110080)))];
            tensor<fp16, [1, 512, 3]> x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16, x = input_475_cast_fp16)[name = tensor<string, []>("x_233_cast_fp16")];
            tensor<int32, [3]> input_477_perm_0 = const()[name = tensor<string, []>("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = tensor<string, []>("transpose_158")];
            tensor<fp16, [1, 3, 512]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
            tensor<int32, [1]> input_481_axes_0 = const()[name = tensor<string, []>("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110634432)))];
            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110635520)))];
            tensor<fp16, [1, 3, 512]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = tensor<string, []>("input_481_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110636608)))];
            tensor<fp16, [1, 3, 2048]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16, x = input_481_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = tensor<string, []>("input_485_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112733824)))];
            tensor<fp16, [1, 3, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16, x = input_485_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")];
            tensor<fp16, []> var_2131_to_fp16 = const()[name = tensor<string, []>("op_2131_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2132_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2131_to_fp16)[name = tensor<string, []>("op_2132_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2132_cast_fp16)[name = tensor<string, []>("input_491_cast_fp16")];
            tensor<int32, [1]> input_493_axes_0 = const()[name = tensor<string, []>("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114831040)))];
            tensor<fp16, [512]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114832128)))];
            tensor<fp16, [1, 3, 512]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
            tensor<int32, [4]> cache_37_begin_0 = const()[name = tensor<string, []>("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> cache_37_end_0 = const()[name = tensor<string, []>("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 70, 512])];
            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = tensor<string, []>("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = tensor<string, []>("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_37_cast_fp16")];
            tensor<int32, [4]> cache_39_begin_0 = const()[name = tensor<string, []>("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> cache_39_end_0 = const()[name = tensor<string, []>("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 512, 8])];
            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = tensor<string, []>("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = tensor<string, []>("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_39_cast_fp16")];
            tensor<int32, [1]> input_495_axes_0 = const()[name = tensor<string, []>("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114833216)))];
            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114834304)))];
            tensor<fp16, [1, 3, 512]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<string, []>("input_495_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114835392)))];
            tensor<fp16, [1, 3, 2048]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16, x = input_495_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = tensor<string, []>("input_499_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116932608)))];
            tensor<fp16, [1, 3, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16, x = input_499_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")];
            tensor<fp16, []> var_2166_to_fp16 = const()[name = tensor<string, []>("op_2166_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2167_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2166_to_fp16)[name = tensor<string, []>("op_2167_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2167_cast_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
            tensor<int32, [1]> key_19_axes_0 = const()[name = tensor<string, []>("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119029824)))];
            tensor<fp16, [512]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119030912)))];
            tensor<fp16, [1, 3, 512]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = tensor<string, []>("key_19_cast_fp16")];
            tensor<bool, []> input_507_interleave_0 = const()[name = tensor<string, []>("input_507_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_507_cast_fp16 = concat(axis = var_64, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = tensor<string, []>("input_507_cast_fp16")];
            tensor<int32, [3]> var_2189_begin_0 = const()[name = tensor<string, []>("op_2189_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2189_end_0 = const()[name = tensor<string, []>("op_2189_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_2189_end_mask_0 = const()[name = tensor<string, []>("op_2189_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_2189_cast_fp16 = slice_by_index(begin = var_2189_begin_0, end = var_2189_end_0, end_mask = var_2189_end_mask_0, x = cache_37_cast_fp16)[name = tensor<string, []>("op_2189_cast_fp16")];
            tensor<int32, [3]> var_2192_begin_0 = const()[name = tensor<string, []>("op_2192_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2192_end_0 = const()[name = tensor<string, []>("op_2192_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_2192_end_mask_0 = const()[name = tensor<string, []>("op_2192_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_2192_cast_fp16 = slice_by_index(begin = var_2192_begin_0, end = var_2192_end_0, end_mask = var_2192_end_mask_0, x = key_19_cast_fp16)[name = tensor<string, []>("op_2192_cast_fp16")];
            tensor<bool, []> var_2195_interleave_0 = const()[name = tensor<string, []>("op_2195_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_2195_cast_fp16 = concat(axis = var_64, interleave = var_2195_interleave_0, values = (var_2189_cast_fp16, var_2192_cast_fp16))[name = tensor<string, []>("op_2195_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119032000)))];
            tensor<fp16, [1, 3, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16, x = key_19_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")];
            tensor<int32, [4]> var_2199 = const()[name = tensor<string, []>("op_2199"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_55_cast_fp16 = reshape(shape = var_2199, x = linear_84_cast_fp16)[name = tensor<string, []>("q_55_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119556352)))];
            tensor<fp16, [1, 73, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")];
            tensor<int32, [4]> var_2203 = const()[name = tensor<string, []>("op_2203"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_37_cast_fp16 = reshape(shape = var_2203, x = linear_85_cast_fp16)[name = tensor<string, []>("k_37_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120080704)))];
            tensor<fp16, [1, 73, 512]> linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")];
            tensor<int32, [4]> var_2207 = const()[name = tensor<string, []>("op_2207"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_19_cast_fp16 = reshape(shape = var_2207, x = linear_86_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")];
            tensor<int32, [4]> value_21_perm_0 = const()[name = tensor<string, []>("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120605056)))];
            tensor<fp16, [1, 3, 8, 64]> var_2219_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2219_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120606144)))];
            tensor<fp16, [1, 3, 8, 64]> var_2221_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2221_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_241_transpose_x_0 = const()[name = tensor<string, []>("x_241_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_241_transpose_y_0 = const()[name = tensor<string, []>("x_241_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_2223_to_fp16 = const()[name = tensor<string, []>("op_2223_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120607232)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2221_cast_fp16)[name = tensor<string, []>("transpose_156")];
            tensor<fp16, [1, 8, 3, 145]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = var_2223_to_fp16)[name = tensor<string, []>("x_241_cast_fp16")];
            tensor<int32, [8]> x_243_pad_0 = const()[name = tensor<string, []>("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_243_mode_0 = const()[name = tensor<string, []>("x_243_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_140_to_fp16 = const()[name = tensor<string, []>("const_140_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_243_cast_fp16 = pad(constant_val = const_140_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = tensor<string, []>("x_243_cast_fp16")];
            tensor<int32, [4]> var_2231 = const()[name = tensor<string, []>("op_2231"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_245_cast_fp16 = reshape(shape = var_2231, x = x_243_cast_fp16)[name = tensor<string, []>("x_245_cast_fp16")];
            tensor<int32, [4]> var_2235_begin_0 = const()[name = tensor<string, []>("op_2235_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2235_end_0 = const()[name = tensor<string, []>("op_2235_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_2235_end_mask_0 = const()[name = tensor<string, []>("op_2235_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_2235_cast_fp16 = slice_by_index(begin = var_2235_begin_0, end = var_2235_end_0, end_mask = var_2235_end_mask_0, x = x_245_cast_fp16)[name = tensor<string, []>("op_2235_cast_fp16")];
            tensor<int32, [4]> var_2236 = const()[name = tensor<string, []>("op_2236"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_37_cast_fp16 = reshape(shape = var_2236, x = var_2235_cast_fp16)[name = tensor<string, []>("matrix_bd_37_cast_fp16")];
            tensor<bool, []> matrix_ac_19_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_19_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_69_perm_0 = const()[name = tensor<string, []>("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = tensor<string, []>("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_70 = transpose(perm = transpose_70_perm_0, x = k_37_cast_fp16)[name = tensor<string, []>("transpose_154")];
            tensor<fp16, [1, 8, 3, 64]> transpose_69 = transpose(perm = transpose_69_perm_0, x = var_2219_cast_fp16)[name = tensor<string, []>("transpose_155")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor<string, []>("matrix_ac_19_cast_fp16")];
            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = tensor<string, []>("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = tensor<string, []>("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = tensor<string, []>("matrix_bd_39_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2245_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = tensor<string, []>("op_2245_cast_fp16")];
            tensor<fp16, []> _inversed_scores_37_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_37_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_37_cast_fp16 = mul(x = var_2245_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_37_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_39_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_39_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2251_cast_fp16 = softmax(axis = var_62, x = scores_39_cast_fp16)[name = tensor<string, []>("op_2251_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_509_cast_fp16 = select(a = var_40_to_fp16, b = var_2251_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_509_cast_fp16")];
            tensor<bool, []> x_247_transpose_x_0 = const()[name = tensor<string, []>("x_247_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_247_transpose_y_0 = const()[name = tensor<string, []>("x_247_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_19_cast_fp16)[name = tensor<string, []>("transpose_157")];
            tensor<fp16, [1, 8, 3, 64]> x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_21_cast_fp16)[name = tensor<string, []>("x_247_cast_fp16")];
            tensor<int32, [4]> var_2255_perm_0 = const()[name = tensor<string, []>("op_2255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2256 = const()[name = tensor<string, []>("op_2256"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_2255_cast_fp16 = transpose(perm = var_2255_perm_0, x = x_247_cast_fp16)[name = tensor<string, []>("transpose_153")];
            tensor<fp16, [1, 3, 512]> input_511_cast_fp16 = reshape(shape = var_2256, x = var_2255_cast_fp16)[name = tensor<string, []>("input_511_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120755776)))];
            tensor<fp16, [1, 3, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16, x = input_511_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = tensor<string, []>("input_515_cast_fp16")];
            tensor<int32, [1]> x_251_axes_0 = const()[name = tensor<string, []>("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121280128)))];
            tensor<fp16, [512]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121281216)))];
            tensor<fp16, [1, 3, 512]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = tensor<string, []>("x_251_cast_fp16")];
            tensor<int32, [3]> input_517_perm_0 = const()[name = tensor<string, []>("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_519_pad_type_0 = const()[name = tensor<string, []>("input_519_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_519_strides_0 = const()[name = tensor<string, []>("input_519_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_519_pad_0 = const()[name = tensor<string, []>("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_519_dilations_0 = const()[name = tensor<string, []>("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_519_groups_0 = const()[name = tensor<string, []>("input_519_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121282304)))];
            tensor<fp16, [1, 512, 3]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = tensor<string, []>("transpose_152")];
            tensor<fp16, [1, 1024, 3]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16, x = input_517_cast_fp16)[name = tensor<string, []>("input_519_cast_fp16")];
            tensor<int32, []> x_253_split_num_splits_0 = const()[name = tensor<string, []>("x_253_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_253_split_axis_0 = const()[name = tensor<string, []>("x_253_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_253_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = tensor<string, []>("x_253_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = tensor<string, []>("x_253_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_253_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_521_cast_fp16 = select(a = var_40_to_fp16, b = x_253_cast_fp16, cond = var_418)[name = tensor<string, []>("input_521_cast_fp16")];
            tensor<bool, []> new_x_39_interleave_0 = const()[name = tensor<string, []>("new_x_39_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_39_cast_fp16 = concat(axis = var_62, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = tensor<string, []>("new_x_39_cast_fp16")];
            tensor<int32, [3]> next_cache_19_begin_0 = const()[name = tensor<string, []>("next_cache_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_19_end_0 = const()[name = tensor<string, []>("next_cache_19_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_19_end_mask_0 = const()[name = tensor<string, []>("next_cache_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_19_cast_fp16 = slice_by_index(begin = next_cache_19_begin_0, end = next_cache_19_end_0, end_mask = next_cache_19_end_mask_0, x = new_x_39_cast_fp16)[name = tensor<string, []>("next_cache_19_cast_fp16")];
            tensor<int32, [3]> var_2297_begin_0 = const()[name = tensor<string, []>("op_2297_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_2297_end_0 = const()[name = tensor<string, []>("op_2297_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_2297_end_mask_0 = const()[name = tensor<string, []>("op_2297_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = next_cache_19_cast_fp16)[name = tensor<string, []>("op_2297_cast_fp16")];
            tensor<string, []> x_255_pad_type_0 = const()[name = tensor<string, []>("x_255_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_255_groups_0 = const()[name = tensor<string, []>("x_255_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_255_strides_0 = const()[name = tensor<string, []>("x_255_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_255_pad_0 = const()[name = tensor<string, []>("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_255_dilations_0 = const()[name = tensor<string, []>("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122330944)))];
            tensor<fp16, [1, 512, 3]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16, x = new_x_39_cast_fp16)[name = tensor<string, []>("x_255_cast_fp16")];
            tensor<int32, [3]> input_523_perm_0 = const()[name = tensor<string, []>("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_257_axes_0 = const()[name = tensor<string, []>("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122340224)))];
            tensor<fp16, [512]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122341312)))];
            tensor<fp16, [1, 3, 512]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = tensor<string, []>("transpose_151")];
            tensor<fp16, [1, 3, 512]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = tensor<string, []>("x_257_cast_fp16")];
            tensor<int32, [3]> input_525_perm_0 = const()[name = tensor<string, []>("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = tensor<string, []>("transpose_150")];
            tensor<fp16, [1, 512, 3]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
            tensor<string, []> x_259_pad_type_0 = const()[name = tensor<string, []>("x_259_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_259_strides_0 = const()[name = tensor<string, []>("x_259_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_259_pad_0 = const()[name = tensor<string, []>("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_259_dilations_0 = const()[name = tensor<string, []>("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_259_groups_0 = const()[name = tensor<string, []>("x_259_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122342400)))];
            tensor<fp16, [1, 512, 3]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16, x = input_527_cast_fp16)[name = tensor<string, []>("x_259_cast_fp16")];
            tensor<int32, [3]> input_529_perm_0 = const()[name = tensor<string, []>("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = tensor<string, []>("transpose_149")];
            tensor<fp16, [1, 3, 512]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = tensor<string, []>("input_531_cast_fp16")];
            tensor<int32, [1]> input_533_axes_0 = const()[name = tensor<string, []>("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122866752)))];
            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122867840)))];
            tensor<fp16, [1, 3, 512]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = tensor<string, []>("input_533_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122868928)))];
            tensor<fp16, [1, 3, 2048]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16, x = input_533_cast_fp16)[name = tensor<string, []>("linear_89_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = tensor<string, []>("input_537_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124966144)))];
            tensor<fp16, [1, 3, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor<string, []>("linear_90_cast_fp16")];
            tensor<fp16, []> var_2338_to_fp16 = const()[name = tensor<string, []>("op_2338_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2339_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2338_to_fp16)[name = tensor<string, []>("op_2339_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2339_cast_fp16)[name = tensor<string, []>("input_543_cast_fp16")];
            tensor<int32, [1]> input_545_axes_0 = const()[name = tensor<string, []>("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127063360)))];
            tensor<fp16, [512]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127064448)))];
            tensor<fp16, [1, 3, 512]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
            tensor<int32, [4]> cache_41_begin_0 = const()[name = tensor<string, []>("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> cache_41_end_0 = const()[name = tensor<string, []>("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 70, 512])];
            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = tensor<string, []>("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = tensor<string, []>("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_41_cast_fp16")];
            tensor<int32, [4]> cache_43_begin_0 = const()[name = tensor<string, []>("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> cache_43_end_0 = const()[name = tensor<string, []>("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 512, 8])];
            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = tensor<string, []>("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = tensor<string, []>("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_43_cast_fp16")];
            tensor<int32, [1]> input_547_axes_0 = const()[name = tensor<string, []>("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127065536)))];
            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127066624)))];
            tensor<fp16, [1, 3, 512]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = tensor<string, []>("input_547_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127067712)))];
            tensor<fp16, [1, 3, 2048]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16, x = input_547_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor<string, []>("input_551_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129164928)))];
            tensor<fp16, [1, 3, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16, x = input_551_cast_fp16)[name = tensor<string, []>("linear_92_cast_fp16")];
            tensor<fp16, []> var_2373_to_fp16 = const()[name = tensor<string, []>("op_2373_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2374_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2373_to_fp16)[name = tensor<string, []>("op_2374_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2374_cast_fp16)[name = tensor<string, []>("input_557_cast_fp16")];
            tensor<int32, [1]> key_21_axes_0 = const()[name = tensor<string, []>("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131262144)))];
            tensor<fp16, [512]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131263232)))];
            tensor<fp16, [1, 3, 512]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
            tensor<bool, []> input_559_interleave_0 = const()[name = tensor<string, []>("input_559_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_559_cast_fp16 = concat(axis = var_64, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = tensor<string, []>("input_559_cast_fp16")];
            tensor<int32, [3]> var_2396_begin_0 = const()[name = tensor<string, []>("op_2396_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2396_end_0 = const()[name = tensor<string, []>("op_2396_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_2396_end_mask_0 = const()[name = tensor<string, []>("op_2396_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_2396_cast_fp16 = slice_by_index(begin = var_2396_begin_0, end = var_2396_end_0, end_mask = var_2396_end_mask_0, x = cache_41_cast_fp16)[name = tensor<string, []>("op_2396_cast_fp16")];
            tensor<int32, [3]> var_2399_begin_0 = const()[name = tensor<string, []>("op_2399_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2399_end_0 = const()[name = tensor<string, []>("op_2399_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_2399_end_mask_0 = const()[name = tensor<string, []>("op_2399_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_2399_cast_fp16 = slice_by_index(begin = var_2399_begin_0, end = var_2399_end_0, end_mask = var_2399_end_mask_0, x = key_21_cast_fp16)[name = tensor<string, []>("op_2399_cast_fp16")];
            tensor<bool, []> var_2402_interleave_0 = const()[name = tensor<string, []>("op_2402_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_2402_cast_fp16 = concat(axis = var_64, interleave = var_2402_interleave_0, values = (var_2396_cast_fp16, var_2399_cast_fp16))[name = tensor<string, []>("op_2402_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131264320)))];
            tensor<fp16, [1, 3, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16, x = key_21_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2406 = const()[name = tensor<string, []>("op_2406"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_61_cast_fp16 = reshape(shape = var_2406, x = linear_93_cast_fp16)[name = tensor<string, []>("q_61_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131788672)))];
            tensor<fp16, [1, 73, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")];
            tensor<int32, [4]> var_2410 = const()[name = tensor<string, []>("op_2410"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_41_cast_fp16 = reshape(shape = var_2410, x = linear_94_cast_fp16)[name = tensor<string, []>("k_41_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132313024)))];
            tensor<fp16, [1, 73, 512]> linear_95_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")];
            tensor<int32, [4]> var_2414 = const()[name = tensor<string, []>("op_2414"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_21_cast_fp16 = reshape(shape = var_2414, x = linear_95_cast_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
            tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<string, []>("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132837376)))];
            tensor<fp16, [1, 3, 8, 64]> var_2426_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2426_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132838464)))];
            tensor<fp16, [1, 3, 8, 64]> var_2428_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2428_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_267_transpose_x_0 = const()[name = tensor<string, []>("x_267_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_267_transpose_y_0 = const()[name = tensor<string, []>("x_267_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_2430_to_fp16 = const()[name = tensor<string, []>("op_2430_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132839552)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2428_cast_fp16)[name = tensor<string, []>("transpose_147")];
            tensor<fp16, [1, 8, 3, 145]> x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = var_2430_to_fp16)[name = tensor<string, []>("x_267_cast_fp16")];
            tensor<int32, [8]> x_269_pad_0 = const()[name = tensor<string, []>("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_269_mode_0 = const()[name = tensor<string, []>("x_269_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_153_to_fp16 = const()[name = tensor<string, []>("const_153_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_269_cast_fp16 = pad(constant_val = const_153_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = tensor<string, []>("x_269_cast_fp16")];
            tensor<int32, [4]> var_2438 = const()[name = tensor<string, []>("op_2438"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_271_cast_fp16 = reshape(shape = var_2438, x = x_269_cast_fp16)[name = tensor<string, []>("x_271_cast_fp16")];
            tensor<int32, [4]> var_2442_begin_0 = const()[name = tensor<string, []>("op_2442_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2442_end_0 = const()[name = tensor<string, []>("op_2442_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_2442_end_mask_0 = const()[name = tensor<string, []>("op_2442_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_2442_cast_fp16 = slice_by_index(begin = var_2442_begin_0, end = var_2442_end_0, end_mask = var_2442_end_mask_0, x = x_271_cast_fp16)[name = tensor<string, []>("op_2442_cast_fp16")];
            tensor<int32, [4]> var_2443 = const()[name = tensor<string, []>("op_2443"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_41_cast_fp16 = reshape(shape = var_2443, x = var_2442_cast_fp16)[name = tensor<string, []>("matrix_bd_41_cast_fp16")];
            tensor<bool, []> matrix_ac_21_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_21_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_71_perm_0 = const()[name = tensor<string, []>("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = tensor<string, []>("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_72 = transpose(perm = transpose_72_perm_0, x = k_41_cast_fp16)[name = tensor<string, []>("transpose_145")];
            tensor<fp16, [1, 8, 3, 64]> transpose_71 = transpose(perm = transpose_71_perm_0, x = var_2426_cast_fp16)[name = tensor<string, []>("transpose_146")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor<string, []>("matrix_ac_21_cast_fp16")];
            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = tensor<string, []>("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = tensor<string, []>("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = tensor<string, []>("matrix_bd_43_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2452_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = tensor<string, []>("op_2452_cast_fp16")];
            tensor<fp16, []> _inversed_scores_41_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_41_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_41_cast_fp16 = mul(x = var_2452_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_41_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_43_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_43_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2458_cast_fp16 = softmax(axis = var_62, x = scores_43_cast_fp16)[name = tensor<string, []>("op_2458_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_561_cast_fp16 = select(a = var_40_to_fp16, b = var_2458_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_561_cast_fp16")];
            tensor<bool, []> x_273_transpose_x_0 = const()[name = tensor<string, []>("x_273_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_273_transpose_y_0 = const()[name = tensor<string, []>("x_273_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_21_cast_fp16)[name = tensor<string, []>("transpose_148")];
            tensor<fp16, [1, 8, 3, 64]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_23_cast_fp16)[name = tensor<string, []>("x_273_cast_fp16")];
            tensor<int32, [4]> var_2462_perm_0 = const()[name = tensor<string, []>("op_2462_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2463 = const()[name = tensor<string, []>("op_2463"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_2462_cast_fp16 = transpose(perm = var_2462_perm_0, x = x_273_cast_fp16)[name = tensor<string, []>("transpose_144")];
            tensor<fp16, [1, 3, 512]> input_563_cast_fp16 = reshape(shape = var_2463, x = var_2462_cast_fp16)[name = tensor<string, []>("input_563_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132988096)))];
            tensor<fp16, [1, 3, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16, x = input_563_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = tensor<string, []>("input_567_cast_fp16")];
            tensor<int32, [1]> x_277_axes_0 = const()[name = tensor<string, []>("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133512448)))];
            tensor<fp16, [512]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133513536)))];
            tensor<fp16, [1, 3, 512]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = tensor<string, []>("x_277_cast_fp16")];
            tensor<int32, [3]> input_569_perm_0 = const()[name = tensor<string, []>("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_571_pad_type_0 = const()[name = tensor<string, []>("input_571_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_571_strides_0 = const()[name = tensor<string, []>("input_571_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_571_pad_0 = const()[name = tensor<string, []>("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_571_dilations_0 = const()[name = tensor<string, []>("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_571_groups_0 = const()[name = tensor<string, []>("input_571_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133514624)))];
            tensor<fp16, [1, 512, 3]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = tensor<string, []>("transpose_143")];
            tensor<fp16, [1, 1024, 3]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16, x = input_569_cast_fp16)[name = tensor<string, []>("input_571_cast_fp16")];
            tensor<int32, []> x_279_split_num_splits_0 = const()[name = tensor<string, []>("x_279_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_279_split_axis_0 = const()[name = tensor<string, []>("x_279_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_279_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = tensor<string, []>("x_279_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = tensor<string, []>("x_279_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_279_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_573_cast_fp16 = select(a = var_40_to_fp16, b = x_279_cast_fp16, cond = var_418)[name = tensor<string, []>("input_573_cast_fp16")];
            tensor<bool, []> new_x_43_interleave_0 = const()[name = tensor<string, []>("new_x_43_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_43_cast_fp16 = concat(axis = var_62, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = tensor<string, []>("new_x_43_cast_fp16")];
            tensor<int32, [3]> next_cache_21_begin_0 = const()[name = tensor<string, []>("next_cache_21_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_21_end_0 = const()[name = tensor<string, []>("next_cache_21_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_21_end_mask_0 = const()[name = tensor<string, []>("next_cache_21_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_21_cast_fp16 = slice_by_index(begin = next_cache_21_begin_0, end = next_cache_21_end_0, end_mask = next_cache_21_end_mask_0, x = new_x_43_cast_fp16)[name = tensor<string, []>("next_cache_21_cast_fp16")];
            tensor<int32, [3]> var_2504_begin_0 = const()[name = tensor<string, []>("op_2504_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_2504_end_0 = const()[name = tensor<string, []>("op_2504_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_2504_end_mask_0 = const()[name = tensor<string, []>("op_2504_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2504_cast_fp16 = slice_by_index(begin = var_2504_begin_0, end = var_2504_end_0, end_mask = var_2504_end_mask_0, x = next_cache_21_cast_fp16)[name = tensor<string, []>("op_2504_cast_fp16")];
            tensor<string, []> x_281_pad_type_0 = const()[name = tensor<string, []>("x_281_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_281_groups_0 = const()[name = tensor<string, []>("x_281_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_281_strides_0 = const()[name = tensor<string, []>("x_281_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_281_pad_0 = const()[name = tensor<string, []>("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_281_dilations_0 = const()[name = tensor<string, []>("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134563264)))];
            tensor<fp16, [1, 512, 3]> x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16, x = new_x_43_cast_fp16)[name = tensor<string, []>("x_281_cast_fp16")];
            tensor<int32, [3]> input_575_perm_0 = const()[name = tensor<string, []>("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_283_axes_0 = const()[name = tensor<string, []>("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134572544)))];
            tensor<fp16, [512]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134573632)))];
            tensor<fp16, [1, 3, 512]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = tensor<string, []>("transpose_142")];
            tensor<fp16, [1, 3, 512]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = tensor<string, []>("x_283_cast_fp16")];
            tensor<int32, [3]> input_577_perm_0 = const()[name = tensor<string, []>("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = tensor<string, []>("transpose_141")];
            tensor<fp16, [1, 512, 3]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
            tensor<string, []> x_285_pad_type_0 = const()[name = tensor<string, []>("x_285_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_285_strides_0 = const()[name = tensor<string, []>("x_285_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_285_pad_0 = const()[name = tensor<string, []>("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_285_dilations_0 = const()[name = tensor<string, []>("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_285_groups_0 = const()[name = tensor<string, []>("x_285_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134574720)))];
            tensor<fp16, [1, 512, 3]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16, x = input_579_cast_fp16)[name = tensor<string, []>("x_285_cast_fp16")];
            tensor<int32, [3]> input_581_perm_0 = const()[name = tensor<string, []>("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = tensor<string, []>("transpose_140")];
            tensor<fp16, [1, 3, 512]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = tensor<string, []>("input_583_cast_fp16")];
            tensor<int32, [1]> input_585_axes_0 = const()[name = tensor<string, []>("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135099072)))];
            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135100160)))];
            tensor<fp16, [1, 3, 512]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = tensor<string, []>("input_585_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135101248)))];
            tensor<fp16, [1, 3, 2048]> linear_98_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16, x = input_585_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = tensor<string, []>("input_589_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137198464)))];
            tensor<fp16, [1, 3, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16, x = input_589_cast_fp16)[name = tensor<string, []>("linear_99_cast_fp16")];
            tensor<fp16, []> var_2545_to_fp16 = const()[name = tensor<string, []>("op_2545_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2546_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2545_to_fp16)[name = tensor<string, []>("op_2546_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2546_cast_fp16)[name = tensor<string, []>("input_595_cast_fp16")];
            tensor<int32, [1]> input_597_axes_0 = const()[name = tensor<string, []>("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139295680)))];
            tensor<fp16, [512]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139296768)))];
            tensor<fp16, [1, 3, 512]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
            tensor<int32, [4]> cache_45_begin_0 = const()[name = tensor<string, []>("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> cache_45_end_0 = const()[name = tensor<string, []>("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 70, 512])];
            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = tensor<string, []>("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = tensor<string, []>("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_45_cast_fp16")];
            tensor<int32, [4]> cache_47_begin_0 = const()[name = tensor<string, []>("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> cache_47_end_0 = const()[name = tensor<string, []>("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 512, 8])];
            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = tensor<string, []>("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = tensor<string, []>("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_47_cast_fp16")];
            tensor<int32, [1]> input_599_axes_0 = const()[name = tensor<string, []>("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139297856)))];
            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139298944)))];
            tensor<fp16, [1, 3, 512]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = tensor<string, []>("input_599_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139300032)))];
            tensor<fp16, [1, 3, 2048]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16, x = input_599_cast_fp16)[name = tensor<string, []>("linear_100_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor<string, []>("input_603_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141397248)))];
            tensor<fp16, [1, 3, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16, x = input_603_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")];
            tensor<fp16, []> var_2580_to_fp16 = const()[name = tensor<string, []>("op_2580_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2581_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2580_to_fp16)[name = tensor<string, []>("op_2581_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2581_cast_fp16)[name = tensor<string, []>("input_609_cast_fp16")];
            tensor<int32, [1]> key_23_axes_0 = const()[name = tensor<string, []>("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143494464)))];
            tensor<fp16, [512]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143495552)))];
            tensor<fp16, [1, 3, 512]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = tensor<string, []>("key_23_cast_fp16")];
            tensor<bool, []> input_611_interleave_0 = const()[name = tensor<string, []>("input_611_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_611_cast_fp16 = concat(axis = var_64, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = tensor<string, []>("input_611_cast_fp16")];
            tensor<int32, [3]> var_2603_begin_0 = const()[name = tensor<string, []>("op_2603_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2603_end_0 = const()[name = tensor<string, []>("op_2603_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_2603_end_mask_0 = const()[name = tensor<string, []>("op_2603_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_2603_cast_fp16 = slice_by_index(begin = var_2603_begin_0, end = var_2603_end_0, end_mask = var_2603_end_mask_0, x = cache_45_cast_fp16)[name = tensor<string, []>("op_2603_cast_fp16")];
            tensor<int32, [3]> var_2606_begin_0 = const()[name = tensor<string, []>("op_2606_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2606_end_0 = const()[name = tensor<string, []>("op_2606_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_2606_end_mask_0 = const()[name = tensor<string, []>("op_2606_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_2606_cast_fp16 = slice_by_index(begin = var_2606_begin_0, end = var_2606_end_0, end_mask = var_2606_end_mask_0, x = key_23_cast_fp16)[name = tensor<string, []>("op_2606_cast_fp16")];
            tensor<bool, []> var_2609_interleave_0 = const()[name = tensor<string, []>("op_2609_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_2609_cast_fp16 = concat(axis = var_64, interleave = var_2609_interleave_0, values = (var_2603_cast_fp16, var_2606_cast_fp16))[name = tensor<string, []>("op_2609_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143496640)))];
            tensor<fp16, [1, 3, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16, x = key_23_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")];
            tensor<int32, [4]> var_2613 = const()[name = tensor<string, []>("op_2613"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_67_cast_fp16 = reshape(shape = var_2613, x = linear_102_cast_fp16)[name = tensor<string, []>("q_67_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144020992)))];
            tensor<fp16, [1, 73, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")];
            tensor<int32, [4]> var_2617 = const()[name = tensor<string, []>("op_2617"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_45_cast_fp16 = reshape(shape = var_2617, x = linear_103_cast_fp16)[name = tensor<string, []>("k_45_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144545344)))];
            tensor<fp16, [1, 73, 512]> linear_104_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")];
            tensor<int32, [4]> var_2621 = const()[name = tensor<string, []>("op_2621"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_23_cast_fp16 = reshape(shape = var_2621, x = linear_104_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")];
            tensor<int32, [4]> value_25_perm_0 = const()[name = tensor<string, []>("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145069696)))];
            tensor<fp16, [1, 3, 8, 64]> var_2633_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2633_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145070784)))];
            tensor<fp16, [1, 3, 8, 64]> var_2635_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2635_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_293_transpose_x_0 = const()[name = tensor<string, []>("x_293_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_293_transpose_y_0 = const()[name = tensor<string, []>("x_293_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_2637_to_fp16 = const()[name = tensor<string, []>("op_2637_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145071872)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2635_cast_fp16)[name = tensor<string, []>("transpose_138")];
            tensor<fp16, [1, 8, 3, 145]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = var_2637_to_fp16)[name = tensor<string, []>("x_293_cast_fp16")];
            tensor<int32, [8]> x_295_pad_0 = const()[name = tensor<string, []>("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_295_mode_0 = const()[name = tensor<string, []>("x_295_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_166_to_fp16 = const()[name = tensor<string, []>("const_166_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_295_cast_fp16 = pad(constant_val = const_166_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = tensor<string, []>("x_295_cast_fp16")];
            tensor<int32, [4]> var_2645 = const()[name = tensor<string, []>("op_2645"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_297_cast_fp16 = reshape(shape = var_2645, x = x_295_cast_fp16)[name = tensor<string, []>("x_297_cast_fp16")];
            tensor<int32, [4]> var_2649_begin_0 = const()[name = tensor<string, []>("op_2649_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2649_end_0 = const()[name = tensor<string, []>("op_2649_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_2649_end_mask_0 = const()[name = tensor<string, []>("op_2649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_2649_cast_fp16 = slice_by_index(begin = var_2649_begin_0, end = var_2649_end_0, end_mask = var_2649_end_mask_0, x = x_297_cast_fp16)[name = tensor<string, []>("op_2649_cast_fp16")];
            tensor<int32, [4]> var_2650 = const()[name = tensor<string, []>("op_2650"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_45_cast_fp16 = reshape(shape = var_2650, x = var_2649_cast_fp16)[name = tensor<string, []>("matrix_bd_45_cast_fp16")];
            tensor<bool, []> matrix_ac_23_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_23_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_73_perm_0 = const()[name = tensor<string, []>("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = tensor<string, []>("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_74 = transpose(perm = transpose_74_perm_0, x = k_45_cast_fp16)[name = tensor<string, []>("transpose_136")];
            tensor<fp16, [1, 8, 3, 64]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_2633_cast_fp16)[name = tensor<string, []>("transpose_137")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor<string, []>("matrix_ac_23_cast_fp16")];
            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = tensor<string, []>("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = tensor<string, []>("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = tensor<string, []>("matrix_bd_47_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2659_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = tensor<string, []>("op_2659_cast_fp16")];
            tensor<fp16, []> _inversed_scores_45_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_45_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_45_cast_fp16 = mul(x = var_2659_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_45_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_47_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_47_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2665_cast_fp16 = softmax(axis = var_62, x = scores_47_cast_fp16)[name = tensor<string, []>("op_2665_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_613_cast_fp16 = select(a = var_40_to_fp16, b = var_2665_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_613_cast_fp16")];
            tensor<bool, []> x_299_transpose_x_0 = const()[name = tensor<string, []>("x_299_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_299_transpose_y_0 = const()[name = tensor<string, []>("x_299_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_23_cast_fp16)[name = tensor<string, []>("transpose_139")];
            tensor<fp16, [1, 8, 3, 64]> x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_25_cast_fp16)[name = tensor<string, []>("x_299_cast_fp16")];
            tensor<int32, [4]> var_2669_perm_0 = const()[name = tensor<string, []>("op_2669_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2670 = const()[name = tensor<string, []>("op_2670"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_2669_cast_fp16 = transpose(perm = var_2669_perm_0, x = x_299_cast_fp16)[name = tensor<string, []>("transpose_135")];
            tensor<fp16, [1, 3, 512]> input_615_cast_fp16 = reshape(shape = var_2670, x = var_2669_cast_fp16)[name = tensor<string, []>("input_615_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145220416)))];
            tensor<fp16, [1, 3, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16, x = input_615_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = tensor<string, []>("input_619_cast_fp16")];
            tensor<int32, [1]> x_303_axes_0 = const()[name = tensor<string, []>("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145744768)))];
            tensor<fp16, [512]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145745856)))];
            tensor<fp16, [1, 3, 512]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = tensor<string, []>("x_303_cast_fp16")];
            tensor<int32, [3]> input_621_perm_0 = const()[name = tensor<string, []>("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_623_pad_type_0 = const()[name = tensor<string, []>("input_623_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_623_strides_0 = const()[name = tensor<string, []>("input_623_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_623_pad_0 = const()[name = tensor<string, []>("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_623_dilations_0 = const()[name = tensor<string, []>("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_623_groups_0 = const()[name = tensor<string, []>("input_623_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145746944)))];
            tensor<fp16, [1, 512, 3]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = tensor<string, []>("transpose_134")];
            tensor<fp16, [1, 1024, 3]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16, x = input_621_cast_fp16)[name = tensor<string, []>("input_623_cast_fp16")];
            tensor<int32, []> x_305_split_num_splits_0 = const()[name = tensor<string, []>("x_305_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_305_split_axis_0 = const()[name = tensor<string, []>("x_305_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_305_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = tensor<string, []>("x_305_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = tensor<string, []>("x_305_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_305_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_625_cast_fp16 = select(a = var_40_to_fp16, b = x_305_cast_fp16, cond = var_418)[name = tensor<string, []>("input_625_cast_fp16")];
            tensor<bool, []> new_x_47_interleave_0 = const()[name = tensor<string, []>("new_x_47_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_47_cast_fp16 = concat(axis = var_62, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = tensor<string, []>("new_x_47_cast_fp16")];
            tensor<int32, [3]> next_cache_23_begin_0 = const()[name = tensor<string, []>("next_cache_23_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_23_end_0 = const()[name = tensor<string, []>("next_cache_23_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_23_end_mask_0 = const()[name = tensor<string, []>("next_cache_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_23_cast_fp16 = slice_by_index(begin = next_cache_23_begin_0, end = next_cache_23_end_0, end_mask = next_cache_23_end_mask_0, x = new_x_47_cast_fp16)[name = tensor<string, []>("next_cache_23_cast_fp16")];
            tensor<int32, [3]> var_2711_begin_0 = const()[name = tensor<string, []>("op_2711_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_2711_end_0 = const()[name = tensor<string, []>("op_2711_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_2711_end_mask_0 = const()[name = tensor<string, []>("op_2711_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2711_cast_fp16 = slice_by_index(begin = var_2711_begin_0, end = var_2711_end_0, end_mask = var_2711_end_mask_0, x = next_cache_23_cast_fp16)[name = tensor<string, []>("op_2711_cast_fp16")];
            tensor<string, []> x_307_pad_type_0 = const()[name = tensor<string, []>("x_307_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_307_groups_0 = const()[name = tensor<string, []>("x_307_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_307_strides_0 = const()[name = tensor<string, []>("x_307_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_307_pad_0 = const()[name = tensor<string, []>("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_307_dilations_0 = const()[name = tensor<string, []>("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146795584)))];
            tensor<fp16, [1, 512, 3]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16, x = new_x_47_cast_fp16)[name = tensor<string, []>("x_307_cast_fp16")];
            tensor<int32, [3]> input_627_perm_0 = const()[name = tensor<string, []>("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_309_axes_0 = const()[name = tensor<string, []>("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146804864)))];
            tensor<fp16, [512]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146805952)))];
            tensor<fp16, [1, 3, 512]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = tensor<string, []>("transpose_133")];
            tensor<fp16, [1, 3, 512]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = tensor<string, []>("x_309_cast_fp16")];
            tensor<int32, [3]> input_629_perm_0 = const()[name = tensor<string, []>("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = tensor<string, []>("transpose_132")];
            tensor<fp16, [1, 512, 3]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = tensor<string, []>("input_631_cast_fp16")];
            tensor<string, []> x_311_pad_type_0 = const()[name = tensor<string, []>("x_311_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_311_strides_0 = const()[name = tensor<string, []>("x_311_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_311_pad_0 = const()[name = tensor<string, []>("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_311_dilations_0 = const()[name = tensor<string, []>("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_311_groups_0 = const()[name = tensor<string, []>("x_311_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146807040)))];
            tensor<fp16, [1, 512, 3]> x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16, x = input_631_cast_fp16)[name = tensor<string, []>("x_311_cast_fp16")];
            tensor<int32, [3]> input_633_perm_0 = const()[name = tensor<string, []>("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = tensor<string, []>("transpose_131")];
            tensor<fp16, [1, 3, 512]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
            tensor<int32, [1]> input_637_axes_0 = const()[name = tensor<string, []>("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147331392)))];
            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147332480)))];
            tensor<fp16, [1, 3, 512]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = tensor<string, []>("input_637_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147333568)))];
            tensor<fp16, [1, 3, 2048]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16, x = input_637_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = tensor<string, []>("input_641_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149430784)))];
            tensor<fp16, [1, 3, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16, x = input_641_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")];
            tensor<fp16, []> var_2752_to_fp16 = const()[name = tensor<string, []>("op_2752_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2753_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2752_to_fp16)[name = tensor<string, []>("op_2753_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2753_cast_fp16)[name = tensor<string, []>("input_647_cast_fp16")];
            tensor<int32, [1]> input_649_axes_0 = const()[name = tensor<string, []>("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151528000)))];
            tensor<fp16, [512]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151529088)))];
            tensor<fp16, [1, 3, 512]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = tensor<string, []>("input_649_cast_fp16")];
            tensor<int32, [4]> cache_49_begin_0 = const()[name = tensor<string, []>("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> cache_49_end_0 = const()[name = tensor<string, []>("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 70, 512])];
            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = tensor<string, []>("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = tensor<string, []>("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_49_cast_fp16")];
            tensor<int32, [4]> cache_51_begin_0 = const()[name = tensor<string, []>("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> cache_51_end_0 = const()[name = tensor<string, []>("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 512, 8])];
            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = tensor<string, []>("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = tensor<string, []>("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_51_cast_fp16")];
            tensor<int32, [1]> input_651_axes_0 = const()[name = tensor<string, []>("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151530176)))];
            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151531264)))];
            tensor<fp16, [1, 3, 512]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = tensor<string, []>("input_651_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151532352)))];
            tensor<fp16, [1, 3, 2048]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16, x = input_651_cast_fp16)[name = tensor<string, []>("linear_109_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = tensor<string, []>("input_655_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153629568)))];
            tensor<fp16, [1, 3, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16, x = input_655_cast_fp16)[name = tensor<string, []>("linear_110_cast_fp16")];
            tensor<fp16, []> var_2787_to_fp16 = const()[name = tensor<string, []>("op_2787_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2788_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_2787_to_fp16)[name = tensor<string, []>("op_2788_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_2788_cast_fp16)[name = tensor<string, []>("input_661_cast_fp16")];
            tensor<int32, [1]> key_25_axes_0 = const()[name = tensor<string, []>("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155726784)))];
            tensor<fp16, [512]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155727872)))];
            tensor<fp16, [1, 3, 512]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
            tensor<bool, []> input_663_interleave_0 = const()[name = tensor<string, []>("input_663_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_663_cast_fp16 = concat(axis = var_64, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = tensor<string, []>("input_663_cast_fp16")];
            tensor<int32, [3]> var_2810_begin_0 = const()[name = tensor<string, []>("op_2810_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2810_end_0 = const()[name = tensor<string, []>("op_2810_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_2810_end_mask_0 = const()[name = tensor<string, []>("op_2810_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_2810_cast_fp16 = slice_by_index(begin = var_2810_begin_0, end = var_2810_end_0, end_mask = var_2810_end_mask_0, x = cache_49_cast_fp16)[name = tensor<string, []>("op_2810_cast_fp16")];
            tensor<int32, [3]> var_2813_begin_0 = const()[name = tensor<string, []>("op_2813_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2813_end_0 = const()[name = tensor<string, []>("op_2813_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_2813_end_mask_0 = const()[name = tensor<string, []>("op_2813_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_2813_cast_fp16 = slice_by_index(begin = var_2813_begin_0, end = var_2813_end_0, end_mask = var_2813_end_mask_0, x = key_25_cast_fp16)[name = tensor<string, []>("op_2813_cast_fp16")];
            tensor<bool, []> var_2816_interleave_0 = const()[name = tensor<string, []>("op_2816_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_2816_cast_fp16 = concat(axis = var_64, interleave = var_2816_interleave_0, values = (var_2810_cast_fp16, var_2813_cast_fp16))[name = tensor<string, []>("op_2816_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155728960)))];
            tensor<fp16, [1, 3, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16, x = key_25_cast_fp16)[name = tensor<string, []>("linear_111_cast_fp16")];
            tensor<int32, [4]> var_2820 = const()[name = tensor<string, []>("op_2820"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_73_cast_fp16 = reshape(shape = var_2820, x = linear_111_cast_fp16)[name = tensor<string, []>("q_73_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156253312)))];
            tensor<fp16, [1, 73, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16, x = input_663_cast_fp16)[name = tensor<string, []>("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2824 = const()[name = tensor<string, []>("op_2824"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_49_cast_fp16 = reshape(shape = var_2824, x = linear_112_cast_fp16)[name = tensor<string, []>("k_49_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156777664)))];
            tensor<fp16, [1, 73, 512]> linear_113_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16, x = input_663_cast_fp16)[name = tensor<string, []>("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2828 = const()[name = tensor<string, []>("op_2828"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_25_cast_fp16 = reshape(shape = var_2828, x = linear_113_cast_fp16)[name = tensor<string, []>("v_25_cast_fp16")];
            tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<string, []>("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157302016)))];
            tensor<fp16, [1, 3, 8, 64]> var_2840_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2840_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157303104)))];
            tensor<fp16, [1, 3, 8, 64]> var_2842_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2842_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_319_transpose_x_0 = const()[name = tensor<string, []>("x_319_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_319_transpose_y_0 = const()[name = tensor<string, []>("x_319_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_2844_to_fp16 = const()[name = tensor<string, []>("op_2844_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157304192)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_2842_cast_fp16)[name = tensor<string, []>("transpose_129")];
            tensor<fp16, [1, 8, 3, 145]> x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = var_2844_to_fp16)[name = tensor<string, []>("x_319_cast_fp16")];
            tensor<int32, [8]> x_321_pad_0 = const()[name = tensor<string, []>("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_321_mode_0 = const()[name = tensor<string, []>("x_321_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_179_to_fp16 = const()[name = tensor<string, []>("const_179_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_321_cast_fp16 = pad(constant_val = const_179_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = tensor<string, []>("x_321_cast_fp16")];
            tensor<int32, [4]> var_2852 = const()[name = tensor<string, []>("op_2852"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_323_cast_fp16 = reshape(shape = var_2852, x = x_321_cast_fp16)[name = tensor<string, []>("x_323_cast_fp16")];
            tensor<int32, [4]> var_2856_begin_0 = const()[name = tensor<string, []>("op_2856_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2856_end_0 = const()[name = tensor<string, []>("op_2856_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_2856_end_mask_0 = const()[name = tensor<string, []>("op_2856_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_2856_cast_fp16 = slice_by_index(begin = var_2856_begin_0, end = var_2856_end_0, end_mask = var_2856_end_mask_0, x = x_323_cast_fp16)[name = tensor<string, []>("op_2856_cast_fp16")];
            tensor<int32, [4]> var_2857 = const()[name = tensor<string, []>("op_2857"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_49_cast_fp16 = reshape(shape = var_2857, x = var_2856_cast_fp16)[name = tensor<string, []>("matrix_bd_49_cast_fp16")];
            tensor<bool, []> matrix_ac_25_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_25_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_75_perm_0 = const()[name = tensor<string, []>("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = tensor<string, []>("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_76 = transpose(perm = transpose_76_perm_0, x = k_49_cast_fp16)[name = tensor<string, []>("transpose_127")];
            tensor<fp16, [1, 8, 3, 64]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_2840_cast_fp16)[name = tensor<string, []>("transpose_128")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor<string, []>("matrix_ac_25_cast_fp16")];
            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = tensor<string, []>("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = tensor<string, []>("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = tensor<string, []>("matrix_bd_51_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2866_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = tensor<string, []>("op_2866_cast_fp16")];
            tensor<fp16, []> _inversed_scores_49_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_49_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_49_cast_fp16 = mul(x = var_2866_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_49_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_51_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_51_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_2872_cast_fp16 = softmax(axis = var_62, x = scores_51_cast_fp16)[name = tensor<string, []>("op_2872_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_665_cast_fp16 = select(a = var_40_to_fp16, b = var_2872_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_665_cast_fp16")];
            tensor<bool, []> x_325_transpose_x_0 = const()[name = tensor<string, []>("x_325_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_325_transpose_y_0 = const()[name = tensor<string, []>("x_325_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_25_cast_fp16)[name = tensor<string, []>("transpose_130")];
            tensor<fp16, [1, 8, 3, 64]> x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_27_cast_fp16)[name = tensor<string, []>("x_325_cast_fp16")];
            tensor<int32, [4]> var_2876_perm_0 = const()[name = tensor<string, []>("op_2876_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2877 = const()[name = tensor<string, []>("op_2877"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_2876_cast_fp16 = transpose(perm = var_2876_perm_0, x = x_325_cast_fp16)[name = tensor<string, []>("transpose_126")];
            tensor<fp16, [1, 3, 512]> input_667_cast_fp16 = reshape(shape = var_2877, x = var_2876_cast_fp16)[name = tensor<string, []>("input_667_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157452736)))];
            tensor<fp16, [1, 3, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16, x = input_667_cast_fp16)[name = tensor<string, []>("linear_115_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = tensor<string, []>("input_671_cast_fp16")];
            tensor<int32, [1]> x_329_axes_0 = const()[name = tensor<string, []>("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157977088)))];
            tensor<fp16, [512]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157978176)))];
            tensor<fp16, [1, 3, 512]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
            tensor<int32, [3]> input_673_perm_0 = const()[name = tensor<string, []>("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_675_pad_type_0 = const()[name = tensor<string, []>("input_675_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_675_strides_0 = const()[name = tensor<string, []>("input_675_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_675_pad_0 = const()[name = tensor<string, []>("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_675_dilations_0 = const()[name = tensor<string, []>("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_675_groups_0 = const()[name = tensor<string, []>("input_675_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157979264)))];
            tensor<fp16, [1, 512, 3]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = tensor<string, []>("transpose_125")];
            tensor<fp16, [1, 1024, 3]> input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16, x = input_673_cast_fp16)[name = tensor<string, []>("input_675_cast_fp16")];
            tensor<int32, []> x_331_split_num_splits_0 = const()[name = tensor<string, []>("x_331_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_331_split_axis_0 = const()[name = tensor<string, []>("x_331_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_331_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = tensor<string, []>("x_331_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = tensor<string, []>("x_331_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_331_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_677_cast_fp16 = select(a = var_40_to_fp16, b = x_331_cast_fp16, cond = var_418)[name = tensor<string, []>("input_677_cast_fp16")];
            tensor<bool, []> new_x_51_interleave_0 = const()[name = tensor<string, []>("new_x_51_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_51_cast_fp16 = concat(axis = var_62, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = tensor<string, []>("new_x_51_cast_fp16")];
            tensor<int32, [3]> next_cache_25_begin_0 = const()[name = tensor<string, []>("next_cache_25_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_25_end_0 = const()[name = tensor<string, []>("next_cache_25_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_25_end_mask_0 = const()[name = tensor<string, []>("next_cache_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_25_cast_fp16 = slice_by_index(begin = next_cache_25_begin_0, end = next_cache_25_end_0, end_mask = next_cache_25_end_mask_0, x = new_x_51_cast_fp16)[name = tensor<string, []>("next_cache_25_cast_fp16")];
            tensor<int32, [3]> var_2918_begin_0 = const()[name = tensor<string, []>("op_2918_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_2918_end_0 = const()[name = tensor<string, []>("op_2918_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_2918_end_mask_0 = const()[name = tensor<string, []>("op_2918_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2918_cast_fp16 = slice_by_index(begin = var_2918_begin_0, end = var_2918_end_0, end_mask = var_2918_end_mask_0, x = next_cache_25_cast_fp16)[name = tensor<string, []>("op_2918_cast_fp16")];
            tensor<string, []> x_333_pad_type_0 = const()[name = tensor<string, []>("x_333_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_333_groups_0 = const()[name = tensor<string, []>("x_333_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_333_strides_0 = const()[name = tensor<string, []>("x_333_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_333_pad_0 = const()[name = tensor<string, []>("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_333_dilations_0 = const()[name = tensor<string, []>("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159027904)))];
            tensor<fp16, [1, 512, 3]> x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16, x = new_x_51_cast_fp16)[name = tensor<string, []>("x_333_cast_fp16")];
            tensor<int32, [3]> input_679_perm_0 = const()[name = tensor<string, []>("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_335_axes_0 = const()[name = tensor<string, []>("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159037184)))];
            tensor<fp16, [512]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159038272)))];
            tensor<fp16, [1, 3, 512]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = tensor<string, []>("transpose_124")];
            tensor<fp16, [1, 3, 512]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = tensor<string, []>("x_335_cast_fp16")];
            tensor<int32, [3]> input_681_perm_0 = const()[name = tensor<string, []>("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = tensor<string, []>("transpose_123")];
            tensor<fp16, [1, 512, 3]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = tensor<string, []>("input_683_cast_fp16")];
            tensor<string, []> x_337_pad_type_0 = const()[name = tensor<string, []>("x_337_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_337_strides_0 = const()[name = tensor<string, []>("x_337_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_337_pad_0 = const()[name = tensor<string, []>("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_337_dilations_0 = const()[name = tensor<string, []>("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_337_groups_0 = const()[name = tensor<string, []>("x_337_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159039360)))];
            tensor<fp16, [1, 512, 3]> x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16, x = input_683_cast_fp16)[name = tensor<string, []>("x_337_cast_fp16")];
            tensor<int32, [3]> input_685_perm_0 = const()[name = tensor<string, []>("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = tensor<string, []>("transpose_122")];
            tensor<fp16, [1, 3, 512]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = tensor<string, []>("input_687_cast_fp16")];
            tensor<int32, [1]> input_689_axes_0 = const()[name = tensor<string, []>("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159563712)))];
            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159564800)))];
            tensor<fp16, [1, 3, 512]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = tensor<string, []>("input_689_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159565888)))];
            tensor<fp16, [1, 3, 2048]> linear_116_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16, x = input_689_cast_fp16)[name = tensor<string, []>("linear_116_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = tensor<string, []>("input_693_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161663104)))];
            tensor<fp16, [1, 3, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16, x = input_693_cast_fp16)[name = tensor<string, []>("linear_117_cast_fp16")];
            tensor<fp16, []> var_2959_to_fp16 = const()[name = tensor<string, []>("op_2959_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2960_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_2959_to_fp16)[name = tensor<string, []>("op_2960_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_2960_cast_fp16)[name = tensor<string, []>("input_699_cast_fp16")];
            tensor<int32, [1]> input_701_axes_0 = const()[name = tensor<string, []>("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163760320)))];
            tensor<fp16, [512]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163761408)))];
            tensor<fp16, [1, 3, 512]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = tensor<string, []>("input_701_cast_fp16")];
            tensor<int32, [4]> cache_53_begin_0 = const()[name = tensor<string, []>("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> cache_53_end_0 = const()[name = tensor<string, []>("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 70, 512])];
            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = tensor<string, []>("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = tensor<string, []>("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_53_cast_fp16")];
            tensor<int32, [4]> cache_55_begin_0 = const()[name = tensor<string, []>("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> cache_55_end_0 = const()[name = tensor<string, []>("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 512, 8])];
            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = tensor<string, []>("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = tensor<string, []>("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_55_cast_fp16")];
            tensor<int32, [1]> input_703_axes_0 = const()[name = tensor<string, []>("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163762496)))];
            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163763584)))];
            tensor<fp16, [1, 3, 512]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = tensor<string, []>("input_703_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163764672)))];
            tensor<fp16, [1, 3, 2048]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16, x = input_703_cast_fp16)[name = tensor<string, []>("linear_118_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = tensor<string, []>("input_707_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165861888)))];
            tensor<fp16, [1, 3, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16, x = input_707_cast_fp16)[name = tensor<string, []>("linear_119_cast_fp16")];
            tensor<fp16, []> var_2994_to_fp16 = const()[name = tensor<string, []>("op_2994_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_2995_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_2994_to_fp16)[name = tensor<string, []>("op_2995_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_2995_cast_fp16)[name = tensor<string, []>("input_713_cast_fp16")];
            tensor<int32, [1]> key_27_axes_0 = const()[name = tensor<string, []>("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167959104)))];
            tensor<fp16, [512]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167960192)))];
            tensor<fp16, [1, 3, 512]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = tensor<string, []>("key_27_cast_fp16")];
            tensor<bool, []> input_715_interleave_0 = const()[name = tensor<string, []>("input_715_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_715_cast_fp16 = concat(axis = var_64, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = tensor<string, []>("input_715_cast_fp16")];
            tensor<int32, [3]> var_3017_begin_0 = const()[name = tensor<string, []>("op_3017_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3017_end_0 = const()[name = tensor<string, []>("op_3017_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_3017_end_mask_0 = const()[name = tensor<string, []>("op_3017_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_3017_cast_fp16 = slice_by_index(begin = var_3017_begin_0, end = var_3017_end_0, end_mask = var_3017_end_mask_0, x = cache_53_cast_fp16)[name = tensor<string, []>("op_3017_cast_fp16")];
            tensor<int32, [3]> var_3020_begin_0 = const()[name = tensor<string, []>("op_3020_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3020_end_0 = const()[name = tensor<string, []>("op_3020_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_3020_end_mask_0 = const()[name = tensor<string, []>("op_3020_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_3020_cast_fp16 = slice_by_index(begin = var_3020_begin_0, end = var_3020_end_0, end_mask = var_3020_end_mask_0, x = key_27_cast_fp16)[name = tensor<string, []>("op_3020_cast_fp16")];
            tensor<bool, []> var_3023_interleave_0 = const()[name = tensor<string, []>("op_3023_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_3023_cast_fp16 = concat(axis = var_64, interleave = var_3023_interleave_0, values = (var_3017_cast_fp16, var_3020_cast_fp16))[name = tensor<string, []>("op_3023_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167961280)))];
            tensor<fp16, [1, 3, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16, x = key_27_cast_fp16)[name = tensor<string, []>("linear_120_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = tensor<string, []>("op_3027"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_79_cast_fp16 = reshape(shape = var_3027, x = linear_120_cast_fp16)[name = tensor<string, []>("q_79_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168485632)))];
            tensor<fp16, [1, 73, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16, x = input_715_cast_fp16)[name = tensor<string, []>("linear_121_cast_fp16")];
            tensor<int32, [4]> var_3031 = const()[name = tensor<string, []>("op_3031"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_53_cast_fp16 = reshape(shape = var_3031, x = linear_121_cast_fp16)[name = tensor<string, []>("k_53_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169009984)))];
            tensor<fp16, [1, 73, 512]> linear_122_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16, x = input_715_cast_fp16)[name = tensor<string, []>("linear_122_cast_fp16")];
            tensor<int32, [4]> var_3035 = const()[name = tensor<string, []>("op_3035"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_27_cast_fp16 = reshape(shape = var_3035, x = linear_122_cast_fp16)[name = tensor<string, []>("v_27_cast_fp16")];
            tensor<int32, [4]> value_29_perm_0 = const()[name = tensor<string, []>("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169534336)))];
            tensor<fp16, [1, 3, 8, 64]> var_3047_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3047_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169535424)))];
            tensor<fp16, [1, 3, 8, 64]> var_3049_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3049_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_345_transpose_x_0 = const()[name = tensor<string, []>("x_345_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_345_transpose_y_0 = const()[name = tensor<string, []>("x_345_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_3051_to_fp16 = const()[name = tensor<string, []>("op_3051_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169536512)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3049_cast_fp16)[name = tensor<string, []>("transpose_120")];
            tensor<fp16, [1, 8, 3, 145]> x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = var_3051_to_fp16)[name = tensor<string, []>("x_345_cast_fp16")];
            tensor<int32, [8]> x_347_pad_0 = const()[name = tensor<string, []>("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_347_mode_0 = const()[name = tensor<string, []>("x_347_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_192_to_fp16 = const()[name = tensor<string, []>("const_192_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_347_cast_fp16 = pad(constant_val = const_192_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = tensor<string, []>("x_347_cast_fp16")];
            tensor<int32, [4]> var_3059 = const()[name = tensor<string, []>("op_3059"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_349_cast_fp16 = reshape(shape = var_3059, x = x_347_cast_fp16)[name = tensor<string, []>("x_349_cast_fp16")];
            tensor<int32, [4]> var_3063_begin_0 = const()[name = tensor<string, []>("op_3063_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3063_end_0 = const()[name = tensor<string, []>("op_3063_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_3063_end_mask_0 = const()[name = tensor<string, []>("op_3063_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_3063_cast_fp16 = slice_by_index(begin = var_3063_begin_0, end = var_3063_end_0, end_mask = var_3063_end_mask_0, x = x_349_cast_fp16)[name = tensor<string, []>("op_3063_cast_fp16")];
            tensor<int32, [4]> var_3064 = const()[name = tensor<string, []>("op_3064"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_53_cast_fp16 = reshape(shape = var_3064, x = var_3063_cast_fp16)[name = tensor<string, []>("matrix_bd_53_cast_fp16")];
            tensor<bool, []> matrix_ac_27_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_27_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_77_perm_0 = const()[name = tensor<string, []>("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = tensor<string, []>("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_78 = transpose(perm = transpose_78_perm_0, x = k_53_cast_fp16)[name = tensor<string, []>("transpose_118")];
            tensor<fp16, [1, 8, 3, 64]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_3047_cast_fp16)[name = tensor<string, []>("transpose_119")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor<string, []>("matrix_ac_27_cast_fp16")];
            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = tensor<string, []>("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = tensor<string, []>("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = tensor<string, []>("matrix_bd_55_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_3073_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = tensor<string, []>("op_3073_cast_fp16")];
            tensor<fp16, []> _inversed_scores_53_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_53_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_53_cast_fp16 = mul(x = var_3073_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_53_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_55_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_55_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_3079_cast_fp16 = softmax(axis = var_62, x = scores_55_cast_fp16)[name = tensor<string, []>("op_3079_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_717_cast_fp16 = select(a = var_40_to_fp16, b = var_3079_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_717_cast_fp16")];
            tensor<bool, []> x_351_transpose_x_0 = const()[name = tensor<string, []>("x_351_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_351_transpose_y_0 = const()[name = tensor<string, []>("x_351_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_27_cast_fp16)[name = tensor<string, []>("transpose_121")];
            tensor<fp16, [1, 8, 3, 64]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_29_cast_fp16)[name = tensor<string, []>("x_351_cast_fp16")];
            tensor<int32, [4]> var_3083_perm_0 = const()[name = tensor<string, []>("op_3083_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3084 = const()[name = tensor<string, []>("op_3084"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_3083_cast_fp16 = transpose(perm = var_3083_perm_0, x = x_351_cast_fp16)[name = tensor<string, []>("transpose_117")];
            tensor<fp16, [1, 3, 512]> input_719_cast_fp16 = reshape(shape = var_3084, x = var_3083_cast_fp16)[name = tensor<string, []>("input_719_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169685056)))];
            tensor<fp16, [1, 3, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16, x = input_719_cast_fp16)[name = tensor<string, []>("linear_124_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = tensor<string, []>("input_723_cast_fp16")];
            tensor<int32, [1]> x_355_axes_0 = const()[name = tensor<string, []>("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170209408)))];
            tensor<fp16, [512]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170210496)))];
            tensor<fp16, [1, 3, 512]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = tensor<string, []>("x_355_cast_fp16")];
            tensor<int32, [3]> input_725_perm_0 = const()[name = tensor<string, []>("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_727_pad_type_0 = const()[name = tensor<string, []>("input_727_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_727_strides_0 = const()[name = tensor<string, []>("input_727_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_727_pad_0 = const()[name = tensor<string, []>("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_727_dilations_0 = const()[name = tensor<string, []>("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_727_groups_0 = const()[name = tensor<string, []>("input_727_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170211584)))];
            tensor<fp16, [1, 512, 3]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = tensor<string, []>("transpose_116")];
            tensor<fp16, [1, 1024, 3]> input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16, x = input_725_cast_fp16)[name = tensor<string, []>("input_727_cast_fp16")];
            tensor<int32, []> x_357_split_num_splits_0 = const()[name = tensor<string, []>("x_357_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_357_split_axis_0 = const()[name = tensor<string, []>("x_357_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_357_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = tensor<string, []>("x_357_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = tensor<string, []>("x_357_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_357_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_729_cast_fp16 = select(a = var_40_to_fp16, b = x_357_cast_fp16, cond = var_418)[name = tensor<string, []>("input_729_cast_fp16")];
            tensor<bool, []> new_x_55_interleave_0 = const()[name = tensor<string, []>("new_x_55_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_55_cast_fp16 = concat(axis = var_62, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = tensor<string, []>("new_x_55_cast_fp16")];
            tensor<int32, [3]> next_cache_27_begin_0 = const()[name = tensor<string, []>("next_cache_27_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_27_end_0 = const()[name = tensor<string, []>("next_cache_27_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_27_end_mask_0 = const()[name = tensor<string, []>("next_cache_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_27_cast_fp16 = slice_by_index(begin = next_cache_27_begin_0, end = next_cache_27_end_0, end_mask = next_cache_27_end_mask_0, x = new_x_55_cast_fp16)[name = tensor<string, []>("next_cache_27_cast_fp16")];
            tensor<int32, [3]> var_3125_begin_0 = const()[name = tensor<string, []>("op_3125_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_3125_end_0 = const()[name = tensor<string, []>("op_3125_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_3125_end_mask_0 = const()[name = tensor<string, []>("op_3125_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_3125_cast_fp16 = slice_by_index(begin = var_3125_begin_0, end = var_3125_end_0, end_mask = var_3125_end_mask_0, x = next_cache_27_cast_fp16)[name = tensor<string, []>("op_3125_cast_fp16")];
            tensor<string, []> x_359_pad_type_0 = const()[name = tensor<string, []>("x_359_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_359_groups_0 = const()[name = tensor<string, []>("x_359_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_359_strides_0 = const()[name = tensor<string, []>("x_359_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_359_pad_0 = const()[name = tensor<string, []>("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_359_dilations_0 = const()[name = tensor<string, []>("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171260224)))];
            tensor<fp16, [1, 512, 3]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16, x = new_x_55_cast_fp16)[name = tensor<string, []>("x_359_cast_fp16")];
            tensor<int32, [3]> input_731_perm_0 = const()[name = tensor<string, []>("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_361_axes_0 = const()[name = tensor<string, []>("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171269504)))];
            tensor<fp16, [512]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171270592)))];
            tensor<fp16, [1, 3, 512]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = tensor<string, []>("transpose_115")];
            tensor<fp16, [1, 3, 512]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = tensor<string, []>("x_361_cast_fp16")];
            tensor<int32, [3]> input_733_perm_0 = const()[name = tensor<string, []>("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = tensor<string, []>("transpose_114")];
            tensor<fp16, [1, 512, 3]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = tensor<string, []>("input_735_cast_fp16")];
            tensor<string, []> x_363_pad_type_0 = const()[name = tensor<string, []>("x_363_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_363_strides_0 = const()[name = tensor<string, []>("x_363_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_363_pad_0 = const()[name = tensor<string, []>("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_363_dilations_0 = const()[name = tensor<string, []>("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_363_groups_0 = const()[name = tensor<string, []>("x_363_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171271680)))];
            tensor<fp16, [1, 512, 3]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16, x = input_735_cast_fp16)[name = tensor<string, []>("x_363_cast_fp16")];
            tensor<int32, [3]> input_737_perm_0 = const()[name = tensor<string, []>("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = tensor<string, []>("transpose_113")];
            tensor<fp16, [1, 3, 512]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = tensor<string, []>("input_739_cast_fp16")];
            tensor<int32, [1]> input_741_axes_0 = const()[name = tensor<string, []>("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171796032)))];
            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171797120)))];
            tensor<fp16, [1, 3, 512]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = tensor<string, []>("input_741_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171798208)))];
            tensor<fp16, [1, 3, 2048]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16, x = input_741_cast_fp16)[name = tensor<string, []>("linear_125_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = tensor<string, []>("input_745_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173895424)))];
            tensor<fp16, [1, 3, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16, x = input_745_cast_fp16)[name = tensor<string, []>("linear_126_cast_fp16")];
            tensor<fp16, []> var_3166_to_fp16 = const()[name = tensor<string, []>("op_3166_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_3167_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3166_to_fp16)[name = tensor<string, []>("op_3167_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3167_cast_fp16)[name = tensor<string, []>("input_751_cast_fp16")];
            tensor<int32, [1]> input_753_axes_0 = const()[name = tensor<string, []>("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175992640)))];
            tensor<fp16, [512]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175993728)))];
            tensor<fp16, [1, 3, 512]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = tensor<string, []>("input_753_cast_fp16")];
            tensor<int32, [4]> cache_57_begin_0 = const()[name = tensor<string, []>("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> cache_57_end_0 = const()[name = tensor<string, []>("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 70, 512])];
            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = tensor<string, []>("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = tensor<string, []>("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_57_cast_fp16")];
            tensor<int32, [4]> cache_59_begin_0 = const()[name = tensor<string, []>("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> cache_59_end_0 = const()[name = tensor<string, []>("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 512, 8])];
            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = tensor<string, []>("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = tensor<string, []>("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_59_cast_fp16")];
            tensor<int32, [1]> input_755_axes_0 = const()[name = tensor<string, []>("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175994816)))];
            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175995904)))];
            tensor<fp16, [1, 3, 512]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = tensor<string, []>("input_755_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175996992)))];
            tensor<fp16, [1, 3, 2048]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16, x = input_755_cast_fp16)[name = tensor<string, []>("linear_127_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = tensor<string, []>("input_759_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178094208)))];
            tensor<fp16, [1, 3, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16, x = input_759_cast_fp16)[name = tensor<string, []>("linear_128_cast_fp16")];
            tensor<fp16, []> var_3201_to_fp16 = const()[name = tensor<string, []>("op_3201_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_3202_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3201_to_fp16)[name = tensor<string, []>("op_3202_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3202_cast_fp16)[name = tensor<string, []>("input_765_cast_fp16")];
            tensor<int32, [1]> key_29_axes_0 = const()[name = tensor<string, []>("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180191424)))];
            tensor<fp16, [512]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180192512)))];
            tensor<fp16, [1, 3, 512]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
            tensor<bool, []> input_767_interleave_0 = const()[name = tensor<string, []>("input_767_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_767_cast_fp16 = concat(axis = var_64, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = tensor<string, []>("input_767_cast_fp16")];
            tensor<int32, [3]> var_3224_begin_0 = const()[name = tensor<string, []>("op_3224_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3224_end_0 = const()[name = tensor<string, []>("op_3224_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_3224_end_mask_0 = const()[name = tensor<string, []>("op_3224_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_3224_cast_fp16 = slice_by_index(begin = var_3224_begin_0, end = var_3224_end_0, end_mask = var_3224_end_mask_0, x = cache_57_cast_fp16)[name = tensor<string, []>("op_3224_cast_fp16")];
            tensor<int32, [3]> var_3227_begin_0 = const()[name = tensor<string, []>("op_3227_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3227_end_0 = const()[name = tensor<string, []>("op_3227_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_3227_end_mask_0 = const()[name = tensor<string, []>("op_3227_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_3227_cast_fp16 = slice_by_index(begin = var_3227_begin_0, end = var_3227_end_0, end_mask = var_3227_end_mask_0, x = key_29_cast_fp16)[name = tensor<string, []>("op_3227_cast_fp16")];
            tensor<bool, []> var_3230_interleave_0 = const()[name = tensor<string, []>("op_3230_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_3230_cast_fp16 = concat(axis = var_64, interleave = var_3230_interleave_0, values = (var_3224_cast_fp16, var_3227_cast_fp16))[name = tensor<string, []>("op_3230_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180193600)))];
            tensor<fp16, [1, 3, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16, x = key_29_cast_fp16)[name = tensor<string, []>("linear_129_cast_fp16")];
            tensor<int32, [4]> var_3234 = const()[name = tensor<string, []>("op_3234"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_85_cast_fp16 = reshape(shape = var_3234, x = linear_129_cast_fp16)[name = tensor<string, []>("q_85_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180717952)))];
            tensor<fp16, [1, 73, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16, x = input_767_cast_fp16)[name = tensor<string, []>("linear_130_cast_fp16")];
            tensor<int32, [4]> var_3238 = const()[name = tensor<string, []>("op_3238"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_57_cast_fp16 = reshape(shape = var_3238, x = linear_130_cast_fp16)[name = tensor<string, []>("k_57_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181242304)))];
            tensor<fp16, [1, 73, 512]> linear_131_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16, x = input_767_cast_fp16)[name = tensor<string, []>("linear_131_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = tensor<string, []>("op_3242"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_29_cast_fp16 = reshape(shape = var_3242, x = linear_131_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")];
            tensor<int32, [4]> value_31_perm_0 = const()[name = tensor<string, []>("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181766656)))];
            tensor<fp16, [1, 3, 8, 64]> var_3254_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3254_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181767744)))];
            tensor<fp16, [1, 3, 8, 64]> var_3256_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3256_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_371_transpose_x_0 = const()[name = tensor<string, []>("x_371_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_371_transpose_y_0 = const()[name = tensor<string, []>("x_371_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_3258_to_fp16 = const()[name = tensor<string, []>("op_3258_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181768832)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3256_cast_fp16)[name = tensor<string, []>("transpose_111")];
            tensor<fp16, [1, 8, 3, 145]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = var_3258_to_fp16)[name = tensor<string, []>("x_371_cast_fp16")];
            tensor<int32, [8]> x_373_pad_0 = const()[name = tensor<string, []>("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_373_mode_0 = const()[name = tensor<string, []>("x_373_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_205_to_fp16 = const()[name = tensor<string, []>("const_205_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_373_cast_fp16 = pad(constant_val = const_205_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = tensor<string, []>("x_373_cast_fp16")];
            tensor<int32, [4]> var_3266 = const()[name = tensor<string, []>("op_3266"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_375_cast_fp16 = reshape(shape = var_3266, x = x_373_cast_fp16)[name = tensor<string, []>("x_375_cast_fp16")];
            tensor<int32, [4]> var_3270_begin_0 = const()[name = tensor<string, []>("op_3270_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3270_end_0 = const()[name = tensor<string, []>("op_3270_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_3270_end_mask_0 = const()[name = tensor<string, []>("op_3270_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_3270_cast_fp16 = slice_by_index(begin = var_3270_begin_0, end = var_3270_end_0, end_mask = var_3270_end_mask_0, x = x_375_cast_fp16)[name = tensor<string, []>("op_3270_cast_fp16")];
            tensor<int32, [4]> var_3271 = const()[name = tensor<string, []>("op_3271"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_57_cast_fp16 = reshape(shape = var_3271, x = var_3270_cast_fp16)[name = tensor<string, []>("matrix_bd_57_cast_fp16")];
            tensor<bool, []> matrix_ac_29_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_29_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_79_perm_0 = const()[name = tensor<string, []>("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = tensor<string, []>("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_80 = transpose(perm = transpose_80_perm_0, x = k_57_cast_fp16)[name = tensor<string, []>("transpose_109")];
            tensor<fp16, [1, 8, 3, 64]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_3254_cast_fp16)[name = tensor<string, []>("transpose_110")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor<string, []>("matrix_ac_29_cast_fp16")];
            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = tensor<string, []>("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = tensor<string, []>("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = tensor<string, []>("matrix_bd_59_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_3280_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = tensor<string, []>("op_3280_cast_fp16")];
            tensor<fp16, []> _inversed_scores_57_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_57_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_57_cast_fp16 = mul(x = var_3280_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_57_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_59_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_59_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_3286_cast_fp16 = softmax(axis = var_62, x = scores_59_cast_fp16)[name = tensor<string, []>("op_3286_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_769_cast_fp16 = select(a = var_40_to_fp16, b = var_3286_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_769_cast_fp16")];
            tensor<bool, []> x_377_transpose_x_0 = const()[name = tensor<string, []>("x_377_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_377_transpose_y_0 = const()[name = tensor<string, []>("x_377_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_29_cast_fp16)[name = tensor<string, []>("transpose_112")];
            tensor<fp16, [1, 8, 3, 64]> x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_31_cast_fp16)[name = tensor<string, []>("x_377_cast_fp16")];
            tensor<int32, [4]> var_3290_perm_0 = const()[name = tensor<string, []>("op_3290_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3291 = const()[name = tensor<string, []>("op_3291"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_3290_cast_fp16 = transpose(perm = var_3290_perm_0, x = x_377_cast_fp16)[name = tensor<string, []>("transpose_108")];
            tensor<fp16, [1, 3, 512]> input_771_cast_fp16 = reshape(shape = var_3291, x = var_3290_cast_fp16)[name = tensor<string, []>("input_771_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181917376)))];
            tensor<fp16, [1, 3, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16, x = input_771_cast_fp16)[name = tensor<string, []>("linear_133_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = tensor<string, []>("input_775_cast_fp16")];
            tensor<int32, [1]> x_381_axes_0 = const()[name = tensor<string, []>("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182441728)))];
            tensor<fp16, [512]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182442816)))];
            tensor<fp16, [1, 3, 512]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = tensor<string, []>("x_381_cast_fp16")];
            tensor<int32, [3]> input_777_perm_0 = const()[name = tensor<string, []>("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_779_pad_type_0 = const()[name = tensor<string, []>("input_779_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_779_strides_0 = const()[name = tensor<string, []>("input_779_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_779_pad_0 = const()[name = tensor<string, []>("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_779_dilations_0 = const()[name = tensor<string, []>("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_779_groups_0 = const()[name = tensor<string, []>("input_779_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182443904)))];
            tensor<fp16, [1, 512, 3]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = tensor<string, []>("transpose_107")];
            tensor<fp16, [1, 1024, 3]> input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16, x = input_777_cast_fp16)[name = tensor<string, []>("input_779_cast_fp16")];
            tensor<int32, []> x_383_split_num_splits_0 = const()[name = tensor<string, []>("x_383_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_383_split_axis_0 = const()[name = tensor<string, []>("x_383_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_383_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = tensor<string, []>("x_383_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = tensor<string, []>("x_383_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_383_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_781_cast_fp16 = select(a = var_40_to_fp16, b = x_383_cast_fp16, cond = var_418)[name = tensor<string, []>("input_781_cast_fp16")];
            tensor<bool, []> new_x_59_interleave_0 = const()[name = tensor<string, []>("new_x_59_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_59_cast_fp16 = concat(axis = var_62, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = tensor<string, []>("new_x_59_cast_fp16")];
            tensor<int32, [3]> next_cache_29_begin_0 = const()[name = tensor<string, []>("next_cache_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_29_end_0 = const()[name = tensor<string, []>("next_cache_29_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_29_end_mask_0 = const()[name = tensor<string, []>("next_cache_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_29_cast_fp16 = slice_by_index(begin = next_cache_29_begin_0, end = next_cache_29_end_0, end_mask = next_cache_29_end_mask_0, x = new_x_59_cast_fp16)[name = tensor<string, []>("next_cache_29_cast_fp16")];
            tensor<int32, [3]> var_3332_begin_0 = const()[name = tensor<string, []>("op_3332_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_3332_end_0 = const()[name = tensor<string, []>("op_3332_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_3332_end_mask_0 = const()[name = tensor<string, []>("op_3332_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_3332_cast_fp16 = slice_by_index(begin = var_3332_begin_0, end = var_3332_end_0, end_mask = var_3332_end_mask_0, x = next_cache_29_cast_fp16)[name = tensor<string, []>("op_3332_cast_fp16")];
            tensor<string, []> x_385_pad_type_0 = const()[name = tensor<string, []>("x_385_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_385_groups_0 = const()[name = tensor<string, []>("x_385_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_385_strides_0 = const()[name = tensor<string, []>("x_385_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_385_pad_0 = const()[name = tensor<string, []>("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_385_dilations_0 = const()[name = tensor<string, []>("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183492544)))];
            tensor<fp16, [1, 512, 3]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16, x = new_x_59_cast_fp16)[name = tensor<string, []>("x_385_cast_fp16")];
            tensor<int32, [3]> input_783_perm_0 = const()[name = tensor<string, []>("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_387_axes_0 = const()[name = tensor<string, []>("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183501824)))];
            tensor<fp16, [512]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183502912)))];
            tensor<fp16, [1, 3, 512]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = tensor<string, []>("transpose_106")];
            tensor<fp16, [1, 3, 512]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = tensor<string, []>("x_387_cast_fp16")];
            tensor<int32, [3]> input_785_perm_0 = const()[name = tensor<string, []>("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = tensor<string, []>("transpose_105")];
            tensor<fp16, [1, 512, 3]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = tensor<string, []>("input_787_cast_fp16")];
            tensor<string, []> x_389_pad_type_0 = const()[name = tensor<string, []>("x_389_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_389_strides_0 = const()[name = tensor<string, []>("x_389_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_389_pad_0 = const()[name = tensor<string, []>("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_389_dilations_0 = const()[name = tensor<string, []>("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_389_groups_0 = const()[name = tensor<string, []>("x_389_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183504000)))];
            tensor<fp16, [1, 512, 3]> x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16, x = input_787_cast_fp16)[name = tensor<string, []>("x_389_cast_fp16")];
            tensor<int32, [3]> input_789_perm_0 = const()[name = tensor<string, []>("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = tensor<string, []>("transpose_104")];
            tensor<fp16, [1, 3, 512]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = tensor<string, []>("input_791_cast_fp16")];
            tensor<int32, [1]> input_793_axes_0 = const()[name = tensor<string, []>("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184028352)))];
            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184029440)))];
            tensor<fp16, [1, 3, 512]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = tensor<string, []>("input_793_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184030528)))];
            tensor<fp16, [1, 3, 2048]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16, x = input_793_cast_fp16)[name = tensor<string, []>("linear_134_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = tensor<string, []>("input_797_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186127744)))];
            tensor<fp16, [1, 3, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16, x = input_797_cast_fp16)[name = tensor<string, []>("linear_135_cast_fp16")];
            tensor<fp16, []> var_3373_to_fp16 = const()[name = tensor<string, []>("op_3373_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_3374_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3373_to_fp16)[name = tensor<string, []>("op_3374_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3374_cast_fp16)[name = tensor<string, []>("input_803_cast_fp16")];
            tensor<int32, [1]> input_805_axes_0 = const()[name = tensor<string, []>("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188224960)))];
            tensor<fp16, [512]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188226048)))];
            tensor<fp16, [1, 3, 512]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = tensor<string, []>("input_805_cast_fp16")];
            tensor<int32, [4]> cache_61_begin_0 = const()[name = tensor<string, []>("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> cache_61_end_0 = const()[name = tensor<string, []>("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 70, 512])];
            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = tensor<string, []>("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = tensor<string, []>("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_61_cast_fp16")];
            tensor<int32, [4]> cache_63_begin_0 = const()[name = tensor<string, []>("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> cache_63_end_0 = const()[name = tensor<string, []>("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 512, 8])];
            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = tensor<string, []>("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = tensor<string, []>("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_63_cast_fp16")];
            tensor<int32, [1]> input_807_axes_0 = const()[name = tensor<string, []>("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188227136)))];
            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188228224)))];
            tensor<fp16, [1, 3, 512]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = tensor<string, []>("input_807_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188229312)))];
            tensor<fp16, [1, 3, 2048]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16, x = input_807_cast_fp16)[name = tensor<string, []>("linear_136_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = tensor<string, []>("input_811_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190326528)))];
            tensor<fp16, [1, 3, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16, x = input_811_cast_fp16)[name = tensor<string, []>("linear_137_cast_fp16")];
            tensor<fp16, []> var_3408_to_fp16 = const()[name = tensor<string, []>("op_3408_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_3409_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3408_to_fp16)[name = tensor<string, []>("op_3409_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3409_cast_fp16)[name = tensor<string, []>("input_817_cast_fp16")];
            tensor<int32, [1]> key_31_axes_0 = const()[name = tensor<string, []>("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192423744)))];
            tensor<fp16, [512]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192424832)))];
            tensor<fp16, [1, 3, 512]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = tensor<string, []>("key_31_cast_fp16")];
            tensor<bool, []> input_819_interleave_0 = const()[name = tensor<string, []>("input_819_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_819_cast_fp16 = concat(axis = var_64, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = tensor<string, []>("input_819_cast_fp16")];
            tensor<int32, [3]> var_3431_begin_0 = const()[name = tensor<string, []>("op_3431_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3431_end_0 = const()[name = tensor<string, []>("op_3431_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_3431_end_mask_0 = const()[name = tensor<string, []>("op_3431_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_3431_cast_fp16 = slice_by_index(begin = var_3431_begin_0, end = var_3431_end_0, end_mask = var_3431_end_mask_0, x = cache_61_cast_fp16)[name = tensor<string, []>("op_3431_cast_fp16")];
            tensor<int32, [3]> var_3434_begin_0 = const()[name = tensor<string, []>("op_3434_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3434_end_0 = const()[name = tensor<string, []>("op_3434_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_3434_end_mask_0 = const()[name = tensor<string, []>("op_3434_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_3434_cast_fp16 = slice_by_index(begin = var_3434_begin_0, end = var_3434_end_0, end_mask = var_3434_end_mask_0, x = key_31_cast_fp16)[name = tensor<string, []>("op_3434_cast_fp16")];
            tensor<bool, []> var_3437_interleave_0 = const()[name = tensor<string, []>("op_3437_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_3437_cast_fp16 = concat(axis = var_64, interleave = var_3437_interleave_0, values = (var_3431_cast_fp16, var_3434_cast_fp16))[name = tensor<string, []>("op_3437_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192425920)))];
            tensor<fp16, [1, 3, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16, x = key_31_cast_fp16)[name = tensor<string, []>("linear_138_cast_fp16")];
            tensor<int32, [4]> var_3441 = const()[name = tensor<string, []>("op_3441"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_91_cast_fp16 = reshape(shape = var_3441, x = linear_138_cast_fp16)[name = tensor<string, []>("q_91_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192950272)))];
            tensor<fp16, [1, 73, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16, x = input_819_cast_fp16)[name = tensor<string, []>("linear_139_cast_fp16")];
            tensor<int32, [4]> var_3445 = const()[name = tensor<string, []>("op_3445"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_61_cast_fp16 = reshape(shape = var_3445, x = linear_139_cast_fp16)[name = tensor<string, []>("k_61_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193474624)))];
            tensor<fp16, [1, 73, 512]> linear_140_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16, x = input_819_cast_fp16)[name = tensor<string, []>("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3449 = const()[name = tensor<string, []>("op_3449"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_31_cast_fp16 = reshape(shape = var_3449, x = linear_140_cast_fp16)[name = tensor<string, []>("v_31_cast_fp16")];
            tensor<int32, [4]> value_33_perm_0 = const()[name = tensor<string, []>("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193998976)))];
            tensor<fp16, [1, 3, 8, 64]> var_3461_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3461_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194000064)))];
            tensor<fp16, [1, 3, 8, 64]> var_3463_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3463_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_397_transpose_x_0 = const()[name = tensor<string, []>("x_397_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_397_transpose_y_0 = const()[name = tensor<string, []>("x_397_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_3465_to_fp16 = const()[name = tensor<string, []>("op_3465_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194001152)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3463_cast_fp16)[name = tensor<string, []>("transpose_102")];
            tensor<fp16, [1, 8, 3, 145]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = var_3465_to_fp16)[name = tensor<string, []>("x_397_cast_fp16")];
            tensor<int32, [8]> x_399_pad_0 = const()[name = tensor<string, []>("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_399_mode_0 = const()[name = tensor<string, []>("x_399_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_218_to_fp16 = const()[name = tensor<string, []>("const_218_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_399_cast_fp16 = pad(constant_val = const_218_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = tensor<string, []>("x_399_cast_fp16")];
            tensor<int32, [4]> var_3473 = const()[name = tensor<string, []>("op_3473"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_401_cast_fp16 = reshape(shape = var_3473, x = x_399_cast_fp16)[name = tensor<string, []>("x_401_cast_fp16")];
            tensor<int32, [4]> var_3477_begin_0 = const()[name = tensor<string, []>("op_3477_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3477_end_0 = const()[name = tensor<string, []>("op_3477_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_3477_end_mask_0 = const()[name = tensor<string, []>("op_3477_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_3477_cast_fp16 = slice_by_index(begin = var_3477_begin_0, end = var_3477_end_0, end_mask = var_3477_end_mask_0, x = x_401_cast_fp16)[name = tensor<string, []>("op_3477_cast_fp16")];
            tensor<int32, [4]> var_3478 = const()[name = tensor<string, []>("op_3478"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_61_cast_fp16 = reshape(shape = var_3478, x = var_3477_cast_fp16)[name = tensor<string, []>("matrix_bd_61_cast_fp16")];
            tensor<bool, []> matrix_ac_31_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_31_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_81_perm_0 = const()[name = tensor<string, []>("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = tensor<string, []>("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_82 = transpose(perm = transpose_82_perm_0, x = k_61_cast_fp16)[name = tensor<string, []>("transpose_100")];
            tensor<fp16, [1, 8, 3, 64]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_3461_cast_fp16)[name = tensor<string, []>("transpose_101")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor<string, []>("matrix_ac_31_cast_fp16")];
            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = tensor<string, []>("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = tensor<string, []>("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = tensor<string, []>("matrix_bd_63_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_3487_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = tensor<string, []>("op_3487_cast_fp16")];
            tensor<fp16, []> _inversed_scores_61_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_61_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_61_cast_fp16 = mul(x = var_3487_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_61_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_63_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_63_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_3493_cast_fp16 = softmax(axis = var_62, x = scores_63_cast_fp16)[name = tensor<string, []>("op_3493_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_821_cast_fp16 = select(a = var_40_to_fp16, b = var_3493_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_821_cast_fp16")];
            tensor<bool, []> x_403_transpose_x_0 = const()[name = tensor<string, []>("x_403_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_403_transpose_y_0 = const()[name = tensor<string, []>("x_403_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_31_cast_fp16)[name = tensor<string, []>("transpose_103")];
            tensor<fp16, [1, 8, 3, 64]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_33_cast_fp16)[name = tensor<string, []>("x_403_cast_fp16")];
            tensor<int32, [4]> var_3497_perm_0 = const()[name = tensor<string, []>("op_3497_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3498 = const()[name = tensor<string, []>("op_3498"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_3497_cast_fp16 = transpose(perm = var_3497_perm_0, x = x_403_cast_fp16)[name = tensor<string, []>("transpose_99")];
            tensor<fp16, [1, 3, 512]> input_823_cast_fp16 = reshape(shape = var_3498, x = var_3497_cast_fp16)[name = tensor<string, []>("input_823_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194149696)))];
            tensor<fp16, [1, 3, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16, x = input_823_cast_fp16)[name = tensor<string, []>("linear_142_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = tensor<string, []>("input_827_cast_fp16")];
            tensor<int32, [1]> x_407_axes_0 = const()[name = tensor<string, []>("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194674048)))];
            tensor<fp16, [512]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194675136)))];
            tensor<fp16, [1, 3, 512]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = tensor<string, []>("x_407_cast_fp16")];
            tensor<int32, [3]> input_829_perm_0 = const()[name = tensor<string, []>("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_831_pad_type_0 = const()[name = tensor<string, []>("input_831_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_831_strides_0 = const()[name = tensor<string, []>("input_831_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_831_pad_0 = const()[name = tensor<string, []>("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_831_dilations_0 = const()[name = tensor<string, []>("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_831_groups_0 = const()[name = tensor<string, []>("input_831_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194676224)))];
            tensor<fp16, [1, 512, 3]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = tensor<string, []>("transpose_98")];
            tensor<fp16, [1, 1024, 3]> input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16, x = input_829_cast_fp16)[name = tensor<string, []>("input_831_cast_fp16")];
            tensor<int32, []> x_409_split_num_splits_0 = const()[name = tensor<string, []>("x_409_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_409_split_axis_0 = const()[name = tensor<string, []>("x_409_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_409_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = tensor<string, []>("x_409_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = tensor<string, []>("x_409_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_409_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_833_cast_fp16 = select(a = var_40_to_fp16, b = x_409_cast_fp16, cond = var_418)[name = tensor<string, []>("input_833_cast_fp16")];
            tensor<bool, []> new_x_63_interleave_0 = const()[name = tensor<string, []>("new_x_63_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_63_cast_fp16 = concat(axis = var_62, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = tensor<string, []>("new_x_63_cast_fp16")];
            tensor<int32, [3]> next_cache_31_begin_0 = const()[name = tensor<string, []>("next_cache_31_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_31_end_0 = const()[name = tensor<string, []>("next_cache_31_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_31_end_mask_0 = const()[name = tensor<string, []>("next_cache_31_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_31_cast_fp16 = slice_by_index(begin = next_cache_31_begin_0, end = next_cache_31_end_0, end_mask = next_cache_31_end_mask_0, x = new_x_63_cast_fp16)[name = tensor<string, []>("next_cache_31_cast_fp16")];
            tensor<int32, [3]> var_3539_begin_0 = const()[name = tensor<string, []>("op_3539_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> var_3539_end_0 = const()[name = tensor<string, []>("op_3539_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> var_3539_end_mask_0 = const()[name = tensor<string, []>("op_3539_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_3539_cast_fp16 = slice_by_index(begin = var_3539_begin_0, end = var_3539_end_0, end_mask = var_3539_end_mask_0, x = next_cache_31_cast_fp16)[name = tensor<string, []>("op_3539_cast_fp16")];
            tensor<string, []> x_411_pad_type_0 = const()[name = tensor<string, []>("x_411_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_411_groups_0 = const()[name = tensor<string, []>("x_411_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_411_strides_0 = const()[name = tensor<string, []>("x_411_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_411_pad_0 = const()[name = tensor<string, []>("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_411_dilations_0 = const()[name = tensor<string, []>("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195724864)))];
            tensor<fp16, [1, 512, 3]> x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16, x = new_x_63_cast_fp16)[name = tensor<string, []>("x_411_cast_fp16")];
            tensor<int32, [3]> input_835_perm_0 = const()[name = tensor<string, []>("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_413_axes_0 = const()[name = tensor<string, []>("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195734144)))];
            tensor<fp16, [512]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195735232)))];
            tensor<fp16, [1, 3, 512]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = tensor<string, []>("transpose_97")];
            tensor<fp16, [1, 3, 512]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = tensor<string, []>("x_413_cast_fp16")];
            tensor<int32, [3]> input_837_perm_0 = const()[name = tensor<string, []>("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = tensor<string, []>("transpose_96")];
            tensor<fp16, [1, 512, 3]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = tensor<string, []>("input_839_cast_fp16")];
            tensor<string, []> x_415_pad_type_0 = const()[name = tensor<string, []>("x_415_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_415_strides_0 = const()[name = tensor<string, []>("x_415_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_415_pad_0 = const()[name = tensor<string, []>("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_415_dilations_0 = const()[name = tensor<string, []>("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_415_groups_0 = const()[name = tensor<string, []>("x_415_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195736320)))];
            tensor<fp16, [1, 512, 3]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16, x = input_839_cast_fp16)[name = tensor<string, []>("x_415_cast_fp16")];
            tensor<int32, [3]> input_841_perm_0 = const()[name = tensor<string, []>("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = tensor<string, []>("transpose_95")];
            tensor<fp16, [1, 3, 512]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = tensor<string, []>("input_843_cast_fp16")];
            tensor<int32, [1]> input_845_axes_0 = const()[name = tensor<string, []>("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196260672)))];
            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196261760)))];
            tensor<fp16, [1, 3, 512]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = tensor<string, []>("input_845_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196262848)))];
            tensor<fp16, [1, 3, 2048]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16, x = input_845_cast_fp16)[name = tensor<string, []>("linear_143_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor<string, []>("input_849_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198360064)))];
            tensor<fp16, [1, 3, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16, x = input_849_cast_fp16)[name = tensor<string, []>("linear_144_cast_fp16")];
            tensor<fp16, []> var_3580_to_fp16 = const()[name = tensor<string, []>("op_3580_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_3581_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3580_to_fp16)[name = tensor<string, []>("op_3581_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3581_cast_fp16)[name = tensor<string, []>("input_855_cast_fp16")];
            tensor<int32, [1]> input_857_axes_0 = const()[name = tensor<string, []>("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200457280)))];
            tensor<fp16, [512]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200458368)))];
            tensor<fp16, [1, 3, 512]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = tensor<string, []>("input_857_cast_fp16")];
            tensor<int32, [4]> cache_65_begin_0 = const()[name = tensor<string, []>("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> cache_65_end_0 = const()[name = tensor<string, []>("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 70, 512])];
            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = tensor<string, []>("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = tensor<string, []>("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 70, 512]> cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_65_cast_fp16")];
            tensor<int32, [4]> cache_begin_0 = const()[name = tensor<string, []>("cache_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> cache_end_0 = const()[name = tensor<string, []>("cache_end_0"), val = tensor<int32, [4]>([17, 1, 512, 8])];
            tensor<bool, [4]> cache_end_mask_0 = const()[name = tensor<string, []>("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = tensor<string, []>("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 512, 8]> cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_cast_fp16")];
            tensor<int32, [1]> input_859_axes_0 = const()[name = tensor<string, []>("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200459456)))];
            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200460544)))];
            tensor<fp16, [1, 3, 512]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = tensor<string, []>("input_859_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200461632)))];
            tensor<fp16, [1, 3, 2048]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16, x = input_859_cast_fp16)[name = tensor<string, []>("linear_145_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = tensor<string, []>("input_863_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202558848)))];
            tensor<fp16, [1, 3, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16, x = input_863_cast_fp16)[name = tensor<string, []>("linear_146_cast_fp16")];
            tensor<fp16, []> var_3615_to_fp16 = const()[name = tensor<string, []>("op_3615_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_3616_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3615_to_fp16)[name = tensor<string, []>("op_3616_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3616_cast_fp16)[name = tensor<string, []>("input_869_cast_fp16")];
            tensor<int32, [1]> key_axes_0 = const()[name = tensor<string, []>("key_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204656064)))];
            tensor<fp16, [512]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204657152)))];
            tensor<fp16, [1, 3, 512]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
            tensor<bool, []> input_871_interleave_0 = const()[name = tensor<string, []>("input_871_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 73, 512]> input_871_cast_fp16 = concat(axis = var_64, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_cast_fp16))[name = tensor<string, []>("input_871_cast_fp16")];
            tensor<int32, [3]> var_3638_begin_0 = const()[name = tensor<string, []>("op_3638_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3638_end_0 = const()[name = tensor<string, []>("op_3638_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_3638_end_mask_0 = const()[name = tensor<string, []>("op_3638_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 69, 512]> var_3638_cast_fp16 = slice_by_index(begin = var_3638_begin_0, end = var_3638_end_0, end_mask = var_3638_end_mask_0, x = cache_65_cast_fp16)[name = tensor<string, []>("op_3638_cast_fp16")];
            tensor<int32, [3]> var_3641_begin_0 = const()[name = tensor<string, []>("op_3641_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3641_end_0 = const()[name = tensor<string, []>("op_3641_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
            tensor<bool, [3]> var_3641_end_mask_0 = const()[name = tensor<string, []>("op_3641_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, 512]> var_3641_cast_fp16 = slice_by_index(begin = var_3641_begin_0, end = var_3641_end_0, end_mask = var_3641_end_mask_0, x = key_cast_fp16)[name = tensor<string, []>("op_3641_cast_fp16")];
            tensor<bool, []> cache_last_channel_cur_interleave_0 = const()[name = tensor<string, []>("cache_last_channel_cur_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> cache_last_channel_cur_cast_fp16 = concat(axis = var_64, interleave = cache_last_channel_cur_interleave_0, values = (var_3638_cast_fp16, var_3641_cast_fp16))[name = tensor<string, []>("cache_last_channel_cur_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204658240)))];
            tensor<fp16, [1, 3, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = tensor<string, []>("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3648 = const()[name = tensor<string, []>("op_3648"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 3, 8, 64]> q_97_cast_fp16 = reshape(shape = var_3648, x = linear_147_cast_fp16)[name = tensor<string, []>("q_97_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205182592)))];
            tensor<fp16, [1, 73, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16, x = input_871_cast_fp16)[name = tensor<string, []>("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3652 = const()[name = tensor<string, []>("op_3652"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> k_65_cast_fp16 = reshape(shape = var_3652, x = linear_148_cast_fp16)[name = tensor<string, []>("k_65_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205706944)))];
            tensor<fp16, [1, 73, 512]> linear_149_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16, x = input_871_cast_fp16)[name = tensor<string, []>("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3656 = const()[name = tensor<string, []>("op_3656"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 73, 8, 64]> v_cast_fp16 = reshape(shape = var_3656, x = linear_149_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
            tensor<int32, [4]> value_perm_0 = const()[name = tensor<string, []>("value_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [8, 64]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206231296)))];
            tensor<fp16, [1, 3, 8, 64]> var_3668_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3668_cast_fp16")];
            tensor<fp16, [8, 64]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206232384)))];
            tensor<fp16, [1, 3, 8, 64]> var_3670_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3670_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> x_423_transpose_x_0 = const()[name = tensor<string, []>("x_423_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_423_transpose_y_0 = const()[name = tensor<string, []>("x_423_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 64, 145]> var_3672_to_fp16 = const()[name = tensor<string, []>("op_3672_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206233472)))];
            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_3670_cast_fp16)[name = tensor<string, []>("transpose_93")];
            tensor<fp16, [1, 8, 3, 145]> x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_cast_fp16, y = var_3672_to_fp16)[name = tensor<string, []>("x_423_cast_fp16")];
            tensor<int32, [8]> x_425_pad_0 = const()[name = tensor<string, []>("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            tensor<string, []> x_425_mode_0 = const()[name = tensor<string, []>("x_425_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_231_to_fp16 = const()[name = tensor<string, []>("const_231_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 3, 146]> x_425_cast_fp16 = pad(constant_val = const_231_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = tensor<string, []>("x_425_cast_fp16")];
            tensor<int32, [4]> var_3680 = const()[name = tensor<string, []>("op_3680"), val = tensor<int32, [4]>([1, 8, -1, 3])];
            tensor<fp16, [1, 8, 146, 3]> x_427_cast_fp16 = reshape(shape = var_3680, x = x_425_cast_fp16)[name = tensor<string, []>("x_427_cast_fp16")];
            tensor<int32, [4]> var_3684_begin_0 = const()[name = tensor<string, []>("op_3684_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3684_end_0 = const()[name = tensor<string, []>("op_3684_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
            tensor<bool, [4]> var_3684_end_mask_0 = const()[name = tensor<string, []>("op_3684_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 145, 3]> var_3684_cast_fp16 = slice_by_index(begin = var_3684_begin_0, end = var_3684_end_0, end_mask = var_3684_end_mask_0, x = x_427_cast_fp16)[name = tensor<string, []>("op_3684_cast_fp16")];
            tensor<int32, [4]> var_3685 = const()[name = tensor<string, []>("op_3685"), val = tensor<int32, [4]>([1, 8, 3, 145])];
            tensor<fp16, [1, 8, 3, 145]> matrix_bd_65_cast_fp16 = reshape(shape = var_3685, x = var_3684_cast_fp16)[name = tensor<string, []>("matrix_bd_65_cast_fp16")];
            tensor<bool, []> matrix_ac_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matrix_ac_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_83_perm_0 = const()[name = tensor<string, []>("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = tensor<string, []>("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 64, 73]> transpose_84 = transpose(perm = transpose_84_perm_0, x = k_65_cast_fp16)[name = tensor<string, []>("transpose_91")];
            tensor<fp16, [1, 8, 3, 64]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_3668_cast_fp16)[name = tensor<string, []>("transpose_92")];
            tensor<fp16, [1, 8, 3, 73]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor<string, []>("matrix_ac_cast_fp16")];
            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = tensor<string, []>("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd_end_0 = const()[name = tensor<string, []>("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 3, 73]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_65_cast_fp16)[name = tensor<string, []>("matrix_bd_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_3694_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = tensor<string, []>("op_3694_cast_fp16")];
            tensor<fp16, []> _inversed_scores_65_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_65_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_65_cast_fp16 = mul(x = var_3694_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_65_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> scores_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> var_3700_cast_fp16 = softmax(axis = var_62, x = scores_cast_fp16)[name = tensor<string, []>("op_3700_cast_fp16")];
            tensor<fp16, [1, 8, 3, 73]> input_873_cast_fp16 = select(a = var_40_to_fp16, b = var_3700_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_873_cast_fp16")];
            tensor<bool, []> x_429_transpose_x_0 = const()[name = tensor<string, []>("x_429_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x_429_transpose_y_0 = const()[name = tensor<string, []>("x_429_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 73, 64]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = tensor<string, []>("transpose_94")];
            tensor<fp16, [1, 8, 3, 64]> x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_cast_fp16)[name = tensor<string, []>("x_429_cast_fp16")];
            tensor<int32, [4]> var_3704_perm_0 = const()[name = tensor<string, []>("op_3704_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3705 = const()[name = tensor<string, []>("op_3705"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 3, 8, 64]> var_3704_cast_fp16 = transpose(perm = var_3704_perm_0, x = x_429_cast_fp16)[name = tensor<string, []>("transpose_90")];
            tensor<fp16, [1, 3, 512]> input_875_cast_fp16 = reshape(shape = var_3705, x = var_3704_cast_fp16)[name = tensor<string, []>("input_875_cast_fp16")];
            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206382016)))];
            tensor<fp16, [1, 3, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16, x = input_875_cast_fp16)[name = tensor<string, []>("linear_151_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = tensor<string, []>("input_879_cast_fp16")];
            tensor<int32, [1]> x_433_axes_0 = const()[name = tensor<string, []>("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206906368)))];
            tensor<fp16, [512]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206907456)))];
            tensor<fp16, [1, 3, 512]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = tensor<string, []>("x_433_cast_fp16")];
            tensor<int32, [3]> input_881_perm_0 = const()[name = tensor<string, []>("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_883_pad_type_0 = const()[name = tensor<string, []>("input_883_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_883_strides_0 = const()[name = tensor<string, []>("input_883_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_883_pad_0 = const()[name = tensor<string, []>("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_883_dilations_0 = const()[name = tensor<string, []>("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_883_groups_0 = const()[name = tensor<string, []>("input_883_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206908544)))];
            tensor<fp16, [1, 512, 3]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = tensor<string, []>("transpose_89")];
            tensor<fp16, [1, 1024, 3]> input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16, x = input_881_cast_fp16)[name = tensor<string, []>("input_883_cast_fp16")];
            tensor<int32, []> x_435_split_num_splits_0 = const()[name = tensor<string, []>("x_435_split_num_splits_0"), val = tensor<int32, []>(2)];
            tensor<int32, []> x_435_split_axis_0 = const()[name = tensor<string, []>("x_435_split_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 512, 3]> x_435_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = tensor<string, []>("x_435_split_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = tensor<string, []>("x_435_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 512, 3]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
            tensor<fp16, [1, 512, 3]> input_885_cast_fp16 = select(a = var_40_to_fp16, b = x_435_cast_fp16, cond = var_418)[name = tensor<string, []>("input_885_cast_fp16")];
            tensor<bool, []> new_x_interleave_0 = const()[name = tensor<string, []>("new_x_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 11]> new_x_cast_fp16 = concat(axis = var_62, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_885_cast_fp16))[name = tensor<string, []>("new_x_cast_fp16")];
            tensor<int32, [3]> next_cache_begin_0 = const()[name = tensor<string, []>("next_cache_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> next_cache_end_0 = const()[name = tensor<string, []>("next_cache_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> next_cache_end_mask_0 = const()[name = tensor<string, []>("next_cache_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 9]> next_cache_cast_fp16 = slice_by_index(begin = next_cache_begin_0, end = next_cache_end_0, end_mask = next_cache_end_mask_0, x = new_x_cast_fp16)[name = tensor<string, []>("next_cache_cast_fp16")];
            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = tensor<string, []>("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = next_cache_cast_fp16)[name = tensor<string, []>("cache_last_time_cur_cast_fp16")];
            tensor<string, []> x_437_pad_type_0 = const()[name = tensor<string, []>("x_437_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_437_groups_0 = const()[name = tensor<string, []>("x_437_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [1]> x_437_strides_0 = const()[name = tensor<string, []>("x_437_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_437_pad_0 = const()[name = tensor<string, []>("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_437_dilations_0 = const()[name = tensor<string, []>("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207957184)))];
            tensor<fp16, [1, 512, 3]> x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16, x = new_x_cast_fp16)[name = tensor<string, []>("x_437_cast_fp16")];
            tensor<int32, [3]> input_887_perm_0 = const()[name = tensor<string, []>("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_439_axes_0 = const()[name = tensor<string, []>("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207966464)))];
            tensor<fp16, [512]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207967552)))];
            tensor<fp16, [1, 3, 512]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = tensor<string, []>("transpose_88")];
            tensor<fp16, [1, 3, 512]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = tensor<string, []>("x_439_cast_fp16")];
            tensor<int32, [3]> input_889_perm_0 = const()[name = tensor<string, []>("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 3]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = tensor<string, []>("transpose_87")];
            tensor<fp16, [1, 512, 3]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = tensor<string, []>("input_891_cast_fp16")];
            tensor<string, []> x_441_pad_type_0 = const()[name = tensor<string, []>("x_441_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_441_strides_0 = const()[name = tensor<string, []>("x_441_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_441_pad_0 = const()[name = tensor<string, []>("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_441_dilations_0 = const()[name = tensor<string, []>("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_441_groups_0 = const()[name = tensor<string, []>("x_441_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207968640)))];
            tensor<fp16, [1, 512, 3]> x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16, x = input_891_cast_fp16)[name = tensor<string, []>("x_441_cast_fp16")];
            tensor<int32, [3]> input_893_perm_0 = const()[name = tensor<string, []>("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 3, 512]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = tensor<string, []>("transpose_86")];
            tensor<fp16, [1, 3, 512]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = tensor<string, []>("input_895_cast_fp16")];
            tensor<int32, [1]> input_897_axes_0 = const()[name = tensor<string, []>("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208492992)))];
            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208494080)))];
            tensor<fp16, [1, 3, 512]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = tensor<string, []>("input_897_cast_fp16")];
            tensor<fp16, [2048, 512]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208495168)))];
            tensor<fp16, [1, 3, 2048]> linear_152_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16, x = input_897_cast_fp16)[name = tensor<string, []>("linear_152_cast_fp16")];
            tensor<fp16, [1, 3, 2048]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor<string, []>("input_901_cast_fp16")];
            tensor<fp16, [512, 2048]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210592384)))];
            tensor<fp16, [1, 3, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16, x = input_901_cast_fp16)[name = tensor<string, []>("linear_153_cast_fp16")];
            tensor<fp16, []> var_3787_to_fp16 = const()[name = tensor<string, []>("op_3787_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 3, 512]> var_3788_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_3787_to_fp16)[name = tensor<string, []>("op_3788_cast_fp16")];
            tensor<fp16, [1, 3, 512]> input_cast_fp16 = add(x = input_895_cast_fp16, y = var_3788_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<int32, [1]> audio_signal_axes_0 = const()[name = tensor<string, []>("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [512]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212689600)))];
            tensor<fp16, [512]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212690688)))];
            tensor<fp16, [1, 3, 512]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("audio_signal_cast_fp16")];
            tensor<int32, [3]> obj_1_perm_0 = const()[name = tensor<string, []>("obj_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> cast_178_dtype_0 = const()[name = tensor<string, []>("cast_178_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> obj_5_axis_0 = const()[name = tensor<string, []>("obj_5_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [17, 1, 70, 512]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_332_cast_fp16, var_539_cast_fp16, var_746_cast_fp16, var_953_cast_fp16, var_1160_cast_fp16, var_1367_cast_fp16, var_1574_cast_fp16, var_1781_cast_fp16, var_1988_cast_fp16, var_2195_cast_fp16, var_2402_cast_fp16, var_2609_cast_fp16, var_2816_cast_fp16, var_3023_cast_fp16, var_3230_cast_fp16, var_3437_cast_fp16, cache_last_channel_cur_cast_fp16))[name = tensor<string, []>("obj_5_cast_fp16")];
            tensor<int32, []> obj_7_axis_0 = const()[name = tensor<string, []>("obj_7_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [17, 1, 512, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_434_cast_fp16, var_641_cast_fp16, var_848_cast_fp16, var_1055_cast_fp16, var_1262_cast_fp16, var_1469_cast_fp16, var_1676_cast_fp16, var_1883_cast_fp16, var_2090_cast_fp16, var_2297_cast_fp16, var_2504_cast_fp16, var_2711_cast_fp16, var_2918_cast_fp16, var_3125_cast_fp16, var_3332_cast_fp16, var_3539_cast_fp16, cache_last_time_cur_cast_fp16))[name = tensor<string, []>("obj_7_cast_fp16")];
            tensor<string, []> obj_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [1]> var_3804 = add(x = cache_last_channel_len, y = cache_keep_size)[name = tensor<string, []>("op_3804")];
            tensor<string, []> var_3804_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3804_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> const_237_to_fp16 = const()[name = tensor<string, []>("const_237_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, []> var_45_promoted_to_fp16 = const()[name = tensor<string, []>("op_45_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
            tensor<fp16, [1]> var_3804_to_fp16 = cast(dtype = var_3804_promoted_to_fp16_dtype_0, x = var_3804)[name = tensor<string, []>("cast_186")];
            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_237_to_fp16, beta = var_45_promoted_to_fp16, x = var_3804_to_fp16)[name = tensor<string, []>("clip_1_cast_fp16")];
            tensor<string, []> var_3831_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3831_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [3]> var_3846_begin_0 = const()[name = tensor<string, []>("op_3846_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3846_end_0 = const()[name = tensor<string, []>("op_3846_end_0"), val = tensor<int32, [3]>([1, 512, 2])];
            tensor<bool, [3]> var_3846_end_mask_0 = const()[name = tensor<string, []>("op_3846_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 512, 3]> obj_1_cast_fp16 = transpose(perm = obj_1_perm_0, x = audio_signal_cast_fp16)[name = tensor<string, []>("transpose_85")];
            tensor<fp16, [1, 512, 2]> var_3846_cast_fp16 = slice_by_index(begin = var_3846_begin_0, end = var_3846_end_0, end_mask = var_3846_end_mask_0, x = obj_1_cast_fp16)[name = tensor<string, []>("op_3846_cast_fp16")];
            tensor<string, []> var_3846_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3846_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<string, []> cast_178_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_178_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> const_238_to_fp16 = const()[name = tensor<string, []>("const_238_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, []> var_3848_promoted_to_fp16 = const()[name = tensor<string, []>("op_3848_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<int32, [1]> clip_0_cast_fp16_to_int32 = cast(dtype = cast_178_dtype_0, x = clip_0_cast_fp16)[name = tensor<string, []>("cast_188")];
            tensor<fp16, [1]> clip_0_cast_fp16_to_int32_to_fp16 = cast(dtype = cast_178_promoted_to_fp16_dtype_0, x = clip_0_cast_fp16_to_int32)[name = tensor<string, []>("cast_183")];
            tensor<fp16, [1]> clip_2_cast_fp16 = clip(alpha = const_238_to_fp16, beta = var_3848_promoted_to_fp16, x = clip_0_cast_fp16_to_int32_to_fp16)[name = tensor<string, []>("clip_2_cast_fp16")];
            tensor<string, []> cast_179_dtype_0 = const()[name = tensor<string, []>("cast_179_dtype_0"), val = tensor<string, []>("int32")];
            tensor<string, []> cast_180_dtype_0 = const()[name = tensor<string, []>("cast_180_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [1]> new_cache_last_channel_len = cast(dtype = cast_180_dtype_0, x = clip_1_cast_fp16)[name = tensor<string, []>("cast_181")];
            tensor<int32, [1]> encoded_length = cast(dtype = cast_179_dtype_0, x = clip_2_cast_fp16)[name = tensor<string, []>("cast_182")];
            tensor<fp32, [1, 512, 2]> encoded_output = cast(dtype = var_3846_cast_fp16_to_fp32_dtype_0, x = var_3846_cast_fp16)[name = tensor<string, []>("cast_184")];
            tensor<fp32, [17, 1, 70, 512]> new_cache_last_channel = cast(dtype = var_3831_cast_fp16_to_fp32_dtype_0, x = obj_5_cast_fp16)[name = tensor<string, []>("cast_185")];
            tensor<fp32, [17, 1, 512, 8]> new_cache_last_time = cast(dtype = obj_7_cast_fp16_to_fp32_dtype_0, x = obj_7_cast_fp16)[name = tensor<string, []>("cast_187")];
            tensor<fp32, [1, 128, 16]> new_pre_cache = cast(dtype = var_28_cast_fp16_to_fp32_dtype_0, x = var_28_cast_fp16)[name = tensor<string, []>("cast_194")];
        } -> (encoded_output, encoded_length, new_pre_cache, new_cache_last_channel, new_cache_last_time, new_cache_last_channel_len);
}