[Pyt][Common] Enabling/Guarding sm120 support (non - attention)#2833

Draft

KshitijLakhani wants to merge 13 commits intoNVIDIA:mainfrom

KshitijLakhani:klakhani/maint/sm120-pyt-cpp-support

Commits on Apr 3, 2026

Do not use fp8::cast_gated_tma for sm120. Instead use the fall back fp8::cast_gated_bwd kernel as sm120 shmem requirements are insufficient
KshitijLakhani
committed

Commits on Apr 9, 2026

Disable SR and fused RHT+case path for sm120
KshitijLakhani
committed

Commits on Apr 10, 2026

Disable SR for sm120
KshitijLakhani
committed
Fallback to unfused quantize, cast RHT instead of the fused op for sm120
KshitijLakhani
committed
Guard cublaslt grouped gemm for sm120 as it does not seem to be supported
KshitijLakhani
committed
Fix: Add a sync after shmem bulk op ro ensure no corruption
KshitijLakhani
committed
Relax test numeric tolerance slightly for sm120 as the backend used is Flash and not Fused
KshitijLakhani
committed
Relax the tolerance slightly for NVFP4 on sm120 as path chose might be different
KshitijLakhani
committed
Use SM120-specific 16-aligned grouped-linear shapes to satisfy FP8 GEMM lda constraints
KshitijLakhani
committed
Add SM120 minor column-parallel tolerance adjustment for distributed debug test activation comparisons
KshitijLakhani
committed
Add SM120 skip guards for grouped GEMM C++ operator tests
KshitijLakhani
committed
Disable cublas lt grouped gemm related PyT tests for sm120
KshitijLakhani
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed