[
  {
    "run_id": 1,
    "timestamp": "2026-07-27 15:59:25",
    "model_name": "qwen3:8b",
    "target_tokens": 150,
    "actual_tokens": 150,
    "decode_tps": 1.39,
    "wall_time_sec": 106.65,
    "load_ms": 200.78,
    "prefill_tps": 33.11,
    "note": "Early baseline before GPU-tuned model routing"
  },
  {
    "run_id": 2,
    "timestamp": "2026-07-27 16:20:48",
    "model_name": "qwen-gpu:latest",
    "target_tokens": 150,
    "actual_tokens": 150,
    "decode_tps": 29.68,
    "wall_time_sec": 11.23,
    "load_ms": 331.08,
    "prefill_tps": 92.47,
    "note": "First GPU-routed run"
  },
  {
    "run_id": 5,
    "timestamp": "2026-07-27 16:27:12",
    "model_name": "qwen-gpu:latest",
    "target_tokens": 150,
    "actual_tokens": 150,
    "decode_tps": 37.47,
    "wall_time_sec": 5.03,
    "load_ms": 243.59,
    "prefill_tps": 110.16,
    "note": "Stabilized GPU decode after routing/tuning"
  }
]