[skyrl-train] Implement loss reduction via advantage normalization and fix `token_mean` reduction strategy#1296

Open

justinvyu wants to merge 20 commits intoNovaSky-AI:mainfrom

justinvyu:token_mean_loss_reduction

Commits on Mar 9, 2026

Move loss reduction normalization to trainer-level advantage scaling, scale loss by dp_size for FSDP/Megatron parity

justinvyu
and
claude
committed
Add token_mean_baseline loss reduction for mean-of-microbatch-means comparison

justinvyu
and
claude
committed
fix assertion
justinvyu
committed

Commits on Mar 10, 2026

Update tests for sum-based reduce_loss and dp_size scaling changes

justinvyu
and
claude
committed

Commits on Mar 17, 2026

Commits on Mar 20, 2026

Commits on Mar 25, 2026

Fix metric reporting: remove dp_size scaling, separate micro-batch vs mini-batch reduction

justinvyu
and
claude
committed

Commits on Mar 27, 2026