honouredBackends method
Explicit accelerator requests this provider honours for spec.
Pure and synchronous; no loading or platform I/O. Core calls this once per request and owns cache identity. An unsupported non-NPU request uses the legacy CPU key and receives a not-applied notice. NPU always has its own key. Automatic requests reuse only non-NPU entries for the same model. A repeated CPU/GPU request can reuse its originating successful fallback.
The default preserves existing CPU/NPU request handling. CPU and NPU membership do not change key normalization: CPU uses the CPU key and an explicit NPU request always reserves the NPU key. GPU membership opts into GPU keys, mandatory loaded-placement reporting, actual-placement recording for non-NPU requests and release-visible fallback warnings. Override for model-dependent capabilities, e.g. {cpu, gpu} for native GGUF. A provider advertising GPU must report the loaded accelerator in EmbeddingPassInfo.activeBackend; a descriptor alone is insufficient.
Implementation
Set<PreferredBackend> honouredBackends(EmbeddingModelSpec spec) => const {
PreferredBackend.cpu,
PreferredBackend.npu,
};