[Common/PyTorch] Grouped weighted-SwiGLU MXFP8 kernel - #3315

Open

cael-ling wants to merge 5 commits into

NVIDIA:mainfrom

cael-ling:feature/mxfp8-group-swiglu-recompute

Commits on Aug 4, 2026

[Common/PyTorch] Grouped weighted-SwiGLU MXFP8 kernel
cael-ling
committed
5ae2ae8
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
2979881

Commits on Aug 5, 2026

[PyTorch] Require contiguous, colocated operands in group_swiglu_quantize
cael-ling
committed
e092890
[PyTorch] Require grouped metadata tensors on the current CUDA device
cael-ling
committed
5bfd791
[PyTorch] Scope the grouped metadata device check to group_swiglu_quantize
cael-ling
committed
8572e3a