Add a backwards linear function to be used with the fused mla q up-proj

[chaseblock]

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

[greptile-apps]

Greptile Summary

The PR adds a reusable linear-backward helper for the fused MLA Q up-projection and exposes dgrad, wgrad, and bias-gradient results.

• Adds an attention-layer delegate into the linear module.
• Builds the backward arguments for BF16 and MXFP8 projection paths.
• The BF16 bias-gradient path remains incomplete because the fused bgrad GEMM is invoked without a bias tensor.

Confidence Score: 4/5

The PR is not yet safe to merge because BF16 fused MLA training still loses the projection bias gradient.

The non-FP8 path delegates bias-gradient computation to the weight-gradient GEMM, but the new wrapper supplies no bias tensor, so the binding does not allocate or return db.

Files Needing Attention: transformer_engine/pytorch/module/linear.py

Important Files Changed

Filename	Overview
transformer_engine/pytorch/module/linear.py	Adds the shared backward helper, but its BF16 use_bias path returns no bias gradient because bias is always configured as None.
transformer_engine/pytorch/attention/fused_mla_q_uproj.py	Adds a thin classmethod that forwards fused MLA backward arguments and return values to the new linear helper.

_{Reviews (3): Last reviewed commit: "Handle biad gradient in lin bwd wrapper." | Re-trigger Greptile}

[greptile-apps]

BF16 bias gradient remains missing

When a BF16 fused MLA backward call sets use_bias=True, this wrapper still passes bias=None; the non-FP8 path delegates bgrad to the weight-gradient GEMM, which does not allocate db without a bias tensor, causing backward_linear to return grad_bias=None and preventing the projection bias from being updated.

Knowledge Base Used: PyTorch Fused Modules (transformer_engine/pytorch/module)