noticeWebEmbedderBackendIgnored function

void noticeWebEmbedderBackendIgnored(
  1. PreferredBackend? requested
)

The web counterpart of noticeEmbedderBackendIgnored, with a different rule as well as different words.

On web NO value of preferredBackend reaches an embedder — cpu included, which the native rule rightly stays quiet about because native embeddings do run on CPU. Here the runtime chooses: LiteRT.js asks for WebGPU and falls back to WASM, and onnxruntime-web tries webgpu then wasm. The native line said "Embeddings run on CPU" to a caller whose embeddings were running on WebGPU — a backend reported that did not run.

Implementation

void noticeWebEmbedderBackendIgnored(PreferredBackend? requested) {
  if (requested == null || _warned) return;
  if (!kDebugMode || gemmaLogLevel == GemmaLogLevel.none) return;
  _warned = true;
  gemmaLog(
    'ℹī¸  preferredBackend ${requested.name} is not applied to embeddings on '
    'web: the runtime picks the accelerator (WebGPU first, WASM as the '
    'fallback). EmbeddingModel.activeBackend is null on web; with LiteRT, '
    'window.getLiteRtEmbeddingAccelerator() names it after the first '
    'embedding.',
  );
}