CONFUSING! TELlamaModel error and torchaudio.save runtime error

[ckgithub2019]

1. this issue happens every time. but the README said "Optional: Install TransformerEngine if using NVIDIA GPU (Linux only), The adaptation of TransformerEngine is currently under development and CANNOT run properly now", I'm using GPU ann Linux, so do I must install it or not? confusing:

use default LlamaModel for importing TELlamaModel error: No module named 'transformer_engine'

2. I tested with a very simple code snippet, but it didn't work, and runtime error here:

We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)
text:   0%|▎                                                                                                  | 1/384(max) [00:00,  6.19it/sttext:  12%|███████████▌                                                                                     | 46/384(max) [00:00, 118.84it/s]
code:   1%|█                                                                                               | 22/2048(max) [00:00, 213.37it/sccode:  20%|██████████████████▊                                                                            | 406/2048(max) [00:01, 213.68it/s]
/home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torchaudio/_backend/ffmpeg.py:245: UserWarning: The use of `x.T` on tensors of dimension other than 2 to reverse their shape is deprecated and it will throw an error in a future release. Consider `x.mT` to transpose batches of matrices or `x.permute(*torch.arange(x.ndim - 1, -1, -1))` to reverse the dimensions of a tensor. (Triggered internally at ../aten/src/ATen/native/TensorShape.cpp:3697.)
  src = src.T
Traceback (most recent call last):
  File "/home/ck/ai_project/tts_server/basic_test.py", line 13, in <module>
    torchaudio.save(f"basic_output{i}.wav", torch.from_numpy(wavs[i]).unsqueeze(0), 24000)
  File "/home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torchaudio/_backend/utils.py", line 313, in save
    return backend.save(
           ^^^^^^^^^^^^^
  File "/home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torchaudio/_backend/ffmpeg.py", line 316, in save
    save_audio(
  File "/home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torchaudio/_backend/ffmpeg.py", line 257, in save_audio
    s.write_audio_chunk(0, src)
  File "/home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torio/io/_streaming_media_encoder.py", line 469, in write_audio_chunk
    self._s.write_audio_chunk(i, chunk, pts)
RuntimeError: Input Tensor has to be 2D.
Exception raised from validate_audio_input at /__w/audio/audio/pytorch/audio/src/libtorio/ffmpeg/stream_writer/tensor_converter.cpp:31 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x96 (0x7def216cbf86 in /home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torch/lib/libc10.so)
frame #1: c10::detail::torchCheckFail(char const*, char const*, unsigned int, char const*) + 0x68 (0x7def2167add9 in /home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torch/lib/libc10.so)
frame #2: <unknown function> + 0x57147 (0x7def20ee0147 in /home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torio/lib/libtorio_ffmpeg4.so)
frame #3: <unknown function> + 0x57afc (0x7def20ee0afc in /home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torio/lib/libtorio_ffmpeg4.so)
frame #4: torio::io::TensorConverter::convert(at::Tensor const&) + 0x33 (0x7def20ee2723 in /home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torio/lib/libtorio_ffmpeg4.so)
frame #5: torio::io::EncodeProcess::process(at::Tensor const&, std::optional<double> const&) + 0xbe (0x7def20ed15ee in /home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torio/lib/libtorio_ffmpeg4.so)
frame #6: torio::io::StreamingMediaEncoder::write_audio_chunk(int, at::Tensor const&, std::optional<double> const&) + 0xa5 (0x7def20edcd85 in /home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torio/lib/libtorio_ffmpeg4.so)
frame #7: <unknown function> + 0x3a306 (0x7dee63d09306 in /home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torio/lib/_torio_ffmpeg4.so)
frame #8: <unknown function> + 0x32bf7 (0x7dee63d01bf7 in /home/ck/anaconda3/envs/tts/lib/python3.11/site-packages/torio/lib/_torio_ffmpeg4.so)
frame #9: python() [0x528767]
<omitting python frames>
frame #12: python() [0x5cbeda]
frame #14: python() [0x5ec6a7]
frame #15: python() [0x5e8240]
frame #16: python() [0x5fd192]
frame #21: <unknown function> + 0x29d90 (0x7def73e29d90 in /lib/x86_64-linux-gnu/libc.so.6)
frame #22: __libc_start_main + 0x80 (0x7def73e29e40 in /lib/x86_64-linux-gnu/libc.so.6)
frame #23: python() [0x5bbac3]

The tested code here:

import ChatTTS
import torch
import torchaudio

chat = ChatTTS.Chat()
chat.load(compile=False) # Set to True for better performance

texts = ["chat T T S is a text to speech model designed for dialogue applications.", "[uv_break]it supports mixed language input [uv_break]"]

wavs = chat.infer(texts)

for i in range(len(wavs)):
    torchaudio.save(f"basic_output{i}.wav", torch.from_numpy(wavs[i]).unsqueeze(0), 24000)

[fumiama]

I'm using GPU ann Linux, so do I must install it or not?

No, you should not install it.

I tested with a very simple code snippet, but it didn't work

It's a known problem, see #635

[ckgithub2019]

I'm using GPU ann Linux, so do I must install it or not?

No, you should not install it.

I tested with a very simple code snippet, but it didn't work

It's a known problem, see #635

it works, thanks. so that means all of "unsqueeze(0)" should be removed by default? the example is wrong or there are some other usages about "unsqueeze(0)"?

From example: torchaudio.save(f"output_sentencelevel{i}.wav", torch.from_numpy(wavs[i]).unsqueeze(0), 24000) torchaudio.save(f"output_wordlevel{i}.wav", torch.from_numpy(wavs[0]).unsqueeze(0), 24000)

[fumiama]

it works, thanks. so that means all of "unsqueeze(0)" should be removed by default?

No. In fact, some version (usually newer versions) of torchaudio will panic if unsqueeze(0) does not exist.

[kevincobain2000]

Tried with sound file instead, which works

import soundfile
soundfile.write(output_filename, wavs[0], 24000)

[fumiama]

Tried with sound file instead, which works

import soundfile
soundfile.write(output_filename, wavs[0], 24000)

Yes. There're many alternatives to save audio.