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 autograd support for the fused MLA query up-projection by applying the inverse RoPE transformation and delegating projection gradients to TE’s linear backward implementation.

• Adds a Triton kernel for the RoPE backward transformation.
• Adds fused projection input and weight gradient computation.
• Adds numerical autograd coverage for input and weight gradients.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/attention/fused_mla_q_uproj.py	Adds the fused MLA autograd function, inverse-RoPE Triton kernel, and delegation to the established linear backward implementation; no blocking issue remains.
tests/pytorch/attention/test_fused_mla_q_uproj.py	Adds numerical coverage for input and weight gradients through the new autograd path.

Sequence Diagram

sequenceDiagram
  participant A as PyTorch autograd
  participant R as RoPE backward kernel
  participant L as TE linear backward
  A->>R: Query output gradient
  R-->>A: Pre-RoPE projection gradient
  A->>L: Gradient, saved input, weight
  L-->>A: Input and weight gradients

Loading

_{Reviews (5): Last reviewed commit: "Add fused MLA Q up-projection backward" | 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)

[sudhakarsingh27]

Do we really need this function as we could directly call _linear_backward? If we absolutely need it, could we move it to fused_mla_q_uproj.py instead then?