• Sources: primary, discussion
  • Summary: Cloudflare says prompts longer than 681,280 tokens overflowed the prefill KV pool, and SGLang in disaggregated prefill and decode mode replaced them with a one-token placeholder, so decode attended over stale pages from earlier requests. A researcher from Knostic reported it on 2026-09-08 and Cloudflare restored the model on 2026-09-09 with a larger pool and a patched engine. The fix is upstream in SGLang v0.5.21 and Cloudflare found no evidence of deliberate exploitation.
  • Why it matters: Anyone self-hosting SGLang with disaggregated serving should upgrade and check KV pool size against the advertised context length, since inference engines do not zero cache pages between tenants.