Add NVFP4 RHT Support for SM120 and SM121 - #3300

Open

new-TonyWang wants to merge 12 commits into

NVIDIA:mainfrom

new-TonyWang:fix/sm121-rht-reference-fallback

Commits on Aug 3, 2026

fix(nvfp4): match RHT reference on SM120 and SM121
new-TonyWang
committed
f40be70
fix(nvfp4): recompute post-RHT amax from fallback output
new-TonyWang
committed
caa2476
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
new-TonyWang
committed
1381291
fix(nvfp4): reduce ATen RHT amax across ranks
new-TonyWang
committed
9142e86
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
70a5da7
test(nvfp4): cover SM12x RHT amax semantics
new-TonyWang
committed
ea87075
perf(nvfp4): reuse SM12x ATen RHT output
new-TonyWang
committed
7a0a0d0
feat(nvfp4): fuse SM12x RHT quantization with TMA
new-TonyWang
committed
434f242
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
fea70f2

Commits on Aug 4, 2026

perf(nvfp4): use warp MMA for SM12x fused RHT
new-TonyWang
committed
f9e28f4
fix(nvfp4): align SM12x fused RHT with native K16 MMA
new-TonyWang
committed
36e1179
perf(nvfp4): quantize SM12x RHT fragments in registers
new-TonyWang
committed
9923d1c