Conversation
Signed-off-by: Jin, Youzhi <youzhi.jin@intel.com>
There was a problem hiding this comment.
Thanks for picking this up! I think this needs a rebase on #8600, the grads come from group_items there now since param.grad isn't the full grad at step time.
Btw NCCL seems fine with uneven all_gather, which is probably why only the gloo job failed.
Thank you for remind! Yes, the ut only failed in Gloo, NCCL better compatible on this case. |
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: Jin, Youzhi <youzhi.jin@intel.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
alanhuangyoo
left a comment
There was a problem hiding this comment.
Thanks, the rebased version looks good to me. The ZeRO-3 Muon tests pass on 2×H20, and ZeRO-3 still matches ZeRO-0 bit for bit with mixed shapes.
|
Thanks! @alanhuangyoo |
Follow up #8381
Fix 3 FAILED
unit/v1/ops/muon/test_per_head_muon_accelerator.pyin https://github.com/deepspeedai/DeepSpeed/actions/runs/35675316100/job/106580600974?pr=8381@delock