Andreas Köpf

I saw a similar error to the one you described without deepspeed ..to run with deepspeed you need to replace python on the command line with deepspeed, e.g. `deepspeed trainer_sft.py...

@EditaNEmilis could you please fix pre-commit & the solve the conflict? thx

OK, I assigned you. Please create a PR. :-)

@olliestanley Could you resolve the conflicts please? What does "installable" mean here (all imports currently seems to happen as relative imports to the worker entry-point)?

The `_ZN3c104cuda9SetDeviceEi` import error seems to be fixed with [v2.5.8](https://github.com/Dao-AILab/flash-attention/commit/9a11f440d3a34f618b4ba814c825b109c6d7e8f5), i.e. latest pip package of `flash-attn==2.5.8` works with PyTorch 2.3.0! thnx a lot @tridao !