[PyTorch] Minor optimizations in fused grouped MLP

[ksivaman]

Description

Small perf/cpu overhead improvements in fused grouped MLP

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
• Performance

Changes

• Reduce number of casts via torch .to wherever possible.
• Use fused kernel to calculate offsets from splits.
• Calculate split points by indexing into offsets to remove an additional cumsum.

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 L0

[greptile-apps]

Greptile Summary

This PR introduces minor CPU-overhead and performance optimizations in the fused grouped MLP forward and backward paths. The key changes are: (1) replacing two torch.cumsum + torch.cat calls with a single tex.splits_to_offsets fused GPU kernel, (2) eliminating redundant no-op .to() casts for split_sizes/split_points in the backward pass since the forward already saves them in the correct dtype, (3) hoisting scales.detach().to(torch.float32) to avoid a duplicate cast in the scale_bias branch, and (4) pulling bias-grad .to(dtype) casts outside if/else branches to apply once on the full tensor.

Confidence Score: 5/5

Safe to merge — all changes are semantically equivalent refactors with no correctness impact.

All optimizations are either no-ops being removed (backward casts on tensors already in the right dtype), or logically equivalent transformations (fused kernel vs. cumsum+cat, hoisted casts). The C++ implementation of splits_to_offsets was verified to produce int64 output matching the original base_offsets computation. No new code paths, no behavior changes for any dtype or shape.

No files require special attention.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/ops/fused/forward_grouped_mlp.py	Replaces two cumsum + cat ops with a single fused splits_to_offsets kernel and derives split_points by indexing into the result; semantically equivalent and correct.
transformer_engine/pytorch/ops/fused/backward_grouped_mlp.py	Removes redundant backward casts (already correct dtype from forward), hoists float32 scale conversion, and consolidates dtype casts for bias grads — all semantically equivalent refactors.

_{Reviews (1): Last reviewed commit: "Merge branch 'main' into minor_opts_fuse..." | Re-trigger Greptile}

[timmoon10]

LGTM