preferredVisionBackend property
Backend for the VISION encoder, independent of preferredBackend (the
text/decoder backend). Null defaults to CPU in the LiteRT-LM engine,
because the vision encoder's STABLEHLO_COMPOSITE ops fail to prepare on
the Metal/WebGPU delegates (LiteRT-LM#2461). Set it (e.g. to
PreferredBackend.gpu) only for a model whose vision section bakes a
gpu-only backend_constraint. Ignored by the MediaPipe engine (whole-model
backend, no per-encoder split).
Implementation
final PreferredBackend? preferredVisionBackend;