Remove bitsandbytes version to fix CUDA version 12x compatible problem

[ArtificialCleverness]

No description provided.

[ArtificialCleverness]

It might fix following issues:
#270
#188

[kohya-ss]

Thank you for this! It is a good idea in general.

However, I worry if this will work in a windows environment. I think rewriting main.py or cextension.py in the latent bitsandbytes may not work.

It might be a good idea to describe the installation instruction in Linux as an option. But I do not have time to test with Linux environment...

[ArtificialCleverness]

I have done my testing on RTX4090 + cu121 + python3.10.11 + Ubuntu 22.04, it works fine for me. I think GUI contributor also have done the testing on Windows. Please check the discussions below.
bmaltais/kohya_ss#686
bmaltais/kohya_ss#684

[ArtificialCleverness]

Also, I do think 'bitsandbytes' is currently under developing, there are many compatible issues. I have read many issues discussing cuda12x problems. I also have done trying many installation/compilation methods they given, but only this one works for me. TimDettmers/bitsandbytes#185

[kohya-ss]

Thank you for update!

It seems that the version of bitsandbytes in the Windows is still 0.35.0 according to requirements.txt. Then I think it will work on Windows without any problem.

But I worry that if some people train on a Linux environment with CUDA 11.x, or on Colab, it may not work if we change bitsandbytes to this version.

I cannot test this myself and am concerned that it will affect these people.

[ArtificialCleverness]

Your concern is definitely correct, I think bitsandbytes is currently working on this according to their plan recently, and I found an open PR adding docs about this.
TimDettmers/bitsandbytes#337

GUI version has already merged this PR regarding b&b version, and we could just observe whether people is getting new errors about env. Also, I do not think we are able to do a comprehensive test on Linux or giving detailed instructions about bitsandbytes installation, I think we just wait bitsandbytes to fix their issue would be a better idea, since they are currently working on cu12x supports.

[kohya-ss]

Thank you for your comment.

So I would like to keep an eye on the GUI repository a little longer. Also, I checked the bitsandbytes repository and it seems that not only CUDA 12 but also 11 are supported, so Colab may not have any problems.

[hemslo]

I'm having a similar issue and was about to raise a PR like this, I'm using bitsandbytes==0.39.0

Tested in RTX4070, cuda 12.2, python 3.10.11, pytorch 1.31.1, Ubuntu 22.04.

My linux install steps are:

# https://pytorch.org/get-started/previous-versions/#linux-and-windows-2
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia

pip install --upgrade -r requirements.txt

# https://github.com/facebookresearch/xformers#installing-xformers
conda install xformers -c xformers