OpenLoRA loads LoRA adapters per request, freeing GPU memory to serve 1000+ behaviors on one GPU.