[Pyt][Common] Enabling/Guarding sm120 support (non - attention) by KshitijLakhani · Pull Request #2833 · NVIDIA/TransformerEngine

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

Disable SR and fused RHT+case path for sm120

KshitijLakhani

committed

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

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

Align grouped fallback layout metadata on SM120

KshitijLakhani

committed

Make grouped scale checks metadata-driven and relax SM120 tolerance

KshitijLakhani

committed

Handle SM120 NVFP4 SR equivalence in stochastic-rounding checks

KshitijLakhani

committed

Fix: Re instate the sm 120 conditional for stats stride and output_s shape that was lost in an earlier PR's merge conflict

KshitijLakhani

committed

Relax tolerance for FP8 CS for sm120 in dist run_layer_with_overlap test

KshitijLakhani

committed

[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]

authored and

KshitijLakhani

committed

For sm120 change tolerance when determinism results in a non fused attn backend

KshitijLakhani

committed

Disable FAv4 on sm120 temporarily due to multiple failure cases

KshitijLakhani

committed

[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]

authored and

KshitijLakhani

committed

Use local quantizer copy intead of modifying the global quantizer state

KshitijLakhani

committed

Code clean via reusability

KshitijLakhani

committed

Clean up test code

KshitijLakhani

committed

[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]

authored and

KshitijLakhani

committed

Feature and test code clean uo

KshitijLakhani

committed

Remove incorrectly pushed files

KshitijLakhani

committed

[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]

authored and

KshitijLakhani

committed

Fix: lint issue

KshitijLakhani

committed

Honor the optimize_for_gemm and instead swizzle the scales after the unfusd cast for sm120 when doing the group quantize for nvfp4

KshitijLakhani

committed

Do not modify the quantizer state

KshitijLakhani

committed

[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]

authored and

KshitijLakhani

committed

Go back to using stricter constraints for sm120 as well

KshitijLakhani

committed

Skip on SM120 whenever the cuBLASLt grouped GEMM path would be selected in fusible-ops tests

KshitijLakhani

committed

Relax sanity atol for SM120 + NVFP4 quantized compute in test_layernorm_mlp

KshitijLakhani

committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

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

[Pyt][Common] Enabling/Guarding sm120 support (non - attention)#2833
KshitijLakhani wants to merge 33 commits into
NVIDIA:mainfrom
KshitijLakhani:klakhani/maint/sm120-pyt-cpp-support

Commits on May 29, 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

Disable SR and fused RHT+case path for sm120

Disable SR for sm120

Fallback to unfused quantize, cast RHT instead of the fused op for sm120

Guard cublaslt grouped gemm for sm120 as it does not seem to be supported

Fix: Add a sync after shmem bulk op ro ensure no corruption

Relax test numeric tolerance slightly for sm120 as the backend used is Flash and not Fused

Use SM120-specific 16-aligned grouped-linear shapes to satisfy FP8 GEMM lda constraints

Add SM120 minor column-parallel tolerance adjustment for distributed debug test activation comparisons

Add SM120 skip guards for grouped GEMM C++ operator tests

Align grouped fallback layout metadata on SM120

Make grouped scale checks metadata-driven and relax SM120 tolerance

Handle SM120 NVFP4 SR equivalence in stochastic-rounding checks

Fix: Re instate the sm 120 conditional for stats stride and output_s shape that was lost in an earlier PR's merge conflict

Relax tolerance for FP8 CS for sm120 in dist run_layer_with_overlap test

[pre-commit.ci] auto fixes from pre-commit.com hooks

For sm120 change tolerance when determinism results in a non fused attn backend

Disable FAv4 on sm120 temporarily due to multiple failure cases

[pre-commit.ci] auto fixes from pre-commit.com hooks

Use local quantizer copy intead of modifying the global quantizer state

Code clean via reusability

Clean up test code

[pre-commit.ci] auto fixes from pre-commit.com hooks

Feature and test code clean uo

Remove incorrectly pushed files

[pre-commit.ci] auto fixes from pre-commit.com hooks

Fix: lint issue

Honor the optimize_for_gemm and instead swizzle the scales after the unfusd cast for sm120 when doing the group quantize for nvfp4

Do not modify the quantizer state

[pre-commit.ci] auto fixes from pre-commit.com hooks

Go back to using stricter constraints for sm120 as well

Skip on SM120 whenever the cuBLASLt grouped GEMM path would be selected in fusible-ops tests

Relax sanity atol for SM120 + NVFP4 quantized compute in test_layernorm_mlp