staylamebro/backburner is a distributed inference tool that runs Qwen3.8-27B locally on an Apple Silicon Mac and an iPhone connected over USB-C. According to the repository description, the two devices use a 10 Gb/s connection, and the purpose of adding the iPhone is to increase input processing throughput and context length. There is no registered latest release.
How it works depends on the context length. At 64k or below, the model layers are split so that the Mac handles layers 1–40 and the iPhone handles layers 41–64, and the prefill, which processes input tokens all at once, is shared between the two devices. Beyond 64k, the iPhone computes attention over older keys to support a longer context. The single-Mac path includes speedups such as SME2 kernels from its own llama.cpp fork, Metal fusion, and DFlash2-based inference decoding. The project is written in Python and is MIT licensed.
The basis for correctness rests only on the limited greedy output comparison reported in the README. Output identity was confirmed under 256/256 and 32/32 token conditions, and the automated test scope is also limited to five security folders. Therefore, while the inference path that uses multiple devices together and the way context is extended are concrete, it is difficult to judge accuracy under various conditions or regression safety from this material alone.