KTransformers

MADSys's CPU/GPU heterogeneous MoE inference framework (17.7K★, SOSP 2025; joint with Approaching.AI): made DeepSeek-R1 671B runnable on a single 24GB GPU and is now integrated into SGLang — the reference system for extreme-offloading inference. (The same lab's Mooncake serving stack files under Moonshot per our attribution rules.)