vLLM Mooncake Connector GPU KV Cache Exhaustion

Executive Summary

This vulnerability involves the vLLM Mooncake connector through version 0.29.0 failing to manage GPU KV cache block ownership correctly in prefill/decode disaggregated deployments. Attackers can exploit this by submitting multiple completion requests with different prompts sharing a single transfer ID, causing orphaned KV cache blocks to accumulate and exhaust GPU memory until the process restarts.

Useful 0 Not Useful 0

Detection Guidance

Monitor GPU memory usage for unexpected spikes or exhaustion during concurrent requests. Check for orphaned KV cache blocks by inspecting vLLM Mooncake connector logs for repeated transfer ID reuse without proper cleanup. Look for processes failing to execute due to memory constraints.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability can lead to GPU memory exhaustion, preventing legitimate requests from executing. Systems relying on vLLM Mooncake connector for GPU-accelerated tasks may experience crashes, degraded performance, or denial of service due to memory overload.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability could lead to denial-of-service conditions by exhausting GPU memory, potentially disrupting service availability for legitimate users. For GDPR, this may impact the right to access or rectify data if services become unavailable. Under HIPAA, it could affect the availability of protected health information systems, though no direct data exposure is indicated.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade vLLM Mooncake connector to a version that properly manages GPU KV cache block ownership. Implement rate limiting to prevent excessive concurrent requests. Restrict access to prefill/decode disaggregated deployments to trusted users only.

Useful 0 Not Useful 0