Cherry-pick optimizer override refactor from #2723#2835

Merged

FDecaYed merged 21 commits intoNVIDIA:devfrom

yaoyu-33:chore/optimizer-override-cherrypick

Jan 14, 2026

Commits on Jan 6, 2026

Implementation of a more flexible optimizer/scheduler override system

jstjohn
authored and
yaoyu-33
committed
Add more comments to get param groups since the behavior is different (no longer 1-1 constraint)

jstjohn
authored and
yaoyu-33
committed
Fix circular import in optimizer_param_scheduler

jstjohn
authored and
yaoyu-33
committed
Add test coverage

jstjohn
authored and
yaoyu-33
committed
Adding test coverage for ParamKey and ParamPredicte matching

jstjohn
authored and
yaoyu-33
committed
Also add clarifying test around what happens if an attribute is there but False

jstjohn
authored and
yaoyu-33
committed
Remove unused import and reformatting

jstjohn
authored and
yaoyu-33
committed
Add/update more comments

jstjohn
authored and
yaoyu-33
committed
Show that ordering does not impact the merging of the param group keys

jstjohn
authored and
yaoyu-33
committed
Add utility function for returning a standard set of config overrides for the user to make use of

jstjohn
authored and
yaoyu-33
committed
Have default behavior for users of mcore who do not provide any skips

jstjohn
authored and
yaoyu-33
committed
Adding missing license headers

jstjohn
authored and
yaoyu-33
committed
Update test to cover new default

jstjohn
authored and
yaoyu-33
committed
Fix issue for fp8 cards that do not support block scaling

jstjohn
authored and
yaoyu-33
committed
Fix issue with consistency checking now that the overrides are partial

jstjohn
authored and
yaoyu-33
committed
Add test coverage for failing consistency check in CI

jstjohn
authored and
yaoyu-33
committed

Commits on Jan 7, 2026

Merge branch 'dev' into chore/optimizer-override-cherrypick
yaoyu-33
authored

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Cherry-pick optimizer override refactor from #2723#2835

Cherry-pick optimizer override refactor from #2723#2835
FDecaYed merged 21 commits intoNVIDIA:devfrom
yaoyu-33:chore/optimizer-override-cherrypick

Commits on Jan 6, 2026

Implementation of a more flexible optimizer/scheduler override system

Add more comments to get param groups since the behavior is different (no longer 1-1 constraint)

Fix circular import in optimizer_param_scheduler

Add test coverage

Adding test coverage for ParamKey and ParamPredicte matching

Also add clarifying test around what happens if an attribute is there but False

Remove unused import and reformatting

Add/update more comments

Show that ordering does not impact the merging of the param group keys

Add utility function for returning a standard set of config overrides for the user to make use of

Have default behavior for users of mcore who do not provide any skips

Adding missing license headers

Update test to cover new default

Fix issue for fp8 cards that do not support block scaling

Fix issue with consistency checking now that the overrides are partial

Add test coverage for failing consistency check in CI

Commits on Jan 7, 2026

Merge branch 'dev' into chore/optimizer-override-cherrypick

Commits on Jan 10, 2026

Fix layerwise CI issue

Lint

Merge branch 'dev' into chore/optimizer-override-cherrypick

fix small issuer