hot-swap

Llamacpp server now supports hot-swapping models in under 30 seconds, a significant speed improvement over previous methods like PyTorch.