{
  "output_tokens": 256,
  "thinking_fields": {
    "chat_template_kwargs": {
      "enable_thinking": false
    }
  },
  "tokens_per_word": 1.028,
  "fixed_tokens": 92.0,
  "samples": [
    {
      "kind": "cold",
      "target_tokens": 1024,
      "prompt_tokens": 901,
      "completion_tokens": 256,
      "ttft_s": 1.276,
      "total_s": 8.53,
      "prefill_tps": 706.1,
      "decode_tps": 35.2,
      "finish_reason": "length",
      "error": null
    },
    {
      "kind": "warm",
      "target_tokens": 1024,
      "prompt_tokens": 1017,
      "completion_tokens": 256,
      "ttft_s": 1.899,
      "total_s": 4.35,
      "prefill_tps": 535.5,
      "decode_tps": 105.6,
      "finish_reason": "length",
      "error": null
    },
    {
      "kind": "warm",
      "target_tokens": 8192,
      "prompt_tokens": 8200,
      "completion_tokens": 256,
      "ttft_s": 4.837,
      "total_s": 7.789,
      "prefill_tps": 1695.3,
      "decode_tps": 87.8,
      "finish_reason": "length",
      "error": null
    },
    {
      "kind": "warm",
      "target_tokens": 32768,
      "prompt_tokens": 32856,
      "completion_tokens": 256,
      "ttft_s": 12.114,
      "total_s": 15.009,
      "prefill_tps": 2712.2,
      "decode_tps": 89.1,
      "finish_reason": "length",
      "error": null
    }
  ],
  "skipped_lengths": [],
  "summary": {
    "cold": {
      "n": 1.0,
      "prompt_tokens": 901,
      "ttft_s": 1.276,
      "prefill_tps": 706.1,
      "decode_tps": 35.2
    },
    "warm_1024": {
      "n": 1.0,
      "prompt_tokens": 1017,
      "ttft_s": 1.899,
      "prefill_tps": 535.5,
      "decode_tps": 105.6
    },
    "warm_8192": {
      "n": 1.0,
      "prompt_tokens": 8200,
      "ttft_s": 4.837,
      "prefill_tps": 1695.3,
      "decode_tps": 87.8
    },
    "warm_32768": {
      "n": 1.0,
      "prompt_tokens": 32856,
      "ttft_s": 12.114,
      "prefill_tps": 2712.2,
      "decode_tps": 89.1
    }
  },
  "memory": {
    "gpu_used_mib_start": 12,
    "gpu_used_mib_peak": 12,
    "ram_used_mib_start": 12046,
    "ram_used_mib_peak": 12088,
    "note": "system-wide numbers, not only the server process"
  }
}