Hipfire brings native LLM inference to AMD GPUs

AMD consumer GPUs finally get a first-class citizen for local LLM inference, completely bypassing the heavy ROCm stack.

• –Built from scratch in Rust and HIP, running as a single binary without Python or ROCm userspace dependencies.
• –Features DFlash speculative decoding, which delivers up to 4.45x speedups on code generation tasks.
• –Consistently outperforms Ollama on AMD hardware, showing up to 2.1x faster decode speeds on the RX 7900 XTX.
• –Introduces "MagnumQuant" (MQ4/MQ6), a custom quantization method aiming for Q8 quality at Q4 bandwidth.