Repository navigation
Conversation
With the hybrid KV cache manager disabled, _promote_local_kv_cache_specs rebuilds each SlidingWindowMLASpec as an MLAAttentionSpec via replace_as, which forwards every field of the source spec. bounded_replay (added in vllm-project#56227) exists only on SlidingWindowMLASpec, so the rebuild raises TypeError: MLAAttentionSpec.__init__() got an unexpected keyword argument 'bounded_replay' and any sliding-window MLA model (DeepSeek-V4) fails at KV cache planning under --disable-hybrid-kv-cache-manager. A promoted spec keeps every block, so the window-bounded replay is moot, as extra_retained_tokens already is. Signed-off-by: Basem Barakat <basem.barakay@amd.com> Co-authored-by: Cursor <cursoragent@cursor.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment Once the PR is approved or has the If you have any questions, please reach out to us on Slack at https://slack.vllm.ai. Agent GuidelinesIMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban. 🚀 |
Overview
With
--disable-hybrid-kv-cache-manager, KV cache planning crashes on any model withSlidingWindowMLASpeclayers (DeepSeek-V4), because promotion toMLAAttentionSpecforwards the SWA-onlybounded_replayfield. Introduced by the interaction of #56227 (addsbounded_replay) and #61183 (moves the promotion intokv_cache_planning.py).Claims
TypeError: MLAAttentionSpec.__init__() got an unexpected keyword argument 'bounded_replay'at startup for sliding-window MLA models with the hybrid KV cache manager disabled.unify_hybrid_kv_cache_specswith aSlidingWindowMLASpeclayer, which no existing test covered.Validation
Unit test, run on MI300X (gfx942) in a ROCm 7.2 image built from this branch's tree; the fix was reverse-applied to the installed package for the "without" column:
-k swa_mlatests/v1/core/test_kv_cache_planning.pyTypeErrorabove)End to end, DeepSeek-V4-Flash and DeepSeek-V4-Pro, disaggregated prefill/decode (MoRIIO connector, 2P/2D and 4P/4D,
-tp 1+ DP + EP), MI300X,--disable-hybrid-kv-cache-manager:_promote_local_kv_cache_specsduringget_kv_cache_configs;Details
When the hybrid KV cache manager is off and full-attention layers are present,
_promote_local_kv_cache_specsrebuilds each sliding-window spec as its full-attention counterpart throughreplace_as, which forwards every field of the source spec.SlidingWindowMLASpechas two fieldsMLAAttentionSpeclacks:extra_retained_tokens, which was already dropped, andbounded_replay, which was not. The field is always present, so the crash happens even at its defaultFalse.Dropping it is semantically right for the same reason as
extra_retained_tokens: a promoted spec allocates blocks for every token and never frees below the window, so window-bounded prefix replay has nothing to bound. The promoted spec reports the ordinary full-attentionprefix_cacheable, which is correct because it retains the full prefix.Pull Request Checklist
I used vLLM's
/pr-checklistskill. (Mandatory for agents, optional for humans).AI assistance was used during the creation of this PR.
Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.
Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.
Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.
Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.