Single grouped weight fixes - #3225

Merged

timmoon10 merged 5 commits into

NVIDIA:mainfrom

CarlosGomes98:cgomes/single-grouped-weight-independent-fixes

Jul 27, 2026

Commits on Jul 24, 2026

[PyTorch] Preserve grouped weight initialization metadata
CarlosGomes98
committed
019738d
[PyTorch] Mark late grouped weights for delayed wgrad
CarlosGomes98
committed
ab4d7e8
[PyTorch] Preserve grouped MXFP8 columnwise usage
CarlosGomes98
committed
b4f2be1

Commits on Jul 27, 2026

Fix bug where grouped MLP used same quantizer in grouped linear op and weight param

timmoon10
authored and
CarlosGomes98
committed
f21e4bc
Fix inconsistent line endings

timmoon10
authored and
CarlosGomes98
committed
19b6b2f