ExLlamaV3 through TabbyAPI

ExLlamaV3 is the inference library; TabbyAPI is its official OpenAI-compatible server. LychD keeps that server outside the daemon dependency graph and registers it as the dynamic animator/exllamav3 Soulstone runtime.

The generated sample pins the TabbyAPI image by digest. The container starts without a model, so its declared capabilities become ACTIVATABLE, not WARM. LychD sends the declared model directory and backend through Tabby's lifecycle API; context, cache, split, and reserve remain TabbyAPI defaults unless a model-local tabby_config.yml overrides them.

Stable LychD [[models]].id values are not Tabby directory names. Each model path must identify a validated direct child of the mounted model directory. The connector derives the runtime name from that path's basename and translates it on both lifecycle and data planes.

Authentication is mandatory even inside the unpublished private pod. The two distinct API and admin keys remain Podman-secret material and are never placed in browser code or ordinary Rune fields. bind --uncaged rejects this Soulstone.

POST /v1/model/load reports stages over server-sent events. LychD consumes the complete stream and verifies the active model with GET /v1/model; a disconnected client alone is not proof that a detached load failed. An indeterminate stream remains contained and is reconciled or reset through the owning Vessel boundary.

Focused contracts cover the runtime, control plane, connector, registration, authentication, and containment behavior. The named GPU/model/runtime receipt remains operator validation.

See ExLlamaV3, TabbyAPI, and Soulstone Disciplines.