[PyTorch] Add grouped linear op and experimental fusion for grouped MLP

[timmoon10]

Description

This PR adds a grouped linear op, which can be used in the grouped MLP block in Mixture-of-Experts models. It also adds an experimental fused operation for a grouped MLP block, using a CuTe DSL kernel that computes an MXFP8 grouped GEMM and SwiGLU.

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

• Add a grouped linear operation
• Add a post-scaled SwiGLU op and add support for interleaving SwiGLU gate and linear units
• Add a fused operation for grouped MLP

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]

_{9 files reviewed, 4 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Incorrect grad-required flags

In ForwardGroupedMLP_CuTeGEMMSwiGLU_MXFP8.fuser_forward, swiglu_ctx.input_requires_grad and swiglu_ctx.extra_input_requires_grad are set to True unconditionally (and input_requires_grad is set to requires_grad unconditionally). This will make ScaledSwiGLU.fuser_backward compute grad_input and grad_extra_input even when neither input_ nor scales require grads, which violates autograd semantics and can raise (e.g., scales.detach() passed into the fused kernel, but extra_input_requires_grad=True forces a gradient).

This should be set based on the actual requirements:

• input_requires_grad = input_.requires_grad
• swiglu_ctx.extra_input_requires_grad = scales.requires_grad
• and for FC weights, check each parameter’s requires_grad (not just weight0).

[greptile-apps]

Broken fusion window scan

Both fuse_backward_ops and fuse_forward_ops have a window/shift loop that can drop or reorder ops when the pattern doesn’t match. In the non-matching branch you do out.extend(window[:-2]); window = window[-2:] and then immediately do out.extend(window[:-3]) (which is a no-op for a 2-element window) before refilling. This causes the scan to advance by 1 op in some cases and by 2 in others, and it never emits window[-1] until the very end. For sequences like [A,B,C,D] where [A,B,C] doesn’t match but [B,C,D] would (or vice versa), this loop will not correctly consider all 3-op windows and can produce an incorrect fused op list.

This needs a standard sliding-window approach (advance by 1 when not matching; replace 3->1 when matching) to ensure no ops are skipped or duplicated.

[greptile-apps]

Incorrect module docstring

transformer_engine/pytorch/ops/basic/grouped_linear.py starts with """Fusible operation for bias.""", which is a copy/paste from basic/bias.py and is incorrect for this file. This impacts generated docs and module-level help text and is user-visible.

Update it to describe GroupedLinear (e.g., “Fusible operation for grouped linear / grouped GEMM”).

[greptile-apps]

Always-saving scales tensor

In ScaledSwiGLU.fuser_forward, you set ctx.extra_input_requires_grad = extra_input.requires_grad, but you always save scales because the conditional is scales if ctx.input_requires_grad else None and ctx.input_requires_grad is forced to True. When scales.requires_grad=False, this needlessly keeps scales alive for backward and increases activation memory; worse, fuser_backward uses scales inside the ctx.input_requires_grad branch, so if someone later changes input_requires_grad to be accurate, the save condition would become wrong.

Save scales based on ctx.extra_input_requires_grad (or save both unconditionally but keep the grad-required flags consistent).