A single inference worker holding backend state.

Started only for models whose backend declares concurrency: :exclusive. Shared backends run in the calling process and never reach this module.

A worker does exactly one thing: call the backend. Cache lookups, input validation and the preprocess/postprocess hooks all run in the caller before dispatch, because a worker slot is the scarce resource — occupying one with an Ecto query while a GPU sits idle is the mistake this design exists to prevent.

Deadlines

Requests carry an absolute monotonic deadline, and the worker checks it before invoking the backend. Work whose caller has already timed out is dropped rather than run.

That matters under overload. A GenServer.call timeout only abandons the caller's side; the worker still grinds through the whole queue, every item arriving later than the last, and the system never recovers. Checking the deadline at the front of the queue turns a death spiral into load shedding.