For llama-server (and possibly other similar applications) you can specify the number of GPU layers (e.g. `--n-gpu-layers`). By default this is set to run the entire model in VRAM, but you can set it to something like 64 or 32 to get it to use less VRAM. This trades speed as it will need to swap layers in and out of VRAM as it runs, but allows you to run a larger model, larger context, or additional models.