Skip to content

Compatibility registry

The registry answers “does model X run on device Y, on which accelerator, how fast, and with which fix?”. It is filled by the nodes themselves: every node files what it measured.

Status Meaning
works loaded on the manifest’s first-choice accelerator
tweaks works, but only after a fallback or a setting (the tweaks field says what)
fails the model failed to load or run (the notes field has the error)
  • Automatic (source: auto). After a model’s first five runs on its backend, a node sends a benchmark frame: p50 and p95 of its last 64 runner times (render, pack, invoke, postprocess; no queueing or load), the backend and the status. At most every 6 h per model, and right after a model goes from failing to working.
  • Curated (source: manual). Measurements filed with the admin key, such as the seed results from the TV and the Pixel.

A newer automatic result for the same model, version, device and backend replaces the older one.

  • Registry page: filter by model, device, SoC, accelerator and status; per-device and per-model pages.
  • API: GET /v1/registry (any key) and the public, key-less GET /public/v1/registry (see Public endpoints).

Example of why the tweaks matter: EfficientNet-Lite4 on the TV’s NPU runs in 22 ms, but only when the model is loaded from a memory-mapped buffer. Loaded by path from shared storage it takes about 9 s, because the vendor driver cannot read the file and NNAPI silently runs on its CPU reference path.