Related approaches¶

vacancyd places an admission gate in front of a dedicated inference server. It checks owner policy and sampled resources before forwarding borrowed work, and explains a refusal with HTTP 503, a reason and Retry-After. This is a bounded product description, not a claim that no other project can implement the same policy.

Model lifetime and admission answer different questions¶

Ollama's keep-alive controls decide how long a model remains loaded after a request. vacancyd can use Ollama's unload API when its reclaim policy permits it. Model lifetime alone does not express vacancyd's owner-presence policy or per-resource admission checks.

A gateway can choose another backend¶

LiteLLM's router supports routing and fallback. vacancyd supplies a separate admission answer for one lending machine; it does not replace a gateway or guarantee how that gateway classifies a refusal. Verify the actual adapter and configuration, including the behavior described in 503 interoperability.

Refusing new work does not stop accepted work¶

A rejected request does not begin its upstream model load. Work already accepted follows its configured cancel, drain or finish policy. A resource-scoped upstream can continue lending while the owner is present. Closing a connection and acknowledging an unload are not proof that computation stopped or memory was recovered. See admission and recovery.

Choose the boundary you need¶

Use vacancyd when borrowers can be directed through its proxy and the supported adapter routes fit the service. Keep the upstream dedicated to borrowed work: direct requests bypass the gate, and process attribution cannot separate owner and borrower requests inside one server process. Operating-system quotas, workload scheduling and network isolation remain the operator's responsibility.