Fix DCP sparse indexer decode metadata

[voipmonitor]

Summary

Fix GLM/Kimi DCP decode sparse-indexer metadata for the uncompressed indexer-cache path (compress_ratio == 1).

DCP writes the indexer KV cache through rank-local pages, but the previous path only built a remapped slot mapping when compress_ratio > 1. For GLM/Kimi DCP4 this meant the sparse indexer could write with global slot ids while B12X attention consumed the selected indices through a DCP-local page table, causing incoherent long-context decode.

This patch:

• Builds DCP-local slot mapping whenever dcp_world_size > 1, not only for compressed KV paths.
• Feeds DCP-local decode seq_lens to the sparse-indexer metadata when available.
• For expanded/native MTP decode rows, computes per-token lengths from global seq_lens first, then converts those expanded lengths to DCP-local lengths so query_start_loc arithmetic remains correct.

Validation

Validated on GLM-5.1 NVFP4 DCP4 no-MTP with B12X_MLA_SPARSE, V2 model runner, DCP4 ag_rs, and FULL+PIECEWISE CUDA graphs.

Commands/results:

• python3 -m py_compile vllm/v1/attention/backends/mla/indexer.py
• git diff --check lil/main..HEAD
• python3 /mnt/test.py --port 5331 --model GLM-5.1 --max-tokens 180 --quiet --json-summary -: coherent, chinese_count=0
• python3 /mnt/test.py --port 5331 --model GLM-5.1 -c 50000 --max-tokens 220 --quiet --json-summary -: coherent, chinese_count=0, prompt tokens 31,876, TTFT 12.43s
• python3 /root/llm-inference-bench/llm_decode_bench.py --port 5331 --concurrency 1 --contexts 0k --max-tokens 2048 --skip-prefill: 58.1 tok/s, TTFT/ITL 248/17 ms

Runtime image used for validation:

voipmonitor/vllm:cu132-vllm611a842-b12xf9226c-a16nativew4a16-20260606

Only vllm/v1/attention/backends/mla/indexer.py was overlaid for the final validation run.

Summary by CodeRabbit

• Bug Fixes
  • Enhanced distributed context processing in Deepseek model attention mechanisms to ensure consistent behavior and correct sequence length computation across various configuration settings.

[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: c94f244b-ef08-4e10-ab13-a2425dc35360

📥 Commits

Reviewing files that changed from the base of the PR and between 8f6c0a7 and 95f53f4.

📒 Files selected for processing (1)

• vllm/v1/attention/backends/mla/indexer.py

---

📝 Walkthrough

Walkthrough

DeepseekV32IndexerMetadataBuilder now routes global decode sequence lengths into decode tensor preparation, enabling each expansion path to apply DCP-local conversion consistently. Buffer allocation conditions broaden to support DCP paging without compression, and downstream DCP conversion is guarded to prevent redundancy.

Changes

DCP-local sequence length support in decode tensor preparation

Layer / File(s)	Summary
Buffer allocation and compression condition broadening vllm/v1/attention/backends/mla/indexer.py	Compressed slot mapping buffer allocation and compressed seq lens computation conditions are expanded from compress_ratio > 1 to compress_ratio > 1 or dcp_world_size > 1, enabling graph-stable slot mapping when DCP paging occurs without compression.
Decode tensor preparation with DCP-local seq lens support vllm/v1/attention/backends/mla/indexer.py	_prepare_decode_tensors gains an optional global_seq_lens parameter. All three expansion paths (uniform decode, variable decode-length, and native MTP) now accept this parameter as the base for expansion and apply get_dcp_local_seq_lens conversion when DCP is enabled before storing into decode_seq_lens_buffer.
Decode metadata build and DCP-local seq lens selection vllm/v1/attention/backends/mla/indexer.py	The decode metadata build logic now computes global_decode_seq_lens and conditionally derives dcp_local_seq_lens from common_attn_metadata.dcp_local_seq_lens when compress_ratio == 1 and DCP is enabled. It then passes the appropriate seq lens to _prepare_decode_tensors via the new global_seq_lens parameter.
DCP conversion guard tightening vllm/v1/attention/backends/mla/indexer.py	A later DCP conversion branch is tightened from elif self.dcp_world_size > 1: to elif self.dcp_world_size > 1 and dcp_local_seq_lens is None:, preventing redundant DCP-local conversion when the build step already selected a DCP-local seq_lens.

🎯 3 (Moderate) | ⏱️ ~25 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 33.33% 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 'Fix DCP sparse indexer decode metadata' directly addresses the main change in the pull request—fixing DCP sparse indexer metadata handling in the MLA indexer.
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.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ 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/dcp4-indexer-clean-20260606

---

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 and usage tips.}

[lukealonso]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.