{
  "format": 1,
  "kind": "vllm-prefix-cache-validation",
  "generatedAt": "2026-10-29T15:55:38.647691+00:00",
  "route": "A (local vLLM paged-KV prefix cache)",
  "hardware": {
    "gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
    "memoryMiB": 97887,
    "driver": "590.44.01",
    "cpuCount": 208,
    "systemMemory": "1.0TiB"
  },
  "software": {
    "python": "3.12.3",
    "torch": "2.8.0+cu128",
    "cuda": "12.8",
    "computeCapability": [
      12,
      0
    ],
    "vllm": "0.10.2",
    "transformers": "4.57.1"
  },
  "server": {
    "model": "Qwen/Qwen2.5-0.5B-Instruct",
    "servedModelName": "qwen",
    "maxModelLen": 4096,
    "enablePrefixCaching": true
  },
  "experiment": {
    "description": "Two requests share a long stable prefix (~2498 tokens) and differ only in the final question.",
    "requests": [
      {
        "promptTokens": 2498,
        "completionTokens": 8,
        "suffix": "Question one: what did we decide about component 7?"
      },
      {
        "promptTokens": 2499,
        "completionTokens": 8,
        "suffix": "Question two: what did we decide about component 42?"
      }
    ]
  },
  "metrics": {
    "prefixCacheQueriesTokens": 4997,
    "prefixCacheHitsTokens": 2480,
    "overallHitRatio": 0.4963,
    "secondRequestPrefixHitRatio": 0.9924
  },
  "interpretation": "vLLM reused the materialized stable prefix through its native paged-KV prefix cache on the second request (2480 cached prompt tokens). This validates the stable-prefix contract of the materializer against a real inference runtime.",
  "caveats": [
    "Small 0.5B model and one synthetic prefix; not a production workload.",
    "This is vLLM-native prefix caching, not an external paged-KV connector with custom block paging.",
    "Single run on one GPU; no statistical claims.",
    "Server credentials and host details are intentionally not stored in the repository."
  ]
}