They need to keep KV cache to avoid prompt reprocessing, so they would need to move it to ram/nvme during longer api calls to use gpu for another request