[eldritch] Fix B12X DCP global-topk MTP path

[voipmonitor]

Summary

Port the validated GLM-5.2 B12X DCP global-topk/MTP fix onto dev/eldritch-enlightenment.

This is the eldritch-targeted equivalent of the already validated dark-devotion PR #31. It keeps the B12X DCP global-topk path rank-consistent during warmup and makes the MTP draft path provide the same top-k score buffer that the target path expects.

Changes

• Preserve empty DCP prefill chunks for B12X global-topk collectives.
• Prewarm the real B12X DCP global-topk merge path.
• Keep no-local-KV chunks in the global merge path with -inf scores and mask invalid results back to -1.
• Prewarm B12X sparse MLA extend kernels using caller-owned scratch via plan.bind(scratch=...).
• Allocate and pass topk_scores_buffer from the MTP draft layer under DCP + B12X sparse indexer.

Notes

This keeps the vLLM B12X binding rule: no cached vLLM workspace/arena ownership; plans bind caller scratch per call.

Validation

• git diff --check lil/dev/eldritch-enlightenment..HEAD
• python3 -m py_compile on changed Python files

Runtime validation will be done in the composed eldritch Docker stack.

Summary by CodeRabbit

• Refactor / Performance
  • Improved sparse attention indexing and distributed top-k handling, including converting rank-local results into global form for B12X.
  • Added smarter DCP-aware synchronization and initialization prewarming to stabilize sparse attention backend startup.
  • Reworked warmup to reuse the real merge path for representative candidate sizes and ensure correct cross-rank behavior.
• Bug Fixes
  • Fixed prefill metadata handling for empty sequence ranges so DCP global top-k can still be preserved when enabled.
• New Features
  • Enabled conditional top-k score buffering for DeepSeek multi-token predictor when using the v32 path with multi-context decoding parallelism.

[coderabbitai]

Caution

Review failed

Pull request was closed or merged during review

📝 Walkthrough

Walkthrough

Adds DCP warmup synchronization and prewarm paths, a local-to-global top-k conversion path, B12X prefill/decode routing changes based on dcp_global_topk, topk_scores_buffer wiring in DeepSeek MTP, and empty DCP chunk preservation in the indexer.

Changes

B12X DCP Top-K Routing, Warmup, and Buffer Wiring

Layer / File(s)	Summary
DCP warmup sync and extend-kernel prewarm vllm/model_executor/layers/sparse_attn_indexer.py, vllm/v1/attention/backends/mla/b12x_mla_sparse.py	Adds _sync_dcp_warmup() in both files. Reworks _prewarm_b12x_dcp_topk_merge() to run _merge_b12x_dcp_topk() over candidate row sizes and synchronize after each run. Adds _EXTEND_PREWARM_DONE and _prewarm_extend_kernels_once() in the MLA backend, called during init before supports_quant_query_input.
Local-to-global DCP top-k conversion and routing vllm/model_executor/layers/sparse_attn_indexer.py	Adds _convert_b12x_dcp_local_topk_to_global() and updates B12X prefill/decode post-processing to choose between _merge_b12x_dcp_topk() and the new local-to-global conversion based on dcp_global_topk.
topk_scores_buffer wiring and empty DCP chunk preservation vllm/v1/attention/backends/mla/indexer.py, vllm/model_executor/layers/sparse_attn_indexer.py, vllm/model_executor/models/deepseek_mtp.py	Changes total_seq_lens == 0 handling to preserve empty DCP chunks when sparse indexing and global top-k are enabled. Updates the zero-length prefill branch to split B12X and non-B12X handling. Allocates and passes topk_scores_buffer through DeepSeek MTP when decode-context parallelism and B12X sparse indexing are enabled.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~60 minutes

Possibly related PRs

• local-inference-lab/vllm#31: Directly overlaps with the same B12X/DCP top-k routing, warmup, empty-chunk handling, and buffer wiring changes across the same files.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 13.33% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title is concise and accurately reflects the main fix to the B12X DCP global-topk MTP path.
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 codex/eldritch-dcp-globaltopk-mtp-20260623

---

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.}

[coderabbitai]

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@vllm/model_executor/layers/sparse_attn_indexer.py`:
- Around line 430-448: The _sync_dcp_warmup function currently catches all
exceptions and silently returns, which masks barrier failures and prevents
proper error surfacing. Replace the broad except Exception clause with a more
specific exception handler that only catches the expected uninitialized DCP
group case (when get_dcp_group() is unavailable or raises an initialization
error), and allow barrier() failures and other runtime exceptions to propagate
up so they surface as real errors instead of being silently ignored.

In `@vllm/v1/attention/backends/mla/b12x_mla_sparse.py`:
- Around line 699-714: The memoization key in the _prewarm_extend_kernels_once
method is being added to _EXTEND_PREWARM_DONE before any kernel warmup succeeds,
and the key tuple does not include the max_batched parameter even though this
parameter controls the largest warmed row shape. Include max_batched as an
additional element in the key tuple construction alongside the other parameters
like q_head_dim, kv_lora_rank, and topk_tokens, and move the
_EXTEND_PREWARM_DONE.add(key) statement to execute only after the kernel warmup
has completed successfully to prevent skipping required prewarm work for failed
attempts or larger scheduler batches.
- Around line 683-697: The current exception handling in the _sync_dcp_warmup
method catches all exceptions including failures from the
get_dcp_group().barrier() call, which hides critical synchronization failures
when dcp_world_size > 1. Restructure the error handling to only catch exceptions
from the import statement itself (the get_dcp_group import), but allow any
exceptions raised by the barrier() call to propagate so that rank-consistency
contract failures are not silently swallowed. This ensures that synchronization
failures are properly surfaced rather than hidden.

🪄 Autofix (Beta)

✅ Autofix completed

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 3670bc94-24c2-4aa9-9c9e-36e3b023ea0a

📥 Commits

Reviewing files that changed from the base of the PR and between 2d040b9 and dcad5d0.

📒 Files selected for processing (4)

• vllm/model_executor/layers/sparse_attn_indexer.py
• vllm/model_executor/models/deepseek_mtp.py
• vllm/v1/attention/backends/mla/b12x_mla_sparse.py
• vllm/v1/attention/backends/mla/indexer.py

[coderabbitai]

Note

Autofix is a beta feature. Expect some limitations and changes as we gather feedback and continue to improve it.

Fixes Applied Successfully

Fixed 2 file(s) based on 3 unresolved review comments.

Files modified:

• vllm/model_executor/layers/sparse_attn_indexer.py
• vllm/v1/attention/backends/mla/b12x_mla_sparse.py

Commit: 3e424016dd6c41063642f822db7cc4be2b5c5bc2

The changes have been pushed to the codex/eldritch-dcp-globaltopk-mtp-20260623 branch.

Time taken: 2m 19s