Pomilon/LEMA ? reverse-engineered prompt

Build me a Python library that lets me fine tune big language models on GPUs with limited VRAM, even when the model would normally not fit.

I want it to work like a memory smart training tool that can move model weights through disk, RAM, and GPU memory automatically, so training stays stable on smaller cards. It should support LoRA fine tuning, Hugging Face safetensors checkpoints, and common models like Llama, Mistral, Mixtral, GPT 2, and similar ones. Please make it easy to use from a simple config, with a trainer object, automatic checkpoint saving, and a quick setup flow that can download a model and start training with minimal code.

If it makes sense, include a Python first version and an optional faster C++ path, plus a small benchmark or auto check step that figures out the best prefetch settings for the current GPU. Add clear docs and a few example scripts so I can try it right away.