The KV-cache quant race just got more interesting. Huawei just open-sourced KVarN , a KV-cache quantization method under Apache 2.0, drops into vLLM with one flag. Posting because the tradeoff it's claiming is genuinely different from what's already in the stack, and I'd like to see it stress-tested. The landscape it's stepping into FP8 ( --kv-cache-dtype fp8 ) is the current default: ~2x KV capacity, BF16-level throughput, near-zero quality loss. Hard to beat, and the bar anything ne...