208x speedup + 5 microsecond prediction, running KMeans directly on H200 reaches 61% peak FLOPS, Berkeley folks have made Triton go all out

Crypto界 claims that the FlashLib open-source machine learning acceleration library disclosed by OneMillion_AI was developed by teams such as Berkeley, covering 15 high-level operators. Based on Triton and Cutedsl, it achieves significant speedups for operators such as KMeans and KNN on H200 GPUs, up to 208x. Compared with cuML 25.10, KMeans is 26x faster, KNN is 19x faster, HDBSCAN is 40x faster, and TruncatedSVD is 208x faster. KMeans peak FLOPS reaches 61%, and KNN bandwidth is 85.2%. In addition, FlashLib also provides a performance prediction API within 5 microseconds to estimate runtime and GPU memory overhead; the code has been open-sourced on GitHub.