lyogavin / airllm

Python library that runs large language model inference layer by layer, so 70B-class models fit on a single 4GB GPU without quantization.