[Common] Support scaled & clamped swiglu, srelu for BF16

[zhongbozhu]

Description

Support Mega-C++ with Cublas BF16 Grouped GEMM backend: #3099

Kernel Perf:

Fixes # (issue)

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

This PR adds new CUDA kernels (scaled_swiglu, scaled_clamped_swiglu, scaled_srelu) that fuse the activation with a per-row scale multiplication and optionally compute the scale gradient in a single pass, supporting BF16/FP16/FP32 data and scale types, including an interleaved GLU layout for compatibility with grouped-GEMM backends.

• Kernel design: Four templated host launchers (launch_scaled_gated_forward/backward, launch_scaled_unary_forward/backward) instantiated for (Empty, silu), (ClampedSwiGLUParam, clamped_silu), and (Empty, srelu). The backward kernels use a one-block-per-row scheme with a shared-memory warp reduction to accumulate grad_act_scales when requested; ComputeScaleGrad is a compile-time template flag that gates the reduction path.
• Vectorization: row_vector_alignment checks pointer alignment and falls back to scalar (nvec=1) when the segment size or pointers are not aligned; the glu_interleave_size % 32 validation ensures vector paths are always reachable for standard interleave values.
• Test coverage: A new test_scaled_activation.cu exercises forward + backward for all three activations across BF16/FP32/FP16 data types, contiguous and interleaved GLU layouts, and optional scale-gradient computation.

Confidence Score: 5/5

The change is purely additive — new source files, a new public header section, and new tests — with no modifications to existing kernels or data paths. The math is correct and the forward/backward formulas are consistent with each other and with the test references.

All new launch code is guarded by early-return dimension checks and type-dispatch macros; the reduction kernel uses correct smem sizing bounded by kReductionThreads. The one finding (ClampedSwiGLU reference hardcoding alpha) is a test-coverage concern rather than a runtime bug.

test_scaled_activation.cu: the ClampedSwiGLU backward reference should explicitly thread kClampedAlpha through qgelu/dqgelu calls rather than relying on the 1.702f default.

Important Files Changed

Filename	Overview
transformer_engine/common/activation/scaled_activation.cu	Core kernel file implementing gated and unary forward/backward kernels for scaled activations; math is correct and block-reduce smem sizing is consistent with kReductionThreads.
transformer_engine/common/activation/scaled_activation.h	Header declaring launch templates and shared utilities; alignment fallback handles non-vectorizable cases cleanly.
transformer_engine/common/activation/scaled_swiglu.cu	Thin wrapper dispatching to launch_scaled_gated_forward/backward for SwiGLU and ClampedSwiGLU; all API functions correctly wired.
transformer_engine/common/activation/scaled_srelu.cu	Thin wrapper dispatching to launch_scaled_unary_forward/backward for SReLU; correct and minimal.
transformer_engine/common/include/transformer_engine/activation.h	Six new C API declarations added inside extern C; documentation is thorough and parameter semantics match the implementation.
tests/cpp/operator/test_scaled_activation.cu	New test suite covering all three activations across BF16/FP32/FP16 types and interleaved layouts. ClampedSwiGLU reference hardcodes alpha=1.702f via test::qgelu rather than using kClampedAlpha explicitly.
transformer_engine/common/CMakeLists.txt	New source files added to both standard and fast-math build lists symmetrically.
tests/cpp/operator/CMakeLists.txt	test_scaled_activation.cu added to the test executable correctly.

_{Reviews (9): Last reviewed commit: "[pre-commit.ci] auto fixes from pre-comm..." | Re-trigger Greptile}

[zhongbozhu]

/te-ci pytorch

[vthumbe1503]

LGTM. Left minor comments

[ptrendx]

A general comment - I think that limiting this to only FP32/FP16 is not needed and you could very
easily have tensor-scaled FP8 support here.

[zhongbozhu]

I agree but it's gonna be challenging to extend to blockwise scaling (when it comes to GLU interleaved gated activation, and non-gated activation should be fine), and I do think that for fp8 fp4, we do expect a fused gemm to be built so it's directly fused into the GEMM epilogue.

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.

[vthumbe1503]

/te-ci

[vthumbe1503]

/te-ci