ZeQLoRA: Efficient Finetuning of Quantized LLMs With ZeRO and LoRA Performance method num_gpus base_model dataset max_steps train_time train_loss eval_loss QLoRA 2 huggyllama-7b belle_0.5M 1875 ~37h38m ~ ~ QLoRA 1 huggyllama-7b belle_0.5M 1875 6h43m 1.4971 1.3185 ZeQLoRA 2 huggyllama-7b belle_0.5M 1875 10h45m 1.4682 1.2966