Any plans to support megatron for GRPO training?

Open sys-reasoner opened this issue 10 months ago • 1 comments

Hi guys,

Thanks for this great job! I am wondering if you have any plans to support megatron for GRPO training?

It would really help when it's needed to full GRPO train large size model, like 72B Qwen2.5VL.

Apr 03 '25 04:04 sys-reasoner

Same request. I am looking forward to using Megatron for training LLM with GRPO. It can help me save GPU resources.

I have seen that Verl added Megatron GRPO support. I am wondering if ms-swift will add this feature in the future. (https://verl.readthedocs.io/en/latest/workers/megatron_workers.html)

Apr 26 '25 10:04 heyubox

stay tuned at https://github.com/modelscope/ms-swift/issues/4561

Jun 26 '25 12:06 hjh0119