Install in Linux without GPU

[timur-g]

This is a question that may have simple answer, but I did not find it. Looked also #1111. I am new to this and modern TTS.

I am in Linux Mint and do not use snap or flatpak. What is the difference between using pip install buzz-captions.. and using https://github.com/chidiwilliams/buzz/blob/main/CONTRIBUTING.md#linux-ubuntu ?
Are there some options to control the installation?
When I tried with pip, I see it downloading multiple nvidia_... packages, but why, when I have no nvidia. It's laptop with AMD CPU without separate GPU: AMD Ryzen 7 PRO 6850U with Radeon Graphics.
Also in the Running Locally instructions, it says Add fix for nvidia cudnn library path to the virtual environment, can I skip that?

[raivisdejus]

If you do not have Nvidia GPU, you can skip everything Nvidia related. Buzz should be smart enough to run the underlying models with the resources you have. Some may run slow.

BUT :)
Latest development code has support for Vulcan for Whisper.cpp, so if you choose this whisper type it will still run pretty fast.

Separation of Nvidia for pip installation is somewhere in plans, and I could use this issue to track and implement it in some not too distant future.

In the meanwhile feel free to post any issues you run into, I will gladly help a fellow Linux user

[timur-g]

Thanks. So pip needs to be adjusted for nvidia/amd/none gpu but how to skip in manual installation,
I see that poetry install also fetches nvidia files.

[raivisdejus]

Just ignore the Nvidia files. They will take up space on your computer but will not do anything. Underlying models do check if Nvidia CUDA is available and if it is they will use it, if not they will use other methods for calculations. Whisper.cpp does need the Nvidia / CUDA at all, other Whipser types may use them.

@timur-g Curious if extra few GB of Nvidia files is an issue for users in real life and the future possible no-Nvidia version would benefit from being much smaller and having only whisper.cpp as transcription engine.

[timur-g]

This is my example: I have 2 TB disk that before installing had only 25 GB free. Avoding any few GB is welcome.

I looked at pyproject.toml but cannot see how to avoid nvidia, maybe that is a dependency for some of these.
IIUC a group may be excluded --without=nvidia, but there is no such a group.
I see nvidia packages in poetry.lock but not aware how to skip them from there.

Also, are these nvidia packages really only for Nvidia or those packages are also for AMD GPU? Which I understood is supported.

Note for above with pip: #1100 says "By default pip will install CPU version and there will be instructions .. to add GPU support."

[timur-g]

As mentioned in bneijt/poetry-lock-package#47 (comment) :

The nvidia package is probably a dependency of one of the packages in required by the buzz project (one of the packages in https://github.com/chidiwilliams/buzz/blob/main/pyproject.toml#L32

[raivisdejus]

Yes current setup of current Buzz version requires the nvidia package and bundles the GPU support. Nvidia packages are needed for torchaudio.

Without torchaudio features, like voice separation from the noisy audio or Huggingface whisper will not work.

In the future we could implement separate versions of buzz with nvidia and without, but some time to implement and test this is needed.