feat(pt_expt): align the training runtime with pt - #5958

Merged

OutisLi merged 6 commits into

deepmodeling:masterfrom

OutisLi:pr/pt-expt-training

Aug 8, 2026

Commits on Aug 4, 2026

feat(pt_expt): close the checkpointing, EMA and training-report gap with pt
OutisLi
committed
58f7301
feat(pt_expt): shard training state across ranks and guard the gradient
OutisLi
committed
6d121cc
fix(train): inherit EMA checkpoint retention by default
OutisLi
committed
775b68f

Commits on Aug 5, 2026

fix(train): correct shared runtime edge cases
OutisLi
committed
c814ddd

Commits on Aug 7, 2026