{
  "format": 1,
  "kind": "vllm-memex-kv-connector-multiprefix",
  "generatedAt": "2026-10-30T14:10:00+08:00",
  "hardware": {
    "gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
    "memoryMiB": 97887
  },
  "software": {
    "vllm": "0.10.2",
    "torch": "2.8.0+cu128",
    "transformers": "4.57.1"
  },
  "model": "Qwen/Qwen2.5-0.5B-Instruct",
  "connector": {
    "class": "MemexKVConnector",
    "prefixTokens": 1500,
    "alignedPrefixTokens": 1488,
    "nativePrefixCaching": false,
    "storage": "/root/memex-kv"
  },
  "workload": {
    "prefixes": 8,
    "requests": 16,
    "promptTokens": 2815,
    "note": "two requests per prefix; second differs only in final question"
  },
  "results": {
    "requests": 16,
    "pairs": [
      {
        "prefix": 0,
        "store_seconds": 0.0812,
        "hit_seconds": 0.0405,
        "prompt_tokens": 2815
      },
      {
        "prefix": 1,
        "store_seconds": 0.0595,
        "hit_seconds": 0.0374,
        "prompt_tokens": 2815
      },
      {
        "prefix": 2,
        "store_seconds": 0.0584,
        "hit_seconds": 0.0384,
        "prompt_tokens": 2815
      },
      {
        "prefix": 3,
        "store_seconds": 0.0567,
        "hit_seconds": 0.0388,
        "prompt_tokens": 2815
      },
      {
        "prefix": 4,
        "store_seconds": 0.0561,
        "hit_seconds": 0.0364,
        "prompt_tokens": 2815
      },
      {
        "prefix": 5,
        "store_seconds": 0.0615,
        "hit_seconds": 0.0392,
        "prompt_tokens": 2815
      },
      {
        "prefix": 6,
        "store_seconds": 0.0602,
        "hit_seconds": 0.0418,
        "prompt_tokens": 2815
      },
      {
        "prefix": 7,
        "store_seconds": 0.0601,
        "hit_seconds": 0.0394,
        "prompt_tokens": 2815
      }
    ],
    "mean_store": 0.0617,
    "mean_hit": 0.039
  },
  "storage": {
    "diskMiB": 141,
    "prefixFolders": 8,
    "kvFilesPerPrefix": 24
  },
  "verification": {
    "externalCacheHitsObserved": 8,
    "secondRequestHitRate": 1.0
  },
  "caveats": [
    "Prototype connector adapted from vLLM SharedStorageConnector; fixed prefix length.",
    "0.5B model, synthetic prefixes, single GPU/process.",
    "Disk safetensors; no eviction/compression beyond directory deletion.",
    "Latency includes API overhead and is not a production benchmark."
  ]
}
