3 ms·QuantumLeap: 2.3× faster MoE inference with intelligent expert caching1 points by ikharoz 8mo ago