{
  "_comment": [
    "多端口/多模型压测配置。defaults 里的键就是命令行参数名（横线换成下划线都行）。",
    "target 里的同名键会覆盖 defaults。cases 支持两种写法：",
    "  字符串  \"输入:输出:并发[:条数]\"，例 \"10k:50k:4\"",
    "  对象    {\"in\":\"10k\",\"out\":\"50k\",\"concurrency\":[1,4,8],\"rounds\":1}",
    "长度单位 K=1024；输入+输出 超过该 target 的 max_model_len 会按 on_overflow 处理。",
    "条数默认 = 并发 × rounds；长序列用例建议 rounds=1。"
  ],
  "defaults": {
    "api": "chat",
    "rounds": 4,
    "concurrency": [
      1,
      4,
      8,
      16
    ],
    "warmup": 2,
    "sleep_between": 5,
    "progress_interval": 30,
    "stall_timeout": 600,
    "case_timeout": 0,
    "on_overflow": "skip",
    "hardware": "",
    "version": "",
    "dtype": "",
    "owner": "vLLM"
  },
  "targets": [
    {
      "name": "Qwen3.6-27B @8000",
      "base_url": "http://127.0.0.1:8000/v1",
      "model": "",
      "model_label": "Qwen3.6-27B",
      "max_model_len": 131072,
      "tp": 1,
      "chips": 1,
      "cases": [
        {
          "in": 256,
          "out": 256
        },
        {
          "in": 512,
          "out": 256
        },
        {
          "in": "1k",
          "out": "1k"
        },
        {
          "in": "2k",
          "out": "1k"
        },
        {
          "in": "4k",
          "out": "1k"
        },
        {
          "in": "10k",
          "out": "50k",
          "concurrency": [
            1,
            4,
            8
          ],
          "rounds": 1
        },
        {
          "in": "32k",
          "out": "32k",
          "concurrency": [
            1,
            4
          ],
          "rounds": 1
        },
        {
          "in": "64k",
          "out": "16k",
          "concurrency": [
            1,
            2
          ],
          "rounds": 1
        }
      ]
    }
  ]
}