Unverified50% confidenceFactExact time
During cross-machine handoff, Decode only asked whether the task was complete and did not check whether the task had been rejected, so after Prefill rejected an overlong prompt and replaced it with a placeholder, Decode still reserved and read pages according to the original plan, exceeding the amount actually sent.
1
Sources
50%
Confidence
Long-term
Relevance
10/31/2026
First Seen
Sources
The KV Cache Leak Incident: The Technical Truth Behind Cloudflare AI's 155 Cross-Conversation Leaks
youtubeFrame By Bytes10/28/2026
Related Entities
Related Claims
UnverifiedThe Decode phase is memory-bandwidth-intensive, requiring model weights and the KV cache to be reloaded from HBM at each step, which determines inter-token latency (ITL).66% similarUnverifiedWhen multiple requests are processed concurrently, prefill tasks and decode tasks compete for GPU resources, creating a prefill-decode interference problem.65% similarVerifiedThe Prefill phase is a compute-intensive task with high demands on compute power; the Decode phase is a memory-access-intensive task that is sensitive to GPU memory bandwidth.65% similarUnverifiedThe Coding Harness focuses on reliable execution of individual coding tasks, including injecting code context, managing file read/write permissions, executing command lines and returning results, and handling error retries.65% similarUnverifiedIn the decoder route, the context can be placed first and cached; under the causal mask, the context only needs one prefill to obtain the KV cache, which is then reused directly for each subsequent question, computing only the newly added question tokens.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1119337API
curl https://kongchang.com/api/v1/knowledge/claims/1119337MCP
get_claim(id=1119337)