mikealanni/gufo is a local inference engine that uses the GPU and the XDNA2 NPU together on AMD Strix Halo (gfx1151). It splits part of the prompt processing across the two devices and supports speculative decoding, which pre-generates the next token for verification. The repository description names Qwen3.6-35B-A3B, Qwen3.8-27B, and Qwen3.8 Flash-Next. It is written in C++ and uses the MIT license.
The feature of this project is its attempt to accelerate prompt processing and token generation by using the graphics processing unit and the NPU together. The README summary explains that NPU-based prefill acceleration is optional and that it offers speculative decoding in MTP, DFlash2, and automatic-selection modes. According to the results recorded in the project, the NPU prefill speedup ranges from 9% to 28%, and speculative decoding speed on structured text reaches up to 159 tok/s. Separately, support for Laguna S 2.1 (118B MoE), a prefill change from 1,145 to 1,277 tok/s, and 44~45 tok/s when using the DFlash drafter are introduced. The execution path per model, its effect on quality, and the measurement methodology are also documented.
These figures should be read as limited to the stated conditions, such as the AMD Strix Halo environment and structured text, and the given material provides no basis for assuming the same speeds on other hardware or workloads. There is also no information on the latest registered release. Therefore, while this record shows the implementation and measurement results on that device, its scope is too narrow to be extended into a general performance comparison.