Backend normalization¶

Topics: architecture, backends, dispatch, generation

Horde requests describe intent shared across workers, while inference backends expose their own sampler names, payload shapes, defaults, and result formats. Dispatch and backend-parsing layers absorb that mismatch.

Conversion happens as close as possible to the backend. API models remain faithful to the Horde wire contract; generation-parameter models retain backend-neutral meaning; adapters resolve backend names and required defaults only when execution is selected.

Costs and alternatives¶

Embedding backend fields in API models would remove some mapping code but would make the public API grow with every backend and allow invalid cross-backend combinations. Passing untyped dictionaries through the worker would maximize flexibility but defer errors until execution and make contracts hard to inspect.

Typed edge adapters add maintenance when either side changes. They provide a single place to reject unsupported values, test exact emitted payloads, and keep backend upgrades from leaking across the package. Reverse conversion uses a separate path because aliases and defaults are rarely invertible.

The backend-mapper procedure gives the implementation steps. The planned dispatch interface reference identifies the formal extension points.