You can memory pool with the right software however pytorch supports spreading large models over multiple GPUs OOTB. Just pass the --gpu-memory parameter with two values (one per GPU) to oobabooga's text-generation-webui for example.