ben1078/hipfire is a fork repository of hipfire, the Rust inference tool for AMD GPUs. The README covers MTP decode work for Qwen3.8-27B, and the repository description states that it records about 100 tok/s on a single RX 7900 XTX. The changes presented alongside are int8 verify GEMV and WMMA verify attention. These point to verification using 8-bit integer matrix-vector operations and to an attention verification path using WMMA, respectively.
The base tool provides model download and execution, chat, and server features. The README distinguishes a stable release from the next version under testing, and also describes image generation and paths for several GPU families. Redline replays the verified path based on the recorded kernel graph and resource dependencies, and falls back to ordinary HIP execution under failure conditions. These descriptions show that the project covers not only decode performance but also execution paths and behavior on failure.
The performance figures are measurements from that fork branch. The README also limits the values in the performance table to a specific task and hardware, and states that they are not a ranking across models. Therefore, the roughly 100 tok/s cannot be read as the performance of an official hipfire release or of the current upstream. The scope confirmed in this repository is a decode path attempted with a specific model and an AMD GPU, and no material backing a broader performance comparison is given.