DeepSpeed
Count each expert once in the fp32 gradient-clipping norm
#8478
Open

Count each expert once in the fp32 gradient-clipping norm #8478

alanhuangyoo
alanhuangyoo Count each expert once in the fp32 gradient-clipping norm
875e8fe8
alanhuangyoo alanhuangyoo requested a review from tjruwase tjruwase 7 days ago
alanhuangyoo alanhuangyoo requested a review from tohtana tohtana 7 days ago
alanhuangyoo alanhuangyoo requested a review from loadams loadams 7 days ago
alanhuangyoo
alanhuangyoo alanhuangyoo marked this pull request as draft 7 days ago
alanhuangyoo Decide the expert path from ownership, not from this step's gradients
338ed514
alanhuangyoo
alanhuangyoo alanhuangyoo marked this pull request as ready for review 7 days ago
ebarkhordar
ebarkhordar commented on 2026-09-13
alanhuangyoo Vote on the expert path across stages, and not at all without MoE
1a761e8d
alanhuangyoo

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone