[Trainer] Preserve MoE router auxiliary loss with label smoothing #51083
Fix Trainer label smoothing dropping MoE router loss
6fbf583e
Simplify MoE label smoothing regression test
ce42eccf
Reuse Qwen3-MoE tester configuration in Trainer regression
ea80a09e
Preserve full router loss during evaluation
7af64c04
Simplify auxiliary loss scaling and regression setup
d2f20afc
Focus label smoothing regression on distinct loss paths
66e037f1
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub