[PyTorch] Enable fused FP8 block-scaling path in GroupedLinear module - #3171

Merged

denera merged 13 commits into

NVIDIA:mainfrom

denera:pytorch/grouped-linear-fused-fp8bs

Jul 22, 2026

Commits on Jul 22, 2026

[Common/PyTorch] Support power-of-2 scales in grouped FP8 block-scaling quantize
denera
committed
3c81a59
[PyTorch] Enable fused grouped FP8 block-scaling path in GroupedLinear module
denera
committed
a3732e8
[PyTorch] Enable FP8 block-scaling in GroupedLinear fusible op
denera
committed
013e42f
[PyTorch] Use persistent workspaces in grouped-tensor GEMM
denera
committed
f149995
[PyTorch] Fix grouped FP8 block-scaling CUDA-graph deadlock via per-role cuBLAS workspaces
denera
committed
d49da23
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
denera
committed
cd8d42b
[PyTorch] Address review: document split-accumulator override, fix stale dbias comment
denera
committed
42e6db7
[PyTorch] Revert fusible-ops FP8 block-scaling; scope PR to GroupedLinear module
denera
committed
cfd8922
[PyTorch] Isolate grouped wgrad cuBLAS workspace by NT layout, not out-discreteness
denera
committed
4798ce3
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
denera
committed
85baa24
[PyTorch] Address review: isolate grouped cuBLAS workspace per layout; drop redundant test spy
denera
committed
98d05bc
updated grouped GEMM workspace comment on stale TMA descriptor related deadlocks
denera
committed
3dc9a8f
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
denera
committed
1393778