fix(ds4): enforce sparse top-k row-length contract

[voipmonitor]

Resulting behavior

DeepSeek V4 sparse-indexer row lengths now obey one contract across metadata production and top-k consumption:

• graph-padding rows produce non-negative per-token context lengths;
• persistent top-k reads at most min(logits.stride(0), max_seq_len) scores from each row;
• cooperative top-k treats a non-positive row length as an empty row;
• a persistent CTA group advances its radix histogram ring only when the assigned row uses the radix path.

Valid row lengths and selected top-k sets are unchanged.

Technical reason

Uniform MTP metadata computed seq_len - max_decode_len + token_offset + 1. A CUDA-graph padding request with seq_len == 0 therefore produced -1 for its first token. The top-k kernels consume row lengths in unsigned indexing decisions, so the negative value could select scores from a non-empty row instead of returning -1 padding.

Persistent CTA groups also indexed their triple-buffered histogram by every assigned row. A short row does not execute radix selection, so counting it advanced the ring without initializing that generation. A later long row assigned to the same group could consume stale histogram state.

The implementation adapts the applicable contracts from upstream vLLM #49139 and #51538 to the Infernal Invocation metadata and B12X integration.

Compatibility

• Scope: DeepSeek V4 sparse MLA metadata and the shared persistent/cooperative top-k kernels.
• CUDA graph compatibility: preserved; buffers and launch shapes are unchanged.
• Model arithmetic: unchanged for rows whose declared lengths are within bounds.
• Performance: the hot path adds integer bounds operations only; no allocation, copy, or synchronization is introduced.

Validation

Status: implemented. Compiled C++ kernel qualification is part of the Infernal Invocation release image build because an existing image contains the previous _C extension.

• ruff check and ruff format: pass.
• CPU metadata tests: 2 passed.
• SM120 Triton producer reproducer on RTX PRO 6000 Blackwell:
  • packaged r16 source: [-1, 0, 4, 5], expected-contract assertion failed;
  • patched source over the same image: [0, 0, 4, 5], passed.
• Added compiled-kernel regression coverage for:
  • persistent histogram reuse after a short row;
  • negative graph-padding lengths;
  • row lengths beyond both the physical stride and active sequence bound.

The compiled-kernel tests must pass against the release-built extension before image publication.

[coderabbitai]

Warning

Review limit reached

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

Next review available in: 50 minutes

Limit details: You’ve used the included review currently available.

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

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 within each organization.

For paid Pro and Pro+ reviews, CodeRabbit uses a developer's included PR review attempts over the past 7 days to set the current hourly allowance. At typical activity levels, the full plan allowance applies. Higher sustained activity can lower the allowance until earlier attempts leave the 7-day window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: b11cfa15-20c7-4b7a-bd25-2ddaf1f66c83

📥 Commits

Reviewing files that changed from the base of the PR and between 6dc2f51 and 03cccf4.

📒 Files selected for processing (5)

• csrc/libtorch_stable/cooperative_topk.cuh
• csrc/libtorch_stable/persistent_topk.cuh
• tests/kernels/test_top_k_per_row.py
• tests/model_executor/layers/test_sparse_attn_indexer_b12x.py
• vllm/v1/attention/backends/mla/indexer.py

---

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.}