DeepSpeed on ROCm 6

[MisterNorwood]

While the project does not support AMD natively, I have seen people run the model with relative ease.
Looking through documentation and repositories, I think it is possible to make DeepSpeed work in rocm, since it is officially supported. My question is if anyone has gotten it to work and if, how can one do so? Default script runs installation for CUDA and expects CUDA libraries.

[rose07]

https://tts.byylook.com/ai/text-to-speech

[jschoch]

here's one way

#739

I was also able to get this working in wsl on windows.

my requirements.txt

tqdm
rotary_embedding_torch
transformers==4.31.0
tokenizers
inflect
progressbar
einops==0.4.1
unidecode
scipy
librosa==0.9.1
ffmpeg
numpy==1.26.4
numba
torchaudio
threadpoolctl
llvmlite
appdirs
--extra-index-url https://download.pytorch.org/whl/rocm6.2
nbconvert==5.3.1
tornado==4.2
pydantic==1.9.1
deepspeed==0.8.3
py-cpuinfo
hjson
psutil
sounddevice
spacy==3.7.5

then i ran

cd <the clone dir>
pip install -e .

I also had to install torchaudio, since i'm on rocm 6.2.3 and torch 2.3 i picked this one:

wget https://download.pytorch.org/whl/rocm6.0/torchaudio-2.3.0%2Brocm6.0-cp310-cp310-linux_x86_64.whl#sha256=dcac7da7d4e19d42a4a48811bab7d35a8cf05b5105f4e05f6ad306c20cf5329e
pip install torchaudio-2.3.0+rocm6.0-cp310-cp310-linux_x86_64.whl

Then i had some issue with more_itertools so this fixed it up

pip install more-itertools==10.3.0