warpfront/hipfire ? reverse-engineered prompt

Build me a fast local LLM app for AMD RDNA GPUs that feels as easy to use as Ollama, but runs as a single Rust binary with HIP and no Python in the hot path.

I want to be able to download models, start a background server, and chat with them from the terminal, plus send one off prompts from the command line. The server should expose an OpenAI compatible API on localhost, and it should auto start if I try to chat before it is running. Make it work well on consumer AMD cards and support both small dense models and larger ones, with a clean model registry and simple pull and run commands.

If anything is unclear, look up current docs online if you need to. I care most about getting the core experience working end to end, model loading, serving, chat, and decent performance on RDNA GPUs.