warpfront/hipfire is a Rust-based local inference engine that runs LLMs on AMD GPUs. The latest release is v0.4.2.1, but no changelog for that version can be found. The v0.4.2 update recorded in the README raised EP2 prefill performance on 2x R9700 from 3,033 to 3,943 tok/s at an 8K context, and from 3,754 to 4,308 tok/s at 16K. That is the record for v0.4.2, and it cannot be read as a change in v0.4.2.1.
The engine is built in Rust and HIP, and places no Python, PyTorch, CUDA conversion layer or third-party inference engine in the hot path. ROCm is loaded dynamically. It includes its own runtime and model implementation, the MQ3–MQ6, MQ4R and MFP4 quantization formats, and HIP kernels, and it also supports speculative decoding that takes advantage of DFlash draft models or native MTP heads, plus multi-GPU expert parallelism. The key point is that this is an independent implementation tailored to local inference on AMD GPUs.
That said, the README notes that the performance figures come from single-fixture measurements and therefore do not represent general throughput. There is also a known issue where sampled MTP produces a different token stream on each run, so results are not reproducible. The default weights of Qwen3.8-Flash are given as roughly 125GB, so a 32GB card alone is not enough. The measurement and model execution conditions given are tied to specific hardware such as the R9700 and Strix Halo, which makes them difficult to interpret as evidence of general-purpose performance.