I think we can help users get more from open source models by helping them fit larger models In Huggingface, these are things such as: - load-in-8bit - offload-dir - low-cpu-mem-usage (or whatever the flag is)