Open jojac47 opened 1 month ago
I added RUN python3 -m pip uninstall torchaudio -y RUN python3 -m pip install torchaudio to the end of the docker file and got the omega conf error that somebody else has gotten recently. I'm rebuilding with the fix they found which is changing the requirement to <= 2.3 ImportError: cannot import name 'SCMode' from 'omegaconf' #136
I still had to leave
RUN python3 -m pip uninstall torchaudio -y RUN python3 -m pip install torchaudio
at the end of the docker file but it made it to download models step inside the running container. I'll keep updating
That seemed to fix it but the model does seem slower
It seems to have started using deep speed properly and is using the gpu for sure now I was seeing a [map to cpu] some where in the output logs when running generate on random voice. I didn't notice much of a speed up, but I seem to be getting a 1:1 Realtime generation speed on random voice. I am using a v100. I was getting about 1.25 Realtime on my local install 4060 on windows.
I also changed the commands listed above to this
RUN python3 -m pip uninstall torchaudio -y RUN python3 -m pip install torch torchvision torchaudio
The Vcoder model is loading on CPU. I wonder if this is because of some of the newer features Ampere architecture and beyond have. I'm going to test on my Windows docker install and see if this works out of the box.
So I found an error when running your docker start.sh. It says os error: unrecognized symbol. This is usually an issue if your torch audio and torch files are either corrupted or un compatible. I noticed in your dockerfile that you use the index for the Windows pytorch whl. I'm rebuilding with the linux specific command, and I'll let you know what I find.