Single GPU training in Multi-GPU system doesn't work.

[Sehyo]

Single GPU training in Multi-GPU system doesn't work even if limited to 1 GPU with os.environ CUDA_VISIBLE_DEVICES before importing unsloth.

Reason:
check_nvidia function spawns new process to check nvidia-smi, thus bypassing the cuda_visible_devices set.

Fix:
#974

Pushed fix here, changed check_nvidia as well as patched the handler to check only the GPU that is set by os environ. If no os environ set, defaults to original behaviour.

[danielhanchen]

Ok will check your PR out - thanks so much!

[Sehyo]

Ok will check your PR out - thanks so much!

Any update? fyi a solution like "CUDA_VISIBLE_DEVICES={x} nvidia-smi" won't work as that still prints out all GPUs anyways.