An implementation of GPT inference in less than ~2000 lines of vanilla Javascript.
GPT-4: WebGPU is set to revolutionize web applications by providing near-native access to the GPU and compute shaders. WebGPT, a vanilla JS and HTML implementation of a transformer model, serves as a proof-of-concept and educational resource. With potential for significant speed increases and optimization, WebGPT demonstrates the capabilities of WebGPU in machine learning applications. Running WebGPT is simple, requiring only compatible browsers like Chrome Canary with “Unsafe WebGPU” enabled.
Read more at GitHub…