Zijie Yan comments

Results 58 comments of


                                            Zijie Yan

Please ensure we have a passing functional test pipeline on GitLab before merging.

/ok to test [1c3e5ce](https://github.com/NVIDIA/Megatron-LM/pull/2388/commits/1c3e5ce1822ebbf3b8172ab3b36e350ee2dcea04)

Hey @tomlifu, could you open a mirror PR to main?

DeepEP is optimized for large topk with cross-node EP(EP>8) scenarios. Based on our experience, for EP

Thanks @nrailg, I'm taking a look and will consult with the PICs.

Thanks for flagging the issue—will add an assertion soon.

Merged at https://github.com/NVIDIA/Megatron-LM/commit/9c1a53515582d826b82ac133de5bc7e0a0ce4142

Thanks for reporting the issue. This should be fixed in commit https://github.com/NVIDIA/Megatron-LM/commit/e6d56d6828c0773f55772b92b2ec0eed5639665e.