NaiveAI-Labs / NaiveRT

git clone https://github.com/NaiveAI-Labs/NaiveRT.git

NaiveRT: a self-contained TP8 inference runtime for the Naive N series on eight Blackwell GPUs with UMMA support. One persistent CUDA graph per rank runs draft, eight-row verify, top-p sampling and cache commit with no host in the loop: 1,700 tokens/s at batch size 1, one-second peaks above 2,000.