Fix Double Application of Softmax for Router Logits in MoE models#45346

Open

ionut-anghelina wants to merge 3 commits intohuggingface:mainfrom

ionut-anghelina:dev/ionut/FixDoubleSoftmax

Commits on Mar 30, 2026

[Bugfix] Fix double softmax in MoE router load-balancing loss

ionut-anghelina
and
claude
committed

Commits on Apr 9, 2026

Add regression tests and fix dtype cast to use raw logits dtype

ionut-anghelina
and
claude
committed

Commits on Apr 13, 2026

Merge branch 'main' into dev/ionut/FixDoubleSoftmax
vasqu
committed