fix(mimo): Scale encoder gradients for heterogeneous DP in multimodule finalization#3197

Closed

kamran-nvidia wants to merge 8 commits intomainfrom

kamran/fix_grad_scale_mimo

Commits on Mar 18, 2026

feat(mimo): Phase 4 — MiMo training, model/provider, data loading, parallelism
aroshanghias-nvd
committed

Commits on Mar 25, 2026

fix(mimo): adapt model layer for MCore submodule bump (#2978)

yashaswikarnati
and
claude
authored
feat(mimo): Phase 5 — checkpoint save/resume, evaluation, e2e tests
aroshanghias-nvd
committed
fix(mimo): update test_pretrain_mimo fixture for Phase 5 checkpoint wiring
aroshanghias-nvd
committed
feat(mimo): add run_mimo_checkpoint_resume.sh wrapper for e2e checkpoint test
aroshanghias-nvd
committed
fix(mimo): Fix for MIMO Slice Batch (#2956)

kamran-nvidia
authored and
aroshanghias-nvd
committed
fix(mimo): adapt training layer for MCore submodule bump (#2979)
yashaswikarnati
authored

Commits on Apr 7, 2026

fix(mimo): add gradient scaling for heterogeneous DP in finalize_model_grads_multimodule
kamran-nvidia
committed