Repository navigation
fix(vllm): stop engine monitors during handler cleanup - #16412
rudrakshkarpe wants to merge 2 commits into
Conversation
Signed-off-by: rudrakshkarpe <rudraksh.karpe@gmail.com>
|
👋 Hi rudrakshkarpe! Thank you for contributing to ai-dynamo/dynamo. Just a reminder: The 🚀 |
Signed-off-by: rudrakshkarpe <rudraksh.karpe@gmail.com>
External PR triageBlocker ( Next action: @rudrakshkarpe: merge the latest Supply-chain audit of Automated triage, 2026-10-19. This comment is edited in place when the PR changes. |
Summary
Stop vLLM health and stats monitoring from generation and embedding handler cleanup through a public, repeatable
VllmEngineMonitor.stop()method. The worker factory retains ownership ofAsyncLLM. The cleanup guide now describes that split.Part of #16284. The transcription handler is in the still-open #16208; its author plans to rebase on this hook after this PR lands.
Validation
Added focused regression tests for running, unstarted, completed, and optional monitor tasks, plus repeated handler cleanup. The new cases fail against main and pass with this fix in local CPU checks using vLLM/runtime import stubs. Python compilation, pre-commit, and docs lint passed. Backend CI remains the integration check.
In the vLLM test image: