nanograft — Gemma-2-2B + custom LoRAs for in-browser WebGPU (MediaPipe)

MediaPipe LLM Inference (Web) build assets for the nanograft demo: a custom LoRA adapter running on Gemma-2-2B fully in the browser on WebGPU, with two adapters hot-swapped at runtime on one loaded base. Live demo: https://nanograft.naklitechie.com

Files

• gemma-2-2b-it-gpu.bin — Gemma-2-2B-it converted to MediaPipe int8 GPU format (~2.4 GB).
• lora_pirate.bin, lora_uwu.bin — two throwaway attention-only (q/k/v/o) style LoRAs (rank 16), converted to MediaPipe LoRA format. Loaded via loadLoraModel() and swapped at runtime with generateResponse(prompt, loraRef).

Provenance & license

Derived from google/gemma-2-2b-it. Use is governed by the Gemma Terms of Use (https://ai.google.dev/gemma/terms) and the Gemma Prohibited Use Policy. "Gemma" is a trademark of Google. These are format-converted / LoRA-adapted weights redistributed with attribution under those terms; no Google endorsement implied. The LoRAs are trivial style patches trained by nanograft for demonstration only.