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, 64]> 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, 9]> 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, 64]> audio_signal_to_fp16 = cast(dtype = audio_signal_to_fp16_dtype_0, x = audio_signal)[name = tensor<string, []>("cast_199")];
            tensor<fp16, [1, 128, 9]> pre_cache_to_fp16 = cast(dtype = pre_cache_to_fp16_dtype_0, x = pre_cache)[name = tensor<string, []>("cast_200")];
            tensor<fp16, [1, 128, 73]> 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, []>(9)];
            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, 64])];
            tensor<int32, [3]> var_28_end_0 = const()[name = tensor<string, []>("op_28_end_0"), val = tensor<int32, [3]>([1, 128, 73])];
            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, 9]> 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_56 = const()[name = tensor<string, []>("op_56"), val = tensor<int32, []>(-1)];
            tensor<int32, []> var_65 = const()[name = tensor<string, []>("op_65"), 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<int32, [1]> tensor_1_axes_0 = const()[name = tensor<string, []>("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 73, 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, 73, 128]> tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = tensor<string, []>("tensor_1_cast_fp16")];
            tensor<int32, [1, 73]> expand_dims_0 = const()[name = tensor<string, []>("expand_dims_0"), 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_120_axes_0 = const()[name = tensor<string, []>("op_120_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_120 = expand_dims(axes = var_120_axes_0, x = value_1)[name = tensor<string, []>("op_120")];
            tensor<bool, [1, 73]> time_mask_1 = less(x = expand_dims_0, y = var_120)[name = tensor<string, []>("time_mask_1")];
            tensor<int32, [1]> var_122_axes_0 = const()[name = tensor<string, []>("op_122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 73, 1]> var_122 = expand_dims(axes = var_122_axes_0, x = time_mask_1)[name = tensor<string, []>("op_122")];
            tensor<int32, [3]> var_124_reps_0 = const()[name = tensor<string, []>("op_124_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
            tensor<bool, [1, 73, 128]> var_124 = tile(reps = var_124_reps_0, x = var_122)[name = tensor<string, []>("op_124")];
            tensor<int32, [1]> var_130_axes_0 = const()[name = tensor<string, []>("op_130_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<string, []> cast_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 73, 128]> var_124_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = var_124)[name = tensor<string, []>("cast_197")];
            tensor<fp16, [1, 1, 73, 128]> var_130_cast_fp16 = expand_dims(axes = var_130_axes_0, x = var_124_to_fp16)[name = tensor<string, []>("op_130_cast_fp16")];
            tensor<fp16, [1, 1, 73, 128]> input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_130_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_9_to_fp16 = const()[name = tensor<string, []>("const_9_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 1, 76, 131]> input_3_cast_fp16 = pad(constant_val = const_9_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, []> tensor_3_pad_type_0 = const()[name = tensor<string, []>("tensor_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> tensor_3_strides_0 = const()[name = tensor<string, []>("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [4]> tensor_3_pad_0 = const()[name = tensor<string, []>("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = tensor<string, []>("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> tensor_3_groups_0 = const()[name = tensor<string, []>("tensor_3_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, 37, 65]> tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("tensor_3_cast_fp16")];
            tensor<string, []> cast_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_143_promoted_to_fp16 = const()[name = tensor<string, []>("op_143_promoted_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_196")];
            tensor<fp16, [1]> var_144_cast_fp16 = add(x = value_1_to_fp16, y = var_143_promoted_to_fp16)[name = tensor<string, []>("op_144_cast_fp16")];
            tensor<fp16, []> var_145_promoted_to_fp16 = const()[name = tensor<string, []>("op_145_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1]> var_146_cast_fp16 = add(x = var_144_cast_fp16, y = var_145_promoted_to_fp16)[name = tensor<string, []>("op_146_cast_fp16")];
            tensor<fp16, []> var_147_promoted_to_fp16 = const()[name = tensor<string, []>("op_147_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
            tensor<fp16, [1]> var_148_cast_fp16 = sub(x = var_146_cast_fp16, y = var_147_promoted_to_fp16)[name = tensor<string, []>("op_148_cast_fp16")];
            tensor<fp16, []> var_53_promoted_to_fp16 = const()[name = tensor<string, []>("op_53_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_148_cast_fp16, y = var_53_promoted_to_fp16)[name = tensor<string, []>("floor_div_0_cast_fp16")];
            tensor<fp16, []> var_150_promoted_to_fp16 = const()[name = tensor<string, []>("op_150_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1]> current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_150_promoted_to_fp16)[name = tensor<string, []>("current_lengths_3_cast_fp16")];
            tensor<string, []> cast_3_dtype_0 = const()[name = tensor<string, []>("cast_3_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [1, 37]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<int32, [1, 37]>([[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]])];
            tensor<int32, [1]> var_159_axes_0 = const()[name = tensor<string, []>("op_159_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> current_lengths_3_cast_fp16_to_int32 = cast(dtype = cast_3_dtype_0, x = current_lengths_3_cast_fp16)[name = tensor<string, []>("cast_195")];
            tensor<int32, [1, 1]> var_159 = expand_dims(axes = var_159_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = tensor<string, []>("op_159")];
            tensor<bool, [1, 37]> time_mask_3 = less(x = expand_dims_1, y = var_159)[name = tensor<string, []>("time_mask_3")];
            tensor<int32, [1]> var_161_axes_0 = const()[name = tensor<string, []>("op_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 37, 1]> var_161 = expand_dims(axes = var_161_axes_0, x = time_mask_3)[name = tensor<string, []>("op_161")];
            tensor<int32, [3]> var_163_reps_0 = const()[name = tensor<string, []>("op_163_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
            tensor<bool, [1, 37, 65]> var_163 = tile(reps = var_163_reps_0, x = var_161)[name = tensor<string, []>("op_163")];
            tensor<int32, [1]> var_169_axes_0 = const()[name = tensor<string, []>("op_169_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<string, []> cast_4_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_4_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 37, 65]> var_163_to_fp16 = cast(dtype = cast_4_to_fp16_dtype_0, x = var_163)[name = tensor<string, []>("cast_194")];
            tensor<fp16, [1, 1, 37, 65]> var_169_cast_fp16 = expand_dims(axes = var_169_axes_0, x = var_163_to_fp16)[name = tensor<string, []>("op_169_cast_fp16")];
            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = tensor<string, []>("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
            tensor<fp16, [1, 256, 37, 65]> expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_169_cast_fp16)[name = tensor<string, []>("expanded_mask_3_cast_fp16")];
            tensor<fp16, [1, 256, 37, 65]> input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<fp16, [1, 256, 37, 65]> tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("tensor_5_cast_fp16")];
            tensor<fp16, [1, 256, 37, 65]> input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_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_23_to_fp16 = const()[name = tensor<string, []>("const_23_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 256, 40, 68]> input_9_cast_fp16 = pad(constant_val = const_23_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, []> tensor_7_pad_type_0 = const()[name = tensor<string, []>("tensor_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> tensor_7_strides_0 = const()[name = tensor<string, []>("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, []> tensor_7_groups_0 = const()[name = tensor<string, []>("tensor_7_groups_0"), val = tensor<int32, []>(256)];
            tensor<int32, [4]> tensor_7_pad_0 = const()[name = tensor<string, []>("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = tensor<string, []>("tensor_7_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, 19, 33]> tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("tensor_7_cast_fp16")];
            tensor<fp16, []> var_191_promoted_to_fp16 = const()[name = tensor<string, []>("op_191_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [1]> var_192_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_191_promoted_to_fp16)[name = tensor<string, []>("op_192_cast_fp16")];
            tensor<fp16, []> var_193_promoted_to_fp16 = const()[name = tensor<string, []>("op_193_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1]> var_194_cast_fp16 = add(x = var_192_cast_fp16, y = var_193_promoted_to_fp16)[name = tensor<string, []>("op_194_cast_fp16")];
            tensor<fp16, []> var_195_promoted_to_fp16 = const()[name = tensor<string, []>("op_195_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
            tensor<fp16, [1]> var_196_cast_fp16 = sub(x = var_194_cast_fp16, y = var_195_promoted_to_fp16)[name = tensor<string, []>("op_196_cast_fp16")];
            tensor<fp16, []> var_53_promoted_1_to_fp16 = const()[name = tensor<string, []>("op_53_promoted_1_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_196_cast_fp16, y = var_53_promoted_1_to_fp16)[name = tensor<string, []>("floor_div_1_cast_fp16")];
            tensor<fp16, []> var_198_promoted_to_fp16 = const()[name = tensor<string, []>("op_198_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1]> current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_198_promoted_to_fp16)[name = tensor<string, []>("current_lengths_5_cast_fp16")];
            tensor<string, []> cast_5_dtype_0 = const()[name = tensor<string, []>("cast_5_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [1, 19]> expand_dims_2 = const()[name = tensor<string, []>("expand_dims_2"), val = tensor<int32, [1, 19]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]])];
            tensor<int32, [1]> var_207_axes_0 = const()[name = tensor<string, []>("op_207_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> current_lengths_5_cast_fp16_to_int32 = cast(dtype = cast_5_dtype_0, x = current_lengths_5_cast_fp16)[name = tensor<string, []>("cast_193")];
            tensor<int32, [1, 1]> var_207 = expand_dims(axes = var_207_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = tensor<string, []>("op_207")];
            tensor<bool, [1, 19]> time_mask_5 = less(x = expand_dims_2, y = var_207)[name = tensor<string, []>("time_mask_5")];
            tensor<int32, [1]> var_209_axes_0 = const()[name = tensor<string, []>("op_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 19, 1]> var_209 = expand_dims(axes = var_209_axes_0, x = time_mask_5)[name = tensor<string, []>("op_209")];
            tensor<int32, [3]> var_211_reps_0 = const()[name = tensor<string, []>("op_211_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
            tensor<bool, [1, 19, 33]> var_211 = tile(reps = var_211_reps_0, x = var_209)[name = tensor<string, []>("op_211")];
            tensor<int32, [1]> var_217_axes_0 = const()[name = tensor<string, []>("op_217_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<string, []> cast_6_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_6_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 19, 33]> var_211_to_fp16 = cast(dtype = cast_6_to_fp16_dtype_0, x = var_211)[name = tensor<string, []>("cast_192")];
            tensor<fp16, [1, 1, 19, 33]> var_217_cast_fp16 = expand_dims(axes = var_217_axes_0, x = var_211_to_fp16)[name = tensor<string, []>("op_217_cast_fp16")];
            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = tensor<string, []>("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
            tensor<fp16, [1, 256, 19, 33]> expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_217_cast_fp16)[name = tensor<string, []>("expanded_mask_7_cast_fp16")];
            tensor<fp16, [1, 256, 19, 33]> input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<string, []> tensor_9_pad_type_0 = const()[name = tensor<string, []>("tensor_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> tensor_9_strides_0 = const()[name = tensor<string, []>("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> tensor_9_pad_0 = const()[name = tensor<string, []>("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = tensor<string, []>("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> tensor_9_groups_0 = const()[name = tensor<string, []>("tensor_9_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, 19, 33]> tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("tensor_9_cast_fp16")];
            tensor<fp16, [1, 256, 19, 33]> input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<fp16, [1, 256, 19, 33]> tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor<string, []>("tensor_11_cast_fp16")];
            tensor<fp16, [1, 256, 19, 33]> input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_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_41_to_fp16 = const()[name = tensor<string, []>("const_41_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 256, 22, 36]> input_17_cast_fp16 = pad(constant_val = const_41_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, []> tensor_13_pad_type_0 = const()[name = tensor<string, []>("tensor_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> tensor_13_strides_0 = const()[name = tensor<string, []>("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, []> tensor_13_groups_0 = const()[name = tensor<string, []>("tensor_13_groups_0"), val = tensor<int32, []>(256)];
            tensor<int32, [4]> tensor_13_pad_0 = const()[name = tensor<string, []>("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = tensor<string, []>("tensor_13_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, 10, 17]> tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("tensor_13_cast_fp16")];
            tensor<fp16, []> var_254_promoted_to_fp16 = const()[name = tensor<string, []>("op_254_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [1]> var_255_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_254_promoted_to_fp16)[name = tensor<string, []>("op_255_cast_fp16")];
            tensor<fp16, []> var_256_promoted_to_fp16 = const()[name = tensor<string, []>("op_256_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1]> var_257_cast_fp16 = add(x = var_255_cast_fp16, y = var_256_promoted_to_fp16)[name = tensor<string, []>("op_257_cast_fp16")];
            tensor<fp16, []> var_258_promoted_to_fp16 = const()[name = tensor<string, []>("op_258_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
            tensor<fp16, [1]> var_259_cast_fp16 = sub(x = var_257_cast_fp16, y = var_258_promoted_to_fp16)[name = tensor<string, []>("op_259_cast_fp16")];
            tensor<fp16, []> var_53_promoted_2_to_fp16 = const()[name = tensor<string, []>("op_53_promoted_2_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_259_cast_fp16, y = var_53_promoted_2_to_fp16)[name = tensor<string, []>("floor_div_2_cast_fp16")];
            tensor<fp16, []> var_261_promoted_to_fp16 = const()[name = tensor<string, []>("op_261_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1]> current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_261_promoted_to_fp16)[name = tensor<string, []>("current_lengths_cast_fp16")];
            tensor<string, []> cast_7_dtype_0 = const()[name = tensor<string, []>("cast_7_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [1, 10]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1, 10]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]])];
            tensor<int32, [1]> var_270_axes_0 = const()[name = tensor<string, []>("op_270_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> current_lengths_cast_fp16_to_int32 = cast(dtype = cast_7_dtype_0, x = current_lengths_cast_fp16)[name = tensor<string, []>("cast_191")];
            tensor<int32, [1, 1]> var_270 = expand_dims(axes = var_270_axes_0, x = current_lengths_cast_fp16_to_int32)[name = tensor<string, []>("op_270")];
            tensor<bool, [1, 10]> time_mask = less(x = expand_dims_3, y = var_270)[name = tensor<string, []>("time_mask")];
            tensor<int32, [1]> var_272_axes_0 = const()[name = tensor<string, []>("op_272_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 10, 1]> var_272 = expand_dims(axes = var_272_axes_0, x = time_mask)[name = tensor<string, []>("op_272")];
            tensor<int32, [3]> var_274_reps_0 = const()[name = tensor<string, []>("op_274_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
            tensor<bool, [1, 10, 17]> var_274 = tile(reps = var_274_reps_0, x = var_272)[name = tensor<string, []>("op_274")];
            tensor<int32, [1]> var_280_axes_0 = const()[name = tensor<string, []>("op_280_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<string, []> cast_8_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_8_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 10, 17]> var_274_to_fp16 = cast(dtype = cast_8_to_fp16_dtype_0, x = var_274)[name = tensor<string, []>("cast_190")];
            tensor<fp16, [1, 1, 10, 17]> var_280_cast_fp16 = expand_dims(axes = var_280_axes_0, x = var_274_to_fp16)[name = tensor<string, []>("op_280_cast_fp16")];
            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = tensor<string, []>("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
            tensor<fp16, [1, 256, 10, 17]> expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_280_cast_fp16)[name = tensor<string, []>("expanded_mask_13_cast_fp16")];
            tensor<fp16, [1, 256, 10, 17]> input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<string, []> tensor_15_pad_type_0 = const()[name = tensor<string, []>("tensor_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> tensor_15_strides_0 = const()[name = tensor<string, []>("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> tensor_15_pad_0 = const()[name = tensor<string, []>("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = tensor<string, []>("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> tensor_15_groups_0 = const()[name = tensor<string, []>("tensor_15_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, 10, 17]> tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("tensor_15_cast_fp16")];
            tensor<fp16, [1, 256, 10, 17]> input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<fp16, [1, 256, 10, 17]> tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = tensor<string, []>("tensor_cast_fp16")];
            tensor<fp16, [1, 256, 10, 17]> x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
            tensor<int32, [4]> var_314_perm_0 = const()[name = tensor<string, []>("op_314_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_315 = const()[name = tensor<string, []>("op_315"), val = tensor<int32, [3]>([1, 10, -1])];
            tensor<fp16, [1, 10, 256, 17]> var_314_cast_fp16 = transpose(perm = var_314_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_240")];
            tensor<fp16, [1, 10, 4352]> input_23_cast_fp16 = reshape(shape = var_315, x = var_314_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, 10, 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_325_begin_0 = const()[name = tensor<string, []>("op_325_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_325_end_0 = const()[name = tensor<string, []>("op_325_end_0"), val = tensor<int32, [3]>([1, 10, 512])];
            tensor<bool, [3]> var_325_end_mask_0 = const()[name = tensor<string, []>("op_325_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 8, 512]> var_325_cast_fp16 = slice_by_index(begin = var_325_begin_0, end = var_325_end_0, end_mask = var_325_end_mask_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_325_cast_fp16")];
            tensor<int32, []> var_327 = const()[name = tensor<string, []>("op_327"), val = tensor<int32, []>(2)];
            tensor<int32, [1]> var_328 = sub(x = current_lengths_cast_fp16_to_int32, y = var_327)[name = tensor<string, []>("op_328")];
            tensor<string, []> var_328_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_328_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_59_promoted_to_fp16 = const()[name = tensor<string, []>("op_59_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, []> const_61_to_fp16 = const()[name = tensor<string, []>("const_61_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<fp16, [1]> var_328_to_fp16 = cast(dtype = var_328_promoted_to_fp16_dtype_0, x = var_328)[name = tensor<string, []>("cast_189")];
            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_59_promoted_to_fp16, beta = const_61_to_fp16, x = var_328_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]>([4])];
            tensor<fp16, []> var_344_promoted_to_fp16 = const()[name = tensor<string, []>("op_344_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_344_promoted_to_fp16)[name = tensor<string, []>("padding_length_cast_fp16")];
            tensor<int32, []> const_63 = const()[name = tensor<string, []>("const_63"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> var_346 = mul(x = cache_last_channel_len, y = const_63)[name = tensor<string, []>("op_346")];
            tensor<int32, []> var_347 = const()[name = tensor<string, []>("op_347"), val = tensor<int32, []>(70)];
            tensor<int32, [1]> offset = add(x = var_346, y = var_347)[name = tensor<string, []>("offset")];
            tensor<int32, [1]> var_387_axes_0 = const()[name = tensor<string, []>("op_387_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> var_387_cast_fp16 = expand_dims(axes = var_387_axes_0, x = padding_length_cast_fp16)[name = tensor<string, []>("op_387_cast_fp16")];
            tensor<fp16, [1, 78]> var_386_promoted_to_fp16 = const()[name = tensor<string, []>("op_386_promoted_to_fp16"), val = tensor<fp16, [1, 78]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736832)))];
            tensor<bool, [1, 78]> pad_mask_1_cast_fp16 = less(x = var_386_promoted_to_fp16, y = var_387_cast_fp16)[name = tensor<string, []>("pad_mask_1_cast_fp16")];
            tensor<int32, [1, 78]> expand_dims_5 = const()[name = tensor<string, []>("expand_dims_5"), val = tensor<int32, [1, 78]>([[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, 73, 74, 75, 76, 77]])];
            tensor<int32, [1]> var_393_axes_0 = const()[name = tensor<string, []>("op_393_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 1]> var_393 = expand_dims(axes = var_393_axes_0, x = offset)[name = tensor<string, []>("op_393")];
            tensor<bool, [1, 78]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_393)[name = tensor<string, []>("pad_mask_off")];
            tensor<bool, [1, 78]> 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_396_axes_0 = const()[name = tensor<string, []>("op_396_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 78]> var_396 = expand_dims(axes = var_396_axes_0, x = pad_mask_3)[name = tensor<string, []>("op_396")];
            tensor<int32, [3]> var_397 = const()[name = tensor<string, []>("op_397"), val = tensor<int32, [3]>([1, 78, 1])];
            tensor<bool, [1, 78, 78]> pad_mask_for_att_mask_1 = tile(reps = var_397, x = var_396)[name = tensor<string, []>("pad_mask_for_att_mask_1")];
            tensor<int32, [3]> var_399_perm_0 = const()[name = tensor<string, []>("op_399_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<bool, [1, 78, 78]> var_399 = transpose(perm = var_399_perm_0, x = pad_mask_for_att_mask_1)[name = tensor<string, []>("transpose_239")];
            tensor<bool, [1, 78, 78]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_399)[name = tensor<string, []>("pad_mask_for_att_mask")];
            tensor<bool, [1, 78, 78]> const_71 = const()[name = tensor<string, []>("const_71"), val = tensor<bool, [1, 78, 78]>([[[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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, 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, true, true, 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, true, true, true, true, 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, true, true, true, true, true, true, 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, true, true, true, true, 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, true, true, true, true, true, true, 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, true, true, true, true, 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, true, true, true, true, true, true], [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, true, true, true, true]]])];
            tensor<bool, [1, 78, 78]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = tensor<string, []>("att_mask_9")];
            tensor<bool, [1, 78, 78]> att_mask = logical_not(x = att_mask_9)[name = tensor<string, []>("att_mask")];
            tensor<bool, [1, 78]> 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, 78])];
            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, 8]> 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_9_begin_0 = const()[name = tensor<string, []>("mask_9_begin_0"), val = tensor<int32, [3]>([0, 70, 0])];
            tensor<int32, [3]> mask_9_end_0 = const()[name = tensor<string, []>("mask_9_end_0"), val = tensor<int32, [3]>([1, 78, 78])];
            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = tensor<string, []>("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<bool, [1, 8, 78]> mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = tensor<string, []>("mask_9")];
            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_188")];
            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_187")];
            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, 8, 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_325_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, 8, 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, 8, 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, 8, 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_436_to_fp16 = const()[name = tensor<string, []>("op_436_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_437_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_436_to_fp16)[name = tensor<string, []>("op_437_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_37_cast_fp16 = add(x = var_325_cast_fp16, y = var_437_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, 8, 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, 78, 512]> input_39_cast_fp16 = concat(axis = var_65, 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_459_begin_0 = const()[name = tensor<string, []>("op_459_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_459_end_0 = const()[name = tensor<string, []>("op_459_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_459_end_mask_0 = const()[name = tensor<string, []>("op_459_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_459_cast_fp16 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = cache_1_cast_fp16)[name = tensor<string, []>("op_459_cast_fp16")];
            tensor<int32, [3]> var_462_begin_0 = const()[name = tensor<string, []>("op_462_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_462_end_0 = const()[name = tensor<string, []>("op_462_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_462_end_mask_0 = const()[name = tensor<string, []>("op_462_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_462_cast_fp16 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = key_1_cast_fp16)[name = tensor<string, []>("op_462_cast_fp16")];
            tensor<bool, []> var_465_interleave_0 = const()[name = tensor<string, []>("op_465_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_465_cast_fp16 = concat(axis = var_65, interleave = var_465_interleave_0, values = (var_459_cast_fp16, var_462_cast_fp16))[name = tensor<string, []>("op_465_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, 8, 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_469 = const()[name = tensor<string, []>("op_469"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_1_cast_fp16 = reshape(shape = var_469, 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, 78, 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_473 = const()[name = tensor<string, []>("op_473"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_1_cast_fp16 = reshape(shape = var_473, 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, 78, 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_477 = const()[name = tensor<string, []>("op_477"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_1_cast_fp16 = reshape(shape = var_477, 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, 8, 8, 64]> var_489_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_489_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, 8, 8, 64]> var_491_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_491_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, 155]> var_493_to_fp16 = const()[name = tensor<string, []>("op_493_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10516352)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_491_cast_fp16)[name = tensor<string, []>("transpose_237")];
            tensor<fp16, [1, 8, 8, 155]> 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_493_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_79_to_fp16 = const()[name = tensor<string, []>("const_79_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_9_cast_fp16 = pad(constant_val = const_79_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_501 = const()[name = tensor<string, []>("op_501"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_11_cast_fp16 = reshape(shape = var_501, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<int32, [4]> var_505_begin_0 = const()[name = tensor<string, []>("op_505_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_505_end_0 = const()[name = tensor<string, []>("op_505_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_505_end_mask_0 = const()[name = tensor<string, []>("op_505_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_505_cast_fp16 = slice_by_index(begin = var_505_begin_0, end = var_505_end_0, end_mask = var_505_end_mask_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_505_cast_fp16")];
            tensor<int32, [4]> var_506 = const()[name = tensor<string, []>("op_506"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_1_cast_fp16 = reshape(shape = var_506, x = var_505_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, 78]> transpose_52 = transpose(perm = transpose_52_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_235")];
            tensor<fp16, [1, 8, 8, 64]> transpose_51 = transpose(perm = transpose_51_perm_0, x = var_489_cast_fp16)[name = tensor<string, []>("transpose_236")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_515_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = tensor<string, []>("op_515_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, 8, 78]> _inversed_scores_1_cast_fp16 = mul(x = var_515_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_1_cast_fp16")];
            tensor<int32, [1]> mask_11_axes_0 = const()[name = tensor<string, []>("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 8, 78]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = tensor<string, []>("mask_11")];
            tensor<fp16, []> var_41_to_fp16 = const()[name = tensor<string, []>("op_41_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
            tensor<fp16, [1, 8, 8, 78]> scores_3_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_3_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_521_cast_fp16 = softmax(axis = var_56, x = scores_3_cast_fp16)[name = tensor<string, []>("op_521_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, 8, 78]> input_41_cast_fp16 = select(a = var_40_to_fp16, b = var_521_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_525_perm_0 = const()[name = tensor<string, []>("op_525_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_526 = const()[name = tensor<string, []>("op_526"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_525_cast_fp16 = transpose(perm = var_525_perm_0, x = x_13_cast_fp16)[name = tensor<string, []>("transpose_234")];
            tensor<fp16, [1, 8, 512]> input_43_cast_fp16 = reshape(shape = var_526, x = var_525_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, []>(10675136)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(11199488)))];
            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, []>(11200576)))];
            tensor<fp16, [1, 8, 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, []>(11201664)))];
            tensor<fp16, [1, 512, 8]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_233")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_19_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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_551_axes_0 = const()[name = tensor<string, []>("op_551_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 8]> var_551 = expand_dims(axes = var_551_axes_0, x = pad_mask)[name = tensor<string, []>("op_551")];
            tensor<fp16, [1, 512, 8]> input_53_cast_fp16 = select(a = var_40_to_fp16, b = x_19_cast_fp16, cond = var_551)[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, 16]> new_x_3_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_567_begin_0 = const()[name = tensor<string, []>("op_567_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_567_end_0 = const()[name = tensor<string, []>("op_567_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_567_end_mask_0 = const()[name = tensor<string, []>("op_567_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_567_cast_fp16 = slice_by_index(begin = var_567_begin_0, end = var_567_end_0, end_mask = var_567_end_mask_0, x = next_cache_1_cast_fp16)[name = tensor<string, []>("op_567_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, []>(12250304)))];
            tensor<fp16, [1, 512, 8]> 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, []>(12259584)))];
            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, []>(12260672)))];
            tensor<fp16, [1, 8, 512]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = tensor<string, []>("transpose_232")];
            tensor<fp16, [1, 8, 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, 8]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_231")];
            tensor<fp16, [1, 512, 8]> 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, []>(12261760)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = tensor<string, []>("transpose_230")];
            tensor<fp16, [1, 8, 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, []>(12786112)))];
            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, []>(12787200)))];
            tensor<fp16, [1, 8, 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, []>(12788288)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(14885504)))];
            tensor<fp16, [1, 8, 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_608_to_fp16 = const()[name = tensor<string, []>("op_608_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_609_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_608_to_fp16)[name = tensor<string, []>("op_609_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_609_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, []>(16982720)))];
            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, []>(16983808)))];
            tensor<fp16, [1, 8, 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, []>(16984896)))];
            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, []>(16985984)))];
            tensor<fp16, [1, 8, 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, []>(16987072)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(19084288)))];
            tensor<fp16, [1, 8, 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_643_to_fp16 = const()[name = tensor<string, []>("op_643_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_644_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_643_to_fp16)[name = tensor<string, []>("op_644_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_644_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, []>(21181504)))];
            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, []>(21182592)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_91_cast_fp16 = concat(axis = var_65, 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_666_begin_0 = const()[name = tensor<string, []>("op_666_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_666_end_0 = const()[name = tensor<string, []>("op_666_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_666_end_mask_0 = const()[name = tensor<string, []>("op_666_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_666_cast_fp16 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = cache_5_cast_fp16)[name = tensor<string, []>("op_666_cast_fp16")];
            tensor<int32, [3]> var_669_begin_0 = const()[name = tensor<string, []>("op_669_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_669_end_0 = const()[name = tensor<string, []>("op_669_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_669_end_mask_0 = const()[name = tensor<string, []>("op_669_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_669_cast_fp16 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = key_3_cast_fp16)[name = tensor<string, []>("op_669_cast_fp16")];
            tensor<bool, []> var_672_interleave_0 = const()[name = tensor<string, []>("op_672_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_672_cast_fp16 = concat(axis = var_65, interleave = var_672_interleave_0, values = (var_666_cast_fp16, var_669_cast_fp16))[name = tensor<string, []>("op_672_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, []>(21183680)))];
            tensor<fp16, [1, 8, 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_676 = const()[name = tensor<string, []>("op_676"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_7_cast_fp16 = reshape(shape = var_676, 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, []>(21708032)))];
            tensor<fp16, [1, 78, 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_680 = const()[name = tensor<string, []>("op_680"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_5_cast_fp16 = reshape(shape = var_680, 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, []>(22232384)))];
            tensor<fp16, [1, 78, 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_684 = const()[name = tensor<string, []>("op_684"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_3_cast_fp16 = reshape(shape = var_684, 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, []>(22756736)))];
            tensor<fp16, [1, 8, 8, 64]> var_696_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_696_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, []>(22757824)))];
            tensor<fp16, [1, 8, 8, 64]> var_698_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_698_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, 155]> var_700_to_fp16 = const()[name = tensor<string, []>("op_700_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22758912)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_698_cast_fp16)[name = tensor<string, []>("transpose_228")];
            tensor<fp16, [1, 8, 8, 155]> 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_700_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_92_to_fp16 = const()[name = tensor<string, []>("const_92_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_35_cast_fp16 = pad(constant_val = const_92_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_708 = const()[name = tensor<string, []>("op_708"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_37_cast_fp16 = reshape(shape = var_708, x = x_35_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<int32, [4]> var_712_begin_0 = const()[name = tensor<string, []>("op_712_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_712_end_0 = const()[name = tensor<string, []>("op_712_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_712_end_mask_0 = const()[name = tensor<string, []>("op_712_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_712_cast_fp16 = slice_by_index(begin = var_712_begin_0, end = var_712_end_0, end_mask = var_712_end_mask_0, x = x_37_cast_fp16)[name = tensor<string, []>("op_712_cast_fp16")];
            tensor<int32, [4]> var_713 = const()[name = tensor<string, []>("op_713"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_5_cast_fp16 = reshape(shape = var_713, x = var_712_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, 78]> transpose_54 = transpose(perm = transpose_54_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_226")];
            tensor<fp16, [1, 8, 8, 64]> transpose_53 = transpose(perm = transpose_53_perm_0, x = var_696_cast_fp16)[name = tensor<string, []>("transpose_227")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_722_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = tensor<string, []>("op_722_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, 8, 78]> _inversed_scores_5_cast_fp16 = mul(x = var_722_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_5_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_7_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_7_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_728_cast_fp16 = softmax(axis = var_56, x = scores_7_cast_fp16)[name = tensor<string, []>("op_728_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_93_cast_fp16 = select(a = var_40_to_fp16, b = var_728_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_732_perm_0 = const()[name = tensor<string, []>("op_732_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_733 = const()[name = tensor<string, []>("op_733"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_732_cast_fp16 = transpose(perm = var_732_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_225")];
            tensor<fp16, [1, 8, 512]> input_95_cast_fp16 = reshape(shape = var_733, x = var_732_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, []>(22917696)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(23442048)))];
            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, []>(23443136)))];
            tensor<fp16, [1, 8, 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, []>(23444224)))];
            tensor<fp16, [1, 512, 8]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_224")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_45_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_105_cast_fp16 = select(a = var_40_to_fp16, b = x_45_cast_fp16, cond = var_551)[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, 16]> new_x_7_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_774_begin_0 = const()[name = tensor<string, []>("op_774_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_774_end_0 = const()[name = tensor<string, []>("op_774_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_774_end_mask_0 = const()[name = tensor<string, []>("op_774_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_774_cast_fp16 = slice_by_index(begin = var_774_begin_0, end = var_774_end_0, end_mask = var_774_end_mask_0, x = next_cache_3_cast_fp16)[name = tensor<string, []>("op_774_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, []>(24492864)))];
            tensor<fp16, [1, 512, 8]> 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, []>(24502144)))];
            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, []>(24503232)))];
            tensor<fp16, [1, 8, 512]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_223")];
            tensor<fp16, [1, 8, 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, 8]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = tensor<string, []>("transpose_222")];
            tensor<fp16, [1, 512, 8]> 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, []>(24504320)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_221")];
            tensor<fp16, [1, 8, 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, []>(25028672)))];
            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, []>(25029760)))];
            tensor<fp16, [1, 8, 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, []>(25030848)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(27128064)))];
            tensor<fp16, [1, 8, 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_815_to_fp16 = const()[name = tensor<string, []>("op_815_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_816_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_815_to_fp16)[name = tensor<string, []>("op_816_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_816_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, []>(29225280)))];
            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, []>(29226368)))];
            tensor<fp16, [1, 8, 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, []>(29227456)))];
            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, []>(29228544)))];
            tensor<fp16, [1, 8, 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, []>(29229632)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(31326848)))];
            tensor<fp16, [1, 8, 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_850_to_fp16 = const()[name = tensor<string, []>("op_850_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_851_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_850_to_fp16)[name = tensor<string, []>("op_851_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_851_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, []>(33424064)))];
            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, []>(33425152)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_143_cast_fp16 = concat(axis = var_65, 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_873_begin_0 = const()[name = tensor<string, []>("op_873_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_873_end_0 = const()[name = tensor<string, []>("op_873_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_873_end_mask_0 = const()[name = tensor<string, []>("op_873_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_873_cast_fp16 = slice_by_index(begin = var_873_begin_0, end = var_873_end_0, end_mask = var_873_end_mask_0, x = cache_9_cast_fp16)[name = tensor<string, []>("op_873_cast_fp16")];
            tensor<int32, [3]> var_876_begin_0 = const()[name = tensor<string, []>("op_876_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_876_end_0 = const()[name = tensor<string, []>("op_876_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_876_end_mask_0 = const()[name = tensor<string, []>("op_876_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_876_cast_fp16 = slice_by_index(begin = var_876_begin_0, end = var_876_end_0, end_mask = var_876_end_mask_0, x = key_5_cast_fp16)[name = tensor<string, []>("op_876_cast_fp16")];
            tensor<bool, []> var_879_interleave_0 = const()[name = tensor<string, []>("op_879_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_879_cast_fp16 = concat(axis = var_65, interleave = var_879_interleave_0, values = (var_873_cast_fp16, var_876_cast_fp16))[name = tensor<string, []>("op_879_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, []>(33426240)))];
            tensor<fp16, [1, 8, 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_883 = const()[name = tensor<string, []>("op_883"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_13_cast_fp16 = reshape(shape = var_883, 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, []>(33950592)))];
            tensor<fp16, [1, 78, 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_887 = const()[name = tensor<string, []>("op_887"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_9_cast_fp16 = reshape(shape = var_887, 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, []>(34474944)))];
            tensor<fp16, [1, 78, 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_891 = const()[name = tensor<string, []>("op_891"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_5_cast_fp16 = reshape(shape = var_891, 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, []>(34999296)))];
            tensor<fp16, [1, 8, 8, 64]> var_903_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_903_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, []>(35000384)))];
            tensor<fp16, [1, 8, 8, 64]> var_905_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_905_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, 155]> var_907_to_fp16 = const()[name = tensor<string, []>("op_907_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35001472)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_905_cast_fp16)[name = tensor<string, []>("transpose_219")];
            tensor<fp16, [1, 8, 8, 155]> 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_907_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_105_to_fp16 = const()[name = tensor<string, []>("const_105_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_61_cast_fp16 = pad(constant_val = const_105_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_915 = const()[name = tensor<string, []>("op_915"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_63_cast_fp16 = reshape(shape = var_915, x = x_61_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
            tensor<int32, [4]> var_919_begin_0 = const()[name = tensor<string, []>("op_919_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_919_end_0 = const()[name = tensor<string, []>("op_919_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_919_end_mask_0 = const()[name = tensor<string, []>("op_919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_919_cast_fp16 = slice_by_index(begin = var_919_begin_0, end = var_919_end_0, end_mask = var_919_end_mask_0, x = x_63_cast_fp16)[name = tensor<string, []>("op_919_cast_fp16")];
            tensor<int32, [4]> var_920 = const()[name = tensor<string, []>("op_920"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_9_cast_fp16 = reshape(shape = var_920, x = var_919_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, 78]> transpose_56 = transpose(perm = transpose_56_perm_0, x = k_9_cast_fp16)[name = tensor<string, []>("transpose_217")];
            tensor<fp16, [1, 8, 8, 64]> transpose_55 = transpose(perm = transpose_55_perm_0, x = var_903_cast_fp16)[name = tensor<string, []>("transpose_218")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_929_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = tensor<string, []>("op_929_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, 8, 78]> _inversed_scores_9_cast_fp16 = mul(x = var_929_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_9_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_11_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_11_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_935_cast_fp16 = softmax(axis = var_56, x = scores_11_cast_fp16)[name = tensor<string, []>("op_935_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_145_cast_fp16 = select(a = var_40_to_fp16, b = var_935_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_939_perm_0 = const()[name = tensor<string, []>("op_939_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_940 = const()[name = tensor<string, []>("op_940"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_939_cast_fp16 = transpose(perm = var_939_perm_0, x = x_65_cast_fp16)[name = tensor<string, []>("transpose_216")];
            tensor<fp16, [1, 8, 512]> input_147_cast_fp16 = reshape(shape = var_940, x = var_939_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, []>(35160256)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(35684608)))];
            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, []>(35685696)))];
            tensor<fp16, [1, 8, 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, []>(35686784)))];
            tensor<fp16, [1, 512, 8]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = tensor<string, []>("transpose_215")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_71_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_157_cast_fp16 = select(a = var_40_to_fp16, b = x_71_cast_fp16, cond = var_551)[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, 16]> new_x_11_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_981_begin_0 = const()[name = tensor<string, []>("op_981_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_981_end_0 = const()[name = tensor<string, []>("op_981_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_981_end_mask_0 = const()[name = tensor<string, []>("op_981_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_981_cast_fp16 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = next_cache_5_cast_fp16)[name = tensor<string, []>("op_981_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, []>(36735424)))];
            tensor<fp16, [1, 512, 8]> 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, []>(36744704)))];
            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, []>(36745792)))];
            tensor<fp16, [1, 8, 512]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = tensor<string, []>("transpose_214")];
            tensor<fp16, [1, 8, 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, 8]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = tensor<string, []>("transpose_213")];
            tensor<fp16, [1, 512, 8]> 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, []>(36746880)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = tensor<string, []>("transpose_212")];
            tensor<fp16, [1, 8, 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, []>(37271232)))];
            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, []>(37272320)))];
            tensor<fp16, [1, 8, 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, []>(37273408)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(39370624)))];
            tensor<fp16, [1, 8, 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_1022_to_fp16 = const()[name = tensor<string, []>("op_1022_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1023_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1022_to_fp16)[name = tensor<string, []>("op_1023_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1023_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, []>(41467840)))];
            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, []>(41468928)))];
            tensor<fp16, [1, 8, 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, []>(41470016)))];
            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, []>(41471104)))];
            tensor<fp16, [1, 8, 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, []>(41472192)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(43569408)))];
            tensor<fp16, [1, 8, 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_1057_to_fp16 = const()[name = tensor<string, []>("op_1057_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1058_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1057_to_fp16)[name = tensor<string, []>("op_1058_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1058_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, []>(45666624)))];
            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, []>(45667712)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_195_cast_fp16 = concat(axis = var_65, 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_1080_begin_0 = const()[name = tensor<string, []>("op_1080_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1080_end_0 = const()[name = tensor<string, []>("op_1080_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1080_end_mask_0 = const()[name = tensor<string, []>("op_1080_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_1080_cast_fp16 = slice_by_index(begin = var_1080_begin_0, end = var_1080_end_0, end_mask = var_1080_end_mask_0, x = cache_13_cast_fp16)[name = tensor<string, []>("op_1080_cast_fp16")];
            tensor<int32, [3]> var_1083_begin_0 = const()[name = tensor<string, []>("op_1083_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1083_end_0 = const()[name = tensor<string, []>("op_1083_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_1083_end_mask_0 = const()[name = tensor<string, []>("op_1083_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_1083_cast_fp16 = slice_by_index(begin = var_1083_begin_0, end = var_1083_end_0, end_mask = var_1083_end_mask_0, x = key_7_cast_fp16)[name = tensor<string, []>("op_1083_cast_fp16")];
            tensor<bool, []> var_1086_interleave_0 = const()[name = tensor<string, []>("op_1086_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1086_cast_fp16 = concat(axis = var_65, interleave = var_1086_interleave_0, values = (var_1080_cast_fp16, var_1083_cast_fp16))[name = tensor<string, []>("op_1086_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, []>(45668800)))];
            tensor<fp16, [1, 8, 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_1090 = const()[name = tensor<string, []>("op_1090"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_19_cast_fp16 = reshape(shape = var_1090, 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, []>(46193152)))];
            tensor<fp16, [1, 78, 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_1094 = const()[name = tensor<string, []>("op_1094"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_13_cast_fp16 = reshape(shape = var_1094, 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, []>(46717504)))];
            tensor<fp16, [1, 78, 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_1098 = const()[name = tensor<string, []>("op_1098"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_7_cast_fp16 = reshape(shape = var_1098, 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, []>(47241856)))];
            tensor<fp16, [1, 8, 8, 64]> var_1110_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1110_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, []>(47242944)))];
            tensor<fp16, [1, 8, 8, 64]> var_1112_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1112_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, 155]> var_1114_to_fp16 = const()[name = tensor<string, []>("op_1114_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47244032)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1112_cast_fp16)[name = tensor<string, []>("transpose_210")];
            tensor<fp16, [1, 8, 8, 155]> 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_1114_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_118_to_fp16 = const()[name = tensor<string, []>("const_118_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_87_cast_fp16 = pad(constant_val = const_118_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_1122 = const()[name = tensor<string, []>("op_1122"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_89_cast_fp16 = reshape(shape = var_1122, x = x_87_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
            tensor<int32, [4]> var_1126_begin_0 = const()[name = tensor<string, []>("op_1126_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1126_end_0 = const()[name = tensor<string, []>("op_1126_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_1126_end_mask_0 = const()[name = tensor<string, []>("op_1126_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_1126_cast_fp16 = slice_by_index(begin = var_1126_begin_0, end = var_1126_end_0, end_mask = var_1126_end_mask_0, x = x_89_cast_fp16)[name = tensor<string, []>("op_1126_cast_fp16")];
            tensor<int32, [4]> var_1127 = const()[name = tensor<string, []>("op_1127"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_13_cast_fp16 = reshape(shape = var_1127, x = var_1126_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, 78]> transpose_58 = transpose(perm = transpose_58_perm_0, x = k_13_cast_fp16)[name = tensor<string, []>("transpose_208")];
            tensor<fp16, [1, 8, 8, 64]> transpose_57 = transpose(perm = transpose_57_perm_0, x = var_1110_cast_fp16)[name = tensor<string, []>("transpose_209")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_1136_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = tensor<string, []>("op_1136_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, 8, 78]> _inversed_scores_13_cast_fp16 = mul(x = var_1136_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_13_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_15_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_15_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_1142_cast_fp16 = softmax(axis = var_56, x = scores_15_cast_fp16)[name = tensor<string, []>("op_1142_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_197_cast_fp16 = select(a = var_40_to_fp16, b = var_1142_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_1146_perm_0 = const()[name = tensor<string, []>("op_1146_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1147 = const()[name = tensor<string, []>("op_1147"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_1146_cast_fp16 = transpose(perm = var_1146_perm_0, x = x_91_cast_fp16)[name = tensor<string, []>("transpose_207")];
            tensor<fp16, [1, 8, 512]> input_199_cast_fp16 = reshape(shape = var_1147, x = var_1146_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, []>(47402816)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(47927168)))];
            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, []>(47928256)))];
            tensor<fp16, [1, 8, 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, []>(47929344)))];
            tensor<fp16, [1, 512, 8]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = tensor<string, []>("transpose_206")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_97_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_209_cast_fp16 = select(a = var_40_to_fp16, b = x_97_cast_fp16, cond = var_551)[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, 16]> new_x_15_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_1188_begin_0 = const()[name = tensor<string, []>("op_1188_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_1188_end_0 = const()[name = tensor<string, []>("op_1188_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_1188_end_mask_0 = const()[name = tensor<string, []>("op_1188_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_1188_cast_fp16 = slice_by_index(begin = var_1188_begin_0, end = var_1188_end_0, end_mask = var_1188_end_mask_0, x = next_cache_7_cast_fp16)[name = tensor<string, []>("op_1188_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, []>(48977984)))];
            tensor<fp16, [1, 512, 8]> 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, []>(48987264)))];
            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, []>(48988352)))];
            tensor<fp16, [1, 8, 512]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_205")];
            tensor<fp16, [1, 8, 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, 8]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = tensor<string, []>("transpose_204")];
            tensor<fp16, [1, 512, 8]> 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, []>(48989440)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = tensor<string, []>("transpose_203")];
            tensor<fp16, [1, 8, 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, []>(49513792)))];
            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, []>(49514880)))];
            tensor<fp16, [1, 8, 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, []>(49515968)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(51613184)))];
            tensor<fp16, [1, 8, 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_1229_to_fp16 = const()[name = tensor<string, []>("op_1229_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1230_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1229_to_fp16)[name = tensor<string, []>("op_1230_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1230_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, []>(53710400)))];
            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, []>(53711488)))];
            tensor<fp16, [1, 8, 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, []>(53712576)))];
            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, []>(53713664)))];
            tensor<fp16, [1, 8, 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, []>(53714752)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(55811968)))];
            tensor<fp16, [1, 8, 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_1264_to_fp16 = const()[name = tensor<string, []>("op_1264_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1265_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1264_to_fp16)[name = tensor<string, []>("op_1265_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1265_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, []>(57909184)))];
            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, []>(57910272)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_247_cast_fp16 = concat(axis = var_65, 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_1287_begin_0 = const()[name = tensor<string, []>("op_1287_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1287_end_0 = const()[name = tensor<string, []>("op_1287_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1287_end_mask_0 = const()[name = tensor<string, []>("op_1287_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_1287_cast_fp16 = slice_by_index(begin = var_1287_begin_0, end = var_1287_end_0, end_mask = var_1287_end_mask_0, x = cache_17_cast_fp16)[name = tensor<string, []>("op_1287_cast_fp16")];
            tensor<int32, [3]> var_1290_begin_0 = const()[name = tensor<string, []>("op_1290_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1290_end_0 = const()[name = tensor<string, []>("op_1290_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_1290_end_mask_0 = const()[name = tensor<string, []>("op_1290_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_1290_cast_fp16 = slice_by_index(begin = var_1290_begin_0, end = var_1290_end_0, end_mask = var_1290_end_mask_0, x = key_9_cast_fp16)[name = tensor<string, []>("op_1290_cast_fp16")];
            tensor<bool, []> var_1293_interleave_0 = const()[name = tensor<string, []>("op_1293_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1293_cast_fp16 = concat(axis = var_65, interleave = var_1293_interleave_0, values = (var_1287_cast_fp16, var_1290_cast_fp16))[name = tensor<string, []>("op_1293_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, []>(57911360)))];
            tensor<fp16, [1, 8, 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_1297 = const()[name = tensor<string, []>("op_1297"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_25_cast_fp16 = reshape(shape = var_1297, 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, []>(58435712)))];
            tensor<fp16, [1, 78, 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_1301 = const()[name = tensor<string, []>("op_1301"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_17_cast_fp16 = reshape(shape = var_1301, 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, []>(58960064)))];
            tensor<fp16, [1, 78, 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_1305 = const()[name = tensor<string, []>("op_1305"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_9_cast_fp16 = reshape(shape = var_1305, 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, []>(59484416)))];
            tensor<fp16, [1, 8, 8, 64]> var_1317_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1317_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, []>(59485504)))];
            tensor<fp16, [1, 8, 8, 64]> var_1319_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1319_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, 155]> var_1321_to_fp16 = const()[name = tensor<string, []>("op_1321_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59486592)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1319_cast_fp16)[name = tensor<string, []>("transpose_201")];
            tensor<fp16, [1, 8, 8, 155]> 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_1321_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_131_to_fp16 = const()[name = tensor<string, []>("const_131_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_113_cast_fp16 = pad(constant_val = const_131_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_1329 = const()[name = tensor<string, []>("op_1329"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_115_cast_fp16 = reshape(shape = var_1329, x = x_113_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")];
            tensor<int32, [4]> var_1333_begin_0 = const()[name = tensor<string, []>("op_1333_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1333_end_0 = const()[name = tensor<string, []>("op_1333_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_1333_end_mask_0 = const()[name = tensor<string, []>("op_1333_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = x_115_cast_fp16)[name = tensor<string, []>("op_1333_cast_fp16")];
            tensor<int32, [4]> var_1334 = const()[name = tensor<string, []>("op_1334"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_17_cast_fp16 = reshape(shape = var_1334, x = var_1333_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, 78]> transpose_60 = transpose(perm = transpose_60_perm_0, x = k_17_cast_fp16)[name = tensor<string, []>("transpose_199")];
            tensor<fp16, [1, 8, 8, 64]> transpose_59 = transpose(perm = transpose_59_perm_0, x = var_1317_cast_fp16)[name = tensor<string, []>("transpose_200")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_1343_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = tensor<string, []>("op_1343_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, 8, 78]> _inversed_scores_17_cast_fp16 = mul(x = var_1343_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_17_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_19_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_19_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_1349_cast_fp16 = softmax(axis = var_56, x = scores_19_cast_fp16)[name = tensor<string, []>("op_1349_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_249_cast_fp16 = select(a = var_40_to_fp16, b = var_1349_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_1353_perm_0 = const()[name = tensor<string, []>("op_1353_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1354 = const()[name = tensor<string, []>("op_1354"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_1353_cast_fp16 = transpose(perm = var_1353_perm_0, x = x_117_cast_fp16)[name = tensor<string, []>("transpose_198")];
            tensor<fp16, [1, 8, 512]> input_251_cast_fp16 = reshape(shape = var_1354, x = var_1353_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, []>(59645376)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(60169728)))];
            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, []>(60170816)))];
            tensor<fp16, [1, 8, 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, []>(60171904)))];
            tensor<fp16, [1, 512, 8]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = tensor<string, []>("transpose_197")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_123_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_261_cast_fp16 = select(a = var_40_to_fp16, b = x_123_cast_fp16, cond = var_551)[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, 16]> new_x_19_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_1395_begin_0 = const()[name = tensor<string, []>("op_1395_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_1395_end_0 = const()[name = tensor<string, []>("op_1395_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_1395_end_mask_0 = const()[name = tensor<string, []>("op_1395_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_1395_cast_fp16 = slice_by_index(begin = var_1395_begin_0, end = var_1395_end_0, end_mask = var_1395_end_mask_0, x = next_cache_9_cast_fp16)[name = tensor<string, []>("op_1395_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, []>(61220544)))];
            tensor<fp16, [1, 512, 8]> 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, []>(61229824)))];
            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, []>(61230912)))];
            tensor<fp16, [1, 8, 512]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = tensor<string, []>("transpose_196")];
            tensor<fp16, [1, 8, 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, 8]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = tensor<string, []>("transpose_195")];
            tensor<fp16, [1, 512, 8]> 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, []>(61232000)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = tensor<string, []>("transpose_194")];
            tensor<fp16, [1, 8, 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, []>(61756352)))];
            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, []>(61757440)))];
            tensor<fp16, [1, 8, 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, []>(61758528)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(63855744)))];
            tensor<fp16, [1, 8, 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_1436_to_fp16 = const()[name = tensor<string, []>("op_1436_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1437_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1436_to_fp16)[name = tensor<string, []>("op_1437_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1437_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, []>(65952960)))];
            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, []>(65954048)))];
            tensor<fp16, [1, 8, 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, []>(65955136)))];
            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, []>(65956224)))];
            tensor<fp16, [1, 8, 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, []>(65957312)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(68054528)))];
            tensor<fp16, [1, 8, 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_1471_to_fp16 = const()[name = tensor<string, []>("op_1471_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1472_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1471_to_fp16)[name = tensor<string, []>("op_1472_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1472_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, []>(70151744)))];
            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, []>(70152832)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_299_cast_fp16 = concat(axis = var_65, 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_1494_begin_0 = const()[name = tensor<string, []>("op_1494_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1494_end_0 = const()[name = tensor<string, []>("op_1494_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1494_end_mask_0 = const()[name = tensor<string, []>("op_1494_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_1494_cast_fp16 = slice_by_index(begin = var_1494_begin_0, end = var_1494_end_0, end_mask = var_1494_end_mask_0, x = cache_21_cast_fp16)[name = tensor<string, []>("op_1494_cast_fp16")];
            tensor<int32, [3]> var_1497_begin_0 = const()[name = tensor<string, []>("op_1497_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1497_end_0 = const()[name = tensor<string, []>("op_1497_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_1497_end_mask_0 = const()[name = tensor<string, []>("op_1497_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_1497_cast_fp16 = slice_by_index(begin = var_1497_begin_0, end = var_1497_end_0, end_mask = var_1497_end_mask_0, x = key_11_cast_fp16)[name = tensor<string, []>("op_1497_cast_fp16")];
            tensor<bool, []> var_1500_interleave_0 = const()[name = tensor<string, []>("op_1500_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1500_cast_fp16 = concat(axis = var_65, interleave = var_1500_interleave_0, values = (var_1494_cast_fp16, var_1497_cast_fp16))[name = tensor<string, []>("op_1500_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, []>(70153920)))];
            tensor<fp16, [1, 8, 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_1504 = const()[name = tensor<string, []>("op_1504"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_31_cast_fp16 = reshape(shape = var_1504, 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, []>(70678272)))];
            tensor<fp16, [1, 78, 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_1508 = const()[name = tensor<string, []>("op_1508"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_21_cast_fp16 = reshape(shape = var_1508, 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, []>(71202624)))];
            tensor<fp16, [1, 78, 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_1512 = const()[name = tensor<string, []>("op_1512"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_11_cast_fp16 = reshape(shape = var_1512, 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, []>(71726976)))];
            tensor<fp16, [1, 8, 8, 64]> var_1524_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1524_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, []>(71728064)))];
            tensor<fp16, [1, 8, 8, 64]> var_1526_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1526_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, 155]> var_1528_to_fp16 = const()[name = tensor<string, []>("op_1528_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71729152)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1526_cast_fp16)[name = tensor<string, []>("transpose_192")];
            tensor<fp16, [1, 8, 8, 155]> 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_1528_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_144_to_fp16 = const()[name = tensor<string, []>("const_144_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_139_cast_fp16 = pad(constant_val = const_144_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_1536 = const()[name = tensor<string, []>("op_1536"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_141_cast_fp16 = reshape(shape = var_1536, x = x_139_cast_fp16)[name = tensor<string, []>("x_141_cast_fp16")];
            tensor<int32, [4]> var_1540_begin_0 = const()[name = tensor<string, []>("op_1540_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1540_end_0 = const()[name = tensor<string, []>("op_1540_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_1540_end_mask_0 = const()[name = tensor<string, []>("op_1540_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_1540_cast_fp16 = slice_by_index(begin = var_1540_begin_0, end = var_1540_end_0, end_mask = var_1540_end_mask_0, x = x_141_cast_fp16)[name = tensor<string, []>("op_1540_cast_fp16")];
            tensor<int32, [4]> var_1541 = const()[name = tensor<string, []>("op_1541"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_21_cast_fp16 = reshape(shape = var_1541, x = var_1540_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, 78]> transpose_62 = transpose(perm = transpose_62_perm_0, x = k_21_cast_fp16)[name = tensor<string, []>("transpose_190")];
            tensor<fp16, [1, 8, 8, 64]> transpose_61 = transpose(perm = transpose_61_perm_0, x = var_1524_cast_fp16)[name = tensor<string, []>("transpose_191")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_1550_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = tensor<string, []>("op_1550_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, 8, 78]> _inversed_scores_21_cast_fp16 = mul(x = var_1550_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_21_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_23_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_23_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_1556_cast_fp16 = softmax(axis = var_56, x = scores_23_cast_fp16)[name = tensor<string, []>("op_1556_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_301_cast_fp16 = select(a = var_40_to_fp16, b = var_1556_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_1560_perm_0 = const()[name = tensor<string, []>("op_1560_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1561 = const()[name = tensor<string, []>("op_1561"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_1560_cast_fp16 = transpose(perm = var_1560_perm_0, x = x_143_cast_fp16)[name = tensor<string, []>("transpose_189")];
            tensor<fp16, [1, 8, 512]> input_303_cast_fp16 = reshape(shape = var_1561, x = var_1560_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, []>(71887936)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(72412288)))];
            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, []>(72413376)))];
            tensor<fp16, [1, 8, 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, []>(72414464)))];
            tensor<fp16, [1, 512, 8]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = tensor<string, []>("transpose_188")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_149_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_313_cast_fp16 = select(a = var_40_to_fp16, b = x_149_cast_fp16, cond = var_551)[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, 16]> new_x_23_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_1602_begin_0 = const()[name = tensor<string, []>("op_1602_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_1602_end_0 = const()[name = tensor<string, []>("op_1602_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_1602_end_mask_0 = const()[name = tensor<string, []>("op_1602_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_1602_cast_fp16 = slice_by_index(begin = var_1602_begin_0, end = var_1602_end_0, end_mask = var_1602_end_mask_0, x = next_cache_11_cast_fp16)[name = tensor<string, []>("op_1602_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, []>(73463104)))];
            tensor<fp16, [1, 512, 8]> 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, []>(73472384)))];
            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, []>(73473472)))];
            tensor<fp16, [1, 8, 512]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = tensor<string, []>("transpose_187")];
            tensor<fp16, [1, 8, 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, 8]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = tensor<string, []>("transpose_186")];
            tensor<fp16, [1, 512, 8]> 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, []>(73474560)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = tensor<string, []>("transpose_185")];
            tensor<fp16, [1, 8, 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, []>(73998912)))];
            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, []>(74000000)))];
            tensor<fp16, [1, 8, 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, []>(74001088)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(76098304)))];
            tensor<fp16, [1, 8, 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_1643_to_fp16 = const()[name = tensor<string, []>("op_1643_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1644_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1643_to_fp16)[name = tensor<string, []>("op_1644_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1644_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, []>(78195520)))];
            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, []>(78196608)))];
            tensor<fp16, [1, 8, 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, []>(78197696)))];
            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, []>(78198784)))];
            tensor<fp16, [1, 8, 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, []>(78199872)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(80297088)))];
            tensor<fp16, [1, 8, 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_1678_to_fp16 = const()[name = tensor<string, []>("op_1678_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1679_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1678_to_fp16)[name = tensor<string, []>("op_1679_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1679_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, []>(82394304)))];
            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, []>(82395392)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_351_cast_fp16 = concat(axis = var_65, 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_1701_begin_0 = const()[name = tensor<string, []>("op_1701_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1701_end_0 = const()[name = tensor<string, []>("op_1701_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1701_end_mask_0 = const()[name = tensor<string, []>("op_1701_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_1701_cast_fp16 = slice_by_index(begin = var_1701_begin_0, end = var_1701_end_0, end_mask = var_1701_end_mask_0, x = cache_25_cast_fp16)[name = tensor<string, []>("op_1701_cast_fp16")];
            tensor<int32, [3]> var_1704_begin_0 = const()[name = tensor<string, []>("op_1704_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1704_end_0 = const()[name = tensor<string, []>("op_1704_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_1704_end_mask_0 = const()[name = tensor<string, []>("op_1704_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_1704_cast_fp16 = slice_by_index(begin = var_1704_begin_0, end = var_1704_end_0, end_mask = var_1704_end_mask_0, x = key_13_cast_fp16)[name = tensor<string, []>("op_1704_cast_fp16")];
            tensor<bool, []> var_1707_interleave_0 = const()[name = tensor<string, []>("op_1707_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1707_cast_fp16 = concat(axis = var_65, interleave = var_1707_interleave_0, values = (var_1701_cast_fp16, var_1704_cast_fp16))[name = tensor<string, []>("op_1707_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, []>(82396480)))];
            tensor<fp16, [1, 8, 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_1711 = const()[name = tensor<string, []>("op_1711"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_37_cast_fp16 = reshape(shape = var_1711, 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, []>(82920832)))];
            tensor<fp16, [1, 78, 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_1715 = const()[name = tensor<string, []>("op_1715"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_25_cast_fp16 = reshape(shape = var_1715, 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, []>(83445184)))];
            tensor<fp16, [1, 78, 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_1719 = const()[name = tensor<string, []>("op_1719"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_13_cast_fp16 = reshape(shape = var_1719, 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, []>(83969536)))];
            tensor<fp16, [1, 8, 8, 64]> var_1731_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1731_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, []>(83970624)))];
            tensor<fp16, [1, 8, 8, 64]> var_1733_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1733_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, 155]> var_1735_to_fp16 = const()[name = tensor<string, []>("op_1735_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83971712)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1733_cast_fp16)[name = tensor<string, []>("transpose_183")];
            tensor<fp16, [1, 8, 8, 155]> 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_1735_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_157_to_fp16 = const()[name = tensor<string, []>("const_157_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_165_cast_fp16 = pad(constant_val = const_157_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_1743 = const()[name = tensor<string, []>("op_1743"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_167_cast_fp16 = reshape(shape = var_1743, x = x_165_cast_fp16)[name = tensor<string, []>("x_167_cast_fp16")];
            tensor<int32, [4]> var_1747_begin_0 = const()[name = tensor<string, []>("op_1747_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1747_end_0 = const()[name = tensor<string, []>("op_1747_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_1747_end_mask_0 = const()[name = tensor<string, []>("op_1747_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_1747_cast_fp16 = slice_by_index(begin = var_1747_begin_0, end = var_1747_end_0, end_mask = var_1747_end_mask_0, x = x_167_cast_fp16)[name = tensor<string, []>("op_1747_cast_fp16")];
            tensor<int32, [4]> var_1748 = const()[name = tensor<string, []>("op_1748"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_25_cast_fp16 = reshape(shape = var_1748, x = var_1747_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, 78]> transpose_64 = transpose(perm = transpose_64_perm_0, x = k_25_cast_fp16)[name = tensor<string, []>("transpose_181")];
            tensor<fp16, [1, 8, 8, 64]> transpose_63 = transpose(perm = transpose_63_perm_0, x = var_1731_cast_fp16)[name = tensor<string, []>("transpose_182")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_1757_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = tensor<string, []>("op_1757_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, 8, 78]> _inversed_scores_25_cast_fp16 = mul(x = var_1757_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_25_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_27_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_27_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_1763_cast_fp16 = softmax(axis = var_56, x = scores_27_cast_fp16)[name = tensor<string, []>("op_1763_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_353_cast_fp16 = select(a = var_40_to_fp16, b = var_1763_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_1767_perm_0 = const()[name = tensor<string, []>("op_1767_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1768 = const()[name = tensor<string, []>("op_1768"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = x_169_cast_fp16)[name = tensor<string, []>("transpose_180")];
            tensor<fp16, [1, 8, 512]> input_355_cast_fp16 = reshape(shape = var_1768, x = var_1767_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, []>(84130496)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(84654848)))];
            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, []>(84655936)))];
            tensor<fp16, [1, 8, 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, []>(84657024)))];
            tensor<fp16, [1, 512, 8]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = tensor<string, []>("transpose_179")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_175_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_365_cast_fp16 = select(a = var_40_to_fp16, b = x_175_cast_fp16, cond = var_551)[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, 16]> new_x_27_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_1809_begin_0 = const()[name = tensor<string, []>("op_1809_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_1809_end_0 = const()[name = tensor<string, []>("op_1809_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_1809_end_mask_0 = const()[name = tensor<string, []>("op_1809_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_1809_cast_fp16 = slice_by_index(begin = var_1809_begin_0, end = var_1809_end_0, end_mask = var_1809_end_mask_0, x = next_cache_13_cast_fp16)[name = tensor<string, []>("op_1809_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, []>(85705664)))];
            tensor<fp16, [1, 512, 8]> 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, []>(85714944)))];
            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, []>(85716032)))];
            tensor<fp16, [1, 8, 512]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = tensor<string, []>("transpose_178")];
            tensor<fp16, [1, 8, 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, 8]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = tensor<string, []>("transpose_177")];
            tensor<fp16, [1, 512, 8]> 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, []>(85717120)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = tensor<string, []>("transpose_176")];
            tensor<fp16, [1, 8, 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, []>(86241472)))];
            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, []>(86242560)))];
            tensor<fp16, [1, 8, 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, []>(86243648)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(88340864)))];
            tensor<fp16, [1, 8, 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_1850_to_fp16 = const()[name = tensor<string, []>("op_1850_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1851_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1850_to_fp16)[name = tensor<string, []>("op_1851_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1851_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, []>(90438080)))];
            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, []>(90439168)))];
            tensor<fp16, [1, 8, 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, []>(90440256)))];
            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, []>(90441344)))];
            tensor<fp16, [1, 8, 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, []>(90442432)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(92539648)))];
            tensor<fp16, [1, 8, 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_1885_to_fp16 = const()[name = tensor<string, []>("op_1885_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_1886_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1885_to_fp16)[name = tensor<string, []>("op_1886_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1886_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, []>(94636864)))];
            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, []>(94637952)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_403_cast_fp16 = concat(axis = var_65, 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_1908_begin_0 = const()[name = tensor<string, []>("op_1908_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1908_end_0 = const()[name = tensor<string, []>("op_1908_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_1908_end_mask_0 = const()[name = tensor<string, []>("op_1908_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_1908_cast_fp16 = slice_by_index(begin = var_1908_begin_0, end = var_1908_end_0, end_mask = var_1908_end_mask_0, x = cache_29_cast_fp16)[name = tensor<string, []>("op_1908_cast_fp16")];
            tensor<int32, [3]> var_1911_begin_0 = const()[name = tensor<string, []>("op_1911_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1911_end_0 = const()[name = tensor<string, []>("op_1911_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_1911_end_mask_0 = const()[name = tensor<string, []>("op_1911_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_1911_cast_fp16 = slice_by_index(begin = var_1911_begin_0, end = var_1911_end_0, end_mask = var_1911_end_mask_0, x = key_15_cast_fp16)[name = tensor<string, []>("op_1911_cast_fp16")];
            tensor<bool, []> var_1914_interleave_0 = const()[name = tensor<string, []>("op_1914_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_1914_cast_fp16 = concat(axis = var_65, interleave = var_1914_interleave_0, values = (var_1908_cast_fp16, var_1911_cast_fp16))[name = tensor<string, []>("op_1914_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, []>(94639040)))];
            tensor<fp16, [1, 8, 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_1918 = const()[name = tensor<string, []>("op_1918"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_43_cast_fp16 = reshape(shape = var_1918, 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, []>(95163392)))];
            tensor<fp16, [1, 78, 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_1922 = const()[name = tensor<string, []>("op_1922"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_29_cast_fp16 = reshape(shape = var_1922, 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, []>(95687744)))];
            tensor<fp16, [1, 78, 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_1926 = const()[name = tensor<string, []>("op_1926"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_15_cast_fp16 = reshape(shape = var_1926, 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, []>(96212096)))];
            tensor<fp16, [1, 8, 8, 64]> var_1938_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1938_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, []>(96213184)))];
            tensor<fp16, [1, 8, 8, 64]> var_1940_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1940_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, 155]> var_1942_to_fp16 = const()[name = tensor<string, []>("op_1942_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96214272)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_1940_cast_fp16)[name = tensor<string, []>("transpose_174")];
            tensor<fp16, [1, 8, 8, 155]> 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_1942_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_170_to_fp16 = const()[name = tensor<string, []>("const_170_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_191_cast_fp16 = pad(constant_val = const_170_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_1950 = const()[name = tensor<string, []>("op_1950"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_193_cast_fp16 = reshape(shape = var_1950, x = x_191_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
            tensor<int32, [4]> var_1954_begin_0 = const()[name = tensor<string, []>("op_1954_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1954_end_0 = const()[name = tensor<string, []>("op_1954_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_1954_end_mask_0 = const()[name = tensor<string, []>("op_1954_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_1954_cast_fp16 = slice_by_index(begin = var_1954_begin_0, end = var_1954_end_0, end_mask = var_1954_end_mask_0, x = x_193_cast_fp16)[name = tensor<string, []>("op_1954_cast_fp16")];
            tensor<int32, [4]> var_1955 = const()[name = tensor<string, []>("op_1955"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_29_cast_fp16 = reshape(shape = var_1955, x = var_1954_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, 78]> transpose_66 = transpose(perm = transpose_66_perm_0, x = k_29_cast_fp16)[name = tensor<string, []>("transpose_172")];
            tensor<fp16, [1, 8, 8, 64]> transpose_65 = transpose(perm = transpose_65_perm_0, x = var_1938_cast_fp16)[name = tensor<string, []>("transpose_173")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_1964_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = tensor<string, []>("op_1964_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, 8, 78]> _inversed_scores_29_cast_fp16 = mul(x = var_1964_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_29_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_31_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_31_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_1970_cast_fp16 = softmax(axis = var_56, x = scores_31_cast_fp16)[name = tensor<string, []>("op_1970_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_405_cast_fp16 = select(a = var_40_to_fp16, b = var_1970_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_1974_perm_0 = const()[name = tensor<string, []>("op_1974_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1975 = const()[name = tensor<string, []>("op_1975"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_1974_cast_fp16 = transpose(perm = var_1974_perm_0, x = x_195_cast_fp16)[name = tensor<string, []>("transpose_171")];
            tensor<fp16, [1, 8, 512]> input_407_cast_fp16 = reshape(shape = var_1975, x = var_1974_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, []>(96373056)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(96897408)))];
            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, []>(96898496)))];
            tensor<fp16, [1, 8, 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, []>(96899584)))];
            tensor<fp16, [1, 512, 8]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = tensor<string, []>("transpose_170")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_201_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_417_cast_fp16 = select(a = var_40_to_fp16, b = x_201_cast_fp16, cond = var_551)[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, 16]> new_x_31_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_2016_begin_0 = const()[name = tensor<string, []>("op_2016_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2016_end_0 = const()[name = tensor<string, []>("op_2016_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_2016_end_mask_0 = const()[name = tensor<string, []>("op_2016_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2016_cast_fp16 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, x = next_cache_15_cast_fp16)[name = tensor<string, []>("op_2016_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, []>(97948224)))];
            tensor<fp16, [1, 512, 8]> 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, []>(97957504)))];
            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, []>(97958592)))];
            tensor<fp16, [1, 8, 512]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = tensor<string, []>("transpose_169")];
            tensor<fp16, [1, 8, 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, 8]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = tensor<string, []>("transpose_168")];
            tensor<fp16, [1, 512, 8]> 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, []>(97959680)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = tensor<string, []>("transpose_167")];
            tensor<fp16, [1, 8, 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, []>(98484032)))];
            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, []>(98485120)))];
            tensor<fp16, [1, 8, 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, []>(98486208)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(100583424)))];
            tensor<fp16, [1, 8, 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_2057_to_fp16 = const()[name = tensor<string, []>("op_2057_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2058_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2057_to_fp16)[name = tensor<string, []>("op_2058_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2058_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, []>(102680640)))];
            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, []>(102681728)))];
            tensor<fp16, [1, 8, 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, []>(102682816)))];
            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, []>(102683904)))];
            tensor<fp16, [1, 8, 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, []>(102684992)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(104782208)))];
            tensor<fp16, [1, 8, 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_2092_to_fp16 = const()[name = tensor<string, []>("op_2092_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2093_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2092_to_fp16)[name = tensor<string, []>("op_2093_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2093_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, []>(106879424)))];
            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, []>(106880512)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_455_cast_fp16 = concat(axis = var_65, 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_2115_begin_0 = const()[name = tensor<string, []>("op_2115_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2115_end_0 = const()[name = tensor<string, []>("op_2115_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_2115_end_mask_0 = const()[name = tensor<string, []>("op_2115_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_2115_cast_fp16 = slice_by_index(begin = var_2115_begin_0, end = var_2115_end_0, end_mask = var_2115_end_mask_0, x = cache_33_cast_fp16)[name = tensor<string, []>("op_2115_cast_fp16")];
            tensor<int32, [3]> var_2118_begin_0 = const()[name = tensor<string, []>("op_2118_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2118_end_0 = const()[name = tensor<string, []>("op_2118_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_2118_end_mask_0 = const()[name = tensor<string, []>("op_2118_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_2118_cast_fp16 = slice_by_index(begin = var_2118_begin_0, end = var_2118_end_0, end_mask = var_2118_end_mask_0, x = key_17_cast_fp16)[name = tensor<string, []>("op_2118_cast_fp16")];
            tensor<bool, []> var_2121_interleave_0 = const()[name = tensor<string, []>("op_2121_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_2121_cast_fp16 = concat(axis = var_65, interleave = var_2121_interleave_0, values = (var_2115_cast_fp16, var_2118_cast_fp16))[name = tensor<string, []>("op_2121_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, []>(106881600)))];
            tensor<fp16, [1, 8, 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_2125 = const()[name = tensor<string, []>("op_2125"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_49_cast_fp16 = reshape(shape = var_2125, 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, []>(107405952)))];
            tensor<fp16, [1, 78, 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_2129 = const()[name = tensor<string, []>("op_2129"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_33_cast_fp16 = reshape(shape = var_2129, 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, []>(107930304)))];
            tensor<fp16, [1, 78, 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_2133 = const()[name = tensor<string, []>("op_2133"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_17_cast_fp16 = reshape(shape = var_2133, 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, []>(108454656)))];
            tensor<fp16, [1, 8, 8, 64]> var_2145_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2145_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, []>(108455744)))];
            tensor<fp16, [1, 8, 8, 64]> var_2147_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2147_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, 155]> var_2149_to_fp16 = const()[name = tensor<string, []>("op_2149_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108456832)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2147_cast_fp16)[name = tensor<string, []>("transpose_165")];
            tensor<fp16, [1, 8, 8, 155]> 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_2149_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_183_to_fp16 = const()[name = tensor<string, []>("const_183_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_217_cast_fp16 = pad(constant_val = const_183_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_2157 = const()[name = tensor<string, []>("op_2157"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_219_cast_fp16 = reshape(shape = var_2157, x = x_217_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
            tensor<int32, [4]> var_2161_begin_0 = const()[name = tensor<string, []>("op_2161_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2161_end_0 = const()[name = tensor<string, []>("op_2161_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_2161_end_mask_0 = const()[name = tensor<string, []>("op_2161_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_2161_cast_fp16 = slice_by_index(begin = var_2161_begin_0, end = var_2161_end_0, end_mask = var_2161_end_mask_0, x = x_219_cast_fp16)[name = tensor<string, []>("op_2161_cast_fp16")];
            tensor<int32, [4]> var_2162 = const()[name = tensor<string, []>("op_2162"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_33_cast_fp16 = reshape(shape = var_2162, x = var_2161_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, 78]> transpose_68 = transpose(perm = transpose_68_perm_0, x = k_33_cast_fp16)[name = tensor<string, []>("transpose_163")];
            tensor<fp16, [1, 8, 8, 64]> transpose_67 = transpose(perm = transpose_67_perm_0, x = var_2145_cast_fp16)[name = tensor<string, []>("transpose_164")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_2171_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = tensor<string, []>("op_2171_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, 8, 78]> _inversed_scores_33_cast_fp16 = mul(x = var_2171_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_33_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_35_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_35_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_2177_cast_fp16 = softmax(axis = var_56, x = scores_35_cast_fp16)[name = tensor<string, []>("op_2177_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_457_cast_fp16 = select(a = var_40_to_fp16, b = var_2177_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_2181_perm_0 = const()[name = tensor<string, []>("op_2181_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2182 = const()[name = tensor<string, []>("op_2182"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_2181_cast_fp16 = transpose(perm = var_2181_perm_0, x = x_221_cast_fp16)[name = tensor<string, []>("transpose_162")];
            tensor<fp16, [1, 8, 512]> input_459_cast_fp16 = reshape(shape = var_2182, x = var_2181_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, []>(108615616)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(109139968)))];
            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, []>(109141056)))];
            tensor<fp16, [1, 8, 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, []>(109142144)))];
            tensor<fp16, [1, 512, 8]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = tensor<string, []>("transpose_161")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_227_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_469_cast_fp16 = select(a = var_40_to_fp16, b = x_227_cast_fp16, cond = var_551)[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, 16]> new_x_35_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_2223_begin_0 = const()[name = tensor<string, []>("op_2223_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2223_end_0 = const()[name = tensor<string, []>("op_2223_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_2223_end_mask_0 = const()[name = tensor<string, []>("op_2223_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2223_cast_fp16 = slice_by_index(begin = var_2223_begin_0, end = var_2223_end_0, end_mask = var_2223_end_mask_0, x = next_cache_17_cast_fp16)[name = tensor<string, []>("op_2223_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, []>(110190784)))];
            tensor<fp16, [1, 512, 8]> 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, []>(110200064)))];
            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, []>(110201152)))];
            tensor<fp16, [1, 8, 512]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = tensor<string, []>("transpose_160")];
            tensor<fp16, [1, 8, 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, 8]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = tensor<string, []>("transpose_159")];
            tensor<fp16, [1, 512, 8]> 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, []>(110202240)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = tensor<string, []>("transpose_158")];
            tensor<fp16, [1, 8, 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, []>(110726592)))];
            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, []>(110727680)))];
            tensor<fp16, [1, 8, 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, []>(110728768)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(112825984)))];
            tensor<fp16, [1, 8, 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_2264_to_fp16 = const()[name = tensor<string, []>("op_2264_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2265_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2264_to_fp16)[name = tensor<string, []>("op_2265_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2265_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, []>(114923200)))];
            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, []>(114924288)))];
            tensor<fp16, [1, 8, 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, []>(114925376)))];
            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, []>(114926464)))];
            tensor<fp16, [1, 8, 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, []>(114927552)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(117024768)))];
            tensor<fp16, [1, 8, 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_2299_to_fp16 = const()[name = tensor<string, []>("op_2299_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2300_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2299_to_fp16)[name = tensor<string, []>("op_2300_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2300_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, []>(119121984)))];
            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, []>(119123072)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_507_cast_fp16 = concat(axis = var_65, 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_2322_begin_0 = const()[name = tensor<string, []>("op_2322_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2322_end_0 = const()[name = tensor<string, []>("op_2322_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_2322_end_mask_0 = const()[name = tensor<string, []>("op_2322_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_2322_cast_fp16 = slice_by_index(begin = var_2322_begin_0, end = var_2322_end_0, end_mask = var_2322_end_mask_0, x = cache_37_cast_fp16)[name = tensor<string, []>("op_2322_cast_fp16")];
            tensor<int32, [3]> var_2325_begin_0 = const()[name = tensor<string, []>("op_2325_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2325_end_0 = const()[name = tensor<string, []>("op_2325_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_2325_end_mask_0 = const()[name = tensor<string, []>("op_2325_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_2325_cast_fp16 = slice_by_index(begin = var_2325_begin_0, end = var_2325_end_0, end_mask = var_2325_end_mask_0, x = key_19_cast_fp16)[name = tensor<string, []>("op_2325_cast_fp16")];
            tensor<bool, []> var_2328_interleave_0 = const()[name = tensor<string, []>("op_2328_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_2328_cast_fp16 = concat(axis = var_65, interleave = var_2328_interleave_0, values = (var_2322_cast_fp16, var_2325_cast_fp16))[name = tensor<string, []>("op_2328_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, []>(119124160)))];
            tensor<fp16, [1, 8, 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_2332 = const()[name = tensor<string, []>("op_2332"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_55_cast_fp16 = reshape(shape = var_2332, 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, []>(119648512)))];
            tensor<fp16, [1, 78, 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_2336 = const()[name = tensor<string, []>("op_2336"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_37_cast_fp16 = reshape(shape = var_2336, 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, []>(120172864)))];
            tensor<fp16, [1, 78, 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_2340 = const()[name = tensor<string, []>("op_2340"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_19_cast_fp16 = reshape(shape = var_2340, 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, []>(120697216)))];
            tensor<fp16, [1, 8, 8, 64]> var_2352_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2352_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, []>(120698304)))];
            tensor<fp16, [1, 8, 8, 64]> var_2354_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2354_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, 155]> var_2356_to_fp16 = const()[name = tensor<string, []>("op_2356_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120699392)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2354_cast_fp16)[name = tensor<string, []>("transpose_156")];
            tensor<fp16, [1, 8, 8, 155]> 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_2356_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_196_to_fp16 = const()[name = tensor<string, []>("const_196_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_243_cast_fp16 = pad(constant_val = const_196_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_2364 = const()[name = tensor<string, []>("op_2364"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_245_cast_fp16 = reshape(shape = var_2364, x = x_243_cast_fp16)[name = tensor<string, []>("x_245_cast_fp16")];
            tensor<int32, [4]> var_2368_begin_0 = const()[name = tensor<string, []>("op_2368_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2368_end_0 = const()[name = tensor<string, []>("op_2368_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_2368_end_mask_0 = const()[name = tensor<string, []>("op_2368_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_2368_cast_fp16 = slice_by_index(begin = var_2368_begin_0, end = var_2368_end_0, end_mask = var_2368_end_mask_0, x = x_245_cast_fp16)[name = tensor<string, []>("op_2368_cast_fp16")];
            tensor<int32, [4]> var_2369 = const()[name = tensor<string, []>("op_2369"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_37_cast_fp16 = reshape(shape = var_2369, x = var_2368_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, 78]> transpose_70 = transpose(perm = transpose_70_perm_0, x = k_37_cast_fp16)[name = tensor<string, []>("transpose_154")];
            tensor<fp16, [1, 8, 8, 64]> transpose_69 = transpose(perm = transpose_69_perm_0, x = var_2352_cast_fp16)[name = tensor<string, []>("transpose_155")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_2378_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = tensor<string, []>("op_2378_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, 8, 78]> _inversed_scores_37_cast_fp16 = mul(x = var_2378_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_37_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_39_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_39_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_2384_cast_fp16 = softmax(axis = var_56, x = scores_39_cast_fp16)[name = tensor<string, []>("op_2384_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_509_cast_fp16 = select(a = var_40_to_fp16, b = var_2384_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_2388_perm_0 = const()[name = tensor<string, []>("op_2388_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2389 = const()[name = tensor<string, []>("op_2389"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_2388_cast_fp16 = transpose(perm = var_2388_perm_0, x = x_247_cast_fp16)[name = tensor<string, []>("transpose_153")];
            tensor<fp16, [1, 8, 512]> input_511_cast_fp16 = reshape(shape = var_2389, x = var_2388_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, []>(120858176)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(121382528)))];
            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, []>(121383616)))];
            tensor<fp16, [1, 8, 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, []>(121384704)))];
            tensor<fp16, [1, 512, 8]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = tensor<string, []>("transpose_152")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_253_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_521_cast_fp16 = select(a = var_40_to_fp16, b = x_253_cast_fp16, cond = var_551)[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, 16]> new_x_39_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_2430_begin_0 = const()[name = tensor<string, []>("op_2430_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2430_end_0 = const()[name = tensor<string, []>("op_2430_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_2430_end_mask_0 = const()[name = tensor<string, []>("op_2430_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2430_cast_fp16 = slice_by_index(begin = var_2430_begin_0, end = var_2430_end_0, end_mask = var_2430_end_mask_0, x = next_cache_19_cast_fp16)[name = tensor<string, []>("op_2430_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, []>(122433344)))];
            tensor<fp16, [1, 512, 8]> 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, []>(122442624)))];
            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, []>(122443712)))];
            tensor<fp16, [1, 8, 512]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = tensor<string, []>("transpose_151")];
            tensor<fp16, [1, 8, 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, 8]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = tensor<string, []>("transpose_150")];
            tensor<fp16, [1, 512, 8]> 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, []>(122444800)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = tensor<string, []>("transpose_149")];
            tensor<fp16, [1, 8, 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, []>(122969152)))];
            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, []>(122970240)))];
            tensor<fp16, [1, 8, 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, []>(122971328)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(125068544)))];
            tensor<fp16, [1, 8, 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_2471_to_fp16 = const()[name = tensor<string, []>("op_2471_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2472_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2471_to_fp16)[name = tensor<string, []>("op_2472_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2472_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, []>(127165760)))];
            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, []>(127166848)))];
            tensor<fp16, [1, 8, 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, []>(127167936)))];
            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, []>(127169024)))];
            tensor<fp16, [1, 8, 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, []>(127170112)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(129267328)))];
            tensor<fp16, [1, 8, 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_2506_to_fp16 = const()[name = tensor<string, []>("op_2506_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2507_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2506_to_fp16)[name = tensor<string, []>("op_2507_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2507_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, []>(131364544)))];
            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, []>(131365632)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_559_cast_fp16 = concat(axis = var_65, 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_2529_begin_0 = const()[name = tensor<string, []>("op_2529_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2529_end_0 = const()[name = tensor<string, []>("op_2529_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_2529_end_mask_0 = const()[name = tensor<string, []>("op_2529_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_2529_cast_fp16 = slice_by_index(begin = var_2529_begin_0, end = var_2529_end_0, end_mask = var_2529_end_mask_0, x = cache_41_cast_fp16)[name = tensor<string, []>("op_2529_cast_fp16")];
            tensor<int32, [3]> var_2532_begin_0 = const()[name = tensor<string, []>("op_2532_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2532_end_0 = const()[name = tensor<string, []>("op_2532_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_2532_end_mask_0 = const()[name = tensor<string, []>("op_2532_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_2532_cast_fp16 = slice_by_index(begin = var_2532_begin_0, end = var_2532_end_0, end_mask = var_2532_end_mask_0, x = key_21_cast_fp16)[name = tensor<string, []>("op_2532_cast_fp16")];
            tensor<bool, []> var_2535_interleave_0 = const()[name = tensor<string, []>("op_2535_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_2535_cast_fp16 = concat(axis = var_65, interleave = var_2535_interleave_0, values = (var_2529_cast_fp16, var_2532_cast_fp16))[name = tensor<string, []>("op_2535_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, []>(131366720)))];
            tensor<fp16, [1, 8, 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_2539 = const()[name = tensor<string, []>("op_2539"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_61_cast_fp16 = reshape(shape = var_2539, 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, []>(131891072)))];
            tensor<fp16, [1, 78, 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_2543 = const()[name = tensor<string, []>("op_2543"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_41_cast_fp16 = reshape(shape = var_2543, 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, []>(132415424)))];
            tensor<fp16, [1, 78, 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_2547 = const()[name = tensor<string, []>("op_2547"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_21_cast_fp16 = reshape(shape = var_2547, 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, []>(132939776)))];
            tensor<fp16, [1, 8, 8, 64]> var_2559_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2559_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, []>(132940864)))];
            tensor<fp16, [1, 8, 8, 64]> var_2561_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2561_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, 155]> var_2563_to_fp16 = const()[name = tensor<string, []>("op_2563_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132941952)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2561_cast_fp16)[name = tensor<string, []>("transpose_147")];
            tensor<fp16, [1, 8, 8, 155]> 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_2563_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_209_to_fp16 = const()[name = tensor<string, []>("const_209_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_269_cast_fp16 = pad(constant_val = const_209_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_2571 = const()[name = tensor<string, []>("op_2571"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_271_cast_fp16 = reshape(shape = var_2571, x = x_269_cast_fp16)[name = tensor<string, []>("x_271_cast_fp16")];
            tensor<int32, [4]> var_2575_begin_0 = const()[name = tensor<string, []>("op_2575_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2575_end_0 = const()[name = tensor<string, []>("op_2575_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_2575_end_mask_0 = const()[name = tensor<string, []>("op_2575_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_2575_cast_fp16 = slice_by_index(begin = var_2575_begin_0, end = var_2575_end_0, end_mask = var_2575_end_mask_0, x = x_271_cast_fp16)[name = tensor<string, []>("op_2575_cast_fp16")];
            tensor<int32, [4]> var_2576 = const()[name = tensor<string, []>("op_2576"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_41_cast_fp16 = reshape(shape = var_2576, x = var_2575_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, 78]> transpose_72 = transpose(perm = transpose_72_perm_0, x = k_41_cast_fp16)[name = tensor<string, []>("transpose_145")];
            tensor<fp16, [1, 8, 8, 64]> transpose_71 = transpose(perm = transpose_71_perm_0, x = var_2559_cast_fp16)[name = tensor<string, []>("transpose_146")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_2585_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = tensor<string, []>("op_2585_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, 8, 78]> _inversed_scores_41_cast_fp16 = mul(x = var_2585_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_41_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_43_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_43_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_2591_cast_fp16 = softmax(axis = var_56, x = scores_43_cast_fp16)[name = tensor<string, []>("op_2591_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_561_cast_fp16 = select(a = var_40_to_fp16, b = var_2591_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_2595_perm_0 = const()[name = tensor<string, []>("op_2595_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2596 = const()[name = tensor<string, []>("op_2596"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_2595_cast_fp16 = transpose(perm = var_2595_perm_0, x = x_273_cast_fp16)[name = tensor<string, []>("transpose_144")];
            tensor<fp16, [1, 8, 512]> input_563_cast_fp16 = reshape(shape = var_2596, x = var_2595_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, []>(133100736)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(133625088)))];
            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, []>(133626176)))];
            tensor<fp16, [1, 8, 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, []>(133627264)))];
            tensor<fp16, [1, 512, 8]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = tensor<string, []>("transpose_143")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_279_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_573_cast_fp16 = select(a = var_40_to_fp16, b = x_279_cast_fp16, cond = var_551)[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, 16]> new_x_43_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_2637_begin_0 = const()[name = tensor<string, []>("op_2637_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2637_end_0 = const()[name = tensor<string, []>("op_2637_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_2637_end_mask_0 = const()[name = tensor<string, []>("op_2637_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2637_cast_fp16 = slice_by_index(begin = var_2637_begin_0, end = var_2637_end_0, end_mask = var_2637_end_mask_0, x = next_cache_21_cast_fp16)[name = tensor<string, []>("op_2637_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, []>(134675904)))];
            tensor<fp16, [1, 512, 8]> 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, []>(134685184)))];
            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, []>(134686272)))];
            tensor<fp16, [1, 8, 512]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = tensor<string, []>("transpose_142")];
            tensor<fp16, [1, 8, 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, 8]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = tensor<string, []>("transpose_141")];
            tensor<fp16, [1, 512, 8]> 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, []>(134687360)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = tensor<string, []>("transpose_140")];
            tensor<fp16, [1, 8, 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, []>(135211712)))];
            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, []>(135212800)))];
            tensor<fp16, [1, 8, 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, []>(135213888)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(137311104)))];
            tensor<fp16, [1, 8, 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_2678_to_fp16 = const()[name = tensor<string, []>("op_2678_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2679_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2678_to_fp16)[name = tensor<string, []>("op_2679_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2679_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, []>(139408320)))];
            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, []>(139409408)))];
            tensor<fp16, [1, 8, 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, []>(139410496)))];
            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, []>(139411584)))];
            tensor<fp16, [1, 8, 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, []>(139412672)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(141509888)))];
            tensor<fp16, [1, 8, 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_2713_to_fp16 = const()[name = tensor<string, []>("op_2713_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2714_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2713_to_fp16)[name = tensor<string, []>("op_2714_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2714_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, []>(143607104)))];
            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, []>(143608192)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_611_cast_fp16 = concat(axis = var_65, 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_2736_begin_0 = const()[name = tensor<string, []>("op_2736_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2736_end_0 = const()[name = tensor<string, []>("op_2736_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_2736_end_mask_0 = const()[name = tensor<string, []>("op_2736_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_2736_cast_fp16 = slice_by_index(begin = var_2736_begin_0, end = var_2736_end_0, end_mask = var_2736_end_mask_0, x = cache_45_cast_fp16)[name = tensor<string, []>("op_2736_cast_fp16")];
            tensor<int32, [3]> var_2739_begin_0 = const()[name = tensor<string, []>("op_2739_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2739_end_0 = const()[name = tensor<string, []>("op_2739_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_2739_end_mask_0 = const()[name = tensor<string, []>("op_2739_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_2739_cast_fp16 = slice_by_index(begin = var_2739_begin_0, end = var_2739_end_0, end_mask = var_2739_end_mask_0, x = key_23_cast_fp16)[name = tensor<string, []>("op_2739_cast_fp16")];
            tensor<bool, []> var_2742_interleave_0 = const()[name = tensor<string, []>("op_2742_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_2742_cast_fp16 = concat(axis = var_65, interleave = var_2742_interleave_0, values = (var_2736_cast_fp16, var_2739_cast_fp16))[name = tensor<string, []>("op_2742_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, []>(143609280)))];
            tensor<fp16, [1, 8, 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_2746 = const()[name = tensor<string, []>("op_2746"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_67_cast_fp16 = reshape(shape = var_2746, 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, []>(144133632)))];
            tensor<fp16, [1, 78, 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_2750 = const()[name = tensor<string, []>("op_2750"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_45_cast_fp16 = reshape(shape = var_2750, 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, []>(144657984)))];
            tensor<fp16, [1, 78, 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_2754 = const()[name = tensor<string, []>("op_2754"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_23_cast_fp16 = reshape(shape = var_2754, 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, []>(145182336)))];
            tensor<fp16, [1, 8, 8, 64]> var_2766_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2766_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, []>(145183424)))];
            tensor<fp16, [1, 8, 8, 64]> var_2768_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2768_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, 155]> var_2770_to_fp16 = const()[name = tensor<string, []>("op_2770_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145184512)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2768_cast_fp16)[name = tensor<string, []>("transpose_138")];
            tensor<fp16, [1, 8, 8, 155]> 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_2770_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_222_to_fp16 = const()[name = tensor<string, []>("const_222_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_295_cast_fp16 = pad(constant_val = const_222_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_2778 = const()[name = tensor<string, []>("op_2778"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_297_cast_fp16 = reshape(shape = var_2778, x = x_295_cast_fp16)[name = tensor<string, []>("x_297_cast_fp16")];
            tensor<int32, [4]> var_2782_begin_0 = const()[name = tensor<string, []>("op_2782_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2782_end_0 = const()[name = tensor<string, []>("op_2782_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_2782_end_mask_0 = const()[name = tensor<string, []>("op_2782_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_2782_cast_fp16 = slice_by_index(begin = var_2782_begin_0, end = var_2782_end_0, end_mask = var_2782_end_mask_0, x = x_297_cast_fp16)[name = tensor<string, []>("op_2782_cast_fp16")];
            tensor<int32, [4]> var_2783 = const()[name = tensor<string, []>("op_2783"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_45_cast_fp16 = reshape(shape = var_2783, x = var_2782_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, 78]> transpose_74 = transpose(perm = transpose_74_perm_0, x = k_45_cast_fp16)[name = tensor<string, []>("transpose_136")];
            tensor<fp16, [1, 8, 8, 64]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_2766_cast_fp16)[name = tensor<string, []>("transpose_137")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_2792_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = tensor<string, []>("op_2792_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, 8, 78]> _inversed_scores_45_cast_fp16 = mul(x = var_2792_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_45_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_47_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_47_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_2798_cast_fp16 = softmax(axis = var_56, x = scores_47_cast_fp16)[name = tensor<string, []>("op_2798_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_613_cast_fp16 = select(a = var_40_to_fp16, b = var_2798_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_2802_perm_0 = const()[name = tensor<string, []>("op_2802_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2803 = const()[name = tensor<string, []>("op_2803"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_2802_cast_fp16 = transpose(perm = var_2802_perm_0, x = x_299_cast_fp16)[name = tensor<string, []>("transpose_135")];
            tensor<fp16, [1, 8, 512]> input_615_cast_fp16 = reshape(shape = var_2803, x = var_2802_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, []>(145343296)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(145867648)))];
            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, []>(145868736)))];
            tensor<fp16, [1, 8, 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, []>(145869824)))];
            tensor<fp16, [1, 512, 8]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = tensor<string, []>("transpose_134")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_305_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_625_cast_fp16 = select(a = var_40_to_fp16, b = x_305_cast_fp16, cond = var_551)[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, 16]> new_x_47_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_2844_begin_0 = const()[name = tensor<string, []>("op_2844_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2844_end_0 = const()[name = tensor<string, []>("op_2844_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_2844_end_mask_0 = const()[name = tensor<string, []>("op_2844_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_2844_cast_fp16 = slice_by_index(begin = var_2844_begin_0, end = var_2844_end_0, end_mask = var_2844_end_mask_0, x = next_cache_23_cast_fp16)[name = tensor<string, []>("op_2844_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, []>(146918464)))];
            tensor<fp16, [1, 512, 8]> 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, []>(146927744)))];
            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, []>(146928832)))];
            tensor<fp16, [1, 8, 512]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = tensor<string, []>("transpose_133")];
            tensor<fp16, [1, 8, 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, 8]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = tensor<string, []>("transpose_132")];
            tensor<fp16, [1, 512, 8]> 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, []>(146929920)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = tensor<string, []>("transpose_131")];
            tensor<fp16, [1, 8, 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, []>(147454272)))];
            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, []>(147455360)))];
            tensor<fp16, [1, 8, 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, []>(147456448)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(149553664)))];
            tensor<fp16, [1, 8, 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_2885_to_fp16 = const()[name = tensor<string, []>("op_2885_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2886_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2885_to_fp16)[name = tensor<string, []>("op_2886_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2886_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, []>(151650880)))];
            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, []>(151651968)))];
            tensor<fp16, [1, 8, 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, []>(151653056)))];
            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, []>(151654144)))];
            tensor<fp16, [1, 8, 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, []>(151655232)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(153752448)))];
            tensor<fp16, [1, 8, 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_2920_to_fp16 = const()[name = tensor<string, []>("op_2920_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_2921_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_2920_to_fp16)[name = tensor<string, []>("op_2921_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_2921_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, []>(155849664)))];
            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, []>(155850752)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_663_cast_fp16 = concat(axis = var_65, 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_2943_begin_0 = const()[name = tensor<string, []>("op_2943_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2943_end_0 = const()[name = tensor<string, []>("op_2943_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_2943_end_mask_0 = const()[name = tensor<string, []>("op_2943_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_2943_cast_fp16 = slice_by_index(begin = var_2943_begin_0, end = var_2943_end_0, end_mask = var_2943_end_mask_0, x = cache_49_cast_fp16)[name = tensor<string, []>("op_2943_cast_fp16")];
            tensor<int32, [3]> var_2946_begin_0 = const()[name = tensor<string, []>("op_2946_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2946_end_0 = const()[name = tensor<string, []>("op_2946_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_2946_end_mask_0 = const()[name = tensor<string, []>("op_2946_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_2946_cast_fp16 = slice_by_index(begin = var_2946_begin_0, end = var_2946_end_0, end_mask = var_2946_end_mask_0, x = key_25_cast_fp16)[name = tensor<string, []>("op_2946_cast_fp16")];
            tensor<bool, []> var_2949_interleave_0 = const()[name = tensor<string, []>("op_2949_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_2949_cast_fp16 = concat(axis = var_65, interleave = var_2949_interleave_0, values = (var_2943_cast_fp16, var_2946_cast_fp16))[name = tensor<string, []>("op_2949_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, []>(155851840)))];
            tensor<fp16, [1, 8, 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_2953 = const()[name = tensor<string, []>("op_2953"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_73_cast_fp16 = reshape(shape = var_2953, 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, []>(156376192)))];
            tensor<fp16, [1, 78, 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_2957 = const()[name = tensor<string, []>("op_2957"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_49_cast_fp16 = reshape(shape = var_2957, 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, []>(156900544)))];
            tensor<fp16, [1, 78, 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_2961 = const()[name = tensor<string, []>("op_2961"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_25_cast_fp16 = reshape(shape = var_2961, 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, []>(157424896)))];
            tensor<fp16, [1, 8, 8, 64]> var_2973_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2973_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, []>(157425984)))];
            tensor<fp16, [1, 8, 8, 64]> var_2975_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2975_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, 155]> var_2977_to_fp16 = const()[name = tensor<string, []>("op_2977_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157427072)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_2975_cast_fp16)[name = tensor<string, []>("transpose_129")];
            tensor<fp16, [1, 8, 8, 155]> 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_2977_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_235_to_fp16 = const()[name = tensor<string, []>("const_235_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_321_cast_fp16 = pad(constant_val = const_235_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_2985 = const()[name = tensor<string, []>("op_2985"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_323_cast_fp16 = reshape(shape = var_2985, x = x_321_cast_fp16)[name = tensor<string, []>("x_323_cast_fp16")];
            tensor<int32, [4]> var_2989_begin_0 = const()[name = tensor<string, []>("op_2989_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2989_end_0 = const()[name = tensor<string, []>("op_2989_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_2989_end_mask_0 = const()[name = tensor<string, []>("op_2989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_2989_cast_fp16 = slice_by_index(begin = var_2989_begin_0, end = var_2989_end_0, end_mask = var_2989_end_mask_0, x = x_323_cast_fp16)[name = tensor<string, []>("op_2989_cast_fp16")];
            tensor<int32, [4]> var_2990 = const()[name = tensor<string, []>("op_2990"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_49_cast_fp16 = reshape(shape = var_2990, x = var_2989_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, 78]> transpose_76 = transpose(perm = transpose_76_perm_0, x = k_49_cast_fp16)[name = tensor<string, []>("transpose_127")];
            tensor<fp16, [1, 8, 8, 64]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_2973_cast_fp16)[name = tensor<string, []>("transpose_128")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_2999_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = tensor<string, []>("op_2999_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, 8, 78]> _inversed_scores_49_cast_fp16 = mul(x = var_2999_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_49_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_51_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_51_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_3005_cast_fp16 = softmax(axis = var_56, x = scores_51_cast_fp16)[name = tensor<string, []>("op_3005_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_665_cast_fp16 = select(a = var_40_to_fp16, b = var_3005_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_3009_perm_0 = const()[name = tensor<string, []>("op_3009_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3010 = const()[name = tensor<string, []>("op_3010"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_3009_cast_fp16 = transpose(perm = var_3009_perm_0, x = x_325_cast_fp16)[name = tensor<string, []>("transpose_126")];
            tensor<fp16, [1, 8, 512]> input_667_cast_fp16 = reshape(shape = var_3010, x = var_3009_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, []>(157585856)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(158110208)))];
            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, []>(158111296)))];
            tensor<fp16, [1, 8, 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, []>(158112384)))];
            tensor<fp16, [1, 512, 8]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = tensor<string, []>("transpose_125")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_331_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_677_cast_fp16 = select(a = var_40_to_fp16, b = x_331_cast_fp16, cond = var_551)[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, 16]> new_x_51_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_3051_begin_0 = const()[name = tensor<string, []>("op_3051_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3051_end_0 = const()[name = tensor<string, []>("op_3051_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_3051_end_mask_0 = const()[name = tensor<string, []>("op_3051_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_3051_cast_fp16 = slice_by_index(begin = var_3051_begin_0, end = var_3051_end_0, end_mask = var_3051_end_mask_0, x = next_cache_25_cast_fp16)[name = tensor<string, []>("op_3051_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, []>(159161024)))];
            tensor<fp16, [1, 512, 8]> 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, []>(159170304)))];
            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, []>(159171392)))];
            tensor<fp16, [1, 8, 512]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = tensor<string, []>("transpose_124")];
            tensor<fp16, [1, 8, 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, 8]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = tensor<string, []>("transpose_123")];
            tensor<fp16, [1, 512, 8]> 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, []>(159172480)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = tensor<string, []>("transpose_122")];
            tensor<fp16, [1, 8, 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, []>(159696832)))];
            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, []>(159697920)))];
            tensor<fp16, [1, 8, 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, []>(159699008)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(161796224)))];
            tensor<fp16, [1, 8, 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_3092_to_fp16 = const()[name = tensor<string, []>("op_3092_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_3093_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3092_to_fp16)[name = tensor<string, []>("op_3093_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3093_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, []>(163893440)))];
            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, []>(163894528)))];
            tensor<fp16, [1, 8, 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, []>(163895616)))];
            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, []>(163896704)))];
            tensor<fp16, [1, 8, 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, []>(163897792)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(165995008)))];
            tensor<fp16, [1, 8, 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_3127_to_fp16 = const()[name = tensor<string, []>("op_3127_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_3128_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3127_to_fp16)[name = tensor<string, []>("op_3128_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3128_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, []>(168092224)))];
            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, []>(168093312)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_715_cast_fp16 = concat(axis = var_65, 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_3150_begin_0 = const()[name = tensor<string, []>("op_3150_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_3150_end_0 = const()[name = tensor<string, []>("op_3150_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_3150_end_mask_0 = const()[name = tensor<string, []>("op_3150_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_3150_cast_fp16 = slice_by_index(begin = var_3150_begin_0, end = var_3150_end_0, end_mask = var_3150_end_mask_0, x = cache_53_cast_fp16)[name = tensor<string, []>("op_3150_cast_fp16")];
            tensor<int32, [3]> var_3153_begin_0 = const()[name = tensor<string, []>("op_3153_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3153_end_0 = const()[name = tensor<string, []>("op_3153_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_3153_end_mask_0 = const()[name = tensor<string, []>("op_3153_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_3153_cast_fp16 = slice_by_index(begin = var_3153_begin_0, end = var_3153_end_0, end_mask = var_3153_end_mask_0, x = key_27_cast_fp16)[name = tensor<string, []>("op_3153_cast_fp16")];
            tensor<bool, []> var_3156_interleave_0 = const()[name = tensor<string, []>("op_3156_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_3156_cast_fp16 = concat(axis = var_65, interleave = var_3156_interleave_0, values = (var_3150_cast_fp16, var_3153_cast_fp16))[name = tensor<string, []>("op_3156_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, []>(168094400)))];
            tensor<fp16, [1, 8, 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_3160 = const()[name = tensor<string, []>("op_3160"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_79_cast_fp16 = reshape(shape = var_3160, 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, []>(168618752)))];
            tensor<fp16, [1, 78, 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_3164 = const()[name = tensor<string, []>("op_3164"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_53_cast_fp16 = reshape(shape = var_3164, 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, []>(169143104)))];
            tensor<fp16, [1, 78, 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_3168 = const()[name = tensor<string, []>("op_3168"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_27_cast_fp16 = reshape(shape = var_3168, 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, []>(169667456)))];
            tensor<fp16, [1, 8, 8, 64]> var_3180_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3180_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, []>(169668544)))];
            tensor<fp16, [1, 8, 8, 64]> var_3182_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3182_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, 155]> var_3184_to_fp16 = const()[name = tensor<string, []>("op_3184_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169669632)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3182_cast_fp16)[name = tensor<string, []>("transpose_120")];
            tensor<fp16, [1, 8, 8, 155]> 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_3184_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_248_to_fp16 = const()[name = tensor<string, []>("const_248_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_347_cast_fp16 = pad(constant_val = const_248_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_3192 = const()[name = tensor<string, []>("op_3192"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_349_cast_fp16 = reshape(shape = var_3192, x = x_347_cast_fp16)[name = tensor<string, []>("x_349_cast_fp16")];
            tensor<int32, [4]> var_3196_begin_0 = const()[name = tensor<string, []>("op_3196_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3196_end_0 = const()[name = tensor<string, []>("op_3196_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_3196_end_mask_0 = const()[name = tensor<string, []>("op_3196_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_3196_cast_fp16 = slice_by_index(begin = var_3196_begin_0, end = var_3196_end_0, end_mask = var_3196_end_mask_0, x = x_349_cast_fp16)[name = tensor<string, []>("op_3196_cast_fp16")];
            tensor<int32, [4]> var_3197 = const()[name = tensor<string, []>("op_3197"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_53_cast_fp16 = reshape(shape = var_3197, x = var_3196_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, 78]> transpose_78 = transpose(perm = transpose_78_perm_0, x = k_53_cast_fp16)[name = tensor<string, []>("transpose_118")];
            tensor<fp16, [1, 8, 8, 64]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_3180_cast_fp16)[name = tensor<string, []>("transpose_119")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_3206_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = tensor<string, []>("op_3206_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, 8, 78]> _inversed_scores_53_cast_fp16 = mul(x = var_3206_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_53_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_55_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_55_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_3212_cast_fp16 = softmax(axis = var_56, x = scores_55_cast_fp16)[name = tensor<string, []>("op_3212_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_717_cast_fp16 = select(a = var_40_to_fp16, b = var_3212_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_3216_perm_0 = const()[name = tensor<string, []>("op_3216_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3217 = const()[name = tensor<string, []>("op_3217"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_3216_cast_fp16 = transpose(perm = var_3216_perm_0, x = x_351_cast_fp16)[name = tensor<string, []>("transpose_117")];
            tensor<fp16, [1, 8, 512]> input_719_cast_fp16 = reshape(shape = var_3217, x = var_3216_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, []>(169828416)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(170352768)))];
            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, []>(170353856)))];
            tensor<fp16, [1, 8, 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, []>(170354944)))];
            tensor<fp16, [1, 512, 8]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = tensor<string, []>("transpose_116")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_357_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_729_cast_fp16 = select(a = var_40_to_fp16, b = x_357_cast_fp16, cond = var_551)[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, 16]> new_x_55_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_3258_begin_0 = const()[name = tensor<string, []>("op_3258_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3258_end_0 = const()[name = tensor<string, []>("op_3258_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_3258_end_mask_0 = const()[name = tensor<string, []>("op_3258_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_3258_cast_fp16 = slice_by_index(begin = var_3258_begin_0, end = var_3258_end_0, end_mask = var_3258_end_mask_0, x = next_cache_27_cast_fp16)[name = tensor<string, []>("op_3258_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, []>(171403584)))];
            tensor<fp16, [1, 512, 8]> 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, []>(171412864)))];
            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, []>(171413952)))];
            tensor<fp16, [1, 8, 512]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = tensor<string, []>("transpose_115")];
            tensor<fp16, [1, 8, 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, 8]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = tensor<string, []>("transpose_114")];
            tensor<fp16, [1, 512, 8]> 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, []>(171415040)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = tensor<string, []>("transpose_113")];
            tensor<fp16, [1, 8, 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, []>(171939392)))];
            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, []>(171940480)))];
            tensor<fp16, [1, 8, 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, []>(171941568)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(174038784)))];
            tensor<fp16, [1, 8, 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_3299_to_fp16 = const()[name = tensor<string, []>("op_3299_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_3300_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3299_to_fp16)[name = tensor<string, []>("op_3300_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3300_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, []>(176136000)))];
            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, []>(176137088)))];
            tensor<fp16, [1, 8, 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, []>(176138176)))];
            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, []>(176139264)))];
            tensor<fp16, [1, 8, 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, []>(176140352)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(178237568)))];
            tensor<fp16, [1, 8, 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_3334_to_fp16 = const()[name = tensor<string, []>("op_3334_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_3335_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3334_to_fp16)[name = tensor<string, []>("op_3335_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3335_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, []>(180334784)))];
            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, []>(180335872)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_767_cast_fp16 = concat(axis = var_65, 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_3357_begin_0 = const()[name = tensor<string, []>("op_3357_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_3357_end_0 = const()[name = tensor<string, []>("op_3357_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_3357_end_mask_0 = const()[name = tensor<string, []>("op_3357_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = cache_57_cast_fp16)[name = tensor<string, []>("op_3357_cast_fp16")];
            tensor<int32, [3]> var_3360_begin_0 = const()[name = tensor<string, []>("op_3360_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3360_end_0 = const()[name = tensor<string, []>("op_3360_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_3360_end_mask_0 = const()[name = tensor<string, []>("op_3360_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_3360_cast_fp16 = slice_by_index(begin = var_3360_begin_0, end = var_3360_end_0, end_mask = var_3360_end_mask_0, x = key_29_cast_fp16)[name = tensor<string, []>("op_3360_cast_fp16")];
            tensor<bool, []> var_3363_interleave_0 = const()[name = tensor<string, []>("op_3363_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_3363_cast_fp16 = concat(axis = var_65, interleave = var_3363_interleave_0, values = (var_3357_cast_fp16, var_3360_cast_fp16))[name = tensor<string, []>("op_3363_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, []>(180336960)))];
            tensor<fp16, [1, 8, 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_3367 = const()[name = tensor<string, []>("op_3367"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_85_cast_fp16 = reshape(shape = var_3367, 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, []>(180861312)))];
            tensor<fp16, [1, 78, 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_3371 = const()[name = tensor<string, []>("op_3371"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_57_cast_fp16 = reshape(shape = var_3371, 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, []>(181385664)))];
            tensor<fp16, [1, 78, 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_3375 = const()[name = tensor<string, []>("op_3375"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_29_cast_fp16 = reshape(shape = var_3375, 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, []>(181910016)))];
            tensor<fp16, [1, 8, 8, 64]> var_3387_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3387_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, []>(181911104)))];
            tensor<fp16, [1, 8, 8, 64]> var_3389_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3389_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, 155]> var_3391_to_fp16 = const()[name = tensor<string, []>("op_3391_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181912192)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3389_cast_fp16)[name = tensor<string, []>("transpose_111")];
            tensor<fp16, [1, 8, 8, 155]> 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_3391_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_261_to_fp16 = const()[name = tensor<string, []>("const_261_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_373_cast_fp16 = pad(constant_val = const_261_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_3399 = const()[name = tensor<string, []>("op_3399"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_375_cast_fp16 = reshape(shape = var_3399, x = x_373_cast_fp16)[name = tensor<string, []>("x_375_cast_fp16")];
            tensor<int32, [4]> var_3403_begin_0 = const()[name = tensor<string, []>("op_3403_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3403_end_0 = const()[name = tensor<string, []>("op_3403_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_3403_end_mask_0 = const()[name = tensor<string, []>("op_3403_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_3403_cast_fp16 = slice_by_index(begin = var_3403_begin_0, end = var_3403_end_0, end_mask = var_3403_end_mask_0, x = x_375_cast_fp16)[name = tensor<string, []>("op_3403_cast_fp16")];
            tensor<int32, [4]> var_3404 = const()[name = tensor<string, []>("op_3404"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_57_cast_fp16 = reshape(shape = var_3404, x = var_3403_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, 78]> transpose_80 = transpose(perm = transpose_80_perm_0, x = k_57_cast_fp16)[name = tensor<string, []>("transpose_109")];
            tensor<fp16, [1, 8, 8, 64]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_3387_cast_fp16)[name = tensor<string, []>("transpose_110")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_3413_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = tensor<string, []>("op_3413_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, 8, 78]> _inversed_scores_57_cast_fp16 = mul(x = var_3413_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_57_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_59_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_59_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_3419_cast_fp16 = softmax(axis = var_56, x = scores_59_cast_fp16)[name = tensor<string, []>("op_3419_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_769_cast_fp16 = select(a = var_40_to_fp16, b = var_3419_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_3423_perm_0 = const()[name = tensor<string, []>("op_3423_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3424 = const()[name = tensor<string, []>("op_3424"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_3423_cast_fp16 = transpose(perm = var_3423_perm_0, x = x_377_cast_fp16)[name = tensor<string, []>("transpose_108")];
            tensor<fp16, [1, 8, 512]> input_771_cast_fp16 = reshape(shape = var_3424, x = var_3423_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, []>(182070976)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(182595328)))];
            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, []>(182596416)))];
            tensor<fp16, [1, 8, 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, []>(182597504)))];
            tensor<fp16, [1, 512, 8]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = tensor<string, []>("transpose_107")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_383_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_781_cast_fp16 = select(a = var_40_to_fp16, b = x_383_cast_fp16, cond = var_551)[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, 16]> new_x_59_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_3465_begin_0 = const()[name = tensor<string, []>("op_3465_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3465_end_0 = const()[name = tensor<string, []>("op_3465_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_3465_end_mask_0 = const()[name = tensor<string, []>("op_3465_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_3465_cast_fp16 = slice_by_index(begin = var_3465_begin_0, end = var_3465_end_0, end_mask = var_3465_end_mask_0, x = next_cache_29_cast_fp16)[name = tensor<string, []>("op_3465_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, []>(183646144)))];
            tensor<fp16, [1, 512, 8]> 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, []>(183655424)))];
            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, []>(183656512)))];
            tensor<fp16, [1, 8, 512]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = tensor<string, []>("transpose_106")];
            tensor<fp16, [1, 8, 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, 8]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = tensor<string, []>("transpose_105")];
            tensor<fp16, [1, 512, 8]> 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, []>(183657600)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = tensor<string, []>("transpose_104")];
            tensor<fp16, [1, 8, 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, []>(184181952)))];
            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, []>(184183040)))];
            tensor<fp16, [1, 8, 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, []>(184184128)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(186281344)))];
            tensor<fp16, [1, 8, 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_3506_to_fp16 = const()[name = tensor<string, []>("op_3506_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_3507_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3506_to_fp16)[name = tensor<string, []>("op_3507_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3507_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, []>(188378560)))];
            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, []>(188379648)))];
            tensor<fp16, [1, 8, 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, []>(188380736)))];
            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, []>(188381824)))];
            tensor<fp16, [1, 8, 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, []>(188382912)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(190480128)))];
            tensor<fp16, [1, 8, 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_3541_to_fp16 = const()[name = tensor<string, []>("op_3541_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_3542_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3541_to_fp16)[name = tensor<string, []>("op_3542_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3542_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, []>(192577344)))];
            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, []>(192578432)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_819_cast_fp16 = concat(axis = var_65, 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_3564_begin_0 = const()[name = tensor<string, []>("op_3564_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_3564_end_0 = const()[name = tensor<string, []>("op_3564_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_3564_end_mask_0 = const()[name = tensor<string, []>("op_3564_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_3564_cast_fp16 = slice_by_index(begin = var_3564_begin_0, end = var_3564_end_0, end_mask = var_3564_end_mask_0, x = cache_61_cast_fp16)[name = tensor<string, []>("op_3564_cast_fp16")];
            tensor<int32, [3]> var_3567_begin_0 = const()[name = tensor<string, []>("op_3567_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3567_end_0 = const()[name = tensor<string, []>("op_3567_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_3567_end_mask_0 = const()[name = tensor<string, []>("op_3567_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_3567_cast_fp16 = slice_by_index(begin = var_3567_begin_0, end = var_3567_end_0, end_mask = var_3567_end_mask_0, x = key_31_cast_fp16)[name = tensor<string, []>("op_3567_cast_fp16")];
            tensor<bool, []> var_3570_interleave_0 = const()[name = tensor<string, []>("op_3570_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 70, 512]> var_3570_cast_fp16 = concat(axis = var_65, interleave = var_3570_interleave_0, values = (var_3564_cast_fp16, var_3567_cast_fp16))[name = tensor<string, []>("op_3570_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, []>(192579520)))];
            tensor<fp16, [1, 8, 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_3574 = const()[name = tensor<string, []>("op_3574"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_91_cast_fp16 = reshape(shape = var_3574, 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, []>(193103872)))];
            tensor<fp16, [1, 78, 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_3578 = const()[name = tensor<string, []>("op_3578"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_61_cast_fp16 = reshape(shape = var_3578, 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, []>(193628224)))];
            tensor<fp16, [1, 78, 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_3582 = const()[name = tensor<string, []>("op_3582"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_31_cast_fp16 = reshape(shape = var_3582, 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, []>(194152576)))];
            tensor<fp16, [1, 8, 8, 64]> var_3594_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3594_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, []>(194153664)))];
            tensor<fp16, [1, 8, 8, 64]> var_3596_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3596_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, 155]> var_3598_to_fp16 = const()[name = tensor<string, []>("op_3598_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194154752)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3596_cast_fp16)[name = tensor<string, []>("transpose_102")];
            tensor<fp16, [1, 8, 8, 155]> 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_3598_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_274_to_fp16 = const()[name = tensor<string, []>("const_274_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_399_cast_fp16 = pad(constant_val = const_274_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_3606 = const()[name = tensor<string, []>("op_3606"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_401_cast_fp16 = reshape(shape = var_3606, x = x_399_cast_fp16)[name = tensor<string, []>("x_401_cast_fp16")];
            tensor<int32, [4]> var_3610_begin_0 = const()[name = tensor<string, []>("op_3610_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3610_end_0 = const()[name = tensor<string, []>("op_3610_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_3610_end_mask_0 = const()[name = tensor<string, []>("op_3610_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_3610_cast_fp16 = slice_by_index(begin = var_3610_begin_0, end = var_3610_end_0, end_mask = var_3610_end_mask_0, x = x_401_cast_fp16)[name = tensor<string, []>("op_3610_cast_fp16")];
            tensor<int32, [4]> var_3611 = const()[name = tensor<string, []>("op_3611"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_61_cast_fp16 = reshape(shape = var_3611, x = var_3610_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, 78]> transpose_82 = transpose(perm = transpose_82_perm_0, x = k_61_cast_fp16)[name = tensor<string, []>("transpose_100")];
            tensor<fp16, [1, 8, 8, 64]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_3594_cast_fp16)[name = tensor<string, []>("transpose_101")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_3620_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = tensor<string, []>("op_3620_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, 8, 78]> _inversed_scores_61_cast_fp16 = mul(x = var_3620_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_61_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_63_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_63_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_3626_cast_fp16 = softmax(axis = var_56, x = scores_63_cast_fp16)[name = tensor<string, []>("op_3626_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_821_cast_fp16 = select(a = var_40_to_fp16, b = var_3626_cast_fp16, cond = mask_11)[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, 78, 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, 8, 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_3630_perm_0 = const()[name = tensor<string, []>("op_3630_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3631 = const()[name = tensor<string, []>("op_3631"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_3630_cast_fp16 = transpose(perm = var_3630_perm_0, x = x_403_cast_fp16)[name = tensor<string, []>("transpose_99")];
            tensor<fp16, [1, 8, 512]> input_823_cast_fp16 = reshape(shape = var_3631, x = var_3630_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, []>(194313536)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(194837888)))];
            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, []>(194838976)))];
            tensor<fp16, [1, 8, 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, []>(194840064)))];
            tensor<fp16, [1, 512, 8]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = tensor<string, []>("transpose_98")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_409_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_833_cast_fp16 = select(a = var_40_to_fp16, b = x_409_cast_fp16, cond = var_551)[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, 16]> new_x_63_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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_3672_begin_0 = const()[name = tensor<string, []>("op_3672_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3672_end_0 = const()[name = tensor<string, []>("op_3672_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
            tensor<bool, [3]> var_3672_end_mask_0 = const()[name = tensor<string, []>("op_3672_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 512, 8]> var_3672_cast_fp16 = slice_by_index(begin = var_3672_begin_0, end = var_3672_end_0, end_mask = var_3672_end_mask_0, x = next_cache_31_cast_fp16)[name = tensor<string, []>("op_3672_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, []>(195888704)))];
            tensor<fp16, [1, 512, 8]> 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, []>(195897984)))];
            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, []>(195899072)))];
            tensor<fp16, [1, 8, 512]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = tensor<string, []>("transpose_97")];
            tensor<fp16, [1, 8, 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, 8]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = tensor<string, []>("transpose_96")];
            tensor<fp16, [1, 512, 8]> 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, []>(195900160)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = tensor<string, []>("transpose_95")];
            tensor<fp16, [1, 8, 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, []>(196424512)))];
            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, []>(196425600)))];
            tensor<fp16, [1, 8, 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, []>(196426688)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(198523904)))];
            tensor<fp16, [1, 8, 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_3713_to_fp16 = const()[name = tensor<string, []>("op_3713_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_3714_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3713_to_fp16)[name = tensor<string, []>("op_3714_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3714_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, []>(200621120)))];
            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, []>(200622208)))];
            tensor<fp16, [1, 8, 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, []>(200623296)))];
            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, []>(200624384)))];
            tensor<fp16, [1, 8, 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, []>(200625472)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(202722688)))];
            tensor<fp16, [1, 8, 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_3748_to_fp16 = const()[name = tensor<string, []>("op_3748_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_3749_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3748_to_fp16)[name = tensor<string, []>("op_3749_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3749_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, []>(204819904)))];
            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, []>(204820992)))];
            tensor<fp16, [1, 8, 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, 78, 512]> input_871_cast_fp16 = concat(axis = var_65, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_cast_fp16))[name = tensor<string, []>("input_871_cast_fp16")];
            tensor<int32, [3]> var_3771_begin_0 = const()[name = tensor<string, []>("op_3771_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_3771_end_0 = const()[name = tensor<string, []>("op_3771_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
            tensor<bool, [3]> var_3771_end_mask_0 = const()[name = tensor<string, []>("op_3771_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 66, 512]> var_3771_cast_fp16 = slice_by_index(begin = var_3771_begin_0, end = var_3771_end_0, end_mask = var_3771_end_mask_0, x = cache_65_cast_fp16)[name = tensor<string, []>("op_3771_cast_fp16")];
            tensor<int32, [3]> var_3774_begin_0 = const()[name = tensor<string, []>("op_3774_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3774_end_0 = const()[name = tensor<string, []>("op_3774_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
            tensor<bool, [3]> var_3774_end_mask_0 = const()[name = tensor<string, []>("op_3774_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 4, 512]> var_3774_cast_fp16 = slice_by_index(begin = var_3774_begin_0, end = var_3774_end_0, end_mask = var_3774_end_mask_0, x = key_cast_fp16)[name = tensor<string, []>("op_3774_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_65, interleave = cache_last_channel_cur_interleave_0, values = (var_3771_cast_fp16, var_3774_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, []>(204822080)))];
            tensor<fp16, [1, 8, 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_3781 = const()[name = tensor<string, []>("op_3781"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 8, 8, 64]> q_97_cast_fp16 = reshape(shape = var_3781, 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, []>(205346432)))];
            tensor<fp16, [1, 78, 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_3785 = const()[name = tensor<string, []>("op_3785"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> k_65_cast_fp16 = reshape(shape = var_3785, 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, []>(205870784)))];
            tensor<fp16, [1, 78, 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_3789 = const()[name = tensor<string, []>("op_3789"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, 78, 8, 64]> v_cast_fp16 = reshape(shape = var_3789, 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, []>(206395136)))];
            tensor<fp16, [1, 8, 8, 64]> var_3801_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3801_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, []>(206396224)))];
            tensor<fp16, [1, 8, 8, 64]> var_3803_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3803_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, 155]> var_3805_to_fp16 = const()[name = tensor<string, []>("op_3805_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206397312)))];
            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_3803_cast_fp16)[name = tensor<string, []>("transpose_93")];
            tensor<fp16, [1, 8, 8, 155]> 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_3805_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_287_to_fp16 = const()[name = tensor<string, []>("const_287_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 8, 156]> x_425_cast_fp16 = pad(constant_val = const_287_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_3813 = const()[name = tensor<string, []>("op_3813"), val = tensor<int32, [4]>([1, 8, -1, 8])];
            tensor<fp16, [1, 8, 156, 8]> x_427_cast_fp16 = reshape(shape = var_3813, x = x_425_cast_fp16)[name = tensor<string, []>("x_427_cast_fp16")];
            tensor<int32, [4]> var_3817_begin_0 = const()[name = tensor<string, []>("op_3817_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3817_end_0 = const()[name = tensor<string, []>("op_3817_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
            tensor<bool, [4]> var_3817_end_mask_0 = const()[name = tensor<string, []>("op_3817_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 155, 8]> var_3817_cast_fp16 = slice_by_index(begin = var_3817_begin_0, end = var_3817_end_0, end_mask = var_3817_end_mask_0, x = x_427_cast_fp16)[name = tensor<string, []>("op_3817_cast_fp16")];
            tensor<int32, [4]> var_3818 = const()[name = tensor<string, []>("op_3818"), val = tensor<int32, [4]>([1, 8, 8, 155])];
            tensor<fp16, [1, 8, 8, 155]> matrix_bd_65_cast_fp16 = reshape(shape = var_3818, x = var_3817_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, 78]> transpose_84 = transpose(perm = transpose_84_perm_0, x = k_65_cast_fp16)[name = tensor<string, []>("transpose_91")];
            tensor<fp16, [1, 8, 8, 64]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_3801_cast_fp16)[name = tensor<string, []>("transpose_92")];
            tensor<fp16, [1, 8, 8, 78]> 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, 8, 78])];
            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, 8, 78]> 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, 8, 78]> var_3827_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = tensor<string, []>("op_3827_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, 8, 78]> _inversed_scores_65_cast_fp16 = mul(x = var_3827_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_65_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> scores_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> var_3833_cast_fp16 = softmax(axis = var_56, x = scores_cast_fp16)[name = tensor<string, []>("op_3833_cast_fp16")];
            tensor<fp16, [1, 8, 8, 78]> input_873_cast_fp16 = select(a = var_40_to_fp16, b = var_3833_cast_fp16, cond = mask_11)[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, 78, 64]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = tensor<string, []>("transpose_94")];
            tensor<fp16, [1, 8, 8, 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_3837_perm_0 = const()[name = tensor<string, []>("op_3837_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3838 = const()[name = tensor<string, []>("op_3838"), val = tensor<int32, [3]>([1, -1, 512])];
            tensor<fp16, [1, 8, 8, 64]> var_3837_cast_fp16 = transpose(perm = var_3837_perm_0, x = x_429_cast_fp16)[name = tensor<string, []>("transpose_90")];
            tensor<fp16, [1, 8, 512]> input_875_cast_fp16 = reshape(shape = var_3838, x = var_3837_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, []>(206556096)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(207080448)))];
            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, []>(207081536)))];
            tensor<fp16, [1, 8, 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, []>(207082624)))];
            tensor<fp16, [1, 512, 8]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = tensor<string, []>("transpose_89")];
            tensor<fp16, [1, 1024, 8]> 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, 8]> x_435_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> 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, 8]> 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, 8]> 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, 8]> input_885_cast_fp16 = select(a = var_40_to_fp16, b = x_435_cast_fp16, cond = var_551)[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, 16]> new_x_cast_fp16 = concat(axis = var_56, 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, 12])];
            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, 12]> 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, 4])];
            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, 12])];
            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, []>(208131264)))];
            tensor<fp16, [1, 512, 8]> 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, []>(208140544)))];
            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, []>(208141632)))];
            tensor<fp16, [1, 8, 512]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = tensor<string, []>("transpose_88")];
            tensor<fp16, [1, 8, 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, 8]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = tensor<string, []>("transpose_87")];
            tensor<fp16, [1, 512, 8]> 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, []>(208142720)))];
            tensor<fp16, [1, 512, 8]> 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, 8, 512]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = tensor<string, []>("transpose_86")];
            tensor<fp16, [1, 8, 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, []>(208667072)))];
            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, []>(208668160)))];
            tensor<fp16, [1, 8, 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, []>(208669248)))];
            tensor<fp16, [1, 8, 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, 8, 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, []>(210766464)))];
            tensor<fp16, [1, 8, 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_3920_to_fp16 = const()[name = tensor<string, []>("op_3920_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
            tensor<fp16, [1, 8, 512]> var_3921_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_3920_to_fp16)[name = tensor<string, []>("op_3921_cast_fp16")];
            tensor<fp16, [1, 8, 512]> input_cast_fp16 = add(x = input_895_cast_fp16, y = var_3921_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, []>(212863680)))];
            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, []>(212864768)))];
            tensor<fp16, [1, 8, 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, []> obj_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            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_465_cast_fp16, var_672_cast_fp16, var_879_cast_fp16, var_1086_cast_fp16, var_1293_cast_fp16, var_1500_cast_fp16, var_1707_cast_fp16, var_1914_cast_fp16, var_2121_cast_fp16, var_2328_cast_fp16, var_2535_cast_fp16, var_2742_cast_fp16, var_2949_cast_fp16, var_3156_cast_fp16, var_3363_cast_fp16, var_3570_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_567_cast_fp16, var_774_cast_fp16, var_981_cast_fp16, var_1188_cast_fp16, var_1395_cast_fp16, var_1602_cast_fp16, var_1809_cast_fp16, var_2016_cast_fp16, var_2223_cast_fp16, var_2430_cast_fp16, var_2637_cast_fp16, var_2844_cast_fp16, var_3051_cast_fp16, var_3258_cast_fp16, var_3465_cast_fp16, var_3672_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_3937 = add(x = cache_last_channel_len, y = cache_keep_size)[name = tensor<string, []>("op_3937")];
            tensor<string, []> var_3937_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3937_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> const_293_to_fp16 = const()[name = tensor<string, []>("const_293_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, []> var_46_promoted_to_fp16 = const()[name = tensor<string, []>("op_46_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
            tensor<fp16, [1]> var_3937_to_fp16 = cast(dtype = var_3937_promoted_to_fp16_dtype_0, x = var_3937)[name = tensor<string, []>("cast_184")];
            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_293_to_fp16, beta = var_46_promoted_to_fp16, x = var_3937_to_fp16)[name = tensor<string, []>("clip_1_cast_fp16")];
            tensor<string, []> var_3964_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3964_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            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_0_cast_fp16)[name = tensor<string, []>("cast_182")];
            tensor<fp32, [17, 1, 70, 512]> new_cache_last_channel = cast(dtype = var_3964_cast_fp16_to_fp32_dtype_0, x = obj_5_cast_fp16)[name = tensor<string, []>("cast_183")];
            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_185")];
            tensor<fp16, [1, 512, 8]> obj_1_cast_fp16 = transpose(perm = obj_1_perm_0, x = audio_signal_cast_fp16)[name = tensor<string, []>("transpose_85")];
            tensor<fp32, [1, 512, 8]> encoded_output = cast(dtype = obj_1_cast_fp16_to_fp32_dtype_0, x = obj_1_cast_fp16)[name = tensor<string, []>("cast_186")];
            tensor<fp32, [1, 128, 9]> new_pre_cache = cast(dtype = var_28_cast_fp16_to_fp32_dtype_0, x = var_28_cast_fp16)[name = tensor<string, []>("cast_198")];
        } -> (encoded_output, encoded_length, new_pre_cache, new_cache_last_channel, new_cache_last_time, new_cache_last_channel_len);
}