Supervisor-side liveness probes for llama-server (completion probe plus SIGKILL)
Parent: Mac local LLMs: Serving ops and multi-model · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
Whether a supervisor should probe each child or only the router port (a router-level probe with the child's `model` name exercises the child).
These notes link each claim to its source. A source may be a research report hosted on this site rather than the primary document. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Facts
- Whether a supervisor should probe each child or only the router port (a router-level probe with the child's `model` name exercises the child). [source]
- The completion-probe plus SIGKILL design, the `/health` blind spot, the SIGTERM-absorbed behaviour and the launchd KeepAlive interaction are already covered by llama-server-readiness-and-health-semantics-afte.md; nothing contradicts it. [source]
- In the router-mode report the poisoned child was never seen to decode once in its life, so a probe that waits for recovery never succeeds; the trigger was a third model swap-in overlapping two residents under `--models-max 2` on a 64 GB M1 Max, pushing wired memory past the limit. [source]
Children
- No children recorded.