verl [fix] Communicating for non_tensor_batch in fsdp

[fix] Communicating for non_tensor_batch in fsdp_vllm.py

Open UbeCc opened this issue 9 months ago • 4 comments

Current verl only perform communication for non_tensor_batch at fsdp_ulysses.py. I support the communication for fsdp_vllm.py

Feb 27 '25 06:02 UbeCc

Actually, could you move this functionality as a class method of DataProto? Thanks.

Feb 27 '25 07:02 vermouth1992

All committers have signed the CLA.

Feb 27 '25 07:02 CLAassistant

Actually, could you move this functionality as a class method of DataProto? Thanks.

Fine, it's already done

Feb 27 '25 07:02 UbeCc

I think this feature already exists in https://github.com/volcengine/verl/pull/386

Feb 27 '25 09:02 hiyouga

This feature is already implemented in #497

Mar 07 '25 12:03 vermouth1992