matteiuspi

Hardware notes, vLLM work and benchmarks from an unusual inference machine built around cheap Ascend cards, which are not CUDA drop-ins.