nanograft — custom LoRA hot-swap on Gemma-2-2B, in your browser

This page runs Google's Gemma-2-2B entirely in your browser on WebGPU, then loads custom LoRA adapters — small fine-tuned "personality" patches — and hot-swaps between them at runtime on a single loaded base model, with no page or model reload.

As far as we can tell, a public demo that runtime-swaps two custom developer-trained LoRAs on one in-browser Gemma base hasn't been shown before. The pieces exist in isolation (Gemma-on-WebGPU is common; adapter hot-swap is common on servers) — doing the swap on-device in a tab is the new bit.

Heads up: the base model is a ~2.4 GB download the first time (cached after). Needs a WebGPU browser (recent Chrome/Edge; Safari 18+). Two throwaway style adapters are loaded: pirate and uwu — deliberately silly so the swap is obvious.