# HaloFPX

A local model server for AMD Strix Halo and Radeon GPUs that keeps command-line compatibility with an existing tool while adding day-zero support for newer large mixture-of-experts models and hardware-tuned quantization.

HaloFPX is a model server and model catalog engineered specifically for AMD's Strix Halo APUs and Radeon GPUs, built as a compatible alternative to an existing tool called Lemonade rather than a replacement requiring retraining. It keeps full command and model-cache compatibility with that tool, including reading its already-downloaded models and sharing its cache, while adding support for newer, larger mixture-of-experts models on the day they are released.

Its performance work is tied to this specific hardware: cooperative-matrix kernels tuned for the RDNA 3.5 architecture, a custom four-bit quantization format mapped directly to the hardware's compute blocks, and silicon-tuned server profiles that set speculative-decoding and context settings without the user guessing at flags. The project reports a validated 262,000-token context window across multiple concurrent request slots in unified memory, and automatic discovery of a model's vision projector for multimodal use over the standard OpenAI-compatible chat endpoint.

Website: https://github.com/julianmb/halofpx