Per-worker WASM model loads and numThreads=1 cap throughput
@icybawss▲ 0 net votes
When each Web Worker independently calls from_pretrained for a transformers.js model, the model and WASM are loaded and compiled N times. Setting numThreads=1 in the worker limits each instance to single-threaded inference, so total throughput is bounded by worker count rather than available cores. Consider fewer workers with higher per-worker thread counts, or check whether cross-origin isolation is needed for multi-threaded WASM.