[GG] Revert unconditional MLA BMM contiguity workaround

[voipmonitor]

Reverts #136 as a whole before replacing it with a producer-side DCP buffer fix.

Why

The backend contiguity contract introduced per-layer input and output copies for every B12X MLA invocation, including DCP1. Release validation measured A16 DCP1/MTP0 decode falling from 87.14 to 82.52 tok/s (~5.3%), while 64K prefill remained near parity.

The Xid31 report itself is valid: a guarded VMM reproducer proves that the selected cuBLAS strided BMM reads to base +64 KiB for the affected shape. The safe and fast fix belongs at DCP output allocation: provide mapped tail padding from B12X A2A and NCCL fallback producers instead of copying every MLA BMM operand.

This PR intentionally contains only the exact revert. A separate clean PR with tail-padded DCP outputs and unit/E2E coverage follows from this reverted base.

Summary by CodeRabbit

• Performance

  • Simplified MLA attention processing by removing unnecessary tensor contiguity conversions.
  • Streamlined weight loading and projection operations while preserving existing runtime behavior.

• Tests

  • Removed tests covering the retired contiguity-handling behavior.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: b16d92d1-c2a5-450e-9f7f-c59a670bb6a1

📥 Commits

Reviewing files that changed from the base of the PR and between 84920d4 and a7413ee.

📒 Files selected for processing (2)

• tests/v1/attention/test_mla_backends.py
• vllm/model_executor/layers/attention/mla_attention.py

---

📝 Walkthrough

Walkthrough

MLA attention removes backend BMM contiguity flag handling across initialization, decoding, weight processing, and V up-projection. Corresponding contiguity tests are deleted, while runtime weight-address preservation assertions remain.

Changes

MLA BMM contiguity behavior

Layer / File(s)	Summary
Remove runtime contiguity paths vllm/model_executor/layers/attention/mla_attention.py	MLA attention no longer caches BMM contiguity flags, forces MQA inputs or transformed weights contiguous, or allocates a temporary V up-projection output.
Update contiguity coverage tests/v1/attention/test_mla_backends.py	Backend contiguity propagation and contiguous BMM weight/output tests are removed; runtime weight-address preservation checks remain.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• local-inference-lab/vllm#136: Adds or validates the MLA BMM contiguity behavior and tests removed here.

Suggested reviewers: lukealonso, matthewbonanni, aoshen02

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch revert/gg-pr136-mla-bmm-20260720

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}