---
title: '56429'
pageLayout: default
slug: thoughts/inference/vllm-prs/56429
permalink: https://aarnphm.xyz/thoughts/inference/vllm-prs/56429.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
# [#56429](https://github.com/aarnphm/aarnphm.github.io/issues/56429) Revert “\[Rocm\]\[Kimi-k3\] Fix pipeline_parallel support for the kimik3 DCP mode ([#53664](https://github.com/aarnphm/aarnphm.github.io/issues/53664))”

[vllm-project/vllm#56429](https://github.com/vllm-project/vllm/pull/56429) · shen-shanshan · MERGED · created 2026-09-11T08:48:23Z · merged 2026-09-11T16:17:26Z

## Purpose

This PR reverts [#53664](https://github.com/aarnphm/aarnphm.github.io/issues/53664) (“\[Rocm\]\[Kimi-k3\] Fix pipeline_parallel support for the kimik3 DCP mode”). That PR added a fallback in `MLAAttentionLayer.build()` that derives `dcp_local_seq_lens` from `seq_lens` whenever `CommonAttentionMetadata` does not carry it, together with an assert and two test-registry entries. The fallback is dead code: in vLLM’s in-tree paths, `dcp_local_seq_lens` is always populated before the MLA decode path runs whenever DCP is enabled.

Evidence:

- The monolithic V1 model runner populates `CommonAttentionMetadata.dcp_local_seq_lens` unconditionally when `decode_context_parallel_size > 1` (`vllm/v1/worker/gpu_model_runner.py:2456-2466`), and this has been in place since before [#53664](https://github.com/aarnphm/aarnphm.github.io/issues/53664) was authored.
- The modular model runner (MRV2) does the same via `maybe_prepare_dcp_local_seq_lens()` right before attention metadata is built, after PCP batch partitioning (`vllm/v1/worker/gpu/model_runner.py:1728`, helper in `vllm/v1/worker/gpu/cp_utils.py`). The helper returns `None` only when `dcp_size == 1`, in which case the MLA layer’s `dcp_world_size` is also 1 and the branch is never taken, so the fallback is unreachable.
- Both runners always set `seq_lens` on `CommonAttentionMetadata`, so the assert added by [#53664](https://github.com/aarnphm/aarnphm.github.io/issues/53664) is likewise unreachable.

Keeping the fallback would also mask future regressions in runner-side population: if a path ever stopped populating `dcp_local_seq_lens`, the layer would silently derive it instead of failing loudly. Removing it restores a single source of truth for DCP-local sequence lengths (the model runner). The two test-registry entries added by [#53664](https://github.com/aarnphm/aarnphm.github.io/issues/53664) are removed along with the revert.

## Test Plan

## Test Result

---

<details>
<summary> Essential Elements of an Effective PR Description Checklist </summary>

- [x] The purpose of the PR, such as “Fix some issue (link existing issues this PR will resolve)”.
- [x] The test plan, such as providing test command.
- [x] The test results, such as pasting the results comparison before and after, or e2e results
- [ ] (Optional) The necessary documentation update, such as updating `supported_models.md` and `examples` for a new model.

</details>

---

## comments

**shen-shanshan** (2026-09-11T09:23:59Z):

/ci run

**github-actions** (2026-09-11T09:25:22Z):

✅ Triggered [Buildkite CI #88299](https://buildkite.com/vllm/ci/builds/88299) for commit `1e49aad0ff3b`.

**shen-shanshan** (2026-09-11T15:32:25Z):

/ci retry

**github-actions** (2026-09-11T15:32:43Z):

✅ Queued 3 failed job(s) for retry in [Buildkite CI #88299](https://buildkite.com/vllm/ci/builds/88299).

