KVarN: Native vLLM KV-cache quantization back end by Huawei

Huawei's KVarN is a new KV-cache quantization method for vLLM that achieves 3-5x greater cache capacity while maintaining FP16-level accuracy and throughput—a significant improvement over existing quantization approaches that typically sacrifice speed or quality. Implemented as a plug-and-play vLLM backend requiring only a single flag, KVarN uses variance-normalized int4 quantization and supports various model architectures including MLA and hybrid models like Mamba. On tested models like Qwen3-32B, it demonstrates matching accuracy to FP16 with superior throughput and substantially increased context capacity.