Skip to content

Disable the RHT fusion for non-SM100 family devices #2968

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
timmoon10 merged 4 commits into from
May 12, 2026
Merged

Disable the RHT fusion for non-SM100 family devices #2968

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ec73fbe
Disable the RHT fusion for non-SM100 family devices
ptrendx May 8, 2026
d1848a8
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] May 8, 2026
1425a8d
Fix the compilation error
ptrendx May 11, 2026
44ca5ea
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] May 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion transformer_engine/pytorch/csrc/quantizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@
#include <pybind.h>

#include "common.h"
#include "common/util/cuda_runtime.h"
#include "common/util/system.h"
#include "pybind.h"
#include "torch/torch.h"
Expand Down Expand Up @@ -2243,7 +2244,8 @@ void NVFP4Quantizer::quantize_impl(const TensorWrapper& input, TensorWrapper& ou

// Restriction for the RHT cast fusion kernel because we are using MMA hardware for computing RHT
bool eligible_for_rht_cast_fusion =
input.dtype() == DType::kBFloat16 && rows % 64 == 0 && cols % 128 == 0;
input.dtype() == DType::kBFloat16 && rows % 64 == 0 && cols % 128 == 0 &&
transformer_engine::cuda::sm_arch() >= 100 && transformer_engine::cuda::sm_arch() <= 110;
Copy link
Copy Markdown
Contributor

@greptile-apps greptile-apps Bot May 8, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 The upper bound <= 110 is tighter than the stated intent ("non-SM100 family"). Using < 120 more precisely captures "anything below SM120" and avoids silently disabling the fusion for hypothetical SM111/SM112 variants that belong to the same Blackwell compute family. The codebase already uses 120 as the implicit dividing line (SM120 = GB20x, which is the architecture that triggered the bug), so < 120 reads as clearly intentional.

Suggested change
transformer_engine::cuda::sm_arch() >= 100 && transformer_engine::cuda::sm_arch() <= 110;
transformer_engine::cuda::sm_arch() >= 100 && transformer_engine::cuda::sm_arch() < 120;


// Stochastic rounding
// When both rowwise and columnwise quantization are used with RHT,
Expand Down
Loading