DeepSpeed
Give Muon's momentum the dtype of the gradient it is combined with
#8483

Merged

Give Muon's momentum the dtype of the gradient it is combined with #8483

delock merged 3 commits into deepspeedai:master from alanhuangyoo:fix/muon-momentum-dtype-both-paths

Give Muon's momentum the dtype of the gradient it is combined with

37028218

alanhuangyoo requested a review from

tjruwase 6 days ago

alanhuangyoo requested a review from

tohtana 6 days ago

alanhuangyoo requested a review from

loadams 6 days ago

delock requested a review from

delock 4 days ago

delock commented on 2026-09-15

delock requested changes on 2026-09-15

Merge master into fix/muon-momentum-dtype-both-paths

42095e61

Address review: drop the dead device check, cover the padded path, mo…

862174ae

delock approved these changes on 2026-09-15

delock enabled auto-merge 2 days ago

delock merged 431f8ace into master 2 days ago

Reviewers

delock

tjruwase

tohtana

loadams

Assignees

No one assigned

Labels

None yet

Milestone

No milestone