dcp: make head-sliced attention views contiguous for the B12X PCIe pool

[voipmonitor]

Problem

TP6 + DCP>1 with the B12X A2A fast path dies at CUDA graph capture:

ValueError: partial_lse must be contiguous
  pool.lse_reduce_scatter(cp_attn_out, cp_attn_lse, ...)   # pcie_dcp_a2a._validate

Root cause: GLM TP6 virtual-TP pads attention heads 64 → 66, and the B12X sparse MLA kernel pads its head dim further, returning the decode output/LSE as head-sliced views (b12x_mla_sparse.py: lse = lse[:, :input_num_heads]) — non-contiguous whenever kernel head count > input head count. TP8 shapes never hit this (counts match), which is why all DCP2/DCP4 runs to date were fine. The NCCL packers take explicit stride arguments and the query all-gather wrapper already calls .contiguous(); only the PCIe pool's lse-reduce entry lacked the fix-up.

Fix

Copy the operands to contiguous in _try_b12x_dcp_lse_reduce, placed after the cheap reject checks so NCCL fallback paths pay nothing. The LSE copy is a tiny [B, H] fp32; the output copy only happens on padded head counts. Graph-capture safe (same in-graph allocation pattern as the packed A2A send/recv buffers).

Matches the minimal patch Martin drafted (blackwell-llm-docker/patches/vllm-dcp-b12x-contiguous-lse-20260707.patch) — this PR adds the placement inside the try-helper plus a unit test that simulates the sliced-view pattern and asserts the pool receives contiguous tensors (31/31 in test_dcp_a2a.py on the v6 image, distributed cases deselected).

Verification

Diagnosed on TP6/MXFP4/A8/DCP2, v6 image (vllm49bed029-b12x26144c0): previously died at capture; E2E with this file mounted boots and serves (numbers in comment below).

Stacked on fable/w4a8mx-qmma-pad-20260707 (#80) — the v6 image pin.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Improved distributed attention processing for head-sliced tensors.
  • Ensured partial attention outputs and LSE data are contiguous before reduction, preventing failures with non-contiguous inputs.
• Tests
  • Added CUDA coverage validating correct handling of non-contiguous attention tensor views.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: d6e8baca-16ad-497c-82c6-130d613eae0d

📥 Commits

Reviewing files that changed from the base of the PR and between dc47f72 and e95da31.

📒 Files selected for processing (2)

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

---

📝 Walkthrough

Walkthrough

The B12X PCIe DCP LSE reduction path now makes non-contiguous attention output and LSE views contiguous before dispatch. A CUDA-only test constructs non-contiguous slices and verifies the pool receives contiguous tensors.

Changes

B12X LSE contiguity handling

Layer / File(s)	Summary
Contiguous reduction dispatch and validation vllm/v1/attention/ops/dcp_alltoall.py, tests/distributed/test_dcp_a2a.py	_try_b12x_dcp_lse_reduce copies non-contiguous output and LSE tensors before lse_reduce_scatter; the CUDA test verifies this behavior with padded tensor views.

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

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: making head-sliced attention views contiguous for the B12X PCIe pool.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ 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 fable/dcp-b12x-contiguous-lse-20260707

---

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.}

[voipmonitor]

E2E verified on the exact failing config — v6 image (vllm49bed029-b12x26144c0), GLM-5.2 AMD MXFP4 experts, TP6 / force-A8 / online MXFP8 dense / DCP2, GMU 0.950, with only this PR's dcp_alltoall.py mounted:

• previously: died at CUDA graph capture with ValueError: partial_lse must be contiguous
• now: boots through capture, KV cache 639,616 tokens, coding smoke c0 = 67.1 tok/s and c3000 = 66.9 tok/s, both 0 CJK (i.e. the copied views also produce correct DCP-reduced outputs, not just a passing boot)

Martin's build patch blackwell-llm-docker/patches/vllm-dcp-b12x-contiguous-lse-20260707.patch is content-equivalent to this PR (same guarded copies, same placement) — safe to unblock the image build with it; this PR adds the explanatory comment and the unit test.

[voipmonitor]

Rebased to a single commit on current dev/eldritch-enlightenment (post #78/#79/#83 merges — those landed while this one was skipped, which had pushed it into conflict). mergeable=true again; suite run on the v6 image with this file mounted: 31 passed + the same 1 pre-existing environment-bound failure (TestDCPCommBackendConfig::test_a2a_with_dcp_valid) that pure dev shows in the same container.

Still needed: dev has no equivalent fix — the sparse-MLA head-sliced LSE views reach the PCIe pool non-contiguous on TP6×DCP≥2 and capture dies with partial_lse must be contiguous. Production images only survive because the content-equivalent build patch (blackwell-llm-docker/patches/vllm-dcp-b12x-contiguous-lse-20260707.patch) rides along; merging this retires that patch.