vLLM x Novita AI: Chord, Faster INT4 MoE for Kimi K2.x. Up to 1.3x on H200, 2.15x on Untuned B300
Novita AI open-sourced Chord, a W4A16 MoE CUDA operator for vLLM that optimizes Kimi K2.x serving. Benchmarks report 1.11x to 1.33x speedups on H200 hardware and up to 2.15x on B300 decode. The release includes code and performance tables, though grouped operator integration remains a work in progress.