Running the training script on CPU-only machines

With the current default settings, running the training script on a CPU-only machine can raise the following error:

ValueError: fp16 mixed precision requires a GPU (not 'mps').

This is caused by an issue in the accelerate package that has been fixed in this Pull Request

Update accelerate to the latest version to get code of the above PR and (optionally) upgrade torch to the required version as shown below.

uv remove accelerate
uv add git+https://github.com/huggingface/accelerate.git
uv add torch==2.8.0

After updating, the training script should run on CPU-only environments without the fp16/GPU error.

NOTE: Running the training script on CPU-only machines is intended for demonstration purposes only. For full training runs, it is recommended to use a GPU-enabled machine.