GGEMM+srelu kernels for MxFP8 Nemotron

[sraman-rgb]

Description

Please include a brief summary of the changes, relevant motivation and context.

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

[ksivaman]

/te-ci pytorch

[ksivaman]

Please sign-off your commits @sraman-rgb

[greptile-apps]

Greptile Summary

This PR refactors the MXFP8 fused grouped-MLP kernel infrastructure to support both GLU-style activations (SwiGLU, QGeGLU) and unary activations (SReLU), then adds the new ScaledSReLU op and corresponding ForwardGroupedMLP_CuTeGEMMUnary_MXFP8 / BackwardGroupedMLP_CuTeGEMMDUnary_MXFP8 fused-kernel classes backed by cuDNN FE ≥ 1.24.0.

• The concrete ForwardGroupedMLP_CuTeGEMMSwiGLU_MXFP8 class is renamed to ForwardGroupedMLP_CuTeGEMMGLU_MXFP8 and extracted into a base class that both the GLU and new Unary subclasses share; the same pattern applies to the backward.
• ScaledSReLU is a new BasicOperation for squared-ReLU with per-row post-scaling supporting an activation_recompute_in_mlp knob.
• fuse_grouped_mlp_ops is generalised to accept an activation_op_types tuple, and validate_grouped_mlp_dims is extended to handle both GLU (2× out-features) and unary (1× out-features) cases.

Confidence Score: 4/5

The new SReLU fused kernel path is logically sound but carries unresolved concerns from prior review rounds in the fused forward and backward files.

The base-class refactoring is clean and ScaledSReLU is correctly wired through both unfused and fused paths. The activation_recompute_in_mlp mechanism is self-consistent. The unreachable grad_scales guard is the main new finding — it cannot cause wrong results today but would silently discard the scale gradient if the kernel interface changed. Three prior-thread concerns remain open.

transformer_engine/pytorch/ops/fused/backward_grouped_mlp.py and forward_grouped_mlp.py deserve a closer look before merging.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/ops/fused/forward_grouped_mlp.py	Introduces base class + GLU/Unary subclasses for fused forward; adds SReLU recompute path and fuse_forward_srelu_ops; has a dead _grouped_gemm_dsrelu_backward_supported helper (flagged in prior review).
transformer_engine/pytorch/ops/fused/backward_grouped_mlp.py	Mirrors forward refactoring for backward; adds BackwardGroupedMLP_CuTeGEMMDUnary_MXFP8; fuser_backward reconstructs grouped_fc2_x from the DSReLU kernel when recompute is enabled; contains an unreachable if grad_scales is not None guard.
transformer_engine/pytorch/ops/basic/activation.py	Adds ScaledSReLU op with correct fuser_forward/fuser_backward overrides, activation-recompute guard, and proper ctx.save_for_backward layout consistent with the fused path.
transformer_engine/pytorch/ops/_common.py	Generalises validate_grouped_mlp_dims, fuse_grouped_mlp_ops, and version-check helpers; extracts is_glu_activation.
transformer_engine/pytorch/ops/basic/swiglu.py	Adds activation_recompute_in_mlp parameter to _ScaledGLU and ScaledClampedQGeGLU; raises informatively in non-fused paths when the flag is set.
tests/pytorch/test_fusible_ops.py	Adds test_scaled_srelu, test_scaled_activation_recompute_in_mlp_config, and extends test_grouped_mlp for scaled_srelu; correctly skips when the unary fusion is unavailable.
transformer_engine/pytorch/ops/fused/init.py	Re-exports renamed GLU and new Unary fused-op classes; straightforward change.
transformer_engine/pytorch/ops/basic/init.py	Exports the new ScaledSReLU symbol; one-line change, no issues.

_{Reviews (12): Last reviewed commit: "Merge branch 'main' into fc1-srelu-main" | Re-trigger Greptile}

[timmoon10]

Overall looks good, but we've gotten to the point where we need to start thinking about how to gracefully handle adding new activations. It seems that every model has a different activation function.

[greptile-apps]

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

[vthumbe1503]

LGTM. We might want to wait on the cudnn release and apt cudnn guards are added.

[vthumbe1503]

/te-ci pytorch

[vthumbe1503]

LGTM.