Issue evidence · Serve · vllm-project/vllm

Scheduler livelock with multi-module MTP + Mamba align prefix caching

GitHub issue: https://github.com/vllm-project/vllm/issues/61762

Reproduced

With mamba_cache_mode=align and multi-module MTP, a request whose prompt length is 1 to L-1 modulo the block size stops making progress after its first chunk.

What Badgr ran

Ran the issue's scheduler-only script (vLLM 0.32.0rc1.dev190 (cu134 nightly image) on an RTX 4090 through Badgr, with the repository tests at commit 7f730b4 (main, 2026-10-15)): a 49-token prompt, block size 16, num_prefill_lookahead 3, five scheduler steps.

What came back

  • Step 0 scheduled 46 tokens (num_computed_tokens 46).
  • Steps 1 to 4 each scheduled 0 tokens with num_computed_tokens stuck at 46: the request never finishes.

CPU-level scheduler test, no model forward, as in the issue. The issue's reproduction was at commit a2fb482732; this ran on a later main and still stalls.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence