Evidence · DiggingBeagle record
KVGuard: Request-level memory safety for GPU KV caches
GPU KV-cache allocations can be valid at the device-memory level while individual inference requests read uninitialized or cross-request cache regions. A request-level sanitizer annotates ownership and initialization state. KVGuard reports finding 29 historic vLLM/SGLang violations and eight new cases under its harness, with measured token-generation overhead. A detected memory-safety condition does not establish a production cross-tenant prompt leak. Research/disclosure claims are specific to the documented settings; do not infer unreported victims or additional compromise.
- Source role
- primary disclosure
Evidence record
GPU KV-cache allocations can be valid at the device-memory level while individual inference requests read uninitialized or cross-request cache regions. A request-level sanitizer annotates ownership and initialization state. KVGuard reports finding 29 historic vLLM/SGLang violations and eight new cases under its harness, with measured token-generation overhead. A detected memory-safety condition does not establish a production cross-tenant prompt leak. Research/disclosure claims are specific to the documented settings; do not infer unreported victims or additional compromise.
Claim-level citations (3)
- supportsKVGuard identifies request-level GPU cache ownership failures missed by allocation checks: KVGuard reports finding 29 historic vLLM/SGLang violations and eight new cases under its harness, with measured token-generation overhead. A detected memory-safety condition does not establish a production cross-tenant prompt leak.
KVGuard design; evaluation of 29 historic and eight new cases
- supportsKVGuard identifies request-level GPU cache ownership failures missed by allocation checks: GPU KV-cache allocations can be valid at the device-memory level while individual inference requests read uninitialized or cross-request cache regions. A request-level sanitizer annotates ownership and initialization state.
KVGuard design; evaluation of 29 historic and eight new cases
- supportsKVGuard identifies request-level GPU cache ownership failures missed by allocation checks: KVGuard reported 29 previously known and eight additional request-level KV-cache violations in vLLM and SGLang, with measured per-token overhead of about 3.0% and 7.4% in selected Qwen2.5-14B configurations. A violation finding does not itself prove a cross-tenant leak.
KVGuard experiments: bug coverage, memory overhead and serving evaluation