FlexKV came up the most in yesterday's replies, so here's the longer version： It's a KV cache layer that sits under your inference engine, open sourced last year. works across SGLang, vLLM, TensorRT-LLM and Dynamo. If you serve long contexts you know the shape of this: cache fills VRAM, gets evicted, and you re-prefill a prefix you already computed once. A cache is only worth what it hits, most of FlexKV's design goes into keeping that number high: — maximizes effective cache capacity through cross-rank deduplication and lossless nvCOMP ANS compression, across CPU, SSD, Mooncake Store, and GDS — maintains a cluster-wide distributed RadixTree for cross-node prefix discovery and reuse — exposes cache residency and prefix matches to Dynamo for cache-aware routing DeepSeek-V4 support landed last month too, heterogeneous KV groups and compress-state sidecars included, with layerwise loading to cut read latency. SGLang path is ready now, other engines coming Repo→