Transformers.js v4 released on NPM with WebGPU runtime, ModelRegistry, and monorepo refactor

AI Impact Summary

Transformers.js v4 introduces a rewritten WebGPU runtime in C++, enabling local, browser, and server-side inference with broader operator support via ONNX Runtime Contrib. The release also restructures the codebase into a pnpm-based monorepo, switches build tooling to esbuild, and introduces the ModelRegistry API to manage pipeline assets and per-file metadata, which can optimize downloads and caching. Enhanced environment controls (env.useWasmCache, env.fetch) and configurable logging, along with new architectures and operators (e.g., com.microsoft.MultiHeadAttention, QMoE) expand model support and performance, including faster runtimes and smaller bundles (notably transformers.web.js). Upgrading will improve performance and deployment simplicity, but teams should plan for API/import changes and adopt ModelRegistry to realize the download/caching benefits.