dcp: make head-sliced attention views contiguous for the B12X PCIe pool (fathomless-firmament)

[voipmonitor]

Same single commit as #81, retargeted at dev/fathomless-firmament (which already carries the #78/#79 re-picks but not this fix — _try_b12x_dcp_lse_reduce still passes head-sliced views straight to the pool).

Problem

Sparse MLA backends can return head-sliced views: GLM TP6 pads 64 → 66 attention heads and slices the kernel output back (lse = lse[:, :input_num_heads]), producing non-contiguous cp_attn_out/cp_attn_lse. The B12X PCIe DCP pool validates contiguity, so TP6 × DCP≥2 dies at CUDA-graph capture with ValueError: partial_lse must be contiguous. TP8 never hits it (kernel head count == input head count); the NCCL packers take explicit strides and are unaffected — only this fast path needs the copy.

Fix

Guarded .contiguous() on both operands in _try_b12x_dcp_lse_reduce, after the cheap reject checks and before the pool call. LSE is a tiny [B, H] fp32 tensor and the output copy only happens on padded head counts; capture-time allocations go to the graph pool like the existing send/recv buffers.

Verification

• Unit test simulates the head66 slicing pattern (padded buffers sliced to [:, :16], asserts the pool receives contiguous tensors). Suite on the v6 image with this file mounted: 31 passed + 1 pre-existing environment-bound failure (TestDCPCommBackendConfig::test_a2a_with_dcp_valid, identical on the unpatched branch in the same container).
• E2E on the eldritch lineage (v6 image, GLM-5.2 AMD MXFP4, TP6/A8/online-MXFP8/DCP2, GMU 0.95): previously died at capture; with the fix boots (KV 639,616), c0 67.1 / c3000 66.9 tok/s, 0 CJK.
• Content-equivalent to the build patch blackwell-llm-docker/patches/vllm-dcp-b12x-contiguous-lse-20260707.patch — merging retires it.

🤖 Generated with Claude Code

[coderabbitai]

Warning

Review limit reached

@voipmonitor, you've reached your PR review limit, so we couldn't start this review.

Next review available in: 29 minutes

Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available.
You're only billed for reviews past your plan's rate limits ($0.25/file).

How can I continue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews.

How do review limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1ecbb4be-4efa-4e61-be0d-1a765a342a80

📥 Commits

Reviewing files that changed from the base of the PR and between 407ace1 and 210a68a.

📒 Files selected for processing (2)

• tests/distributed/test_dcp_a2a.py
• vllm/v1/attention/ops/dcp_alltoall.py

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fable/dcp-b12x-contiguous-lse-ff-20260709

---

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.}