kvcachescope uncovers hidden vLLM KV cache leaks that nvidia-smi misses
kvcachescope plugs into vLLM’s BlockSpaceManager to expose logical KV‑cache allocations that standard GPU tools ignore. It visualizes block‑to‑token mappings, flags orphaned sequences, and lets CI pipelines catch silent memory leaks, saving GPU costs and avoiding out‑of‑memory crashes.