ROCm isnt supported

[MidnightKittenCat]

Is it possible to add ROCm support for amd gpus?

[absadiki]

I think Pytorch supports ROCm, so it should be supported out of the box basically, but I didn't test it so I couldn't tell.
You can give it a try, and let me know how it goes :

1. create a new virtual environment.
2. install Pytorch ROCm compatible

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2

3. Then install subsai (as described in the README)

[MidnightKittenCat]

Doesn't seem like it, complains about cuda even if I've used the rocm one.

Seems to be hardcoded and there isnt a rocm version for the device type.

[MidnightKittenCat]

Pytorch does support rocm but this just complains about cuda libraries

[MidnightKittenCat]

RuntimeError: CUDA failed with error CUDA driver version is insufficient for CUDA runtime version

[absadiki]

Doesn't seem like it, complains about cuda even if I've used the rocm one.

Seems to be hardcoded and there isnt a rocm version for the device type.

AFAIK, Pytorch is using cuda device name even for ROCm devices, so this is not a problem.

[absadiki]

Pytorch does support rocm but this just complains about cuda libraries

if it complains about cuda libraries, then maybe your GPU chip is not supported by Pytorch.
Please make sure first that your GPU is supported from the ROCm docs, then run some tests so you can know if it is working well or not!

[MidnightKittenCat]

It is indeed supported by rocm, like I said it's not pytorch that's complaining about it.

[MidnightKittenCat]

I've tried a conda env that didn't work, tried doing it globally also didn't work.

My GPU Chip is supported because I can use it on S.D. Next (Stable diffusion webui) just fine

[MidnightKittenCat]

Could you attempt to add ROCm to the device list and I can test it out for you or?

[absadiki]

Weird! cause it should be working with the same device name of cuda as well!
See this comment here as well.
So you think this is bug in the package ? Do you have any idea where this is coming from ? what model are you trying to use ? what is the full error you are getting ?

[MidnightKittenCat]

I don’t think it’s an issue with PyTorch itself, it seems to me that the cuda device could be hardcoding cuda or something similar, as PyTorch with the same version on rocm works completely fine for me on other projects.

The full error I was getting was mainly what I’ve sent you, however I’ll go ahead and get you it.

Perhaps take a look at the code and see if cuda is hardcoded.

[absadiki]

Yes cuda is hardcoded if torch.cuda.is_available() is True.
To what a value should I change it so that ROCm device should be supported ?

[MidnightKittenCat]

I believe I've found the issue, "EXPORT HSA_OVERRIDE_GFX_VERSION=10.3.0" seems to of fixed the issue for me, so for future cases I'd redirect people to this fix and see if that works for them.

Unfortunately it's just AMD making it hard to identify their own hardware.

Sorry for the troubles <3

[MidnightKittenCat]

I would however like a feature that embeds the subtitles into the video, allowing us to download it afterwards (hopefully to get around the 200mb limit)

Or even just making a flag to disable it. Unless there is one?

[absadiki]

I believe I've found the issue, "EXPORT HSA_OVERRIDE_GFX_VERSION=10.3.0" seems to of fixed the issue for me, so for future cases I'd redirect people to this fix and see if that works for them.

Unfortunately it's just AMD making it hard to identify their own hardware.

Sorry for the troubles <3

Great to hear that you found the source of the issue and thanks for posting the solution, It will certainly help other AMD users facing similar issues. I will add a reference to this issue in the README file as well.

[absadiki]

I would however like a feature that embeds the subtitles into the video, allowing us to download it afterwards (hopefully to get around the 200mb limit)

Or even just making a flag to disable it. Unless there is one?

You mean to merge the video with the subtitles directly ? without exporting just the srt file ?

[absadiki]

Yeah the 200mb limit is imposed by the video component of streamlit, I tried to look for alternatives that support subtitles but unfortunaltey I couldn't find any! I will see if I there is any other solution.
you can always use the CLI, it is easy and there is no limit!

[MidnightKittenCat]

Sounds great! Also is it possible to have multiple files in a queue?

[absadiki]

Yes using the CLI, you can provide a text file containing the absolute path of the files, it will run them one by one,

[MidnightKittenCat]

A little update on this, seems only "whisper" by OpenAI and "whisper timestamped" detects rocm (cuda:0) the rest do not

[absadiki]

I've re-checked the device attribute of all models, I have fixed WhsiperX and hopefully it should be working now, please give it a try!
For fatser-whisper, I am not sure if supports ROCm, because the implementation is in C++ and they are not using Pytorch AFAIK.

[MidnightKittenCat]

For WhisperX, I'm getting this error:

ValueError: unsupported device cuda:0 Traceback: File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/streamlit/runtime/scriptrunner/script_runner.py", line 565, in _run_script exec(code, module.__dict__) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/webui.py", line 523, in <module> run() File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/webui.py", line 516, in run webui() File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/webui.py", line 316, in webui subs = _transcribe(file_path, stt_model_name, model_config) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/streamlit/runtime/caching/cache_utils.py", line 194, in wrapper return cached_func(*args, **kwargs) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/streamlit/runtime/caching/cache_utils.py", line 223, in __call__ return self._get_or_create_cached_value(args, kwargs) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/streamlit/runtime/caching/cache_utils.py", line 248, in _get_or_create_cached_value return self._handle_cache_miss(cache, value_key, func_args, func_kwargs) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/streamlit/runtime/caching/cache_utils.py", line 302, in _handle_cache_miss computed_value = self._info.func(*func_args, **func_kwargs) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/webui.py", line 189, in _transcribe model = subs_ai.create_model(model_name, model_config=model_config) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/main.py", line 95, in create_model return AVAILABLE_MODELS[model_name]['class'](model_config) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/models/whisperX_model.py", line 123, in __init__ self.model = whisperx.load_model(self.model_type, File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/whisperx/asr.py", line 50, in load_model model = WhisperModel(whisper_arch, File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/faster_whisper/transcribe.py", line 120, in __init__ self.model = ctranslate2.models.Whisper(

However it does say cuda:0 (which usually indicates that the gpu is detected) so something is wrong here.

[absadiki]

Oh Yeah, WhisperX is using Faster-whisper as its backend! so I doubt if it will work in your case!

[MidnightKittenCat]

Ah that’s very unfortunate, thank you for trying though!

I would however like a feature that embeds the subtitles into the video, allowing us to download it afterwards (hopefully to get around the 200mb limit)
Or even just making a flag to disable it. Unless there is one?

You mean to merge the video with the subtitles directly ? without exporting just the srt file ?

And yes, I do mean this.

[absadiki]

Ah that’s very unfortunate, thank you for trying though!

I would however like a feature that embeds the subtitles into the video, allowing us to download it afterwards (hopefully to get around the 200mb limit)
Or even just making a flag to disable it. Unless there is one?

You mean to merge the video with the subtitles directly ? without exporting just the srt file ?

And yes, I do mean this.

Ok, I've added this feature.
If you are using the webui, you can find in the export section.
Please give it a try and let me know if you find any issues ?

[MidnightKittenCat]

Ah that’s very unfortunate, thank you for trying though!

I would however like a feature that embeds the subtitles into the video, allowing us to download it afterwards (hopefully to get around the 200mb limit)
Or even just making a flag to disable it. Unless there is one?

You mean to merge the video with the subtitles directly ? without exporting just the srt file ?

And yes, I do mean this.

Ok, I've added this feature. If you are using the webui, you can find in the export section. Please give it a try and let me know if you find any issues ?

Very sorry I was very busy.

I'm now getting this error:

Error: ffmpeg error (see stderr output for detail) Traceback: File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/streamlit/runtime/scriptrunner/script_runner.py", line 565, in _run_script exec(code, module.__dict__) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/webui.py", line 535, in <module> run() File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/webui.py", line 528, in run webui() File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/webui.py", line 518, in webui exported_file_path = tools.merge_subs_with_video({subs_lang: subs}, str(media_file.resolve()), exported_video_filename) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/subsai/main.py", line 299, in merge_subs_with_video ffmpeg.run(output_ffmpeg) File "/home/midnight/miniconda3/envs/caption/lib/python3.10/site-packages/ffmpeg/_run.py", line 325, in run raise Error('ffmpeg', out, err)

[MidnightKittenCat]

Hopefully this helps.

[MidnightKittenCat]

Could we also perhaps look into this? https://huggingface.co/facebook/nllb-200-3.3B

[absadiki]

Hopefully this helps.

Seems like FFmpeg cannot infer the encodec format of your video file. Are you perhaps applying the function on a an audio file ? if not could you please share a sample so I can test on my end ?

[MidnightKittenCat]

Very sorry something came up again, I seemed to of gotten it to work by using an mp4 this time.

When it comes to the 200mb limit, is this just for displaying the video itself? If so can we have an option that if the video is >200mb that it doesn’t embed it and only gives u the merge video with subtitles/download video button? Thanks!

[absadiki]

@MidnightKittenCat, it is already the case I think, if the video exceeds 200mb you just can't view it, but you can transcribe and merge as well.

[MidnightKittenCat]

Finally got around to trying this for myself, this doesn't seem to be the case.\

@MidnightKittenCat, it is already the case I think, if the video exceeds 200mb you just can't view it, but you can transcribe and merge as well.

"File must be 200.0MB or smaller." when inputting a file over 200mb, I was hopping it can be transcribed/merged without showing the video if >200mb

[matusnovak]

I think Pytorch supports ROCm, so it should be supported out of the box basically, but I didn't test it so I couldn't tell. You can give it a try, and let me know how it goes :

1. create a new virtual environment.

2. install Pytorch ROCm compatible

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2

3. Then install `subsai` (as described in the README)

Thank you!

I had to use rocm5.7 instead of rocm5.4.2 and that did the trick for me on Manjaro Linux 6.2.16 with AMD Radeon 6800XT

[insberr]

I am getting RuntimeError: CUDA failed with error CUDA driver version is insufficient for CUDA runtime version. I installed pytorch with rocm6.0 and I have a RX 6800. I am also using the m-bain/whisperX model. Any ideas on how to fix this?

[absadiki]

I am getting RuntimeError: CUDA failed with error CUDA driver version is insufficient for CUDA runtime version. I installed pytorch with rocm6.0 and I have a RX 6800. I am also using the m-bain/whisperX model. Any ideas on how to fix this?

@insberr, maybe you should downgrade rocm to a lower version as I described in the comment before, rocm5.7 worked for @matusnovak, so maybe try that one.

[aidan-gibson]

I also have an RX6800; the following setup worked for me:

micromamba create -n whispergpu python=3.11.9
micromamba activate whispergpu
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.1
pip install git+https://github.com/abdeladim-s/subsai

And my conda channels are:

• conda-forge
• defaults
• pypi

[Snuupy]

There is now a fork of CTranslate2 that works with whisperX OpenNMT/CTranslate2#1072 (comment)

Can this be included in subsai? @abdeladim-s

[absadiki]

@Snuupy, Good to know that there is finally a fork that works for ROCm.
From the installation instructions it doesn't seem straightforward to include it as a package in subsai's dependencies, but you could give it a try. Just install the fork in the same virtual environment where subsai is installed—it should work.