{
  "format": 1,
  "kind": "vllm-memex-kv-connector-validation",
  "generatedAt": "2026-10-30T06:00:29.281513+00:00",
  "route": "A (custom vLLM KV connector, external disk KV)",
  "hardware": {
    "gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
    "memoryMiB": 97887,
    "driver": "590.44.01"
  },
  "software": {
    "torch": "2.8.0+cu128",
    "cuda": "12.8",
    "vllm": "0.10.2",
    "transformers": "4.57.1"
  },
  "model": "Qwen/Qwen2.5-0.5B-Instruct (local /root/models/qwen)",
  "connector": {
    "class": "MemexKVConnector",
    "module": "memex_kv_connector",
    "storage": "/root/memex-kv",
    "prefixTokens": 1500,
    "alignedPrefixTokens": 1488,
    "layers": 24,
    "blockSize": 16
  },
  "experiment": {
    "description": "Two requests share the same materialized prefix and differ only in the final question; prefix caching disabled so the connector is the only external KV path.",
    "requests": [
      {
        "promptTokens": 2498,
        "completionTokens": 8,
        "result": "store 1,488 KV tokens per layer to disk"
      },
      {
        "promptTokens": 2499,
        "completionTokens": 8,
        "result": "external cache hit; inject 1,488 KV tokens"
      }
    ],
    "logs": [
      "MEMEX store KV tokens=1488 layer=model.layers.0..23",
      "MEMEX EXTERNAL CACHE HIT prefix_tokens=1500",
      "MEMEX inject KV tokens=1488"
    ]
  },
  "storage": {
    "kvFiles": 24,
    "diskMiB": 29
  },
  "verification": {
    "externalCacheHitLog": true,
    "injectedTokens": 1488,
    "nativePrefixCaching": false
  },
  "caveats": [
    "Prototype connector adapted from vLLM SharedStorageConnector; fixed prefix length configured explicitly.",
    "0.5B model and synthetic materialized prefix; not a production workload.",
    "KV is stored as safetensors on local disk, not in a Memex memory pool; block-level eviction/compression not yet implemented.",
    "Single request pair; no concurrency/preemption tests."
  ]
}