@techreport{li-kvcache-efficient-transmission-00,
    number =    {draft-li-kvcache-efficient-transmission-00},
    type =      {Internet-Draft},
    institution =   {Internet Engineering Task Force},
    publisher = {Internet Engineering Task Force},
    note =      {Work in Progress},
    url =       {https://datatracker.ietf.org/doc/draft-li-kvcache-efficient-transmission/00/},
    author =    {Zhiqiang Li and Zongpeng Du and Junjie Wang and Wei Cheng and Ying Zhao and Chunhao Zhao and Xun Sun},
    title =     {{Layer-Aligned Key-Value (KV) Cache Streaming over Media over QUIC Transport (MOQT)}},
    pagetotal = 19,
    year =      2026,
    month =     nov,
    day =       2,
    abstract =  {The key-value (KV) cache of a language model can be computed once and reused by many inference instances, possibly at many sites. Media over QUIC Transport (MOQT) provides publish/subscribe delivery through relays that cache objects without understanding their content. This document specifies KV Cache Streaming (KVCS), an application of MOQT for delivering KV caches. KVCS maps the identity of a token prefix to a Track Namespace built from keyed identifiers that do not reveal the prompt, carries the key and value tensors (or a jointly encoded form) in separate Tracks, aligns each Group with one model layer and each Object with one block of tokens, and defines a self-describing object payload and a manifest track. These choices let consumers start using early layers while later layers are still being produced, retrieve only the layers and token blocks they need, and verify what they receive.},
}