vLLM's P2pNcclConnector leaks KV-cache tensors when callers plant a host in X-Request-Id
A tracing header steers prefill workers to attacker ZMQ peers and ships the KV cache; three companion DoS bugs want 0.31.0.
Read the pieceOne published Frontier piece on this subject — open it below.
1 article · Full archive · Topics
This tag currently points at a single story.
A tracing header steers prefill workers to attacker ZMQ peers and ships the KV cache; three companion DoS bugs want 0.31.0.
Read the pieceNew releases, then the constraint.