Airframe
A pure-Rust inference engine that runs GGUF language models directly on the GPU through WebGPU compute shaders, working across NVIDIA, AMD, Intel, integrated, and Apple Silicon GPUs without a C++ toolchain or llama.cpp.
Airframe runs full transformer inference on the GPU via WGSL compute shaders, avoiding the C++ toolchain, native library linking, and cross-compilation difficulty that come with wrapping llama.cpp. It builds with plain cargo build and targets WebGPU (wgpu) as its backend, so the same code runs on NVIDIA, AMD, Intel, integrated GPUs, and Apple Silicon.
It is the GPU inference core behind the Shimmy project, with a certification pipeline covering math, inference, and determinism across dozens of model and quantization combinations, and metadata-driven architecture support derived from GGUF tensor presence rather than hardcoded architecture strings.
The WebGPU inference runtime -- attention, the GGUF loader, and quantization -- is MIT licensed. A separate Fused Semantic Execution subsystem included in the repository is covered by a pending patent and licensed differently.