Strix Halo Quant Lab
Measured benchmarks, build recipes and more than a hundred ready-to-run GGUF models for running large language models locally on AMD's Ryzen AI Max+ 395 (Strix Halo) hardware with 4-bit ROCmFP4 quantization.
Strix Halo Quant Lab documents which model architectures run faster after being quantized to the ROCmFP4 4-bit format on AMD Ryzen AI Max+ 395 (Strix Halo, gfx1151) hardware, by how much, and which ones do not. The ROCmFP4 format itself comes from a separate llama.cpp fork; this repository provides the measurement work, build recipes, the architecture ports needed to make specific models load, and every published quantization, with every number measured on the same machine and the same binary, including failures.
Because ROCmFP4 introduces a new tensor type that upstream llama.cpp, Ollama, LM Studio and similar runtimes do not recognize, these builds only load in the fork, though AMD's own Lemonade Server can also be pointed at that engine. The project targets gfx1151 hardware specifically, with other AMD ROCm GPUs likely to work after retargeting, and points NVIDIA users to separate NVFP4 repositories instead.