webgpu-llm

Train BPE tokenizers and small transformer LLMs entirely in the browser on WebGPU — forward pass, backward pass, AdamW, all in hand-written WGSL. No server, no Python.