srelus/llama.cpp-ML ? reverse-engineered prompt

Build me a version of llama.cpp that can train LoRA and QLoRA adapters directly on GGUF models without using PyTorch. I want it to keep normal llama.cpp inference working, but add a separate training flow that runs on Linux first, with CPU and CUDA support if possible. It should be able to read simple JSONL training data, including prompt and completion style data and chat style messages, then train an adapter, save it as a GGUF LoRA file, and load it back for inference with the existing CLI.

Please also include a basic quality check after training with a few holdout prompts so I can see if the adapter is doing anything useful. If anything needs current upstream knowledge or build details, look up the latest docs online and keep the setup as simple as possible.