GpuDelegateV2 class

GPU delegate for Android Prefer PerformanceConfig over constructing this directly: it picks a sensible delegate per platform and handles the fallback when one declines a model. This class stays available for callers that need to configure a delegate themselves.

Not deprecated in favour of CompiledModel. LiteRT Next is the direction upstream is taking, but as of flutter_litert 3.7.0 it miscomputes models whose output tensor ends up dynamic, returning kLiteRtStatusOk while the output buffer is never written (see doc/delegate_verification.md and verifyCompiledModel). Until that is fixed upstream, delegate-backed Interpreter inference is the correct choice for those models, so pointing callers away from it would be wrong.