Issue evidence · Serve · vllm-project/vllm

Abort before load-failure processing leaves invalid prefix-cache blocks reusable

GitHub issue: https://github.com/vllm-project/vllm/issues/61768

Reproduced

If a request is aborted before a KV load failure is processed, the failed blocks stay hashed in the prefix cache and a later request reuses them.

What Badgr ran

Ran the issue's isolated mock-connector 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)).

What came back

  • After the abort and the invalid-block report, all 8 blocks still had a block hash: [True, True, True, True, True, True, True, True].
  • A second request then reused 112 cached tokens (7 blocks, 4 of them never loaded); the expected value is 48.

Mock KV connector at the scheduler level; no real remote KV transfer was involved.

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

← All issue evidence