{
  "format": 1,
  "kind": "vllm-memex-kv-connector-3b",
  "generatedAt": "2026-10-30T14:25:00+08:00",
  "hardware": {
    "gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
    "memoryMiB": 97887,
    "usedMiB": 88933
  },
  "software": {
    "vllm": "0.10.2",
    "torch": "2.8.0+cu128",
    "transformers": "4.57.1"
  },
  "model": "Qwen/Qwen2.5-3B-Instruct",
  "connector": {
    "class": "MemexKVConnector",
    "prefixTokens": 1500,
    "alignedPrefixTokens": 1488,
    "nativePrefixCaching": false,
    "layers": 36
  },
  "workload": {
    "prefixes": 8,
    "requests": 16,
    "promptTokens": 2815
  },
  "results": {
    "externalCacheHits": 8,
    "diskMiB": 420,
    "meanStoreAll": 8.3617,
    "meanHit": 0.0941,
    "warmupStoreSeconds": 65.8642,
    "steadyMeanStore": 0.147,
    "steadyMeanHit": 0.0937,
    "pairs": [
      {
        "prefix": 0,
        "store_seconds": 65.8642,
        "hit_seconds": 0.0964,
        "prompt_tokens": 2815
      },
      {
        "prefix": 1,
        "store_seconds": 0.1553,
        "hit_seconds": 0.0916,
        "prompt_tokens": 2815
      },
      {
        "prefix": 2,
        "store_seconds": 0.1486,
        "hit_seconds": 0.0971,
        "prompt_tokens": 2815
      },
      {
        "prefix": 3,
        "store_seconds": 0.1481,
        "hit_seconds": 0.0988,
        "prompt_tokens": 2815
      },
      {
        "prefix": 4,
        "store_seconds": 0.1456,
        "hit_seconds": 0.0935,
        "prompt_tokens": 2815
      },
      {
        "prefix": 5,
        "store_seconds": 0.1445,
        "hit_seconds": 0.0918,
        "prompt_tokens": 2815
      },
      {
        "prefix": 6,
        "store_seconds": 0.1424,
        "hit_seconds": 0.0919,
        "prompt_tokens": 2815
      },
      {
        "prefix": 7,
        "store_seconds": 0.1447,
        "hit_seconds": 0.0914,
        "prompt_tokens": 2815
      }
    ]
  },
  "caveats": [
    "First request includes torch.compile/warmup (65.9s); steady-state means exclude it.",
    "Prototype connector, fixed prefix length, disk safetensors, no eviction/compression.",
    "Single process, 3B model; not a production throughput benchmark."
  ]
}
