Whisper crashes with SIGSEGV error

[HristovCodes]

Using whisper with the --device cuda option or emitting it because by default it uses cuda throws the following error:

fish: Job 1, 'whisper pytorch_env/test.mp3 --…' terminated by signal SIGSEGV (Address boundary error)

I tested it in cpu mode with --device cpu and it works flawlessly (it seems) but it's very slow, and when running it even in the console it tells me UserWarning: Performing inference on CPU when CUDA is available.

The gpu I'm using is RX 6700 XT if it matters.

Here's the output of pip list:

Package             Version
------------------- --------------------------
certifi             2022.12.7
charset-normalizer  2.1.1
cmake               3.25.0
ffmpeg-python       0.2.0
filelock            3.9.0
fsspec              2023.4.0
future              0.18.3
idna                3.4
Jinja2              3.1.2
lit                 15.0.7
llvmlite            0.40.1
MarkupSafe          2.1.2
more-itertools      10.1.0
mpmath              1.2.1
networkx            3.0
numba               0.57.1
numpy               1.24.1
openai-whisper      20230314
Pillow              9.3.0
pip                 22.3.1
pytorch-triton-rocm 2.1.0+05d67b9418
regex               2023.8.8
requests            2.28.1
semantic-version    2.10.0
setuptools          65.5.0
setuptools-rust     1.7.0
sympy               1.11.1
tiktoken            0.3.1
torch               2.1.0.dev20230828+rocm5.6
torchaudio          2.1.0.dev20230828+rocm5.6
torchvision         0.16.0.dev20230828+rocm5.6
tqdm                4.66.1
triton              2.0.0
typing_extensions   4.4.0
urllib3             1.26.13

Running torch.cuda.is_available() also returns True.

I'm running latest version of Arch. And yes I tried using both the stable version of pytorch and the nightly version but both give the same error. I also tried using the rocm/pytorch docker image but that also gives the same error.

Any help will be appreciated because it feels like I've been hitting my head against the wall for days and nothing I try and no ammount of googling seems to help.

[glangford]

See this AMD community discussion thread

• https://community.amd.com/t5/infinity-hub-container/radeon-rx6600xt-ubuntu-22-04-and-rocm-woes/m-p/584733

and this whisper discussion (I highlighted a couple of comments within that thread). I am not an AMD user so I can't assist beyond this, but fyi in case it is helpful.

• ROCm support #55
• ROCm support #55 (reply in thread)
• ROCm support #55 (comment)

[glangford]

"In case it helps anyone else, I needed to install rocm-libs and set environment variable HSA_OVERRIDE_GFX_VERSION=10.3.0 to make this work."

[HristovCodes]

Setting the correct environment variable seemed to do the job for me.
For anyone else reading this you can do it with the command export like so export HSA_OVERRIDE_GFX_VERSION=10.3.0
I did already have amdgpu drivers installed so that migth have been why I didn't need additional configuration other than setting the variable.
Thanks for pointing me in the rigth direction!