Daniel Han

@Erland366 Could you check if vLLM works still if no LoRA adapters are added? I think you also had a PR on moving `load_lora` outside of `get_peft_model`

@wdlctc Thanks a lot again!! I'll test it and verify all losses match! Appreciate it!

Sorry on the delay - was planning to add this together with Vision support :) It might take a few more days!

Oh lol I noticed I accidentally deleted this PR after I deleted the nightly branch - whoops so sorry!

Interesting so I looked through the paper and code, essentially you're proposing to essentially do gradient accumulation inside of each sequence length? Ie the first is normally chunking the CE...

@Erland366 Could you confirm if my latest fix allows multi GPU to work OK? Thanks. I think it's CCE related but unsure

Do you mean the logging that I provided? Or as in you want to use a reward model? GRPO in general doesn't use a reward model - it calculates advantages...

Yes working on it!

Hopefully out soon.

@nottrz @DiTo97 @Summer142857 @Summer142857 Apologies just fixed! For Colab / Kaggle, please restart and run all. For local machines, please do: ``` pip install --force-reinstall --upgrade --no-cache-dir --no-deps unsloth unsloth_zoo...