[BUG] GPU id error when restarting on a different instance

[maxjeblick]

🐛 Bug

Hi :D

This is a follow up on #99.

I'm currently using LLM Studio on different instances that have varying number of GPUs. data and output folder are stored persistently and I attach them when starting a new instance.

On a single GPU instance, when I start from a previous experiment that was using gpu id 2, I get the following error:

ImportError: Using `bitsandbytes` 8-bit quantization requires Accelerate: `pip install accelerate` and the latest version of bitsandbytes: `pip install -i https://pypi.org/simple/ bitsandbytes`

rather than an error like "Gpu id 2 does not exist on this machine, please change your configuration".

It seems that

• config checks do not account for gpu id mismatch
• train.py either silently switches to cpu or raises the error above

To Reproduce

1. Run an experiment on a machine with at least 2 gpus, select gpu 1. Can be default settings.
2. Restart llm studio, with only gpu 0 as visible device.
3. Click on New experiment from current experiment.
4. The following message should appear ImportError: Using bitsandbytes8-bit quantization requires Accelerate:pip install accelerateand the latest version of bitsandbytes:pip install -i https://pypi.org/simple/ bitsandbytes``
5. If using float16, etc. no error should occur, but model will train on cpu.

LLM Studio version

bc2ff4f