pipelined-decoding

Moondream's Photon inference engine eliminates GPU bubbles through pipelined decoding, achieving near-realtime VLM inference with up to 35% higher decode throughput.