Hierarchos-Native trains 143 Transformer architectures on Rust and Vulkan with no CUDA or PyTorch.

necat101/Hierarchos-Native is a native Rust plus Vulkan backend for training and inference across 143 canonical architectures, with Hugging Face model_type aliases bringing the advertised registry to 224 spellings. The stated goal is testing how much of the modern Transformer ecosystem runs on Vulkan compute without assuming CUDA and PyTorch. Single-source and unbenchmarked against a PyTorch baseline, so the architecture count is a claim rather than a result, but a vendor-neutral training path is rare enough to track.