Unable to load unsloth models in just single GPU in a multi GPU system

[karthik-codex]

I have two RTX A6000 and I understand the opensource unsloth version only supports single GPU. Is there anyway I can config to use only one GPU? I currently get this error when I import unsloth and load models

[coenvdgrinten]

In the command line, you can try setting the CUDA_VISIBLE_DEVICES environment variable to a single number, e.g.:
export CUDA_VISIBLE_DEVICES=0

You could look at nvidia-smi for which GPU you want to use.

[karthik-codex]

I still get the same error

[mahiatlinux]

Try using the CUDA_VISIBLE_DEVICES environment variable in the script itself.

import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'  # Use GPU 0

or

import os
os.environ['CUDA_VISIBLE_DEVICES'] = '1'  # Use GPU 1

Put it at the top of the script. By the way, the first GPU is "0" and second one is "1".

[Sehyo]

I have implemented a fix for this here:
#974

[danielhanchen]

Will look into @Sehyo's PR and implement a fix - sorry everyone!

[jlin816]

I'm still getting this issue, is there a fix planned? I patched with @Sehyo 's fix locally, but 50% of runs still fail:

trainer.train()
  File "<string>", line 154, in train
  File "<string>", line 219, in _fast_inner_training_loop
RuntimeError: Unsloth currently does not support multi GPU setups - but we are working on it!

[giuliabaldini]

Any updates on this? Also having the same problem. Would it be possible to put it as a warning to be able to ignore it?