[
 {
  "metadata": {
   "timestamp": "2026-07-27 18:02:16",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "bloom-1b7",
   "hf_model_id": "bigscience/bloom-1b7",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-1b7",
   "param_count": 1722408960
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 3.198,
   "TTFT": 0.556,
   "TBT_avg": 0.377,
   "Throughput": 2.502
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 3197.729,
     "ttft": 555.985,
     "tbt_avg": 377.392,
     "throughput": 2.502,
     "steps": [
      555.985,
      427.194,
      364.233,
      363.553,
      383.166,
      378.944,
      362.682,
      361.972
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 3197.729,
     "min": 3197.729,
     "max": 3197.729,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 555.985,
     "min": 555.985,
     "max": 555.985,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 377.392,
     "min": 377.392,
     "max": 377.392,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 2.502,
     "min": 2.502,
     "max": 2.502,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 377.392,
     "min": 361.972,
     "max": 427.194,
     "std": 21.827,
     "cv_pct": 5.784
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 3153.646,
   "total_nodes": 2168,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 3153.646,
     "pct": 100.0,
     "nodes": 2168,
     "calls": 17344
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 3065.833,
     "pct": 97.216,
     "by_device": {
      "CPU": {
       "time_ms": 3065.833,
       "nodes": 145,
       "calls": 1160
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 17.328,
     "pct": 0.549,
     "by_device": {
      "CPU": {
       "time_ms": 17.328,
       "nodes": 245,
       "calls": 1960
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 12.954,
     "pct": 0.411,
     "by_device": {
      "CPU": {
       "time_ms": 12.954,
       "nodes": 50,
       "calls": 400
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 12.306,
     "pct": 0.39,
     "by_device": {
      "CPU": {
       "time_ms": 12.306,
       "nodes": 245,
       "calls": 1960
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 10.511,
     "pct": 0.333,
     "by_device": {
      "CPU": {
       "time_ms": 10.511,
       "nodes": 225,
       "calls": 1800
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 8.777,
     "pct": 0.278,
     "by_device": {
      "CPU": {
       "time_ms": 8.777,
       "nodes": 354,
       "calls": 2832
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 6.74,
     "pct": 0.214,
     "by_device": {
      "CPU": {
       "time_ms": 6.74,
       "nodes": 250,
       "calls": 2000
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 5.821,
     "pct": 0.185,
     "by_device": {
      "CPU": {
       "time_ms": 5.821,
       "nodes": 120,
       "calls": 960
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 4.258,
     "pct": 0.135,
     "by_device": {
      "CPU": {
       "time_ms": 4.258,
       "nodes": 199,
       "calls": 1592
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.727,
     "pct": 0.086,
     "by_device": {
      "CPU": {
       "time_ms": 2.727,
       "nodes": 185,
       "calls": 1480
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 1.723,
     "pct": 0.055,
     "by_device": {
      "CPU": {
       "time_ms": 1.723,
       "nodes": 30,
       "calls": 240
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 1.714,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 1.714,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.066,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 1.066,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 0.614,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.614,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.252,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.252,
       "nodes": 13,
       "calls": 104
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.247,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.247,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.198,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.198,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.162,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.162,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.133,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.133,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.107,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.107,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.049,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.049,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.034,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "CumSum",
     "total_ms": 0.034,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sub",
     "total_ms": 0.031,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.031,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.027,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_bloom-1b7_cpu.simplified.json",
   "source_profile": "ort_bloom-1b7_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:54:39",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "bloom-560m",
   "hf_model_id": "bigscience/bloom-560m",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-560m",
   "param_count": 559214592
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.049,
   "TTFT": 0.177,
   "TBT_avg": 0.125,
   "Throughput": 7.626
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1049.102,
     "ttft": 177.23,
     "tbt_avg": 124.553,
     "throughput": 7.626,
     "steps": [
      177.23,
      124.841,
      124.476,
      124.539,
      124.214,
      124.244,
      125.051,
      124.507
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1049.102,
     "min": 1049.102,
     "max": 1049.102,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 177.23,
     "min": 177.23,
     "max": 177.23,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 124.553,
     "min": 124.553,
     "max": 124.553,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 7.626,
     "min": 7.626,
     "max": 7.626,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 124.553,
     "min": 124.214,
     "max": 125.051,
     "std": 0.28,
     "cv_pct": 0.225
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1007.498,
   "total_nodes": 2168,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1007.498,
     "pct": 100.0,
     "nodes": 2168,
     "calls": 17344
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 940.835,
     "pct": 93.383,
     "by_device": {
      "CPU": {
       "time_ms": 940.835,
       "nodes": 145,
       "calls": 1160
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 12.481,
     "pct": 1.239,
     "by_device": {
      "CPU": {
       "time_ms": 12.481,
       "nodes": 245,
       "calls": 1960
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 9.286,
     "pct": 0.922,
     "by_device": {
      "CPU": {
       "time_ms": 9.286,
       "nodes": 245,
       "calls": 1960
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 7.91,
     "pct": 0.785,
     "by_device": {
      "CPU": {
       "time_ms": 7.91,
       "nodes": 225,
       "calls": 1800
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 7.878,
     "pct": 0.782,
     "by_device": {
      "CPU": {
       "time_ms": 7.878,
       "nodes": 354,
       "calls": 2832
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 7.248,
     "pct": 0.719,
     "by_device": {
      "CPU": {
       "time_ms": 7.248,
       "nodes": 50,
       "calls": 400
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 6.002,
     "pct": 0.596,
     "by_device": {
      "CPU": {
       "time_ms": 6.002,
       "nodes": 250,
       "calls": 2000
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.002,
     "pct": 0.397,
     "by_device": {
      "CPU": {
       "time_ms": 4.002,
       "nodes": 120,
       "calls": 960
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 3.929,
     "pct": 0.39,
     "by_device": {
      "CPU": {
       "time_ms": 3.929,
       "nodes": 199,
       "calls": 1592
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.651,
     "pct": 0.263,
     "by_device": {
      "CPU": {
       "time_ms": 2.651,
       "nodes": 185,
       "calls": 1480
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 1.446,
     "pct": 0.144,
     "by_device": {
      "CPU": {
       "time_ms": 1.446,
       "nodes": 30,
       "calls": 240
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 1.052,
     "pct": 0.104,
     "by_device": {
      "CPU": {
       "time_ms": 1.052,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.038,
     "pct": 0.103,
     "by_device": {
      "CPU": {
       "time_ms": 1.038,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 0.604,
     "pct": 0.06,
     "by_device": {
      "CPU": {
       "time_ms": 0.604,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.224,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.224,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.218,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.218,
       "nodes": 13,
       "calls": 104
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.167,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.167,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.147,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.147,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.121,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.121,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.104,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.104,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.047,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.047,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.032,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "CumSum",
     "total_ms": 0.027,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sub",
     "total_ms": 0.025,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.024,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_bloom-560m_cpu.simplified.json",
   "source_profile": "ort_bloom-560m_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:57:00",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "Falcon3-1B-Base",
   "hf_model_id": "tiiuae/Falcon3-1B-Base",
   "hf_model_url": "https://huggingface.co/tiiuae/Falcon3-1B-Base",
   "param_count": 1669000000
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.637,
   "TTFT": 0.472,
   "TBT_avg": 0.309,
   "Throughput": 3.033
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2637.372,
     "ttft": 472.051,
     "tbt_avg": 309.332,
     "throughput": 3.033,
     "steps": [
      472.051,
      298.451,
      336.034,
      299.452,
      298.077,
      299.371,
      330.834,
      303.102
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2637.372,
     "min": 2637.372,
     "max": 2637.372,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 472.051,
     "min": 472.051,
     "max": 472.051,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 309.332,
     "min": 309.332,
     "max": 309.332,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 3.033,
     "min": 3.033,
     "max": 3.033,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 309.332,
     "min": 298.077,
     "max": 336.034,
     "std": 15.381,
     "cv_pct": 4.972
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2583.124,
   "total_nodes": 2790,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 2583.124,
     "pct": 100.0,
     "nodes": 2790,
     "calls": 22320
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 2507.157,
     "pct": 97.059,
     "by_device": {
      "CPU": {
       "time_ms": 2507.157,
       "nodes": 164,
       "calls": 1312
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 11.96,
     "pct": 0.463,
     "by_device": {
      "CPU": {
       "time_ms": 11.96,
       "nodes": 574,
       "calls": 4592
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 10.655,
     "pct": 0.412,
     "by_device": {
      "CPU": {
       "time_ms": 10.655,
       "nodes": 220,
       "calls": 1760
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 9.884,
     "pct": 0.383,
     "by_device": {
      "CPU": {
       "time_ms": 9.884,
       "nodes": 295,
       "calls": 2360
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 5.868,
     "pct": 0.227,
     "by_device": {
      "CPU": {
       "time_ms": 5.868,
       "nodes": 242,
       "calls": 1936
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.848,
     "pct": 0.188,
     "by_device": {
      "CPU": {
       "time_ms": 4.848,
       "nodes": 91,
       "calls": 728
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 4.328,
     "pct": 0.168,
     "by_device": {
      "CPU": {
       "time_ms": 4.328,
       "nodes": 91,
       "calls": 728
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 4.167,
     "pct": 0.161,
     "by_device": {
      "CPU": {
       "time_ms": 4.167,
       "nodes": 130,
       "calls": 1040
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 3.473,
     "pct": 0.134,
     "by_device": {
      "CPU": {
       "time_ms": 3.473,
       "nodes": 279,
       "calls": 2232
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 3.328,
     "pct": 0.129,
     "by_device": {
      "CPU": {
       "time_ms": 3.328,
       "nodes": 148,
       "calls": 1184
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 3.104,
     "pct": 0.12,
     "by_device": {
      "CPU": {
       "time_ms": 3.104,
       "nodes": 38,
       "calls": 304
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 2.94,
     "pct": 0.114,
     "by_device": {
      "CPU": {
       "time_ms": 2.94,
       "nodes": 74,
       "calls": 592
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.887,
     "pct": 0.073,
     "by_device": {
      "CPU": {
       "time_ms": 1.887,
       "nodes": 73,
       "calls": 584
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 1.704,
     "pct": 0.066,
     "by_device": {
      "CPU": {
       "time_ms": 1.704,
       "nodes": 84,
       "calls": 672
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 1.546,
     "pct": 0.06,
     "by_device": {
      "CPU": {
       "time_ms": 1.546,
       "nodes": 37,
       "calls": 296
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.492,
     "pct": 0.058,
     "by_device": {
      "CPU": {
       "time_ms": 1.492,
       "nodes": 18,
       "calls": 144
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 1.146,
     "pct": 0.044,
     "by_device": {
      "CPU": {
       "time_ms": 1.146,
       "nodes": 37,
       "calls": 296
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.74,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.74,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.65,
     "pct": 0.025,
     "by_device": {
      "CPU": {
       "time_ms": 0.65,
       "nodes": 38,
       "calls": 304
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.602,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.602,
       "nodes": 37,
       "calls": 296
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.553,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.553,
       "nodes": 39,
       "calls": 312
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.528,
     "pct": 0.02,
     "by_device": {
      "CPU": {
       "time_ms": 0.528,
       "nodes": 18,
       "calls": 144
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.297,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.297,
       "nodes": 18,
       "calls": 144
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.106,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.106,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.055,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.055,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.034,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.028,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.027,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.017,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_falcon3-1b_cpu.simplified.json",
   "source_profile": "ort_falcon3-1b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 18:04:16",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "gemma-2-2b",
   "hf_model_id": "google/gemma-2-2b",
   "hf_model_url": "https://huggingface.co/google/gemma-2-2b",
   "param_count": 2614341888
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 4.872,
   "TTFT": 0.868,
   "TBT_avg": 0.572,
   "Throughput": 1.642
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 4872.102,
     "ttft": 868.461,
     "tbt_avg": 571.949,
     "throughput": 1.642,
     "steps": [
      868.461,
      579.185,
      567.167,
      579.165,
      554.927,
      589.429,
      554.875,
      578.893
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 4872.102,
     "min": 4872.102,
     "max": 4872.102,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 868.461,
     "min": 868.461,
     "max": 868.461,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 571.949,
     "min": 571.949,
     "max": 571.949,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 1.642,
     "min": 1.642,
     "max": 1.642,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 571.949,
     "min": 554.875,
     "max": 589.429,
     "std": 12.319,
     "cv_pct": 2.154
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 4782.425,
   "total_nodes": 4558,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 4782.425,
     "pct": 100.0,
     "nodes": 4558,
     "calls": 36464
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 4647.622,
     "pct": 97.181,
     "by_device": {
      "CPU": {
       "time_ms": 4647.622,
       "nodes": 236,
       "calls": 1888
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 25.503,
     "pct": 0.533,
     "by_device": {
      "CPU": {
       "time_ms": 25.503,
       "nodes": 660,
       "calls": 5280
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 18.362,
     "pct": 0.384,
     "by_device": {
      "CPU": {
       "time_ms": 18.362,
       "nodes": 825,
       "calls": 6600
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 16.681,
     "pct": 0.349,
     "by_device": {
      "CPU": {
       "time_ms": 16.681,
       "nodes": 317,
       "calls": 2536
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 9.119,
     "pct": 0.191,
     "by_device": {
      "CPU": {
       "time_ms": 9.119,
       "nodes": 290,
       "calls": 2320
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 8.789,
     "pct": 0.184,
     "by_device": {
      "CPU": {
       "time_ms": 8.789,
       "nodes": 346,
       "calls": 2768
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 6.724,
     "pct": 0.141,
     "by_device": {
      "CPU": {
       "time_ms": 6.724,
       "nodes": 131,
       "calls": 1048
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 6.679,
     "pct": 0.14,
     "by_device": {
      "CPU": {
       "time_ms": 6.679,
       "nodes": 131,
       "calls": 1048
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 5.156,
     "pct": 0.108,
     "by_device": {
      "CPU": {
       "time_ms": 5.156,
       "nodes": 400,
       "calls": 3200
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 5.026,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 5.026,
       "nodes": 213,
       "calls": 1704
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 4.974,
     "pct": 0.104,
     "by_device": {
      "CPU": {
       "time_ms": 4.974,
       "nodes": 158,
       "calls": 1264
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 4.558,
     "pct": 0.095,
     "by_device": {
      "CPU": {
       "time_ms": 4.558,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 4.333,
     "pct": 0.091,
     "by_device": {
      "CPU": {
       "time_ms": 4.333,
       "nodes": 107,
       "calls": 856
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 3.759,
     "pct": 0.079,
     "by_device": {
      "CPU": {
       "time_ms": 3.759,
       "nodes": 55,
       "calls": 440
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 3.307,
     "pct": 0.069,
     "by_device": {
      "CPU": {
       "time_ms": 3.307,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 2.921,
     "pct": 0.061,
     "by_device": {
      "CPU": {
       "time_ms": 2.921,
       "nodes": 117,
       "calls": 936
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 2.493,
     "pct": 0.052,
     "by_device": {
      "CPU": {
       "time_ms": 2.493,
       "nodes": 27,
       "calls": 216
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 1.701,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 1.701,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 1.148,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 1.148,
       "nodes": 52,
       "calls": 416
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.992,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.992,
       "nodes": 55,
       "calls": 440
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.883,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.883,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.814,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.814,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.432,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.432,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.16,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.16,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.109,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.109,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Sub",
     "total_ms": 0.037,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.037,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.035,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.028,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.023,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.02,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_gemma2-2b_cpu.simplified.json",
   "source_profile": "ort_gemma2-2b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:56:35",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "gemma-3-1b-it",
   "hf_model_id": "google/gemma-3-1b-it",
   "hf_model_url": "https://huggingface.co/google/gemma-3-1b-it",
   "param_count": 999885952
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.021,
   "TTFT": 0.342,
   "TBT_avg": 0.24,
   "Throughput": 3.958
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2021.182,
     "ttft": 341.754,
     "tbt_avg": 239.918,
     "throughput": 3.958,
     "steps": [
      341.754,
      267.358,
      229.006,
      228.687,
      228.57,
      228.616,
      269.105,
      228.086
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2021.182,
     "min": 2021.182,
     "max": 2021.182,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 341.754,
     "min": 341.754,
     "max": 341.754,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 239.918,
     "min": 239.918,
     "max": 239.918,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 3.958,
     "min": 3.958,
     "max": 3.958,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 239.918,
     "min": 228.086,
     "max": 269.105,
     "std": 17.915,
     "cv_pct": 7.467
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1925.609,
   "total_nodes": 4991,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1925.609,
     "pct": 100.0,
     "nodes": 4991,
     "calls": 39928
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 1802.225,
     "pct": 93.592,
     "by_device": {
      "CPU": {
       "time_ms": 1802.225,
       "nodes": 237,
       "calls": 1896
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 23.363,
     "pct": 1.213,
     "by_device": {
      "CPU": {
       "time_ms": 23.363,
       "nodes": 766,
       "calls": 6128
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 16.824,
     "pct": 0.874,
     "by_device": {
      "CPU": {
       "time_ms": 16.824,
       "nodes": 828,
       "calls": 6624
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 11.911,
     "pct": 0.619,
     "by_device": {
      "CPU": {
       "time_ms": 11.911,
       "nodes": 319,
       "calls": 2552
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 9.288,
     "pct": 0.482,
     "by_device": {
      "CPU": {
       "time_ms": 9.288,
       "nodes": 342,
       "calls": 2736
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 8.354,
     "pct": 0.434,
     "by_device": {
      "CPU": {
       "time_ms": 8.354,
       "nodes": 346,
       "calls": 2768
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 6.015,
     "pct": 0.312,
     "by_device": {
      "CPU": {
       "time_ms": 6.015,
       "nodes": 131,
       "calls": 1048
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 5.331,
     "pct": 0.277,
     "by_device": {
      "CPU": {
       "time_ms": 5.331,
       "nodes": 157,
       "calls": 1256
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 5.15,
     "pct": 0.267,
     "by_device": {
      "CPU": {
       "time_ms": 5.15,
       "nodes": 209,
       "calls": 1672
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 4.972,
     "pct": 0.258,
     "by_device": {
      "CPU": {
       "time_ms": 4.972,
       "nodes": 401,
       "calls": 3208
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 4.725,
     "pct": 0.245,
     "by_device": {
      "CPU": {
       "time_ms": 4.725,
       "nodes": 214,
       "calls": 1712
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 4.566,
     "pct": 0.237,
     "by_device": {
      "CPU": {
       "time_ms": 4.566,
       "nodes": 157,
       "calls": 1256
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.448,
     "pct": 0.231,
     "by_device": {
      "CPU": {
       "time_ms": 4.448,
       "nodes": 132,
       "calls": 1056
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 4.113,
     "pct": 0.214,
     "by_device": {
      "CPU": {
       "time_ms": 4.113,
       "nodes": 108,
       "calls": 864
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 3.559,
     "pct": 0.185,
     "by_device": {
      "CPU": {
       "time_ms": 3.559,
       "nodes": 171,
       "calls": 1368
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 2.589,
     "pct": 0.134,
     "by_device": {
      "CPU": {
       "time_ms": 2.589,
       "nodes": 157,
       "calls": 1256
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 2.335,
     "pct": 0.121,
     "by_device": {
      "CPU": {
       "time_ms": 2.335,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 1.544,
     "pct": 0.08,
     "by_device": {
      "CPU": {
       "time_ms": 1.544,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.974,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 0.974,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.946,
     "pct": 0.049,
     "by_device": {
      "CPU": {
       "time_ms": 0.946,
       "nodes": 52,
       "calls": 416
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.837,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.837,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.625,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.625,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.435,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.435,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.157,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.157,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.108,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.108,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.06,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.051,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.051,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sub",
     "total_ms": 0.034,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.03,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.023,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.017,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_gemma3-1b_cpu.simplified.json",
   "source_profile": "ort_gemma3-1b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:59:30",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "gpt2-large",
   "hf_model_id": "openai-community/gpt2-large",
   "hf_model_url": "https://huggingface.co/openai-community/gpt2-large",
   "param_count": 774030080
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.607,
   "TTFT": 0.319,
   "TBT_avg": 0.184,
   "Throughput": 4.977
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1607.367,
     "ttft": 319.425,
     "tbt_avg": 183.992,
     "throughput": 4.977,
     "steps": [
      319.425,
      189.48,
      181.815,
      180.486,
      180.459,
      183.165,
      182.211,
      190.326
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1607.367,
     "min": 1607.367,
     "max": 1607.367,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 319.425,
     "min": 319.425,
     "max": 319.425,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 183.992,
     "min": 183.992,
     "max": 183.992,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 4.977,
     "min": 4.977,
     "max": 4.977,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 183.992,
     "min": 180.459,
     "max": 190.326,
     "std": 3.847,
     "cv_pct": 2.091
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1523.548,
   "total_nodes": 4722,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1523.548,
     "pct": 100.0,
     "nodes": 4722,
     "calls": 37776
    }
   ],
   "ops": [
    {
     "op": "Gemm",
     "total_ms": 1257.805,
     "pct": 82.558,
     "by_device": {
      "CPU": {
       "time_ms": 1257.805,
       "nodes": 144,
       "calls": 1152
      }
     }
    },
    {
     "op": "MatMul",
     "total_ms": 134.937,
     "pct": 8.857,
     "by_device": {
      "CPU": {
       "time_ms": 134.937,
       "nodes": 73,
       "calls": 584
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 21.88,
     "pct": 1.436,
     "by_device": {
      "CPU": {
       "time_ms": 21.88,
       "nodes": 509,
       "calls": 4072
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 15.995,
     "pct": 1.05,
     "by_device": {
      "CPU": {
       "time_ms": 15.995,
       "nodes": 772,
       "calls": 6176
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 12.543,
     "pct": 0.823,
     "by_device": {
      "CPU": {
       "time_ms": 12.543,
       "nodes": 73,
       "calls": 584
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 11.821,
     "pct": 0.776,
     "by_device": {
      "CPU": {
       "time_ms": 11.821,
       "nodes": 550,
       "calls": 4400
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 10.052,
     "pct": 0.66,
     "by_device": {
      "CPU": {
       "time_ms": 10.052,
       "nodes": 218,
       "calls": 1744
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 8.692,
     "pct": 0.571,
     "by_device": {
      "CPU": {
       "time_ms": 8.692,
       "nodes": 438,
       "calls": 3504
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 8.488,
     "pct": 0.557,
     "by_device": {
      "CPU": {
       "time_ms": 8.488,
       "nodes": 730,
       "calls": 5840
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 8.281,
     "pct": 0.544,
     "by_device": {
      "CPU": {
       "time_ms": 8.281,
       "nodes": 218,
       "calls": 1744
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 6.964,
     "pct": 0.457,
     "by_device": {
      "CPU": {
       "time_ms": 6.964,
       "nodes": 180,
       "calls": 1440
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 6.241,
     "pct": 0.41,
     "by_device": {
      "CPU": {
       "time_ms": 6.241,
       "nodes": 184,
       "calls": 1472
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 3.459,
     "pct": 0.227,
     "by_device": {
      "CPU": {
       "time_ms": 3.459,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 3.37,
     "pct": 0.221,
     "by_device": {
      "CPU": {
       "time_ms": 3.37,
       "nodes": 73,
       "calls": 584
      }
     }
    },
    {
     "op": "Squeeze",
     "total_ms": 3.268,
     "pct": 0.214,
     "by_device": {
      "CPU": {
       "time_ms": 3.268,
       "nodes": 145,
       "calls": 1160
      }
     }
    },
    {
     "op": "Split",
     "total_ms": 2.084,
     "pct": 0.137,
     "by_device": {
      "CPU": {
       "time_ms": 2.084,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 1.754,
     "pct": 0.115,
     "by_device": {
      "CPU": {
       "time_ms": 1.754,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.593,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 1.593,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 1.353,
     "pct": 0.089,
     "by_device": {
      "CPU": {
       "time_ms": 1.353,
       "nodes": 108,
       "calls": 864
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 1.196,
     "pct": 0.079,
     "by_device": {
      "CPU": {
       "time_ms": 1.196,
       "nodes": 80,
       "calls": 640
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 0.849,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 0.849,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.608,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.608,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.105,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.105,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.061,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.061,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.037,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.037,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.029,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.023,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.023,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_gpt2-large_cpu.simplified.json",
   "source_profile": "ort_gpt2-large_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:57:39",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "Llama-3.2-1B",
   "hf_model_id": "meta-llama/Llama-3.2-1B",
   "hf_model_url": "https://huggingface.co/meta-llama/Llama-3.2-1B",
   "param_count": 1235814400
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.335,
   "TTFT": 0.416,
   "TBT_avg": 0.274,
   "Throughput": 3.425
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2335.461,
     "ttft": 415.872,
     "tbt_avg": 274.227,
     "throughput": 3.425,
     "steps": [
      415.872,
      263.167,
      300.102,
      263.843,
      264.203,
      263.945,
      264.267,
      300.062
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2335.461,
     "min": 2335.461,
     "max": 2335.461,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 415.872,
     "min": 415.872,
     "max": 415.872,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 274.227,
     "min": 274.227,
     "max": 274.227,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 3.425,
     "min": 3.425,
     "max": 3.425,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 274.227,
     "min": 263.167,
     "max": 300.102,
     "std": 16.355,
     "cv_pct": 5.964
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2287.546,
   "total_nodes": 2490,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 2287.546,
     "pct": 100.0,
     "nodes": 2490,
     "calls": 19920
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 2217.983,
     "pct": 96.959,
     "by_device": {
      "CPU": {
       "time_ms": 2217.983,
       "nodes": 146,
       "calls": 1168
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 10.61,
     "pct": 0.464,
     "by_device": {
      "CPU": {
       "time_ms": 10.61,
       "nodes": 512,
       "calls": 4096
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 9.085,
     "pct": 0.397,
     "by_device": {
      "CPU": {
       "time_ms": 9.085,
       "nodes": 263,
       "calls": 2104
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 8.564,
     "pct": 0.374,
     "by_device": {
      "CPU": {
       "time_ms": 8.564,
       "nodes": 196,
       "calls": 1568
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 5.932,
     "pct": 0.259,
     "by_device": {
      "CPU": {
       "time_ms": 5.932,
       "nodes": 216,
       "calls": 1728
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.768,
     "pct": 0.208,
     "by_device": {
      "CPU": {
       "time_ms": 4.768,
       "nodes": 81,
       "calls": 648
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 3.936,
     "pct": 0.172,
     "by_device": {
      "CPU": {
       "time_ms": 3.936,
       "nodes": 81,
       "calls": 648
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 3.608,
     "pct": 0.158,
     "by_device": {
      "CPU": {
       "time_ms": 3.608,
       "nodes": 116,
       "calls": 928
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 3.18,
     "pct": 0.139,
     "by_device": {
      "CPU": {
       "time_ms": 3.18,
       "nodes": 34,
       "calls": 272
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 3.149,
     "pct": 0.138,
     "by_device": {
      "CPU": {
       "time_ms": 3.149,
       "nodes": 249,
       "calls": 1992
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.995,
     "pct": 0.131,
     "by_device": {
      "CPU": {
       "time_ms": 2.995,
       "nodes": 132,
       "calls": 1056
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 2.777,
     "pct": 0.121,
     "by_device": {
      "CPU": {
       "time_ms": 2.777,
       "nodes": 66,
       "calls": 528
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.668,
     "pct": 0.073,
     "by_device": {
      "CPU": {
       "time_ms": 1.668,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.517,
     "pct": 0.066,
     "by_device": {
      "CPU": {
       "time_ms": 1.517,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 1.486,
     "pct": 0.065,
     "by_device": {
      "CPU": {
       "time_ms": 1.486,
       "nodes": 76,
       "calls": 608
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 1.361,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 1.361,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 1.07,
     "pct": 0.047,
     "by_device": {
      "CPU": {
       "time_ms": 1.07,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.999,
     "pct": 0.044,
     "by_device": {
      "CPU": {
       "time_ms": 0.999,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.642,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.642,
       "nodes": 32,
       "calls": 256
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.591,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.591,
       "nodes": 34,
       "calls": 272
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.553,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 0.553,
       "nodes": 35,
       "calls": 280
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.55,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 0.55,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.272,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.272,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.102,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.102,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.055,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.055,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.026,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.024,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.018,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_llama32-1b_cpu.simplified.json",
   "source_profile": "ort_llama32-1b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:57:20",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "OLMo-2-0425-1B",
   "hf_model_id": "allenai/OLMo-2-0425-1B",
   "hf_model_url": "https://huggingface.co/allenai/OLMo-2-0425-1B",
   "param_count": 1480000000
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.409,
   "TTFT": 0.435,
   "TBT_avg": 0.282,
   "Throughput": 3.321
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2408.571,
     "ttft": 435.412,
     "tbt_avg": 281.88,
     "throughput": 3.321,
     "steps": [
      435.412,
      312.518,
      269.681,
      269.873,
      270.242,
      289.97,
      290.497,
      270.378
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2408.571,
     "min": 2408.571,
     "max": 2408.571,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 435.412,
     "min": 435.412,
     "max": 435.412,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 281.88,
     "min": 281.88,
     "max": 281.88,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 3.321,
     "min": 3.321,
     "max": 3.321,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 281.88,
     "min": 269.681,
     "max": 312.518,
     "std": 15.302,
     "cv_pct": 5.429
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2371.724,
   "total_nodes": 1908,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 2371.724,
     "pct": 100.0,
     "nodes": 1908,
     "calls": 15264
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 2312.002,
     "pct": 97.482,
     "by_device": {
      "CPU": {
       "time_ms": 2312.002,
       "nodes": 146,
       "calls": 1168
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 10.559,
     "pct": 0.445,
     "by_device": {
      "CPU": {
       "time_ms": 10.559,
       "nodes": 263,
       "calls": 2104
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 9.981,
     "pct": 0.421,
     "by_device": {
      "CPU": {
       "time_ms": 9.981,
       "nodes": 132,
       "calls": 1056
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.584,
     "pct": 0.193,
     "by_device": {
      "CPU": {
       "time_ms": 4.584,
       "nodes": 81,
       "calls": 648
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 4.546,
     "pct": 0.192,
     "by_device": {
      "CPU": {
       "time_ms": 4.546,
       "nodes": 224,
       "calls": 1792
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 3.968,
     "pct": 0.167,
     "by_device": {
      "CPU": {
       "time_ms": 3.968,
       "nodes": 170,
       "calls": 1360
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 3.94,
     "pct": 0.166,
     "by_device": {
      "CPU": {
       "time_ms": 3.94,
       "nodes": 148,
       "calls": 1184
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 3.843,
     "pct": 0.162,
     "by_device": {
      "CPU": {
       "time_ms": 3.843,
       "nodes": 81,
       "calls": 648
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 2.84,
     "pct": 0.12,
     "by_device": {
      "CPU": {
       "time_ms": 2.84,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 2.375,
     "pct": 0.1,
     "by_device": {
      "CPU": {
       "time_ms": 2.375,
       "nodes": 97,
       "calls": 776
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 2.127,
     "pct": 0.09,
     "by_device": {
      "CPU": {
       "time_ms": 2.127,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 2.108,
     "pct": 0.089,
     "by_device": {
      "CPU": {
       "time_ms": 2.108,
       "nodes": 86,
       "calls": 688
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 1.641,
     "pct": 0.069,
     "by_device": {
      "CPU": {
       "time_ms": 1.641,
       "nodes": 34,
       "calls": 272
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 1.594,
     "pct": 0.067,
     "by_device": {
      "CPU": {
       "time_ms": 1.594,
       "nodes": 68,
       "calls": 544
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.402,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 1.402,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 1.197,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 1.197,
       "nodes": 87,
       "calls": 696
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 1.081,
     "pct": 0.046,
     "by_device": {
      "CPU": {
       "time_ms": 1.081,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.644,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.644,
       "nodes": 32,
       "calls": 256
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.642,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.642,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.254,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.254,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.095,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.095,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.069,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.069,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.056,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.056,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.052,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.052,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.04,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.04,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.018,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_olmo2-1b_cpu.simplified.json",
   "source_profile": "ort_olmo2-1b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:58:37",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "opt-1.3b",
   "hf_model_id": "facebook/opt-1.3b",
   "hf_model_url": "https://huggingface.co/facebook/opt-1.3b",
   "param_count": 1315758080
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.448,
   "TTFT": 0.465,
   "TBT_avg": 0.283,
   "Throughput": 3.268
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2448.354,
     "ttft": 464.686,
     "tbt_avg": 283.381,
     "throughput": 3.268,
     "steps": [
      464.686,
      276.081,
      278.987,
      316.137,
      276.692,
      276.69,
      277.991,
      281.09
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2448.354,
     "min": 2448.354,
     "max": 2448.354,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 464.686,
     "min": 464.686,
     "max": 464.686,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 283.381,
     "min": 283.381,
     "max": 283.381,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 3.268,
     "min": 3.268,
     "max": 3.268,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 283.381,
     "min": 276.081,
     "max": 316.137,
     "std": 13.466,
     "cv_pct": 4.752
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2413.457,
   "total_nodes": 1777,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 2413.457,
     "pct": 100.0,
     "nodes": 1777,
     "calls": 14216
    }
   ],
   "ops": [
    {
     "op": "Gemm",
     "total_ms": 1462.629,
     "pct": 60.603,
     "by_device": {
      "CPU": {
       "time_ms": 1462.629,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "MatMul",
     "total_ms": 881.78,
     "pct": 36.536,
     "by_device": {
      "CPU": {
       "time_ms": 881.78,
       "nodes": 145,
       "calls": 1160
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 14.894,
     "pct": 0.617,
     "by_device": {
      "CPU": {
       "time_ms": 14.894,
       "nodes": 197,
       "calls": 1576
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 12.307,
     "pct": 0.51,
     "by_device": {
      "CPU": {
       "time_ms": 12.307,
       "nodes": 49,
       "calls": 392
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 7.671,
     "pct": 0.318,
     "by_device": {
      "CPU": {
       "time_ms": 7.671,
       "nodes": 175,
       "calls": 1400
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 5.95,
     "pct": 0.247,
     "by_device": {
      "CPU": {
       "time_ms": 5.95,
       "nodes": 120,
       "calls": 960
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 5.695,
     "pct": 0.236,
     "by_device": {
      "CPU": {
       "time_ms": 5.695,
       "nodes": 256,
       "calls": 2048
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 5.112,
     "pct": 0.212,
     "by_device": {
      "CPU": {
       "time_ms": 5.112,
       "nodes": 78,
       "calls": 624
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 3.717,
     "pct": 0.154,
     "by_device": {
      "CPU": {
       "time_ms": 3.717,
       "nodes": 155,
       "calls": 1240
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 3.146,
     "pct": 0.13,
     "by_device": {
      "CPU": {
       "time_ms": 3.146,
       "nodes": 151,
       "calls": 1208
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 2.426,
     "pct": 0.101,
     "by_device": {
      "CPU": {
       "time_ms": 2.426,
       "nodes": 54,
       "calls": 432
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.365,
     "pct": 0.098,
     "by_device": {
      "CPU": {
       "time_ms": 2.365,
       "nodes": 185,
       "calls": 1480
      }
     }
    },
    {
     "op": "Relu",
     "total_ms": 1.817,
     "pct": 0.075,
     "by_device": {
      "CPU": {
       "time_ms": 1.817,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.516,
     "pct": 0.063,
     "by_device": {
      "CPU": {
       "time_ms": 1.516,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Identity",
     "total_ms": 0.801,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.801,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Squeeze",
     "total_ms": 0.565,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.565,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.234,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.234,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.181,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.181,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.177,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.177,
       "nodes": 11,
       "calls": 88
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.159,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.159,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.119,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.119,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.097,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.097,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.046,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.046,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.027,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.026,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_opt-1.3b_cpu.simplified.json",
   "source_profile": "ort_opt-1.3b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:54:24",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "opt-350m",
   "hf_model_id": "facebook/opt-350m",
   "hf_model_url": "https://huggingface.co/facebook/opt-350m",
   "param_count": 331196416
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.68,
   "TTFT": 0.128,
   "TBT_avg": 0.079,
   "Throughput": 11.772
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 679.565,
     "ttft": 128.204,
     "tbt_avg": 78.766,
     "throughput": 11.772,
     "steps": [
      128.204,
      77.13,
      78.47,
      76.895,
      76.802,
      77.358,
      80.477,
      84.229
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 679.565,
     "min": 679.565,
     "max": 679.565,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 128.204,
     "min": 128.204,
     "max": 128.204,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 78.766,
     "min": 78.766,
     "max": 78.766,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 11.772,
     "min": 11.772,
     "max": 11.772,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 78.766,
     "min": 76.802,
     "max": 84.229,
     "std": 2.532,
     "cv_pct": 3.215
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 643.605,
   "total_nodes": 1777,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 643.605,
     "pct": 100.0,
     "nodes": 1777,
     "calls": 14216
    }
   ],
   "ops": [
    {
     "op": "Gemm",
     "total_ms": 359.204,
     "pct": 55.811,
     "by_device": {
      "CPU": {
       "time_ms": 359.204,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "MatMul",
     "total_ms": 233.537,
     "pct": 36.286,
     "by_device": {
      "CPU": {
       "time_ms": 233.537,
       "nodes": 147,
       "calls": 1176
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 10.378,
     "pct": 1.612,
     "by_device": {
      "CPU": {
       "time_ms": 10.378,
       "nodes": 197,
       "calls": 1576
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 6.833,
     "pct": 1.062,
     "by_device": {
      "CPU": {
       "time_ms": 6.833,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 6.122,
     "pct": 0.951,
     "by_device": {
      "CPU": {
       "time_ms": 6.122,
       "nodes": 175,
       "calls": 1400
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 5.341,
     "pct": 0.83,
     "by_device": {
      "CPU": {
       "time_ms": 5.341,
       "nodes": 256,
       "calls": 2048
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 3.673,
     "pct": 0.571,
     "by_device": {
      "CPU": {
       "time_ms": 3.673,
       "nodes": 120,
       "calls": 960
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 3.589,
     "pct": 0.558,
     "by_device": {
      "CPU": {
       "time_ms": 3.589,
       "nodes": 155,
       "calls": 1240
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 2.972,
     "pct": 0.462,
     "by_device": {
      "CPU": {
       "time_ms": 2.972,
       "nodes": 78,
       "calls": 624
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.948,
     "pct": 0.458,
     "by_device": {
      "CPU": {
       "time_ms": 2.948,
       "nodes": 151,
       "calls": 1208
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 2.293,
     "pct": 0.356,
     "by_device": {
      "CPU": {
       "time_ms": 2.293,
       "nodes": 54,
       "calls": 432
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.188,
     "pct": 0.34,
     "by_device": {
      "CPU": {
       "time_ms": 2.188,
       "nodes": 185,
       "calls": 1480
      }
     }
    },
    {
     "op": "Relu",
     "total_ms": 1.219,
     "pct": 0.189,
     "by_device": {
      "CPU": {
       "time_ms": 1.219,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.972,
     "pct": 0.151,
     "by_device": {
      "CPU": {
       "time_ms": 0.972,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Identity",
     "total_ms": 0.837,
     "pct": 0.13,
     "by_device": {
      "CPU": {
       "time_ms": 0.837,
       "nodes": 47,
       "calls": 376
      }
     }
    },
    {
     "op": "Squeeze",
     "total_ms": 0.509,
     "pct": 0.079,
     "by_device": {
      "CPU": {
       "time_ms": 0.509,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.218,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.218,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.165,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.165,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.153,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 0.153,
       "nodes": 11,
       "calls": 88
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.143,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.143,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.117,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.117,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.091,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.091,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.046,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.046,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.031,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.031,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.026,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_opt-350m_cpu.simplified.json",
   "source_profile": "ort_opt-350m_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:58:18",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "phi-1_5",
   "hf_model_id": "microsoft/phi-1_5",
   "hf_model_url": "https://huggingface.co/microsoft/phi-1_5",
   "param_count": 1418270720
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.525,
   "TTFT": 0.469,
   "TBT_avg": 0.294,
   "Throughput": 3.168
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2525.242,
     "ttft": 468.767,
     "tbt_avg": 293.782,
     "throughput": 3.168,
     "steps": [
      468.767,
      321.186,
      280.905,
      280.3,
      285.268,
      325.493,
      281.601,
      281.722
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2525.242,
     "min": 2525.242,
     "max": 2525.242,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 468.767,
     "min": 468.767,
     "max": 468.767,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 293.782,
     "min": 293.782,
     "max": 293.782,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 3.168,
     "min": 3.168,
     "max": 3.168,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 293.782,
     "min": 280.3,
     "max": 325.493,
     "std": 18.786,
     "cv_pct": 6.395
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2480.164,
   "total_nodes": 2315,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 2480.164,
     "pct": 100.0,
     "nodes": 2315,
     "calls": 18520
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 2389.766,
     "pct": 96.355,
     "by_device": {
      "CPU": {
       "time_ms": 2389.766,
       "nodes": 194,
       "calls": 1552
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 15.842,
     "pct": 0.639,
     "by_device": {
      "CPU": {
       "time_ms": 15.842,
       "nodes": 244,
       "calls": 1952
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 13.82,
     "pct": 0.557,
     "by_device": {
      "CPU": {
       "time_ms": 13.82,
       "nodes": 316,
       "calls": 2528
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 11.258,
     "pct": 0.454,
     "by_device": {
      "CPU": {
       "time_ms": 11.258,
       "nodes": 245,
       "calls": 1960
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 10.634,
     "pct": 0.429,
     "by_device": {
      "CPU": {
       "time_ms": 10.634,
       "nodes": 217,
       "calls": 1736
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 7.026,
     "pct": 0.283,
     "by_device": {
      "CPU": {
       "time_ms": 7.026,
       "nodes": 328,
       "calls": 2624
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 6.507,
     "pct": 0.262,
     "by_device": {
      "CPU": {
       "time_ms": 6.507,
       "nodes": 121,
       "calls": 968
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 6.384,
     "pct": 0.257,
     "by_device": {
      "CPU": {
       "time_ms": 6.384,
       "nodes": 25,
       "calls": 200
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 3.047,
     "pct": 0.123,
     "by_device": {
      "CPU": {
       "time_ms": 3.047,
       "nodes": 128,
       "calls": 1024
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 2.922,
     "pct": 0.118,
     "by_device": {
      "CPU": {
       "time_ms": 2.922,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 2.673,
     "pct": 0.108,
     "by_device": {
      "CPU": {
       "time_ms": 2.673,
       "nodes": 50,
       "calls": 400
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.299,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 2.299,
       "nodes": 100,
       "calls": 800
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 1.855,
     "pct": 0.075,
     "by_device": {
      "CPU": {
       "time_ms": 1.855,
       "nodes": 129,
       "calls": 1032
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 1.569,
     "pct": 0.063,
     "by_device": {
      "CPU": {
       "time_ms": 1.569,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.485,
     "pct": 0.06,
     "by_device": {
      "CPU": {
       "time_ms": 1.485,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.204,
     "pct": 0.049,
     "by_device": {
      "CPU": {
       "time_ms": 1.204,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.89,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 0.89,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.413,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.413,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.166,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.166,
       "nodes": 10,
       "calls": 80
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.1,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.1,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.074,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.074,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.056,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.056,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.048,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.048,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.04,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.04,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.024,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.021,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_phi-1_5_cpu.simplified.json",
   "source_profile": "ort_phi-1_5_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:56:16",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "pythia-1b",
   "hf_model_id": "EleutherAI/pythia-1b",
   "hf_model_url": "https://huggingface.co/EleutherAI/pythia-1b",
   "param_count": 1011781632
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.733,
   "TTFT": 0.325,
   "TBT_avg": 0.201,
   "Throughput": 4.615
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1733.483,
     "ttft": 325.132,
     "tbt_avg": 201.193,
     "throughput": 4.615,
     "steps": [
      325.132,
      193.905,
      196.537,
      195.464,
      196.978,
      232.012,
      196.258,
      197.197
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1733.483,
     "min": 1733.483,
     "max": 1733.483,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 325.132,
     "min": 325.132,
     "max": 325.132,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 201.193,
     "min": 201.193,
     "max": 201.193,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 4.615,
     "min": 4.615,
     "max": 4.615,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 201.193,
     "min": 193.905,
     "max": 232.012,
     "std": 12.624,
     "cv_pct": 6.274
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1703.402,
   "total_nodes": 1556,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1703.402,
     "pct": 100.0,
     "nodes": 1556,
     "calls": 12448
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 1641.894,
     "pct": 96.389,
     "by_device": {
      "CPU": {
       "time_ms": 1641.894,
       "nodes": 98,
       "calls": 784
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 8.958,
     "pct": 0.526,
     "by_device": {
      "CPU": {
       "time_ms": 8.958,
       "nodes": 132,
       "calls": 1056
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 8.623,
     "pct": 0.506,
     "by_device": {
      "CPU": {
       "time_ms": 8.623,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 8.227,
     "pct": 0.483,
     "by_device": {
      "CPU": {
       "time_ms": 8.227,
       "nodes": 179,
       "calls": 1432
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 8.094,
     "pct": 0.475,
     "by_device": {
      "CPU": {
       "time_ms": 8.094,
       "nodes": 193,
       "calls": 1544
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 6.45,
     "pct": 0.379,
     "by_device": {
      "CPU": {
       "time_ms": 6.45,
       "nodes": 181,
       "calls": 1448
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 4.576,
     "pct": 0.269,
     "by_device": {
      "CPU": {
       "time_ms": 4.576,
       "nodes": 224,
       "calls": 1792
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 3.519,
     "pct": 0.207,
     "by_device": {
      "CPU": {
       "time_ms": 3.519,
       "nodes": 49,
       "calls": 392
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 2.564,
     "pct": 0.151,
     "by_device": {
      "CPU": {
       "time_ms": 2.564,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Erf",
     "total_ms": 2.172,
     "pct": 0.128,
     "by_device": {
      "CPU": {
       "time_ms": 2.172,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 2.171,
     "pct": 0.127,
     "by_device": {
      "CPU": {
       "time_ms": 2.171,
       "nodes": 64,
       "calls": 512
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 1.51,
     "pct": 0.089,
     "by_device": {
      "CPU": {
       "time_ms": 1.51,
       "nodes": 106,
       "calls": 848
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 1.485,
     "pct": 0.087,
     "by_device": {
      "CPU": {
       "time_ms": 1.485,
       "nodes": 34,
       "calls": 272
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 0.792,
     "pct": 0.046,
     "by_device": {
      "CPU": {
       "time_ms": 0.792,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.581,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.581,
       "nodes": 32,
       "calls": 256
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.496,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.496,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Identity",
     "total_ms": 0.414,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 0.414,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.261,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.261,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.179,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.179,
       "nodes": 10,
       "calls": 80
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.104,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.104,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.074,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.074,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.056,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.056,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.052,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.052,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.048,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.048,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.029,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.028,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.026,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.019,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_pythia-1b_cpu.simplified.json",
   "source_profile": "ort_pythia-1b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 18:01:49",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "Qwen2.5-1.5B",
   "hf_model_id": "Qwen/Qwen2.5-1.5B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen2.5-1.5B",
   "param_count": 1543714304
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.973,
   "TTFT": 0.529,
   "TBT_avg": 0.349,
   "Throughput": 2.691
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2973.309,
     "ttft": 529.384,
     "tbt_avg": 349.132,
     "throughput": 2.691,
     "steps": [
      529.384,
      347.635,
      372.691,
      338.086,
      336.849,
      336.368,
      368.784,
      343.512
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2973.309,
     "min": 2973.309,
     "max": 2973.309,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 529.384,
     "min": 529.384,
     "max": 529.384,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 349.132,
     "min": 349.132,
     "max": 349.132,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 2.691,
     "min": 2.691,
     "max": 2.691,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 349.132,
     "min": 336.368,
     "max": 372.691,
     "std": 14.197,
     "cv_pct": 4.066
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2888.099,
   "total_nodes": 4374,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 2888.099,
     "pct": 100.0,
     "nodes": 4374,
     "calls": 34992
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 2773.549,
     "pct": 96.034,
     "by_device": {
      "CPU": {
       "time_ms": 2773.549,
       "nodes": 254,
       "calls": 2032
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 18.769,
     "pct": 0.65,
     "by_device": {
      "CPU": {
       "time_ms": 18.769,
       "nodes": 884,
       "calls": 7072
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 15.072,
     "pct": 0.522,
     "by_device": {
      "CPU": {
       "time_ms": 15.072,
       "nodes": 455,
       "calls": 3640
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 13.483,
     "pct": 0.467,
     "by_device": {
      "CPU": {
       "time_ms": 13.483,
       "nodes": 340,
       "calls": 2720
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 9.238,
     "pct": 0.32,
     "by_device": {
      "CPU": {
       "time_ms": 9.238,
       "nodes": 372,
       "calls": 2976
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 9.178,
     "pct": 0.318,
     "by_device": {
      "CPU": {
       "time_ms": 9.178,
       "nodes": 284,
       "calls": 2272
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 6.784,
     "pct": 0.235,
     "by_device": {
      "CPU": {
       "time_ms": 6.784,
       "nodes": 141,
       "calls": 1128
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 5.947,
     "pct": 0.206,
     "by_device": {
      "CPU": {
       "time_ms": 5.947,
       "nodes": 141,
       "calls": 1128
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 5.52,
     "pct": 0.191,
     "by_device": {
      "CPU": {
       "time_ms": 5.52,
       "nodes": 429,
       "calls": 3432
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 5.041,
     "pct": 0.175,
     "by_device": {
      "CPU": {
       "time_ms": 5.041,
       "nodes": 228,
       "calls": 1824
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 4.569,
     "pct": 0.158,
     "by_device": {
      "CPU": {
       "time_ms": 4.569,
       "nodes": 114,
       "calls": 912
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 3.267,
     "pct": 0.113,
     "by_device": {
      "CPU": {
       "time_ms": 3.267,
       "nodes": 58,
       "calls": 464
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 2.938,
     "pct": 0.102,
     "by_device": {
      "CPU": {
       "time_ms": 2.938,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 2.666,
     "pct": 0.092,
     "by_device": {
      "CPU": {
       "time_ms": 2.666,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 2.476,
     "pct": 0.086,
     "by_device": {
      "CPU": {
       "time_ms": 2.476,
       "nodes": 124,
       "calls": 992
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 2.191,
     "pct": 0.076,
     "by_device": {
      "CPU": {
       "time_ms": 2.191,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 1.8,
     "pct": 0.062,
     "by_device": {
      "CPU": {
       "time_ms": 1.8,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 1.091,
     "pct": 0.038,
     "by_device": {
      "CPU": {
       "time_ms": 1.091,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.993,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.993,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.991,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.991,
       "nodes": 58,
       "calls": 464
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.971,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.971,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.844,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.844,
       "nodes": 59,
       "calls": 472
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.461,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.461,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.108,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.108,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.054,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.054,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.027,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.026,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.02,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_qwen2.5-1.5b_cpu.simplified.json",
   "source_profile": "ort_qwen2.5-1.5b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:54:54",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "Qwen3-0.6B",
   "hf_model_id": "Qwen/Qwen3-0.6B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen3-0.6B",
   "param_count": 596049920
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.292,
   "TTFT": 0.219,
   "TBT_avg": 0.153,
   "Throughput": 6.192
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1291.921,
     "ttft": 219.291,
     "tbt_avg": 153.233,
     "throughput": 6.192,
     "steps": [
      219.291,
      146.5,
      146.701,
      146.228,
      146.73,
      185.646,
      151.616,
      149.209
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1291.921,
     "min": 1291.921,
     "max": 1291.921,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 219.291,
     "min": 219.291,
     "max": 219.291,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 153.233,
     "min": 153.233,
     "max": 153.233,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 6.192,
     "min": 6.192,
     "max": 6.192,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 153.233,
     "min": 146.228,
     "max": 185.646,
     "std": 13.356,
     "cv_pct": 8.716
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1203.202,
   "total_nodes": 4794,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1203.202,
     "pct": 100.0,
     "nodes": 4794,
     "calls": 38352
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 1084.843,
     "pct": 90.163,
     "by_device": {
      "CPU": {
       "time_ms": 1084.843,
       "nodes": 254,
       "calls": 2032
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 17.233,
     "pct": 1.432,
     "by_device": {
      "CPU": {
       "time_ms": 17.233,
       "nodes": 884,
       "calls": 7072
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 15.401,
     "pct": 1.28,
     "by_device": {
      "CPU": {
       "time_ms": 15.401,
       "nodes": 340,
       "calls": 2720
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 15.136,
     "pct": 1.258,
     "by_device": {
      "CPU": {
       "time_ms": 15.136,
       "nodes": 567,
       "calls": 4536
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 8.137,
     "pct": 0.676,
     "by_device": {
      "CPU": {
       "time_ms": 8.137,
       "nodes": 372,
       "calls": 2976
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 7.176,
     "pct": 0.596,
     "by_device": {
      "CPU": {
       "time_ms": 7.176,
       "nodes": 256,
       "calls": 2048
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 6.543,
     "pct": 0.544,
     "by_device": {
      "CPU": {
       "time_ms": 6.543,
       "nodes": 141,
       "calls": 1128
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 6.084,
     "pct": 0.506,
     "by_device": {
      "CPU": {
       "time_ms": 6.084,
       "nodes": 141,
       "calls": 1128
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 4.912,
     "pct": 0.408,
     "by_device": {
      "CPU": {
       "time_ms": 4.912,
       "nodes": 429,
       "calls": 3432
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 4.81,
     "pct": 0.4,
     "by_device": {
      "CPU": {
       "time_ms": 4.81,
       "nodes": 228,
       "calls": 1824
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 4.768,
     "pct": 0.396,
     "by_device": {
      "CPU": {
       "time_ms": 4.768,
       "nodes": 58,
       "calls": 464
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 4.757,
     "pct": 0.395,
     "by_device": {
      "CPU": {
       "time_ms": 4.757,
       "nodes": 236,
       "calls": 1888
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 4.292,
     "pct": 0.357,
     "by_device": {
      "CPU": {
       "time_ms": 4.292,
       "nodes": 114,
       "calls": 912
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 4.202,
     "pct": 0.349,
     "by_device": {
      "CPU": {
       "time_ms": 4.202,
       "nodes": 169,
       "calls": 1352
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 3.942,
     "pct": 0.328,
     "by_device": {
      "CPU": {
       "time_ms": 3.942,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 3.17,
     "pct": 0.263,
     "by_device": {
      "CPU": {
       "time_ms": 3.17,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 1.861,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 1.861,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.353,
     "pct": 0.112,
     "by_device": {
      "CPU": {
       "time_ms": 1.353,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.08,
     "pct": 0.09,
     "by_device": {
      "CPU": {
       "time_ms": 1.08,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 1.042,
     "pct": 0.087,
     "by_device": {
      "CPU": {
       "time_ms": 1.042,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.983,
     "pct": 0.082,
     "by_device": {
      "CPU": {
       "time_ms": 0.983,
       "nodes": 58,
       "calls": 464
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.777,
     "pct": 0.065,
     "by_device": {
      "CPU": {
       "time_ms": 0.777,
       "nodes": 59,
       "calls": 472
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.453,
     "pct": 0.038,
     "by_device": {
      "CPU": {
       "time_ms": 0.453,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.1,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.1,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.05,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.05,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.028,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.026,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.025,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.018,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_qwen3-0.6b_cpu.simplified.json",
   "source_profile": "ort_qwen3-0.6b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:59:53",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "stablelm-2-1_6b",
   "hf_model_id": "stabilityai/stablelm-2-1_6b",
   "hf_model_url": "https://huggingface.co/stabilityai/stablelm-2-1_6b",
   "param_count": 1644417024
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.743,
   "TTFT": 0.539,
   "TBT_avg": 0.315,
   "Throughput": 2.917
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2742.551,
     "ttft": 538.628,
     "tbt_avg": 314.846,
     "throughput": 2.917,
     "steps": [
      538.628,
      305.073,
      305.818,
      309.22,
      346.152,
      307.752,
      305.348,
      324.56
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2742.551,
     "min": 2742.551,
     "max": 2742.551,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 538.628,
     "min": 538.628,
     "max": 538.628,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 314.846,
     "min": 314.846,
     "max": 314.846,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 2.917,
     "min": 2.917,
     "max": 2.917,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 314.846,
     "min": 305.073,
     "max": 346.152,
     "std": 14.261,
     "cv_pct": 4.53
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2696.585,
   "total_nodes": 2391,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 2696.585,
     "pct": 100.0,
     "nodes": 2391,
     "calls": 19128
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 2607.733,
     "pct": 96.705,
     "by_device": {
      "CPU": {
       "time_ms": 2607.733,
       "nodes": 218,
       "calls": 1744
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 16.251,
     "pct": 0.603,
     "by_device": {
      "CPU": {
       "time_ms": 16.251,
       "nodes": 246,
       "calls": 1968
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 12.225,
     "pct": 0.453,
     "by_device": {
      "CPU": {
       "time_ms": 12.225,
       "nodes": 49,
       "calls": 392
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 11.931,
     "pct": 0.442,
     "by_device": {
      "CPU": {
       "time_ms": 11.931,
       "nodes": 246,
       "calls": 1968
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 9.457,
     "pct": 0.351,
     "by_device": {
      "CPU": {
       "time_ms": 9.457,
       "nodes": 201,
       "calls": 1608
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 7.876,
     "pct": 0.292,
     "by_device": {
      "CPU": {
       "time_ms": 7.876,
       "nodes": 197,
       "calls": 1576
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 7.236,
     "pct": 0.268,
     "by_device": {
      "CPU": {
       "time_ms": 7.236,
       "nodes": 331,
       "calls": 2648
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 6.113,
     "pct": 0.227,
     "by_device": {
      "CPU": {
       "time_ms": 6.113,
       "nodes": 121,
       "calls": 968
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 3.27,
     "pct": 0.121,
     "by_device": {
      "CPU": {
       "time_ms": 3.27,
       "nodes": 130,
       "calls": 1040
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.498,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 2.498,
       "nodes": 103,
       "calls": 824
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.319,
     "pct": 0.086,
     "by_device": {
      "CPU": {
       "time_ms": 2.319,
       "nodes": 162,
       "calls": 1296
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.817,
     "pct": 0.067,
     "by_device": {
      "CPU": {
       "time_ms": 1.817,
       "nodes": 72,
       "calls": 576
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.702,
     "pct": 0.063,
     "by_device": {
      "CPU": {
       "time_ms": 1.702,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.472,
     "pct": 0.055,
     "by_device": {
      "CPU": {
       "time_ms": 1.472,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 1.033,
     "pct": 0.038,
     "by_device": {
      "CPU": {
       "time_ms": 1.033,
       "nodes": 72,
       "calls": 576
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.875,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.875,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.863,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.863,
       "nodes": 61,
       "calls": 488
      }
     }
    },
    {
     "op": "Identity",
     "total_ms": 0.81,
     "pct": 0.03,
     "by_device": {
      "CPU": {
       "time_ms": 0.81,
       "nodes": 47,
       "calls": 376
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.297,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.297,
       "nodes": 7,
       "calls": 56
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.188,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.188,
       "nodes": 7,
       "calls": 56
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.18,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.18,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.151,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.151,
       "nodes": 7,
       "calls": 56
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.121,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.121,
       "nodes": 7,
       "calls": 56
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.058,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.058,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.029,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.029,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.027,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.024,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_stablelm2-1.6b_cpu.simplified.json",
   "source_profile": "ort_stablelm2-1.6b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:57:57",
   "framework": "onnxruntime",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "TinyLlama-1.1B-Chat-v1.0",
   "hf_model_id": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "hf_model_url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "param_count": 1100048384
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.046,
   "TTFT": 0.38,
   "TBT_avg": 0.238,
   "Throughput": 3.91
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2045.873,
     "ttft": 379.864,
     "tbt_avg": 238.001,
     "throughput": 3.91,
     "steps": [
      379.864,
      263.04,
      226.337,
      226.812,
      227.54,
      227.185,
      229.527,
      265.568
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2045.873,
     "min": 2045.873,
     "max": 2045.873,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 379.864,
     "min": 379.864,
     "max": 379.864,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 238.001,
     "min": 238.001,
     "max": 238.001,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 3.91,
     "min": 3.91,
     "max": 3.91,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 238.001,
     "min": 226.337,
     "max": 265.568,
     "std": 16.675,
     "cv_pct": 7.006
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1982.742,
   "total_nodes": 3390,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1982.742,
     "pct": 100.0,
     "nodes": 3390,
     "calls": 27120
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 1898.155,
     "pct": 95.734,
     "by_device": {
      "CPU": {
       "time_ms": 1898.155,
       "nodes": 200,
       "calls": 1600
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 13.437,
     "pct": 0.678,
     "by_device": {
      "CPU": {
       "time_ms": 13.437,
       "nodes": 698,
       "calls": 5584
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 10.774,
     "pct": 0.543,
     "by_device": {
      "CPU": {
       "time_ms": 10.774,
       "nodes": 359,
       "calls": 2872
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 9.69,
     "pct": 0.489,
     "by_device": {
      "CPU": {
       "time_ms": 9.69,
       "nodes": 268,
       "calls": 2144
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 6.409,
     "pct": 0.323,
     "by_device": {
      "CPU": {
       "time_ms": 6.409,
       "nodes": 294,
       "calls": 2352
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 5.807,
     "pct": 0.293,
     "by_device": {
      "CPU": {
       "time_ms": 5.807,
       "nodes": 111,
       "calls": 888
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 4.891,
     "pct": 0.247,
     "by_device": {
      "CPU": {
       "time_ms": 4.891,
       "nodes": 111,
       "calls": 888
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 4.763,
     "pct": 0.24,
     "by_device": {
      "CPU": {
       "time_ms": 4.763,
       "nodes": 158,
       "calls": 1264
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 4.07,
     "pct": 0.205,
     "by_device": {
      "CPU": {
       "time_ms": 4.07,
       "nodes": 339,
       "calls": 2712
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 3.701,
     "pct": 0.187,
     "by_device": {
      "CPU": {
       "time_ms": 3.701,
       "nodes": 46,
       "calls": 368
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 3.678,
     "pct": 0.186,
     "by_device": {
      "CPU": {
       "time_ms": 3.678,
       "nodes": 90,
       "calls": 720
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 3.672,
     "pct": 0.185,
     "by_device": {
      "CPU": {
       "time_ms": 3.672,
       "nodes": 180,
       "calls": 1440
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 2.188,
     "pct": 0.11,
     "by_device": {
      "CPU": {
       "time_ms": 2.188,
       "nodes": 89,
       "calls": 712
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 1.932,
     "pct": 0.097,
     "by_device": {
      "CPU": {
       "time_ms": 1.932,
       "nodes": 100,
       "calls": 800
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 1.83,
     "pct": 0.092,
     "by_device": {
      "CPU": {
       "time_ms": 1.83,
       "nodes": 45,
       "calls": 360
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.41,
     "pct": 0.071,
     "by_device": {
      "CPU": {
       "time_ms": 1.41,
       "nodes": 22,
       "calls": 176
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 1.362,
     "pct": 0.069,
     "by_device": {
      "CPU": {
       "time_ms": 1.362,
       "nodes": 45,
       "calls": 360
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.341,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 1.341,
       "nodes": 22,
       "calls": 176
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.831,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.831,
       "nodes": 44,
       "calls": 352
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.773,
     "pct": 0.039,
     "by_device": {
      "CPU": {
       "time_ms": 0.773,
       "nodes": 46,
       "calls": 368
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.729,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.729,
       "nodes": 45,
       "calls": 360
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.672,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.672,
       "nodes": 47,
       "calls": 376
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.38,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.38,
       "nodes": 22,
       "calls": 176
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.101,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.101,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.053,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.053,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.026,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.024,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.018,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/cpu/ort_tinyllama-1.1b_cpu.simplified.json",
   "source_profile": "ort_tinyllama-1.1b_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 18:02:24",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "bloom-1b7",
   "hf_model_id": "bigscience/bloom-1b7",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-1b7",
   "param_count": 1722408960
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.126,
   "TTFT": 0.339,
   "TBT_avg": 0.112,
   "Throughput": 7.103
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1126.304,
     "ttft": 339.447,
     "tbt_avg": 112.408,
     "throughput": 7.103,
     "steps": [
      339.447,
      113.214,
      111.693,
      111.74,
      112.018,
      113.741,
      111.897,
      112.554
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1126.304,
     "min": 1126.304,
     "max": 1126.304,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 339.447,
     "min": 339.447,
     "max": 339.447,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 112.408,
     "min": 112.408,
     "max": 112.408,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 7.103,
     "min": 7.103,
     "max": 7.103,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 112.408,
     "min": 111.693,
     "max": 113.741,
     "std": 0.739,
     "cv_pct": 0.657
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1087.829,
   "total_nodes": 2361,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 1022.198,
     "pct": 93.967,
     "nodes": 386,
     "calls": 3088
    },
    {
     "device": "CPU",
     "time_ms": 65.631,
     "pct": 6.033,
     "nodes": 1975,
     "calls": 15800
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 949.314,
     "pct": 87.267,
     "by_device": {
      "PIM": {
       "time_ms": 946.244,
       "nodes": 97,
       "calls": 776
      },
      "CPU": {
       "time_ms": 3.07,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 30.716,
     "pct": 2.824,
     "by_device": {
      "PIM": {
       "time_ms": 26.618,
       "nodes": 96,
       "calls": 768
      },
      "CPU": {
       "time_ms": 4.098,
       "nodes": 149,
       "calls": 1192
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 27.151,
     "pct": 2.496,
     "by_device": {
      "PIM": {
       "time_ms": 27.151,
       "nodes": 96,
       "calls": 768
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 22.185,
     "pct": 2.039,
     "by_device": {
      "PIM": {
       "time_ms": 22.185,
       "nodes": 97,
       "calls": 776
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 13.295,
     "pct": 1.222,
     "by_device": {
      "CPU": {
       "time_ms": 13.295,
       "nodes": 245,
       "calls": 1960
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 11.784,
     "pct": 1.083,
     "by_device": {
      "CPU": {
       "time_ms": 11.784,
       "nodes": 50,
       "calls": 400
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 8.103,
     "pct": 0.745,
     "by_device": {
      "CPU": {
       "time_ms": 8.103,
       "nodes": 225,
       "calls": 1800
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 6.026,
     "pct": 0.554,
     "by_device": {
      "CPU": {
       "time_ms": 6.026,
       "nodes": 354,
       "calls": 2832
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 4.786,
     "pct": 0.44,
     "by_device": {
      "CPU": {
       "time_ms": 4.786,
       "nodes": 250,
       "calls": 2000
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.047,
     "pct": 0.372,
     "by_device": {
      "CPU": {
       "time_ms": 4.047,
       "nodes": 120,
       "calls": 960
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 3.357,
     "pct": 0.309,
     "by_device": {
      "CPU": {
       "time_ms": 3.357,
       "nodes": 199,
       "calls": 1592
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.101,
     "pct": 0.193,
     "by_device": {
      "CPU": {
       "time_ms": 2.101,
       "nodes": 185,
       "calls": 1480
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 1.658,
     "pct": 0.152,
     "by_device": {
      "CPU": {
       "time_ms": 1.658,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 1.064,
     "pct": 0.098,
     "by_device": {
      "CPU": {
       "time_ms": 1.064,
       "nodes": 30,
       "calls": 240
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.95,
     "pct": 0.087,
     "by_device": {
      "CPU": {
       "time_ms": 0.95,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 0.434,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.434,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.177,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.177,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.171,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.171,
       "nodes": 13,
       "calls": 104
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.12,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.12,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.109,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.109,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.093,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.093,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.07,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.07,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.035,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.023,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sub",
     "total_ms": 0.021,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "CumSum",
     "total_ms": 0.02,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.019,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_bloom-1b7_pim.simplified.json",
   "source_profile": "ort_bloom-1b7_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:54:45",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "bloom-560m",
   "hf_model_id": "bigscience/bloom-560m",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-560m",
   "param_count": 559214592
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.493,
   "TTFT": 0.105,
   "TBT_avg": 0.055,
   "Throughput": 16.211
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 493.493,
     "ttft": 105.426,
     "tbt_avg": 55.438,
     "throughput": 16.211,
     "steps": [
      105.426,
      55.092,
      55.472,
      55.404,
      55.431,
      55.947,
      55.205,
      55.516
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 493.493,
     "min": 493.493,
     "max": 493.493,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 105.426,
     "min": 105.426,
     "max": 105.426,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 55.438,
     "min": 55.438,
     "max": 55.438,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 16.211,
     "min": 16.211,
     "max": 16.211,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 55.438,
     "min": 55.092,
     "max": 55.947,
     "std": 0.251,
     "cv_pct": 0.453
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 462.905,
   "total_nodes": 2361,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 421.569,
     "pct": 91.07,
     "nodes": 386,
     "calls": 3088
    },
    {
     "device": "CPU",
     "time_ms": 41.336,
     "pct": 8.93,
     "nodes": 1975,
     "calls": 15800
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 357.741,
     "pct": 77.282,
     "by_device": {
      "PIM": {
       "time_ms": 355.748,
       "nodes": 97,
       "calls": 776
      },
      "CPU": {
       "time_ms": 1.993,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 25.626,
     "pct": 5.536,
     "by_device": {
      "PIM": {
       "time_ms": 22.74,
       "nodes": 96,
       "calls": 768
      },
      "CPU": {
       "time_ms": 2.886,
       "nodes": 149,
       "calls": 1192
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 23.416,
     "pct": 5.058,
     "by_device": {
      "PIM": {
       "time_ms": 23.416,
       "nodes": 96,
       "calls": 768
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 19.665,
     "pct": 4.248,
     "by_device": {
      "PIM": {
       "time_ms": 19.665,
       "nodes": 97,
       "calls": 776
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 6.326,
     "pct": 1.367,
     "by_device": {
      "CPU": {
       "time_ms": 6.326,
       "nodes": 245,
       "calls": 1960
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 6.213,
     "pct": 1.342,
     "by_device": {
      "CPU": {
       "time_ms": 6.213,
       "nodes": 50,
       "calls": 400
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 4.855,
     "pct": 1.049,
     "by_device": {
      "CPU": {
       "time_ms": 4.855,
       "nodes": 225,
       "calls": 1800
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 4.541,
     "pct": 0.981,
     "by_device": {
      "CPU": {
       "time_ms": 4.541,
       "nodes": 354,
       "calls": 2832
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 3.91,
     "pct": 0.845,
     "by_device": {
      "CPU": {
       "time_ms": 3.91,
       "nodes": 250,
       "calls": 2000
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.877,
     "pct": 0.622,
     "by_device": {
      "CPU": {
       "time_ms": 2.877,
       "nodes": 199,
       "calls": 1592
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 2.403,
     "pct": 0.519,
     "by_device": {
      "CPU": {
       "time_ms": 2.403,
       "nodes": 120,
       "calls": 960
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 1.607,
     "pct": 0.347,
     "by_device": {
      "CPU": {
       "time_ms": 1.607,
       "nodes": 185,
       "calls": 1480
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 0.95,
     "pct": 0.205,
     "by_device": {
      "CPU": {
       "time_ms": 0.95,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.865,
     "pct": 0.187,
     "by_device": {
      "CPU": {
       "time_ms": 0.865,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 0.8,
     "pct": 0.173,
     "by_device": {
      "CPU": {
       "time_ms": 0.8,
       "nodes": 30,
       "calls": 240
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 0.361,
     "pct": 0.078,
     "by_device": {
      "CPU": {
       "time_ms": 0.361,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.157,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.157,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.151,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.151,
       "nodes": 13,
       "calls": 104
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.104,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.104,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.099,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.099,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.08,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.08,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.059,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.059,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.028,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "CumSum",
     "total_ms": 0.019,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.018,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sub",
     "total_ms": 0.018,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.016,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_bloom-560m_pim.simplified.json",
   "source_profile": "ort_bloom-560m_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:57:07",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "Falcon3-1B-Base",
   "hf_model_id": "tiiuae/Falcon3-1B-Base",
   "hf_model_url": "https://huggingface.co/tiiuae/Falcon3-1B-Base",
   "param_count": 1669000000
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.039,
   "TTFT": 0.29,
   "TBT_avg": 0.107,
   "Throughput": 7.697
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1039.405,
     "ttft": 290.129,
     "tbt_avg": 107.039,
     "throughput": 7.697,
     "steps": [
      290.129,
      105.777,
      107.892,
      109.603,
      108.605,
      106.024,
      105.803,
      105.572
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1039.405,
     "min": 1039.405,
     "max": 1039.405,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 290.129,
     "min": 290.129,
     "max": 290.129,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 107.039,
     "min": 107.039,
     "max": 107.039,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 7.697,
     "min": 7.697,
     "max": 7.697,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 107.039,
     "min": 105.572,
     "max": 109.603,
     "std": 1.515,
     "cv_pct": 1.415
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 995.11,
   "total_nodes": 2989,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 942.386,
     "pct": 94.702,
     "nodes": 326,
     "calls": 2608
    },
    {
     "device": "CPU",
     "time_ms": 52.724,
     "pct": 5.298,
     "nodes": 2663,
     "calls": 21304
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 888.442,
     "pct": 89.281,
     "by_device": {
      "PIM": {
       "time_ms": 886.299,
       "nodes": 127,
       "calls": 1016
      },
      "CPU": {
       "time_ms": 2.143,
       "nodes": 37,
       "calls": 296
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 36.331,
     "pct": 3.651,
     "by_device": {
      "PIM": {
       "time_ms": 36.331,
       "nodes": 126,
       "calls": 1008
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 19.756,
     "pct": 1.985,
     "by_device": {
      "PIM": {
       "time_ms": 19.756,
       "nodes": 73,
       "calls": 584
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 7.682,
     "pct": 0.772,
     "by_device": {
      "CPU": {
       "time_ms": 7.682,
       "nodes": 574,
       "calls": 4592
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 7.17,
     "pct": 0.721,
     "by_device": {
      "CPU": {
       "time_ms": 7.17,
       "nodes": 295,
       "calls": 2360
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 7.072,
     "pct": 0.711,
     "by_device": {
      "CPU": {
       "time_ms": 7.072,
       "nodes": 220,
       "calls": 1760
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 3.848,
     "pct": 0.387,
     "by_device": {
      "CPU": {
       "time_ms": 3.848,
       "nodes": 242,
       "calls": 1936
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 2.848,
     "pct": 0.286,
     "by_device": {
      "CPU": {
       "time_ms": 2.848,
       "nodes": 91,
       "calls": 728
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 2.676,
     "pct": 0.269,
     "by_device": {
      "CPU": {
       "time_ms": 2.676,
       "nodes": 91,
       "calls": 728
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 2.558,
     "pct": 0.257,
     "by_device": {
      "CPU": {
       "time_ms": 2.558,
       "nodes": 130,
       "calls": 1040
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.497,
     "pct": 0.251,
     "by_device": {
      "CPU": {
       "time_ms": 2.497,
       "nodes": 279,
       "calls": 2232
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.276,
     "pct": 0.229,
     "by_device": {
      "CPU": {
       "time_ms": 2.276,
       "nodes": 148,
       "calls": 1184
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 2.097,
     "pct": 0.211,
     "by_device": {
      "CPU": {
       "time_ms": 2.097,
       "nodes": 74,
       "calls": 592
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 1.539,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 1.539,
       "nodes": 38,
       "calls": 304
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.271,
     "pct": 0.128,
     "by_device": {
      "CPU": {
       "time_ms": 1.271,
       "nodes": 73,
       "calls": 584
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.26,
     "pct": 0.127,
     "by_device": {
      "CPU": {
       "time_ms": 1.26,
       "nodes": 18,
       "calls": 144
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 1.139,
     "pct": 0.114,
     "by_device": {
      "CPU": {
       "time_ms": 1.139,
       "nodes": 37,
       "calls": 296
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 1.13,
     "pct": 0.114,
     "by_device": {
      "CPU": {
       "time_ms": 1.13,
       "nodes": 84,
       "calls": 672
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 0.814,
     "pct": 0.082,
     "by_device": {
      "CPU": {
       "time_ms": 0.814,
       "nodes": 37,
       "calls": 296
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.511,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 0.511,
       "nodes": 18,
       "calls": 144
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.503,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 0.503,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.445,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.445,
       "nodes": 38,
       "calls": 304
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.42,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.42,
       "nodes": 37,
       "calls": 296
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.352,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.352,
       "nodes": 39,
       "calls": 312
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.287,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.287,
       "nodes": 18,
       "calls": 144
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.064,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.064,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.037,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.03,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.021,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.018,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.016,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_falcon3-1b_pim.simplified.json",
   "source_profile": "ort_falcon3-1b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 18:04:30",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "gemma-2-2b",
   "hf_model_id": "google/gemma-2-2b",
   "hf_model_url": "https://huggingface.co/google/gemma-2-2b",
   "param_count": 2614341888
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.043,
   "TTFT": 0.582,
   "TBT_avg": 0.209,
   "Throughput": 3.916
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2042.698,
     "ttft": 581.723,
     "tbt_avg": 208.711,
     "throughput": 3.916,
     "steps": [
      581.723,
      208.613,
      208.232,
      207.834,
      206.407,
      210.02,
      207.065,
      212.804
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2042.698,
     "min": 2042.698,
     "max": 2042.698,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 581.723,
     "min": 581.723,
     "max": 581.723,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 208.711,
     "min": 208.711,
     "max": 208.711,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 3.916,
     "min": 3.916,
     "max": 3.916,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 208.711,
     "min": 206.407,
     "max": 212.804,
     "std": 1.982,
     "cv_pct": 0.95
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1965.18,
   "total_nodes": 4846,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 1863.788,
     "pct": 94.841,
     "nodes": 471,
     "calls": 3768
    },
    {
     "device": "CPU",
     "time_ms": 101.392,
     "pct": 5.159,
     "nodes": 4375,
     "calls": 35000
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 1782.907,
     "pct": 90.725,
     "by_device": {
      "PIM": {
       "time_ms": 1779.609,
       "nodes": 183,
       "calls": 1464
      },
      "CPU": {
       "time_ms": 3.298,
       "nodes": 53,
       "calls": 424
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 56.602,
     "pct": 2.88,
     "by_device": {
      "PIM": {
       "time_ms": 56.602,
       "nodes": 183,
       "calls": 1464
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 27.577,
     "pct": 1.403,
     "by_device": {
      "PIM": {
       "time_ms": 27.577,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 20.09,
     "pct": 1.022,
     "by_device": {
      "CPU": {
       "time_ms": 20.09,
       "nodes": 660,
       "calls": 5280
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 12.911,
     "pct": 0.657,
     "by_device": {
      "CPU": {
       "time_ms": 12.911,
       "nodes": 825,
       "calls": 6600
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 11.505,
     "pct": 0.585,
     "by_device": {
      "CPU": {
       "time_ms": 11.505,
       "nodes": 317,
       "calls": 2536
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 6.83,
     "pct": 0.348,
     "by_device": {
      "CPU": {
       "time_ms": 6.83,
       "nodes": 290,
       "calls": 2320
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 5.944,
     "pct": 0.302,
     "by_device": {
      "CPU": {
       "time_ms": 5.944,
       "nodes": 346,
       "calls": 2768
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.406,
     "pct": 0.224,
     "by_device": {
      "CPU": {
       "time_ms": 4.406,
       "nodes": 131,
       "calls": 1048
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 4.337,
     "pct": 0.221,
     "by_device": {
      "CPU": {
       "time_ms": 4.337,
       "nodes": 131,
       "calls": 1048
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 3.795,
     "pct": 0.193,
     "by_device": {
      "CPU": {
       "time_ms": 3.795,
       "nodes": 400,
       "calls": 3200
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 3.603,
     "pct": 0.183,
     "by_device": {
      "CPU": {
       "time_ms": 3.603,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 3.566,
     "pct": 0.181,
     "by_device": {
      "CPU": {
       "time_ms": 3.566,
       "nodes": 213,
       "calls": 1704
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 3.359,
     "pct": 0.171,
     "by_device": {
      "CPU": {
       "time_ms": 3.359,
       "nodes": 158,
       "calls": 1264
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 3.329,
     "pct": 0.169,
     "by_device": {
      "CPU": {
       "time_ms": 3.329,
       "nodes": 107,
       "calls": 856
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 2.461,
     "pct": 0.125,
     "by_device": {
      "CPU": {
       "time_ms": 2.461,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 2.437,
     "pct": 0.124,
     "by_device": {
      "CPU": {
       "time_ms": 2.437,
       "nodes": 55,
       "calls": 440
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 2.403,
     "pct": 0.122,
     "by_device": {
      "CPU": {
       "time_ms": 2.403,
       "nodes": 27,
       "calls": 216
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 2.099,
     "pct": 0.107,
     "by_device": {
      "CPU": {
       "time_ms": 2.099,
       "nodes": 117,
       "calls": 936
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 1.248,
     "pct": 0.064,
     "by_device": {
      "CPU": {
       "time_ms": 1.248,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.894,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.894,
       "nodes": 52,
       "calls": 416
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.844,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.844,
       "nodes": 55,
       "calls": 440
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.748,
     "pct": 0.038,
     "by_device": {
      "CPU": {
       "time_ms": 0.748,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.515,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.515,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.46,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.46,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.113,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.113,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.073,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.073,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sub",
     "total_ms": 0.023,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.022,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.022,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_gemma2-2b_pim.simplified.json",
   "source_profile": "ort_gemma2-2b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:56:46",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "gemma-3-1b-it",
   "hf_model_id": "google/gemma-3-1b-it",
   "hf_model_url": "https://huggingface.co/google/gemma-3-1b-it",
   "param_count": 999885952
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.121,
   "TTFT": 0.265,
   "TBT_avg": 0.122,
   "Throughput": 7.138
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1120.708,
     "ttft": 264.514,
     "tbt_avg": 122.313,
     "throughput": 7.138,
     "steps": [
      264.514,
      124.859,
      121.517,
      121.924,
      123.225,
      121.936,
      121.436,
      121.297
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1120.708,
     "min": 1120.708,
     "max": 1120.708,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 264.514,
     "min": 264.514,
     "max": 264.514,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 122.313,
     "min": 122.313,
     "max": 122.313,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 7.138,
     "min": 7.138,
     "max": 7.138,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 122.313,
     "min": 121.297,
     "max": 124.859,
     "std": 1.198,
     "cv_pct": 0.979
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1040.891,
   "total_nodes": 5278,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 956.729,
     "pct": 91.914,
     "nodes": 470,
     "calls": 3760
    },
    {
     "device": "CPU",
     "time_ms": 84.162,
     "pct": 8.086,
     "nodes": 4808,
     "calls": 38464
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 887.994,
     "pct": 85.311,
     "by_device": {
      "PIM": {
       "time_ms": 885.705,
       "nodes": 183,
       "calls": 1464
      },
      "CPU": {
       "time_ms": 2.289,
       "nodes": 54,
       "calls": 432
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 48.495,
     "pct": 4.659,
     "by_device": {
      "PIM": {
       "time_ms": 48.495,
       "nodes": 182,
       "calls": 1456
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 22.529,
     "pct": 2.164,
     "by_device": {
      "PIM": {
       "time_ms": 22.529,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 16.7,
     "pct": 1.604,
     "by_device": {
      "CPU": {
       "time_ms": 16.7,
       "nodes": 766,
       "calls": 6128
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 10.887,
     "pct": 1.046,
     "by_device": {
      "CPU": {
       "time_ms": 10.887,
       "nodes": 828,
       "calls": 6624
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 6.955,
     "pct": 0.668,
     "by_device": {
      "CPU": {
       "time_ms": 6.955,
       "nodes": 319,
       "calls": 2552
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 6.477,
     "pct": 0.622,
     "by_device": {
      "CPU": {
       "time_ms": 6.477,
       "nodes": 342,
       "calls": 2736
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 5.331,
     "pct": 0.512,
     "by_device": {
      "CPU": {
       "time_ms": 5.331,
       "nodes": 346,
       "calls": 2768
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 3.502,
     "pct": 0.336,
     "by_device": {
      "CPU": {
       "time_ms": 3.502,
       "nodes": 401,
       "calls": 3208
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 3.473,
     "pct": 0.334,
     "by_device": {
      "CPU": {
       "time_ms": 3.473,
       "nodes": 157,
       "calls": 1256
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 3.468,
     "pct": 0.333,
     "by_device": {
      "CPU": {
       "time_ms": 3.468,
       "nodes": 131,
       "calls": 1048
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 3.442,
     "pct": 0.331,
     "by_device": {
      "CPU": {
       "time_ms": 3.442,
       "nodes": 209,
       "calls": 1672
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 3.406,
     "pct": 0.327,
     "by_device": {
      "CPU": {
       "time_ms": 3.406,
       "nodes": 214,
       "calls": 1712
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 2.855,
     "pct": 0.274,
     "by_device": {
      "CPU": {
       "time_ms": 2.855,
       "nodes": 108,
       "calls": 864
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 2.844,
     "pct": 0.273,
     "by_device": {
      "CPU": {
       "time_ms": 2.844,
       "nodes": 157,
       "calls": 1256
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 2.788,
     "pct": 0.268,
     "by_device": {
      "CPU": {
       "time_ms": 2.788,
       "nodes": 132,
       "calls": 1056
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 2.121,
     "pct": 0.204,
     "by_device": {
      "CPU": {
       "time_ms": 2.121,
       "nodes": 171,
       "calls": 1368
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 1.577,
     "pct": 0.152,
     "by_device": {
      "CPU": {
       "time_ms": 1.577,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 1.558,
     "pct": 0.15,
     "by_device": {
      "CPU": {
       "time_ms": 1.558,
       "nodes": 157,
       "calls": 1256
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 1.406,
     "pct": 0.135,
     "by_device": {
      "CPU": {
       "time_ms": 1.406,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.703,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 0.703,
       "nodes": 52,
       "calls": 416
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.642,
     "pct": 0.062,
     "by_device": {
      "CPU": {
       "time_ms": 0.642,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.562,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.562,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.489,
     "pct": 0.047,
     "by_device": {
      "CPU": {
       "time_ms": 0.489,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.36,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.36,
       "nodes": 26,
       "calls": 208
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.105,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.105,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.063,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.063,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.044,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.044,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.042,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.042,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sub",
     "total_ms": 0.02,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.019,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.018,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.016,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_gemma3-1b_pim.simplified.json",
   "source_profile": "ort_gemma3-1b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:59:39",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "gpt2-large",
   "hf_model_id": "openai-community/gpt2-large",
   "hf_model_url": "https://huggingface.co/openai-community/gpt2-large",
   "param_count": 774030080
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.506,
   "TTFT": 0.283,
   "TBT_avg": 0.175,
   "Throughput": 5.31
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1506.487,
     "ttft": 282.796,
     "tbt_avg": 174.813,
     "throughput": 5.31,
     "steps": [
      282.796,
      171.188,
      170.827,
      170.944,
      173.798,
      194.682,
      170.824,
      171.428
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1506.487,
     "min": 1506.487,
     "max": 1506.487,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 282.796,
     "min": 282.796,
     "max": 282.796,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 174.813,
     "min": 174.813,
     "max": 174.813,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 5.31,
     "min": 5.31,
     "max": 5.31,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 174.813,
     "min": 170.824,
     "max": 194.682,
     "std": 8.169,
     "cv_pct": 4.673
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1423.141,
   "total_nodes": 4723,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1397.982,
     "pct": 98.232,
     "nodes": 4721,
     "calls": 37768
    },
    {
     "device": "PIM",
     "time_ms": 25.159,
     "pct": 1.768,
     "nodes": 2,
     "calls": 16
    }
   ],
   "ops": [
    {
     "op": "Gemm",
     "total_ms": 1263.252,
     "pct": 88.765,
     "by_device": {
      "CPU": {
       "time_ms": 1263.252,
       "nodes": 144,
       "calls": 1152
      }
     }
    },
    {
     "op": "MatMul",
     "total_ms": 28.797,
     "pct": 2.023,
     "by_device": {
      "CPU": {
       "time_ms": 4.809,
       "nodes": 72,
       "calls": 576
      },
      "PIM": {
       "time_ms": 23.988,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 21.791,
     "pct": 1.531,
     "by_device": {
      "CPU": {
       "time_ms": 21.791,
       "nodes": 509,
       "calls": 4072
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 15.829,
     "pct": 1.112,
     "by_device": {
      "CPU": {
       "time_ms": 15.829,
       "nodes": 772,
       "calls": 6176
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 12.539,
     "pct": 0.881,
     "by_device": {
      "CPU": {
       "time_ms": 12.539,
       "nodes": 73,
       "calls": 584
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 11.751,
     "pct": 0.826,
     "by_device": {
      "CPU": {
       "time_ms": 11.751,
       "nodes": 550,
       "calls": 4400
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 10.136,
     "pct": 0.712,
     "by_device": {
      "CPU": {
       "time_ms": 10.136,
       "nodes": 218,
       "calls": 1744
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 8.635,
     "pct": 0.607,
     "by_device": {
      "CPU": {
       "time_ms": 8.635,
       "nodes": 438,
       "calls": 3504
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 8.357,
     "pct": 0.587,
     "by_device": {
      "CPU": {
       "time_ms": 8.357,
       "nodes": 730,
       "calls": 5840
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 8.297,
     "pct": 0.583,
     "by_device": {
      "CPU": {
       "time_ms": 8.297,
       "nodes": 218,
       "calls": 1744
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 6.843,
     "pct": 0.481,
     "by_device": {
      "CPU": {
       "time_ms": 6.843,
       "nodes": 180,
       "calls": 1440
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 6.183,
     "pct": 0.434,
     "by_device": {
      "CPU": {
       "time_ms": 6.183,
       "nodes": 184,
       "calls": 1472
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 3.43,
     "pct": 0.241,
     "by_device": {
      "CPU": {
       "time_ms": 3.43,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 3.322,
     "pct": 0.233,
     "by_device": {
      "CPU": {
       "time_ms": 3.322,
       "nodes": 73,
       "calls": 584
      }
     }
    },
    {
     "op": "Squeeze",
     "total_ms": 3.245,
     "pct": 0.228,
     "by_device": {
      "CPU": {
       "time_ms": 3.245,
       "nodes": 145,
       "calls": 1160
      }
     }
    },
    {
     "op": "Split",
     "total_ms": 2.069,
     "pct": 0.145,
     "by_device": {
      "CPU": {
       "time_ms": 2.069,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 1.738,
     "pct": 0.122,
     "by_device": {
      "CPU": {
       "time_ms": 1.738,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.601,
     "pct": 0.112,
     "by_device": {
      "CPU": {
       "time_ms": 1.601,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 1.332,
     "pct": 0.094,
     "by_device": {
      "CPU": {
       "time_ms": 1.332,
       "nodes": 108,
       "calls": 864
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 1.171,
     "pct": 0.082,
     "by_device": {
      "PIM": {
       "time_ms": 1.171,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 1.153,
     "pct": 0.081,
     "by_device": {
      "CPU": {
       "time_ms": 1.153,
       "nodes": 80,
       "calls": 640
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 0.804,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 0.804,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.594,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.594,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.094,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.094,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.052,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.052,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.033,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.032,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.023,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.022,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_gpt2-large_pim.simplified.json",
   "source_profile": "ort_gpt2-large_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:57:47",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "Llama-3.2-1B",
   "hf_model_id": "meta-llama/Llama-3.2-1B",
   "hf_model_url": "https://huggingface.co/meta-llama/Llama-3.2-1B",
   "param_count": 1235814400
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.912,
   "TTFT": 0.249,
   "TBT_avg": 0.095,
   "Throughput": 8.772
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 912.038,
     "ttft": 249.186,
     "tbt_avg": 94.693,
     "throughput": 8.772,
     "steps": [
      249.186,
      93.0,
      93.167,
      93.304,
      95.177,
      94.091,
      97.429,
      96.684
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 912.038,
     "min": 912.038,
     "max": 912.038,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 249.186,
     "min": 249.186,
     "max": 249.186,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 94.693,
     "min": 94.693,
     "max": 94.693,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 8.772,
     "min": 8.772,
     "max": 8.772,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 94.693,
     "min": 93.0,
     "max": 97.429,
     "std": 1.655,
     "cv_pct": 1.748
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 874.537,
   "total_nodes": 2667,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 827.344,
     "pct": 94.604,
     "nodes": 290,
     "calls": 2320
    },
    {
     "device": "CPU",
     "time_ms": 47.193,
     "pct": 5.396,
     "nodes": 2377,
     "calls": 19016
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 783.861,
     "pct": 89.632,
     "by_device": {
      "PIM": {
       "time_ms": 781.503,
       "nodes": 113,
       "calls": 904
      },
      "CPU": {
       "time_ms": 2.358,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 30.629,
     "pct": 3.502,
     "by_device": {
      "PIM": {
       "time_ms": 30.629,
       "nodes": 112,
       "calls": 896
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 15.212,
     "pct": 1.739,
     "by_device": {
      "PIM": {
       "time_ms": 15.212,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 6.818,
     "pct": 0.78,
     "by_device": {
      "CPU": {
       "time_ms": 6.818,
       "nodes": 512,
       "calls": 4096
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 6.405,
     "pct": 0.732,
     "by_device": {
      "CPU": {
       "time_ms": 6.405,
       "nodes": 263,
       "calls": 2104
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 5.358,
     "pct": 0.613,
     "by_device": {
      "CPU": {
       "time_ms": 5.358,
       "nodes": 196,
       "calls": 1568
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 3.365,
     "pct": 0.385,
     "by_device": {
      "CPU": {
       "time_ms": 3.365,
       "nodes": 216,
       "calls": 1728
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 2.675,
     "pct": 0.306,
     "by_device": {
      "CPU": {
       "time_ms": 2.675,
       "nodes": 81,
       "calls": 648
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 2.409,
     "pct": 0.275,
     "by_device": {
      "CPU": {
       "time_ms": 2.409,
       "nodes": 81,
       "calls": 648
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 2.34,
     "pct": 0.268,
     "by_device": {
      "CPU": {
       "time_ms": 2.34,
       "nodes": 116,
       "calls": 928
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.143,
     "pct": 0.245,
     "by_device": {
      "CPU": {
       "time_ms": 2.143,
       "nodes": 249,
       "calls": 1992
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.077,
     "pct": 0.237,
     "by_device": {
      "CPU": {
       "time_ms": 2.077,
       "nodes": 132,
       "calls": 1056
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 1.987,
     "pct": 0.227,
     "by_device": {
      "CPU": {
       "time_ms": 1.987,
       "nodes": 66,
       "calls": 528
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 1.362,
     "pct": 0.156,
     "by_device": {
      "CPU": {
       "time_ms": 1.362,
       "nodes": 34,
       "calls": 272
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.121,
     "pct": 0.128,
     "by_device": {
      "CPU": {
       "time_ms": 1.121,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.1,
     "pct": 0.126,
     "by_device": {
      "CPU": {
       "time_ms": 1.1,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.001,
     "pct": 0.114,
     "by_device": {
      "CPU": {
       "time_ms": 1.001,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 0.99,
     "pct": 0.113,
     "by_device": {
      "CPU": {
       "time_ms": 0.99,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.989,
     "pct": 0.113,
     "by_device": {
      "CPU": {
       "time_ms": 0.989,
       "nodes": 76,
       "calls": 608
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 0.702,
     "pct": 0.08,
     "by_device": {
      "CPU": {
       "time_ms": 0.702,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.472,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.472,
       "nodes": 32,
       "calls": 256
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.404,
     "pct": 0.046,
     "by_device": {
      "CPU": {
       "time_ms": 0.404,
       "nodes": 34,
       "calls": 272
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.361,
     "pct": 0.041,
     "by_device": {
      "CPU": {
       "time_ms": 0.361,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.305,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.305,
       "nodes": 35,
       "calls": 280
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.286,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.286,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.065,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.065,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.033,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.018,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.018,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.017,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.014,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_llama32-1b_pim.simplified.json",
   "source_profile": "ort_llama32-1b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:57:27",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "OLMo-2-0425-1B",
   "hf_model_id": "allenai/OLMo-2-0425-1B",
   "hf_model_url": "https://huggingface.co/allenai/OLMo-2-0425-1B",
   "param_count": 1480000000
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.888,
   "TTFT": 0.264,
   "TBT_avg": 0.089,
   "Throughput": 9.008
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 888.058,
     "ttft": 263.871,
     "tbt_avg": 89.17,
     "throughput": 9.008,
     "steps": [
      263.871,
      88.721,
      89.143,
      88.734,
      88.922,
      89.268,
      89.206,
      90.193
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 888.058,
     "min": 888.058,
     "max": 888.058,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 263.871,
     "min": 263.871,
     "max": 263.871,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 89.17,
     "min": 89.17,
     "max": 89.17,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 9.008,
     "min": 9.008,
     "max": 9.008,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 89.17,
     "min": 88.721,
     "max": 90.193,
     "std": 0.465,
     "cv_pct": 0.521
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 856.511,
   "total_nodes": 2085,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 817.089,
     "pct": 95.397,
     "nodes": 290,
     "calls": 2320
    },
    {
     "device": "CPU",
     "time_ms": 39.422,
     "pct": 4.603,
     "nodes": 1795,
     "calls": 14360
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 773.371,
     "pct": 90.293,
     "by_device": {
      "PIM": {
       "time_ms": 771.362,
       "nodes": 113,
       "calls": 904
      },
      "CPU": {
       "time_ms": 2.009,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 30.782,
     "pct": 3.594,
     "by_device": {
      "PIM": {
       "time_ms": 30.782,
       "nodes": 112,
       "calls": 896
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 14.945,
     "pct": 1.745,
     "by_device": {
      "PIM": {
       "time_ms": 14.945,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 6.818,
     "pct": 0.796,
     "by_device": {
      "CPU": {
       "time_ms": 6.818,
       "nodes": 263,
       "calls": 2104
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 5.313,
     "pct": 0.62,
     "by_device": {
      "CPU": {
       "time_ms": 5.313,
       "nodes": 132,
       "calls": 1056
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 2.757,
     "pct": 0.322,
     "by_device": {
      "CPU": {
       "time_ms": 2.757,
       "nodes": 148,
       "calls": 1184
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 2.675,
     "pct": 0.312,
     "by_device": {
      "CPU": {
       "time_ms": 2.675,
       "nodes": 81,
       "calls": 648
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 2.672,
     "pct": 0.312,
     "by_device": {
      "CPU": {
       "time_ms": 2.672,
       "nodes": 224,
       "calls": 1792
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 2.417,
     "pct": 0.282,
     "by_device": {
      "CPU": {
       "time_ms": 2.417,
       "nodes": 170,
       "calls": 1360
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 2.242,
     "pct": 0.262,
     "by_device": {
      "CPU": {
       "time_ms": 2.242,
       "nodes": 81,
       "calls": 648
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 1.973,
     "pct": 0.23,
     "by_device": {
      "CPU": {
       "time_ms": 1.973,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.586,
     "pct": 0.185,
     "by_device": {
      "CPU": {
       "time_ms": 1.586,
       "nodes": 97,
       "calls": 776
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 1.371,
     "pct": 0.16,
     "by_device": {
      "CPU": {
       "time_ms": 1.371,
       "nodes": 86,
       "calls": 688
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 1.327,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 1.327,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 1.141,
     "pct": 0.133,
     "by_device": {
      "CPU": {
       "time_ms": 1.141,
       "nodes": 34,
       "calls": 272
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.127,
     "pct": 0.132,
     "by_device": {
      "CPU": {
       "time_ms": 1.127,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 1.034,
     "pct": 0.121,
     "by_device": {
      "CPU": {
       "time_ms": 1.034,
       "nodes": 68,
       "calls": 544
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 0.74,
     "pct": 0.086,
     "by_device": {
      "CPU": {
       "time_ms": 0.74,
       "nodes": 87,
       "calls": 696
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.618,
     "pct": 0.072,
     "by_device": {
      "CPU": {
       "time_ms": 0.618,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.602,
     "pct": 0.07,
     "by_device": {
      "CPU": {
       "time_ms": 0.602,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.496,
     "pct": 0.058,
     "by_device": {
      "CPU": {
       "time_ms": 0.496,
       "nodes": 32,
       "calls": 256
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.244,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.244,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.049,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.049,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.042,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.042,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.039,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.039,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.032,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.027,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.019,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.019,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.017,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.016,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_olmo2-1b_pim.simplified.json",
   "source_profile": "ort_olmo2-1b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:58:47",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "opt-1.3b",
   "hf_model_id": "facebook/opt-1.3b",
   "hf_model_url": "https://huggingface.co/facebook/opt-1.3b",
   "param_count": 1315758080
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.972,
   "TTFT": 0.412,
   "TBT_avg": 0.223,
   "Throughput": 4.056
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1972.263,
     "ttft": 412.315,
     "tbt_avg": 222.85,
     "throughput": 4.056,
     "steps": [
      412.315,
      216.969,
      217.064,
      216.695,
      255.553,
      217.639,
      218.259,
      217.769
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1972.263,
     "min": 1972.263,
     "max": 1972.263,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 412.315,
     "min": 412.315,
     "max": 412.315,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 222.85,
     "min": 222.85,
     "max": 222.85,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 4.056,
     "min": 4.056,
     "max": 4.056,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 222.85,
     "min": 216.695,
     "max": 255.553,
     "std": 13.36,
     "cv_pct": 5.995
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1937.064,
   "total_nodes": 1922,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1521.133,
     "pct": 78.528,
     "nodes": 1584,
     "calls": 12672
    },
    {
     "device": "PIM",
     "time_ms": 415.931,
     "pct": 21.472,
     "nodes": 338,
     "calls": 2704
    }
   ],
   "ops": [
    {
     "op": "Gemm",
     "total_ms": 1457.542,
     "pct": 75.245,
     "by_device": {
      "CPU": {
       "time_ms": 1457.542,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "MatMul",
     "total_ms": 359.285,
     "pct": 18.548,
     "by_device": {
      "PIM": {
       "time_ms": 355.32,
       "nodes": 97,
       "calls": 776
      },
      "CPU": {
       "time_ms": 3.965,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 26.012,
     "pct": 1.343,
     "by_device": {
      "PIM": {
       "time_ms": 26.012,
       "nodes": 96,
       "calls": 768
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 25.883,
     "pct": 1.336,
     "by_device": {
      "CPU": {
       "time_ms": 2.575,
       "nodes": 79,
       "calls": 632
      },
      "PIM": {
       "time_ms": 23.308,
       "nodes": 96,
       "calls": 768
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 13.915,
     "pct": 0.718,
     "by_device": {
      "CPU": {
       "time_ms": 13.915,
       "nodes": 197,
       "calls": 1576
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 12.091,
     "pct": 0.624,
     "by_device": {
      "CPU": {
       "time_ms": 12.091,
       "nodes": 49,
       "calls": 392
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 11.291,
     "pct": 0.583,
     "by_device": {
      "PIM": {
       "time_ms": 11.291,
       "nodes": 49,
       "calls": 392
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 5.471,
     "pct": 0.282,
     "by_device": {
      "CPU": {
       "time_ms": 5.471,
       "nodes": 120,
       "calls": 960
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 4.913,
     "pct": 0.254,
     "by_device": {
      "CPU": {
       "time_ms": 4.913,
       "nodes": 256,
       "calls": 2048
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 4.502,
     "pct": 0.232,
     "by_device": {
      "CPU": {
       "time_ms": 4.502,
       "nodes": 78,
       "calls": 624
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 3.334,
     "pct": 0.172,
     "by_device": {
      "CPU": {
       "time_ms": 3.334,
       "nodes": 155,
       "calls": 1240
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.802,
     "pct": 0.145,
     "by_device": {
      "CPU": {
       "time_ms": 2.802,
       "nodes": 151,
       "calls": 1208
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 2.209,
     "pct": 0.114,
     "by_device": {
      "CPU": {
       "time_ms": 2.209,
       "nodes": 54,
       "calls": 432
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.049,
     "pct": 0.106,
     "by_device": {
      "CPU": {
       "time_ms": 2.049,
       "nodes": 185,
       "calls": 1480
      }
     }
    },
    {
     "op": "Relu",
     "total_ms": 1.774,
     "pct": 0.092,
     "by_device": {
      "CPU": {
       "time_ms": 1.774,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.741,
     "pct": 0.09,
     "by_device": {
      "CPU": {
       "time_ms": 1.741,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Identity",
     "total_ms": 0.755,
     "pct": 0.039,
     "by_device": {
      "CPU": {
       "time_ms": 0.755,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Squeeze",
     "total_ms": 0.495,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.495,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.221,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.221,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.164,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.164,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.159,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.159,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.151,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.151,
       "nodes": 11,
       "calls": 88
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.117,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.117,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.088,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.088,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.049,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.049,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.027,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.024,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_opt-1.3b_pim.simplified.json",
   "source_profile": "ort_opt-1.3b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:54:31",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "opt-350m",
   "hf_model_id": "facebook/opt-350m",
   "hf_model_url": "https://huggingface.co/facebook/opt-350m",
   "param_count": 331196416
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.676,
   "TTFT": 0.123,
   "TBT_avg": 0.079,
   "Throughput": 11.842
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 675.551,
     "ttft": 123.248,
     "tbt_avg": 78.9,
     "throughput": 11.842,
     "steps": [
      123.248,
      74.525,
      74.06,
      73.699,
      76.168,
      105.925,
      73.899,
      74.027
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 675.551,
     "min": 675.551,
     "max": 675.551,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 123.248,
     "min": 123.248,
     "max": 123.248,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 78.9,
     "min": 78.9,
     "max": 78.9,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 11.842,
     "min": 11.842,
     "max": 11.842,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 78.9,
     "min": 73.699,
     "max": 105.925,
     "std": 11.059,
     "cv_pct": 14.017
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 642.208,
   "total_nodes": 1926,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 413.178,
     "pct": 64.337,
     "nodes": 1582,
     "calls": 12656
    },
    {
     "device": "PIM",
     "time_ms": 229.03,
     "pct": 35.663,
     "nodes": 344,
     "calls": 2752
    }
   ],
   "ops": [
    {
     "op": "Gemm",
     "total_ms": 368.857,
     "pct": 57.436,
     "by_device": {
      "CPU": {
       "time_ms": 368.857,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "MatMul",
     "total_ms": 168.897,
     "pct": 26.299,
     "by_device": {
      "PIM": {
       "time_ms": 166.632,
       "nodes": 99,
       "calls": 792
      },
      "CPU": {
       "time_ms": 2.265,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 28.486,
     "pct": 4.436,
     "by_device": {
      "CPU": {
       "time_ms": 1.95,
       "nodes": 79,
       "calls": 632
      },
      "PIM": {
       "time_ms": 26.536,
       "nodes": 96,
       "calls": 768
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 23.387,
     "pct": 3.642,
     "by_device": {
      "PIM": {
       "time_ms": 23.387,
       "nodes": 98,
       "calls": 784
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 12.475,
     "pct": 1.943,
     "by_device": {
      "PIM": {
       "time_ms": 12.475,
       "nodes": 51,
       "calls": 408
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 8.992,
     "pct": 1.4,
     "by_device": {
      "CPU": {
       "time_ms": 8.992,
       "nodes": 197,
       "calls": 1576
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 6.607,
     "pct": 1.029,
     "by_device": {
      "CPU": {
       "time_ms": 6.607,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 4.607,
     "pct": 0.717,
     "by_device": {
      "CPU": {
       "time_ms": 4.607,
       "nodes": 256,
       "calls": 2048
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 3.588,
     "pct": 0.559,
     "by_device": {
      "CPU": {
       "time_ms": 3.588,
       "nodes": 120,
       "calls": 960
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 2.928,
     "pct": 0.456,
     "by_device": {
      "CPU": {
       "time_ms": 2.928,
       "nodes": 155,
       "calls": 1240
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 2.811,
     "pct": 0.438,
     "by_device": {
      "CPU": {
       "time_ms": 2.811,
       "nodes": 78,
       "calls": 624
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.641,
     "pct": 0.411,
     "by_device": {
      "CPU": {
       "time_ms": 2.641,
       "nodes": 151,
       "calls": 1208
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 1.994,
     "pct": 0.31,
     "by_device": {
      "CPU": {
       "time_ms": 1.994,
       "nodes": 185,
       "calls": 1480
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 1.972,
     "pct": 0.307,
     "by_device": {
      "CPU": {
       "time_ms": 1.972,
       "nodes": 54,
       "calls": 432
      }
     }
    },
    {
     "op": "Relu",
     "total_ms": 1.1,
     "pct": 0.171,
     "by_device": {
      "CPU": {
       "time_ms": 1.1,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.931,
     "pct": 0.145,
     "by_device": {
      "CPU": {
       "time_ms": 0.931,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Identity",
     "total_ms": 0.618,
     "pct": 0.096,
     "by_device": {
      "CPU": {
       "time_ms": 0.618,
       "nodes": 47,
       "calls": 376
      }
     }
    },
    {
     "op": "Squeeze",
     "total_ms": 0.445,
     "pct": 0.069,
     "by_device": {
      "CPU": {
       "time_ms": 0.445,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.188,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.188,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.144,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.144,
       "nodes": 11,
       "calls": 88
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.139,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.139,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.138,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.138,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.096,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.096,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.078,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.078,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.042,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.042,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.024,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.023,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_opt-350m_pim.simplified.json",
   "source_profile": "ort_opt-350m_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:58:25",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "phi-1_5",
   "hf_model_id": "microsoft/phi-1_5",
   "hf_model_url": "https://huggingface.co/microsoft/phi-1_5",
   "param_count": 1418270720
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.048,
   "TTFT": 0.306,
   "TBT_avg": 0.106,
   "Throughput": 7.634
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1047.943,
     "ttft": 306.491,
     "tbt_avg": 105.922,
     "throughput": 7.634,
     "steps": [
      306.491,
      104.742,
      105.072,
      104.858,
      106.269,
      104.028,
      104.865,
      111.618
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1047.943,
     "min": 1047.943,
     "max": 1047.943,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 306.491,
     "min": 306.491,
     "max": 306.491,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 105.922,
     "min": 105.922,
     "max": 105.922,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 7.634,
     "min": 7.634,
     "max": 7.634,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 105.922,
     "min": 104.028,
     "max": 111.618,
     "std": 2.406,
     "cv_pct": 2.271
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1008.493,
   "total_nodes": 2535,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 945.995,
     "pct": 93.803,
     "nodes": 512,
     "calls": 4096
    },
    {
     "device": "CPU",
     "time_ms": 62.498,
     "pct": 6.197,
     "nodes": 2023,
     "calls": 16184
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 858.022,
     "pct": 85.08,
     "by_device": {
      "PIM": {
       "time_ms": 854.463,
       "nodes": 145,
       "calls": 1160
      },
      "CPU": {
       "time_ms": 3.559,
       "nodes": 49,
       "calls": 392
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 40.866,
     "pct": 4.052,
     "by_device": {
      "PIM": {
       "time_ms": 36.875,
       "nodes": 146,
       "calls": 1168
      },
      "CPU": {
       "time_ms": 3.991,
       "nodes": 170,
       "calls": 1360
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 37.348,
     "pct": 3.703,
     "by_device": {
      "PIM": {
       "time_ms": 37.348,
       "nodes": 145,
       "calls": 1160
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 17.026,
     "pct": 1.688,
     "by_device": {
      "PIM": {
       "time_ms": 17.026,
       "nodes": 75,
       "calls": 600
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 12.221,
     "pct": 1.212,
     "by_device": {
      "CPU": {
       "time_ms": 12.221,
       "nodes": 244,
       "calls": 1952
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 7.672,
     "pct": 0.761,
     "by_device": {
      "CPU": {
       "time_ms": 7.389,
       "nodes": 244,
       "calls": 1952
      },
      "PIM": {
       "time_ms": 0.283,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 6.516,
     "pct": 0.646,
     "by_device": {
      "CPU": {
       "time_ms": 6.516,
       "nodes": 217,
       "calls": 1736
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 5.854,
     "pct": 0.58,
     "by_device": {
      "CPU": {
       "time_ms": 5.854,
       "nodes": 25,
       "calls": 200
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 4.221,
     "pct": 0.419,
     "by_device": {
      "CPU": {
       "time_ms": 4.221,
       "nodes": 328,
       "calls": 2624
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.068,
     "pct": 0.403,
     "by_device": {
      "CPU": {
       "time_ms": 4.068,
       "nodes": 121,
       "calls": 968
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 2.688,
     "pct": 0.267,
     "by_device": {
      "CPU": {
       "time_ms": 2.688,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 2.096,
     "pct": 0.208,
     "by_device": {
      "CPU": {
       "time_ms": 2.096,
       "nodes": 128,
       "calls": 1024
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 1.771,
     "pct": 0.176,
     "by_device": {
      "CPU": {
       "time_ms": 1.771,
       "nodes": 50,
       "calls": 400
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 1.691,
     "pct": 0.168,
     "by_device": {
      "CPU": {
       "time_ms": 1.691,
       "nodes": 100,
       "calls": 800
      }
     }
    },
    {
     "op": "Tanh",
     "total_ms": 1.576,
     "pct": 0.156,
     "by_device": {
      "CPU": {
       "time_ms": 1.576,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.441,
     "pct": 0.143,
     "by_device": {
      "CPU": {
       "time_ms": 1.441,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 1.181,
     "pct": 0.117,
     "by_device": {
      "CPU": {
       "time_ms": 1.181,
       "nodes": 129,
       "calls": 1032
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 0.836,
     "pct": 0.083,
     "by_device": {
      "CPU": {
       "time_ms": 0.836,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.669,
     "pct": 0.066,
     "by_device": {
      "CPU": {
       "time_ms": 0.669,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.363,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 0.363,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.109,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.109,
       "nodes": 10,
       "calls": 80
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.052,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.052,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.046,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.046,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.036,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.032,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.03,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.018,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.017,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.014,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.013,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.013,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_phi-1_5_pim.simplified.json",
   "source_profile": "ort_phi-1_5_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:56:21",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "pythia-1b",
   "hf_model_id": "EleutherAI/pythia-1b",
   "hf_model_url": "https://huggingface.co/EleutherAI/pythia-1b",
   "param_count": 1011781632
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.656,
   "TTFT": 0.207,
   "TBT_avg": 0.064,
   "Throughput": 12.203
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 655.574,
     "ttft": 206.539,
     "tbt_avg": 64.148,
     "throughput": 12.203,
     "steps": [
      206.539,
      64.599,
      63.695,
      64.141,
      63.924,
      63.939,
      64.042,
      64.695
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 655.574,
     "min": 655.574,
     "max": 655.574,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 206.539,
     "min": 206.539,
     "max": 206.539,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 64.148,
     "min": 64.148,
     "max": 64.148,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 12.203,
     "min": 12.203,
     "max": 12.203,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 64.148,
     "min": 63.695,
     "max": 64.695,
     "std": 0.341,
     "cv_pct": 0.531
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 630.332,
   "total_nodes": 1687,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 589.543,
     "pct": 93.529,
     "nodes": 245,
     "calls": 1960
    },
    {
     "device": "CPU",
     "time_ms": 40.789,
     "pct": 6.471,
     "nodes": 1442,
     "calls": 11536
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 544.834,
     "pct": 86.436,
     "by_device": {
      "CPU": {
       "time_ms": 1.867,
       "nodes": 33,
       "calls": 264
      },
      "PIM": {
       "time_ms": 542.967,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 18.451,
     "pct": 2.927,
     "by_device": {
      "PIM": {
       "time_ms": 18.451,
       "nodes": 65,
       "calls": 520
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 15.835,
     "pct": 2.512,
     "by_device": {
      "PIM": {
       "time_ms": 13.13,
       "nodes": 48,
       "calls": 384
      },
      "CPU": {
       "time_ms": 2.705,
       "nodes": 131,
       "calls": 1048
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 14.686,
     "pct": 2.33,
     "by_device": {
      "PIM": {
       "time_ms": 14.686,
       "nodes": 66,
       "calls": 528
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 7.723,
     "pct": 1.225,
     "by_device": {
      "CPU": {
       "time_ms": 7.723,
       "nodes": 33,
       "calls": 264
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 5.558,
     "pct": 0.882,
     "by_device": {
      "CPU": {
       "time_ms": 5.558,
       "nodes": 132,
       "calls": 1056
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 4.95,
     "pct": 0.785,
     "by_device": {
      "CPU": {
       "time_ms": 4.95,
       "nodes": 193,
       "calls": 1544
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 4.343,
     "pct": 0.689,
     "by_device": {
      "CPU": {
       "time_ms": 4.034,
       "nodes": 180,
       "calls": 1440
      },
      "PIM": {
       "time_ms": 0.309,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 2.53,
     "pct": 0.401,
     "by_device": {
      "CPU": {
       "time_ms": 2.53,
       "nodes": 224,
       "calls": 1792
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 2.149,
     "pct": 0.341,
     "by_device": {
      "CPU": {
       "time_ms": 2.149,
       "nodes": 49,
       "calls": 392
      }
     }
    },
    {
     "op": "Erf",
     "total_ms": 2.067,
     "pct": 0.328,
     "by_device": {
      "CPU": {
       "time_ms": 2.067,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 1.596,
     "pct": 0.253,
     "by_device": {
      "CPU": {
       "time_ms": 1.596,
       "nodes": 105,
       "calls": 840
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.591,
     "pct": 0.252,
     "by_device": {
      "CPU": {
       "time_ms": 1.591,
       "nodes": 64,
       "calls": 512
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.96,
     "pct": 0.152,
     "by_device": {
      "CPU": {
       "time_ms": 0.96,
       "nodes": 34,
       "calls": 272
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 0.917,
     "pct": 0.145,
     "by_device": {
      "CPU": {
       "time_ms": 0.917,
       "nodes": 106,
       "calls": 848
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 0.541,
     "pct": 0.086,
     "by_device": {
      "CPU": {
       "time_ms": 0.541,
       "nodes": 36,
       "calls": 288
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.431,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 0.431,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.369,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 0.369,
       "nodes": 32,
       "calls": 256
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.213,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.213,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Identity",
     "total_ms": 0.21,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.21,
       "nodes": 16,
       "calls": 128
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.112,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.112,
       "nodes": 10,
       "calls": 80
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.054,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.054,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.042,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.042,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.035,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.034,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.032,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.018,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.017,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.017,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.017,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_pythia-1b_pim.simplified.json",
   "source_profile": "ort_pythia-1b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 18:02:00",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "Qwen2.5-1.5B",
   "hf_model_id": "Qwen/Qwen2.5-1.5B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen2.5-1.5B",
   "param_count": 1543714304
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.447,
   "TTFT": 0.398,
   "TBT_avg": 0.15,
   "Throughput": 5.53
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1446.59,
     "ttft": 398.308,
     "tbt_avg": 149.755,
     "throughput": 5.53,
     "steps": [
      398.308,
      150.688,
      150.09,
      149.137,
      148.715,
      151.044,
      148.632,
      149.976
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1446.59,
     "min": 1446.59,
     "max": 1446.59,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 398.308,
     "min": 398.308,
     "max": 398.308,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 149.755,
     "min": 149.755,
     "max": 149.755,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 5.53,
     "min": 5.53,
     "max": 5.53,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 149.755,
     "min": 148.632,
     "max": 151.044,
     "std": 0.88,
     "cv_pct": 0.587
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1374.396,
   "total_nodes": 4683,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 1295.308,
     "pct": 94.246,
     "nodes": 590,
     "calls": 4720
    },
    {
     "device": "CPU",
     "time_ms": 79.088,
     "pct": 5.754,
     "nodes": 4093,
     "calls": 32744
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 1199.258,
     "pct": 87.257,
     "by_device": {
      "PIM": {
       "time_ms": 1196.284,
       "nodes": 197,
       "calls": 1576
      },
      "CPU": {
       "time_ms": 2.974,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 55.337,
     "pct": 4.026,
     "by_device": {
      "PIM": {
       "time_ms": 55.337,
       "nodes": 196,
       "calls": 1568
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 24.4,
     "pct": 1.775,
     "by_device": {
      "PIM": {
       "time_ms": 24.4,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 23.219,
     "pct": 1.689,
     "by_device": {
      "CPU": {
       "time_ms": 3.932,
       "nodes": 200,
       "calls": 1600
      },
      "PIM": {
       "time_ms": 19.287,
       "nodes": 84,
       "calls": 672
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 12.867,
     "pct": 0.936,
     "by_device": {
      "CPU": {
       "time_ms": 12.867,
       "nodes": 884,
       "calls": 7072
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 11.623,
     "pct": 0.846,
     "by_device": {
      "CPU": {
       "time_ms": 11.623,
       "nodes": 455,
       "calls": 3640
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 8.362,
     "pct": 0.608,
     "by_device": {
      "CPU": {
       "time_ms": 8.362,
       "nodes": 340,
       "calls": 2720
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 6.08,
     "pct": 0.442,
     "by_device": {
      "CPU": {
       "time_ms": 6.08,
       "nodes": 372,
       "calls": 2976
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 4.17,
     "pct": 0.303,
     "by_device": {
      "CPU": {
       "time_ms": 4.17,
       "nodes": 141,
       "calls": 1128
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 3.891,
     "pct": 0.283,
     "by_device": {
      "CPU": {
       "time_ms": 3.891,
       "nodes": 429,
       "calls": 3432
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 3.681,
     "pct": 0.268,
     "by_device": {
      "CPU": {
       "time_ms": 3.681,
       "nodes": 141,
       "calls": 1128
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 3.677,
     "pct": 0.268,
     "by_device": {
      "CPU": {
       "time_ms": 3.677,
       "nodes": 228,
       "calls": 1824
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 3.196,
     "pct": 0.233,
     "by_device": {
      "CPU": {
       "time_ms": 3.196,
       "nodes": 114,
       "calls": 912
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 2.262,
     "pct": 0.165,
     "by_device": {
      "CPU": {
       "time_ms": 2.262,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.944,
     "pct": 0.141,
     "by_device": {
      "CPU": {
       "time_ms": 1.944,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 1.847,
     "pct": 0.134,
     "by_device": {
      "CPU": {
       "time_ms": 1.847,
       "nodes": 58,
       "calls": 464
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 1.619,
     "pct": 0.118,
     "by_device": {
      "CPU": {
       "time_ms": 1.619,
       "nodes": 124,
       "calls": 992
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 1.527,
     "pct": 0.111,
     "by_device": {
      "CPU": {
       "time_ms": 1.527,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 1.159,
     "pct": 0.084,
     "by_device": {
      "CPU": {
       "time_ms": 1.159,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 0.921,
     "pct": 0.067,
     "by_device": {
      "CPU": {
       "time_ms": 0.921,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.855,
     "pct": 0.062,
     "by_device": {
      "CPU": {
       "time_ms": 0.855,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.701,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 0.701,
       "nodes": 58,
       "calls": 464
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.638,
     "pct": 0.046,
     "by_device": {
      "CPU": {
       "time_ms": 0.638,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.535,
     "pct": 0.039,
     "by_device": {
      "CPU": {
       "time_ms": 0.535,
       "nodes": 59,
       "calls": 472
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.446,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.446,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.07,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.07,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.034,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.023,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.019,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.019,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_qwen2.5-1.5b_pim.simplified.json",
   "source_profile": "ort_qwen2.5-1.5b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:55:02",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "Qwen3-0.6B",
   "hf_model_id": "Qwen/Qwen3-0.6B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen3-0.6B",
   "param_count": 596049920
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.752,
   "TTFT": 0.149,
   "TBT_avg": 0.086,
   "Throughput": 10.632
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 752.452,
     "ttft": 149.14,
     "tbt_avg": 86.187,
     "throughput": 10.632,
     "steps": [
      149.14,
      86.216,
      86.456,
      86.926,
      88.107,
      84.454,
      85.564,
      85.589
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 752.452,
     "min": 752.452,
     "max": 752.452,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 149.14,
     "min": 149.14,
     "max": 149.14,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 86.187,
     "min": 86.187,
     "max": 86.187,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 10.632,
     "min": 10.632,
     "max": 10.632,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 86.187,
     "min": 84.454,
     "max": 88.107,
     "std": 1.073,
     "cv_pct": 1.245
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 678.52,
   "total_nodes": 5103,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 593.183,
     "pct": 87.423,
     "nodes": 506,
     "calls": 4048
    },
    {
     "device": "CPU",
     "time_ms": 85.337,
     "pct": 12.577,
     "nodes": 4597,
     "calls": 36776
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 527.189,
     "pct": 77.697,
     "by_device": {
      "PIM": {
       "time_ms": 523.782,
       "nodes": 197,
       "calls": 1576
      },
      "CPU": {
       "time_ms": 3.407,
       "nodes": 57,
       "calls": 456
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 47.948,
     "pct": 7.067,
     "by_device": {
      "PIM": {
       "time_ms": 47.948,
       "nodes": 196,
       "calls": 1568
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 21.453,
     "pct": 3.162,
     "by_device": {
      "PIM": {
       "time_ms": 21.453,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 11.824,
     "pct": 1.743,
     "by_device": {
      "CPU": {
       "time_ms": 11.824,
       "nodes": 567,
       "calls": 4536
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 11.698,
     "pct": 1.724,
     "by_device": {
      "CPU": {
       "time_ms": 11.698,
       "nodes": 884,
       "calls": 7072
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 9.575,
     "pct": 1.411,
     "by_device": {
      "CPU": {
       "time_ms": 9.575,
       "nodes": 340,
       "calls": 2720
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 5.79,
     "pct": 0.853,
     "by_device": {
      "CPU": {
       "time_ms": 5.79,
       "nodes": 372,
       "calls": 2976
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 4.833,
     "pct": 0.712,
     "by_device": {
      "CPU": {
       "time_ms": 4.833,
       "nodes": 256,
       "calls": 2048
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.296,
     "pct": 0.633,
     "by_device": {
      "CPU": {
       "time_ms": 4.296,
       "nodes": 141,
       "calls": 1128
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 4.031,
     "pct": 0.594,
     "by_device": {
      "CPU": {
       "time_ms": 4.031,
       "nodes": 141,
       "calls": 1128
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 3.746,
     "pct": 0.552,
     "by_device": {
      "CPU": {
       "time_ms": 3.746,
       "nodes": 429,
       "calls": 3432
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 3.623,
     "pct": 0.534,
     "by_device": {
      "CPU": {
       "time_ms": 3.623,
       "nodes": 228,
       "calls": 1824
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 3.218,
     "pct": 0.474,
     "by_device": {
      "CPU": {
       "time_ms": 3.218,
       "nodes": 114,
       "calls": 912
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 3.018,
     "pct": 0.445,
     "by_device": {
      "CPU": {
       "time_ms": 3.018,
       "nodes": 236,
       "calls": 1888
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 2.958,
     "pct": 0.436,
     "by_device": {
      "CPU": {
       "time_ms": 2.958,
       "nodes": 169,
       "calls": 1352
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 2.81,
     "pct": 0.414,
     "by_device": {
      "CPU": {
       "time_ms": 2.81,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 2.366,
     "pct": 0.349,
     "by_device": {
      "CPU": {
       "time_ms": 2.366,
       "nodes": 58,
       "calls": 464
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 2.18,
     "pct": 0.321,
     "by_device": {
      "CPU": {
       "time_ms": 2.18,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 1.145,
     "pct": 0.169,
     "by_device": {
      "CPU": {
       "time_ms": 1.145,
       "nodes": 113,
       "calls": 904
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.091,
     "pct": 0.161,
     "by_device": {
      "CPU": {
       "time_ms": 1.091,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.076,
     "pct": 0.159,
     "by_device": {
      "CPU": {
       "time_ms": 1.076,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.845,
     "pct": 0.125,
     "by_device": {
      "CPU": {
       "time_ms": 0.845,
       "nodes": 56,
       "calls": 448
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.658,
     "pct": 0.097,
     "by_device": {
      "CPU": {
       "time_ms": 0.658,
       "nodes": 58,
       "calls": 464
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.506,
     "pct": 0.075,
     "by_device": {
      "CPU": {
       "time_ms": 0.506,
       "nodes": 59,
       "calls": 472
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.475,
     "pct": 0.07,
     "by_device": {
      "CPU": {
       "time_ms": 0.475,
       "nodes": 28,
       "calls": 224
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.06,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.034,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.022,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.019,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.017,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.016,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_qwen3-0.6b_pim.simplified.json",
   "source_profile": "ort_qwen3-0.6b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 18:00:01",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "stablelm-2-1_6b",
   "hf_model_id": "stabilityai/stablelm-2-1_6b",
   "hf_model_url": "https://huggingface.co/stabilityai/stablelm-2-1_6b",
   "param_count": 1644417024
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.151,
   "TTFT": 0.322,
   "TBT_avg": 0.118,
   "Throughput": 6.95
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1151.057,
     "ttft": 322.409,
     "tbt_avg": 118.378,
     "throughput": 6.95,
     "steps": [
      322.409,
      117.183,
      116.641,
      116.058,
      118.897,
      124.314,
      117.944,
      117.611
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1151.057,
     "min": 1151.057,
     "max": 1151.057,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 322.409,
     "min": 322.409,
     "max": 322.409,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 118.378,
     "min": 118.378,
     "max": 118.378,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 6.95,
     "min": 6.95,
     "max": 6.95,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 118.378,
     "min": 116.058,
     "max": 124.314,
     "std": 2.567,
     "cv_pct": 2.168
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1109.418,
   "total_nodes": 2656,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 1042.563,
     "pct": 93.974,
     "nodes": 506,
     "calls": 4048
    },
    {
     "device": "CPU",
     "time_ms": 66.855,
     "pct": 6.026,
     "nodes": 2150,
     "calls": 17200
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 955.984,
     "pct": 86.17,
     "by_device": {
      "PIM": {
       "time_ms": 952.376,
       "nodes": 169,
       "calls": 1352
      },
      "CPU": {
       "time_ms": 3.608,
       "nodes": 49,
       "calls": 392
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 45.254,
     "pct": 4.079,
     "by_device": {
      "PIM": {
       "time_ms": 45.254,
       "nodes": 168,
       "calls": 1344
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 24.344,
     "pct": 2.194,
     "by_device": {
      "PIM": {
       "time_ms": 24.344,
       "nodes": 97,
       "calls": 776
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 23.172,
     "pct": 2.089,
     "by_device": {
      "PIM": {
       "time_ms": 20.589,
       "nodes": 72,
       "calls": 576
      },
      "CPU": {
       "time_ms": 2.583,
       "nodes": 125,
       "calls": 1000
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 13.973,
     "pct": 1.259,
     "by_device": {
      "CPU": {
       "time_ms": 13.973,
       "nodes": 246,
       "calls": 1968
      }
     }
    },
    {
     "op": "LayerNormalization",
     "total_ms": 11.489,
     "pct": 1.036,
     "by_device": {
      "CPU": {
       "time_ms": 11.489,
       "nodes": 49,
       "calls": 392
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 7.651,
     "pct": 0.69,
     "by_device": {
      "CPU": {
       "time_ms": 7.651,
       "nodes": 246,
       "calls": 1968
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 5.536,
     "pct": 0.499,
     "by_device": {
      "CPU": {
       "time_ms": 5.536,
       "nodes": 201,
       "calls": 1608
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 4.895,
     "pct": 0.441,
     "by_device": {
      "CPU": {
       "time_ms": 4.895,
       "nodes": 331,
       "calls": 2648
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 4.068,
     "pct": 0.367,
     "by_device": {
      "CPU": {
       "time_ms": 4.068,
       "nodes": 121,
       "calls": 968
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 2.243,
     "pct": 0.202,
     "by_device": {
      "CPU": {
       "time_ms": 2.243,
       "nodes": 130,
       "calls": 1040
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 1.836,
     "pct": 0.165,
     "by_device": {
      "CPU": {
       "time_ms": 1.836,
       "nodes": 103,
       "calls": 824
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 1.647,
     "pct": 0.148,
     "by_device": {
      "CPU": {
       "time_ms": 1.647,
       "nodes": 162,
       "calls": 1296
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.471,
     "pct": 0.133,
     "by_device": {
      "CPU": {
       "time_ms": 1.471,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.349,
     "pct": 0.122,
     "by_device": {
      "CPU": {
       "time_ms": 1.349,
       "nodes": 24,
       "calls": 192
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.253,
     "pct": 0.113,
     "by_device": {
      "CPU": {
       "time_ms": 1.253,
       "nodes": 72,
       "calls": 576
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.713,
     "pct": 0.064,
     "by_device": {
      "CPU": {
       "time_ms": 0.713,
       "nodes": 48,
       "calls": 384
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.644,
     "pct": 0.058,
     "by_device": {
      "CPU": {
       "time_ms": 0.644,
       "nodes": 72,
       "calls": 576
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 0.606,
     "pct": 0.055,
     "by_device": {
      "CPU": {
       "time_ms": 0.606,
       "nodes": 61,
       "calls": 488
      }
     }
    },
    {
     "op": "Identity",
     "total_ms": 0.504,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.504,
       "nodes": 47,
       "calls": 376
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 0.215,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.215,
       "nodes": 7,
       "calls": 56
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 0.141,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.141,
       "nodes": 7,
       "calls": 56
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.127,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.127,
       "nodes": 6,
       "calls": 48
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.103,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.103,
       "nodes": 7,
       "calls": 56
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.082,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.082,
       "nodes": 7,
       "calls": 56
      }
     }
    },
    {
     "op": "ScatterND",
     "total_ms": 0.038,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Greater",
     "total_ms": 0.024,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Trilu",
     "total_ms": 0.023,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.018,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.015,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.015,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_stablelm2-1.6b_pim.simplified.json",
   "source_profile": "ort_stablelm2-1.6b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:58:05",
   "framework": "onnxruntime",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "TinyLlama-1.1B-Chat-v1.0",
   "hf_model_id": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "hf_model_url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "param_count": 1100048384
  },
  "test_conditions": {
   "dtype": "float32",
   "input_tokens": 16,
   "batch_size": 1,
   "output_tokens": 8,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.018,
   "TTFT": 0.258,
   "TBT_avg": 0.109,
   "Throughput": 7.855
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 8,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1018.41,
     "ttft": 258.022,
     "tbt_avg": 108.627,
     "throughput": 7.855,
     "steps": [
      258.022,
      107.682,
      107.663,
      108.182,
      108.01,
      107.356,
      113.154,
      108.341
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1018.41,
     "min": 1018.41,
     "max": 1018.41,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 258.022,
     "min": 258.022,
     "max": 258.022,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 108.627,
     "min": 108.627,
     "max": 108.627,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 7.855,
     "min": 7.855,
     "max": 7.855,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 108.627,
     "min": 107.356,
     "max": 113.154,
     "std": 1.874,
     "cv_pct": 1.725
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 962.956,
   "total_nodes": 3633,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 901.702,
     "pct": 93.639,
     "nodes": 398,
     "calls": 3184
    },
    {
     "device": "CPU",
     "time_ms": 61.254,
     "pct": 6.361,
     "nodes": 3235,
     "calls": 25880
    }
   ],
   "ops": [
    {
     "op": "MatMul",
     "total_ms": 844.71,
     "pct": 87.721,
     "by_device": {
      "PIM": {
       "time_ms": 841.367,
       "nodes": 155,
       "calls": 1240
      },
      "CPU": {
       "time_ms": 3.343,
       "nodes": 45,
       "calls": 360
      }
     }
    },
    {
     "op": "MemcpyToHost",
     "total_ms": 41.421,
     "pct": 4.301,
     "by_device": {
      "PIM": {
       "time_ms": 41.421,
       "nodes": 154,
       "calls": 1232
      }
     }
    },
    {
     "op": "MemcpyFromHost",
     "total_ms": 18.914,
     "pct": 1.964,
     "by_device": {
      "PIM": {
       "time_ms": 18.914,
       "nodes": 89,
       "calls": 712
      }
     }
    },
    {
     "op": "Unsqueeze",
     "total_ms": 8.904,
     "pct": 0.925,
     "by_device": {
      "CPU": {
       "time_ms": 8.904,
       "nodes": 698,
       "calls": 5584
      }
     }
    },
    {
     "op": "Mul",
     "total_ms": 7.978,
     "pct": 0.828,
     "by_device": {
      "CPU": {
       "time_ms": 7.978,
       "nodes": 359,
       "calls": 2872
      }
     }
    },
    {
     "op": "Concat",
     "total_ms": 6.53,
     "pct": 0.678,
     "by_device": {
      "CPU": {
       "time_ms": 6.53,
       "nodes": 268,
       "calls": 2144
      }
     }
    },
    {
     "op": "Gather",
     "total_ms": 4.44,
     "pct": 0.461,
     "by_device": {
      "CPU": {
       "time_ms": 4.44,
       "nodes": 294,
       "calls": 2352
      }
     }
    },
    {
     "op": "Transpose",
     "total_ms": 3.496,
     "pct": 0.363,
     "by_device": {
      "CPU": {
       "time_ms": 3.496,
       "nodes": 111,
       "calls": 888
      }
     }
    },
    {
     "op": "Slice",
     "total_ms": 3.153,
     "pct": 0.327,
     "by_device": {
      "CPU": {
       "time_ms": 3.153,
       "nodes": 111,
       "calls": 888
      }
     }
    },
    {
     "op": "Add",
     "total_ms": 3.13,
     "pct": 0.325,
     "by_device": {
      "CPU": {
       "time_ms": 3.13,
       "nodes": 158,
       "calls": 1264
      }
     }
    },
    {
     "op": "Shape",
     "total_ms": 2.967,
     "pct": 0.308,
     "by_device": {
      "CPU": {
       "time_ms": 2.967,
       "nodes": 339,
       "calls": 2712
      }
     }
    },
    {
     "op": "Reshape",
     "total_ms": 2.81,
     "pct": 0.292,
     "by_device": {
      "CPU": {
       "time_ms": 2.81,
       "nodes": 180,
       "calls": 1440
      }
     }
    },
    {
     "op": "Where",
     "total_ms": 2.707,
     "pct": 0.281,
     "by_device": {
      "CPU": {
       "time_ms": 2.707,
       "nodes": 90,
       "calls": 720
      }
     }
    },
    {
     "op": "Expand",
     "total_ms": 1.694,
     "pct": 0.176,
     "by_device": {
      "CPU": {
       "time_ms": 1.694,
       "nodes": 46,
       "calls": 368
      }
     }
    },
    {
     "op": "Div",
     "total_ms": 1.488,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 1.488,
       "nodes": 89,
       "calls": 712
      }
     }
    },
    {
     "op": "Pow",
     "total_ms": 1.354,
     "pct": 0.141,
     "by_device": {
      "CPU": {
       "time_ms": 1.354,
       "nodes": 45,
       "calls": 360
      }
     }
    },
    {
     "op": "Cast",
     "total_ms": 1.339,
     "pct": 0.139,
     "by_device": {
      "CPU": {
       "time_ms": 1.339,
       "nodes": 100,
       "calls": 800
      }
     }
    },
    {
     "op": "Softmax",
     "total_ms": 1.322,
     "pct": 0.137,
     "by_device": {
      "CPU": {
       "time_ms": 1.322,
       "nodes": 22,
       "calls": 176
      }
     }
    },
    {
     "op": "Sigmoid",
     "total_ms": 1.146,
     "pct": 0.119,
     "by_device": {
      "CPU": {
       "time_ms": 1.146,
       "nodes": 22,
       "calls": 176
      }
     }
    },
    {
     "op": "ReduceMean",
     "total_ms": 0.918,
     "pct": 0.095,
     "by_device": {
      "CPU": {
       "time_ms": 0.918,
       "nodes": 45,
       "calls": 360
      }
     }
    },
    {
     "op": "Neg",
     "total_ms": 0.626,
     "pct": 0.065,
     "by_device": {
      "CPU": {
       "time_ms": 0.626,
       "nodes": 44,
       "calls": 352
      }
     }
    },
    {
     "op": "Equal",
     "total_ms": 0.506,
     "pct": 0.053,
     "by_device": {
      "CPU": {
       "time_ms": 0.506,
       "nodes": 46,
       "calls": 368
      }
     }
    },
    {
     "op": "Sqrt",
     "total_ms": 0.459,
     "pct": 0.048,
     "by_device": {
      "CPU": {
       "time_ms": 0.459,
       "nodes": 45,
       "calls": 360
      }
     }
    },
    {
     "op": "ConstantOfShape",
     "total_ms": 0.416,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.416,
       "nodes": 47,
       "calls": 376
      }
     }
    },
    {
     "op": "IsNaN",
     "total_ms": 0.356,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.356,
       "nodes": 22,
       "calls": 176
      }
     }
    },
    {
     "op": "Range",
     "total_ms": 0.066,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.066,
       "nodes": 3,
       "calls": 24
      }
     }
    },
    {
     "op": "And",
     "total_ms": 0.037,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 2,
       "calls": 16
      }
     }
    },
    {
     "op": "LessOrEqual",
     "total_ms": 0.018,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Sin",
     "total_ms": 0.017,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Cos",
     "total_ms": 0.017,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "Flatten",
     "total_ms": 0.017,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "ort/pim/ort_tinyllama-1.1b_pim.simplified.json",
   "source_profile": "ort_tinyllama-1.1b_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:39:42",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "bloom-1b7",
   "hf_model_id": "bigscience/bloom-1b7",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-1b7",
   "param_count": 1722408960
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.402,
   "TTFT": 0.261,
   "TBT_avg": 0.141,
   "Throughput": 79.633
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 401.843,
     "ttft": 260.889,
     "tbt_avg": 140.954,
     "throughput": 79.633,
     "steps": [
      260.889,
      140.954
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 401.843,
     "min": 401.843,
     "max": 401.843,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 260.889,
     "min": 260.889,
     "max": 260.889,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 140.954,
     "min": 140.954,
     "max": 140.954,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 79.633,
     "min": 79.633,
     "max": 79.633,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 140.954,
     "min": 140.954,
     "max": 140.954,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 757.082,
   "total_nodes": 57,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 757.082,
     "pct": 100.0,
     "nodes": 57,
     "calls": 8292
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 324.663,
     "pct": 42.883,
     "by_device": {
      "CPU": {
       "time_ms": 324.663,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 312.041,
     "pct": 41.216,
     "by_device": {
      "CPU": {
       "time_ms": 312.041,
       "nodes": 1,
       "calls": 192
      }
     }
    },
    {
     "op": "GeLUFunction",
     "total_ms": 19.937,
     "pct": 2.633,
     "by_device": {
      "CPU": {
       "time_ms": 19.937,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 9.979,
     "pct": 1.318,
     "by_device": {
      "CPU": {
       "time_ms": 9.979,
       "nodes": 1,
       "calls": 292
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 8.366,
     "pct": 1.105,
     "by_device": {
      "CPU": {
       "time_ms": 8.366,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 8.313,
     "pct": 1.098,
     "by_device": {
      "CPU": {
       "time_ms": 8.313,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "add",
     "total_ms": 7.992,
     "pct": 1.056,
     "by_device": {
      "CPU": {
       "time_ms": 7.992,
       "nodes": 1,
       "calls": 244
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 6.999,
     "pct": 0.924,
     "by_device": {
      "CPU": {
       "time_ms": 6.999,
       "nodes": 1,
       "calls": 656
      }
     }
    },
    {
     "op": "baddbmm",
     "total_ms": 5.895,
     "pct": 0.779,
     "by_device": {
      "CPU": {
       "time_ms": 5.895,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 5.229,
     "pct": 0.691,
     "by_device": {
      "CPU": {
       "time_ms": 5.229,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 4.856,
     "pct": 0.641,
     "by_device": {
      "CPU": {
       "time_ms": 4.856,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 4.478,
     "pct": 0.591,
     "by_device": {
      "CPU": {
       "time_ms": 4.478,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "to",
     "total_ms": 4.419,
     "pct": 0.584,
     "by_device": {
      "CPU": {
       "time_ms": 4.419,
       "nodes": 1,
       "calls": 406
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 3.757,
     "pct": 0.496,
     "by_device": {
      "CPU": {
       "time_ms": 3.757,
       "nodes": 1,
       "calls": 344
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 3.195,
     "pct": 0.422,
     "by_device": {
      "CPU": {
       "time_ms": 3.195,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 2.965,
     "pct": 0.392,
     "by_device": {
      "CPU": {
       "time_ms": 2.965,
       "nodes": 1,
       "calls": 202
      }
     }
    },
    {
     "op": "softmax",
     "total_ms": 2.666,
     "pct": 0.352,
     "by_device": {
      "CPU": {
       "time_ms": 2.666,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 2.31,
     "pct": 0.305,
     "by_device": {
      "CPU": {
       "time_ms": 2.31,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 2.082,
     "pct": 0.275,
     "by_device": {
      "CPU": {
       "time_ms": 2.082,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.959,
     "pct": 0.259,
     "by_device": {
      "CPU": {
       "time_ms": 1.959,
       "nodes": 1,
       "calls": 903
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.959,
     "pct": 0.259,
     "by_device": {
      "CPU": {
       "time_ms": 1.959,
       "nodes": 1,
       "calls": 386
      }
     }
    },
    {
     "op": "_softmax",
     "total_ms": 1.444,
     "pct": 0.191,
     "by_device": {
      "CPU": {
       "time_ms": 1.444,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.268,
     "pct": 0.168,
     "by_device": {
      "CPU": {
       "time_ms": 1.268,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.084,
     "pct": 0.143,
     "by_device": {
      "CPU": {
       "time_ms": 1.084,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.067,
     "pct": 0.141,
     "by_device": {
      "CPU": {
       "time_ms": 1.067,
       "nodes": 1,
       "calls": 980
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.007,
     "pct": 0.133,
     "by_device": {
      "CPU": {
       "time_ms": 1.007,
       "nodes": 1,
       "calls": 344
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.984,
     "pct": 0.13,
     "by_device": {
      "CPU": {
       "time_ms": 0.984,
       "nodes": 1,
       "calls": 222
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.967,
     "pct": 0.128,
     "by_device": {
      "CPU": {
       "time_ms": 0.967,
       "nodes": 1,
       "calls": 440
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.84,
     "pct": 0.111,
     "by_device": {
      "CPU": {
       "time_ms": 0.84,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.754,
     "pct": 0.1,
     "by_device": {
      "CPU": {
       "time_ms": 0.754,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.645,
     "pct": 0.085,
     "by_device": {
      "CPU": {
       "time_ms": 0.645,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.432,
     "pct": 0.057,
     "by_device": {
      "CPU": {
       "time_ms": 0.432,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.428,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 0.428,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.375,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 0.375,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "permute",
     "total_ms": 0.326,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.326,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "arange",
     "total_ms": 0.203,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.203,
       "nodes": 1,
       "calls": 20
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.172,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.172,
       "nodes": 1,
       "calls": 28
      }
     }
    },
    {
     "op": "index",
     "total_ms": 0.166,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.166,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.158,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.158,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.13,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.13,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.099,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.099,
       "nodes": 1,
       "calls": 580
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.069,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.069,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.068,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.068,
       "nodes": 1,
       "calls": 29
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.06,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "where",
     "total_ms": 0.06,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "new_ones",
     "total_ms": 0.038,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "__and__",
     "total_ms": 0.037,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "cumsum",
     "total_ms": 0.029,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bitwise_and",
     "total_ms": 0.025,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "new_empty",
     "total_ms": 0.022,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sub",
     "total_ms": 0.02,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "le",
     "total_ms": 0.016,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.013,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.013,
       "nodes": 1,
       "calls": 10
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.012,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "scalar_tensor",
     "total_ms": 0.004,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_bloom-1b7_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_bloom-1b7_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:39:34",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "bloom-560m",
   "hf_model_id": "bigscience/bloom-560m",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-560m",
   "param_count": 559214592
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.116,
   "TTFT": 0.073,
   "TBT_avg": 0.044,
   "Throughput": 275.14
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 116.304,
     "ttft": 72.699,
     "tbt_avg": 43.605,
     "throughput": 275.14,
     "steps": [
      72.699,
      43.605
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 116.304,
     "min": 116.304,
     "max": 116.304,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 72.699,
     "min": 72.699,
     "max": 72.699,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 43.605,
     "min": 43.605,
     "max": 43.605,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 275.14,
     "min": 275.14,
     "max": 275.14,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 43.605,
     "min": 43.605,
     "max": 43.605,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 190.683,
   "total_nodes": 57,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 190.683,
     "pct": 100.0,
     "nodes": 57,
     "calls": 8292
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 45.323,
     "pct": 23.769,
     "by_device": {
      "CPU": {
       "time_ms": 45.323,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 34.516,
     "pct": 18.101,
     "by_device": {
      "CPU": {
       "time_ms": 34.516,
       "nodes": 1,
       "calls": 192
      }
     }
    },
    {
     "op": "GeLUFunction",
     "total_ms": 18.474,
     "pct": 9.688,
     "by_device": {
      "CPU": {
       "time_ms": 18.474,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 9.127,
     "pct": 4.786,
     "by_device": {
      "CPU": {
       "time_ms": 9.127,
       "nodes": 1,
       "calls": 292
      }
     }
    },
    {
     "op": "add",
     "total_ms": 7.582,
     "pct": 3.976,
     "by_device": {
      "CPU": {
       "time_ms": 7.582,
       "nodes": 1,
       "calls": 244
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 7.187,
     "pct": 3.769,
     "by_device": {
      "CPU": {
       "time_ms": 7.187,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 7.127,
     "pct": 3.738,
     "by_device": {
      "CPU": {
       "time_ms": 7.127,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "baddbmm",
     "total_ms": 6.018,
     "pct": 3.156,
     "by_device": {
      "CPU": {
       "time_ms": 6.018,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 5.338,
     "pct": 2.8,
     "by_device": {
      "CPU": {
       "time_ms": 5.338,
       "nodes": 1,
       "calls": 656
      }
     }
    },
    {
     "op": "to",
     "total_ms": 4.352,
     "pct": 2.282,
     "by_device": {
      "CPU": {
       "time_ms": 4.352,
       "nodes": 1,
       "calls": 406
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 4.319,
     "pct": 2.265,
     "by_device": {
      "CPU": {
       "time_ms": 4.319,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 3.973,
     "pct": 2.083,
     "by_device": {
      "CPU": {
       "time_ms": 3.973,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 3.941,
     "pct": 2.067,
     "by_device": {
      "CPU": {
       "time_ms": 3.941,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 3.662,
     "pct": 1.921,
     "by_device": {
      "CPU": {
       "time_ms": 3.662,
       "nodes": 1,
       "calls": 344
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 3.636,
     "pct": 1.907,
     "by_device": {
      "CPU": {
       "time_ms": 3.636,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.218,
     "pct": 1.688,
     "by_device": {
      "CPU": {
       "time_ms": 3.218,
       "nodes": 1,
       "calls": 202
      }
     }
    },
    {
     "op": "softmax",
     "total_ms": 2.313,
     "pct": 1.213,
     "by_device": {
      "CPU": {
       "time_ms": 2.313,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 2.151,
     "pct": 1.128,
     "by_device": {
      "CPU": {
       "time_ms": 2.151,
       "nodes": 1,
       "calls": 386
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 2.092,
     "pct": 1.097,
     "by_device": {
      "CPU": {
       "time_ms": 2.092,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.73,
     "pct": 0.907,
     "by_device": {
      "CPU": {
       "time_ms": 1.73,
       "nodes": 1,
       "calls": 903
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 1.67,
     "pct": 0.876,
     "by_device": {
      "CPU": {
       "time_ms": 1.67,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.288,
     "pct": 0.676,
     "by_device": {
      "CPU": {
       "time_ms": 1.288,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.221,
     "pct": 0.64,
     "by_device": {
      "CPU": {
       "time_ms": 1.221,
       "nodes": 1,
       "calls": 980
      }
     }
    },
    {
     "op": "_softmax",
     "total_ms": 1.219,
     "pct": 0.639,
     "by_device": {
      "CPU": {
       "time_ms": 1.219,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 1.117,
     "pct": 0.586,
     "by_device": {
      "CPU": {
       "time_ms": 1.117,
       "nodes": 1,
       "calls": 222
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 1.106,
     "pct": 0.58,
     "by_device": {
      "CPU": {
       "time_ms": 1.106,
       "nodes": 1,
       "calls": 440
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.082,
     "pct": 0.567,
     "by_device": {
      "CPU": {
       "time_ms": 1.082,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.964,
     "pct": 0.506,
     "by_device": {
      "CPU": {
       "time_ms": 0.964,
       "nodes": 1,
       "calls": 344
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.783,
     "pct": 0.41,
     "by_device": {
      "CPU": {
       "time_ms": 0.783,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.759,
     "pct": 0.398,
     "by_device": {
      "CPU": {
       "time_ms": 0.759,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.496,
     "pct": 0.26,
     "by_device": {
      "CPU": {
       "time_ms": 0.496,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.474,
     "pct": 0.249,
     "by_device": {
      "CPU": {
       "time_ms": 0.474,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.425,
     "pct": 0.223,
     "by_device": {
      "CPU": {
       "time_ms": 0.425,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.386,
     "pct": 0.202,
     "by_device": {
      "CPU": {
       "time_ms": 0.386,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "permute",
     "total_ms": 0.312,
     "pct": 0.164,
     "by_device": {
      "CPU": {
       "time_ms": 0.312,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "arange",
     "total_ms": 0.174,
     "pct": 0.091,
     "by_device": {
      "CPU": {
       "time_ms": 0.174,
       "nodes": 1,
       "calls": 20
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.152,
     "pct": 0.08,
     "by_device": {
      "CPU": {
       "time_ms": 0.152,
       "nodes": 1,
       "calls": 28
      }
     }
    },
    {
     "op": "index",
     "total_ms": 0.147,
     "pct": 0.077,
     "by_device": {
      "CPU": {
       "time_ms": 0.147,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.12,
     "pct": 0.063,
     "by_device": {
      "CPU": {
       "time_ms": 0.12,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.116,
     "pct": 0.061,
     "by_device": {
      "CPU": {
       "time_ms": 0.116,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "where",
     "total_ms": 0.098,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 0.098,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.093,
     "pct": 0.049,
     "by_device": {
      "CPU": {
       "time_ms": 0.093,
       "nodes": 1,
       "calls": 580
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.065,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.065,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.061,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.061,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.058,
     "pct": 0.031,
     "by_device": {
      "CPU": {
       "time_ms": 0.058,
       "nodes": 1,
       "calls": 29
      }
     }
    },
    {
     "op": "sub",
     "total_ms": 0.051,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.051,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "__and__",
     "total_ms": 0.04,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.04,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "cumsum",
     "total_ms": 0.031,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.031,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bitwise_and",
     "total_ms": 0.027,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "new_ones",
     "total_ms": 0.022,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "le",
     "total_ms": 0.016,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.01,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.008,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 10
      }
     }
    },
    {
     "op": "new_empty",
     "total_ms": 0.007,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.007,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "scalar_tensor",
     "total_ms": 0.004,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.001,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_bloom-560m_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_bloom-560m_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:38:01",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "Falcon3-1B-Base",
   "hf_model_id": "tiiuae/Falcon3-1B-Base",
   "hf_model_url": "https://huggingface.co/tiiuae/Falcon3-1B-Base",
   "param_count": 1669000000
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.398,
   "TTFT": 0.294,
   "TBT_avg": 0.104,
   "Throughput": 80.399
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 398.015,
     "ttft": 293.911,
     "tbt_avg": 104.104,
     "throughput": 80.399,
     "steps": [
      293.911,
      104.104
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 398.015,
     "min": 398.015,
     "max": 398.015,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 293.911,
     "min": 293.911,
     "max": 293.911,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 104.104,
     "min": 104.104,
     "max": 104.104,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 80.399,
     "min": 80.399,
     "max": 80.399,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 104.104,
     "min": 104.104,
     "max": 104.104,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1094.103,
   "total_nodes": 48,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1094.103,
     "pct": 100.0,
     "nodes": 48,
     "calls": 8125
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 348.189,
     "pct": 31.824,
     "by_device": {
      "CPU": {
       "time_ms": 348.189,
       "nodes": 1,
       "calls": 254
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 346.57,
     "pct": 31.676,
     "by_device": {
      "CPU": {
       "time_ms": 346.57,
       "nodes": 1,
       "calls": 256
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 344.179,
     "pct": 31.458,
     "by_device": {
      "CPU": {
       "time_ms": 344.179,
       "nodes": 1,
       "calls": 254
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 7.819,
     "pct": 0.715,
     "by_device": {
      "CPU": {
       "time_ms": 7.819,
       "nodes": 1,
       "calls": 332
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 5.049,
     "pct": 0.461,
     "by_device": {
      "CPU": {
       "time_ms": 5.049,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "to",
     "total_ms": 4.8,
     "pct": 0.439,
     "by_device": {
      "CPU": {
       "time_ms": 4.8,
       "nodes": 1,
       "calls": 428
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 4.412,
     "pct": 0.403,
     "by_device": {
      "CPU": {
       "time_ms": 4.412,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "add",
     "total_ms": 4.104,
     "pct": 0.375,
     "by_device": {
      "CPU": {
       "time_ms": 4.104,
       "nodes": 1,
       "calls": 218
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 3.279,
     "pct": 0.3,
     "by_device": {
      "CPU": {
       "time_ms": 3.279,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 2.876,
     "pct": 0.263,
     "by_device": {
      "CPU": {
       "time_ms": 2.876,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.832,
     "pct": 0.259,
     "by_device": {
      "CPU": {
       "time_ms": 2.832,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 2.639,
     "pct": 0.241,
     "by_device": {
      "CPU": {
       "time_ms": 2.639,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 2.118,
     "pct": 0.194,
     "by_device": {
      "CPU": {
       "time_ms": 2.118,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 1.892,
     "pct": 0.173,
     "by_device": {
      "CPU": {
       "time_ms": 1.892,
       "nodes": 1,
       "calls": 78
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 1.716,
     "pct": 0.157,
     "by_device": {
      "CPU": {
       "time_ms": 1.716,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.663,
     "pct": 0.152,
     "by_device": {
      "CPU": {
       "time_ms": 1.663,
       "nodes": 1,
       "calls": 616
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 1.373,
     "pct": 0.126,
     "by_device": {
      "CPU": {
       "time_ms": 1.373,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.12,
     "pct": 0.102,
     "by_device": {
      "CPU": {
       "time_ms": 1.12,
       "nodes": 1,
       "calls": 254
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 1.06,
     "pct": 0.097,
     "by_device": {
      "CPU": {
       "time_ms": 1.06,
       "nodes": 1,
       "calls": 399
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.962,
     "pct": 0.088,
     "by_device": {
      "CPU": {
       "time_ms": 0.962,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.862,
     "pct": 0.079,
     "by_device": {
      "CPU": {
       "time_ms": 0.862,
       "nodes": 1,
       "calls": 253
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.71,
     "pct": 0.065,
     "by_device": {
      "CPU": {
       "time_ms": 0.71,
       "nodes": 1,
       "calls": 296
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.606,
     "pct": 0.055,
     "by_device": {
      "CPU": {
       "time_ms": 0.606,
       "nodes": 1,
       "calls": 1105
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.606,
     "pct": 0.055,
     "by_device": {
      "CPU": {
       "time_ms": 0.606,
       "nodes": 1,
       "calls": 346
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.462,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.462,
       "nodes": 1,
       "calls": 403
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.39,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 0.39,
       "nodes": 1,
       "calls": 256
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.284,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.284,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.279,
     "pct": 0.025,
     "by_device": {
      "CPU": {
       "time_ms": 0.279,
       "nodes": 1,
       "calls": 78
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.254,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.254,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.191,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.191,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.184,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.184,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.159,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.159,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.091,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.091,
       "nodes": 1,
       "calls": 76
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.071,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.071,
       "nodes": 2,
       "calls": 72
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.062,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.062,
       "nodes": 1,
       "calls": 512
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.057,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.057,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.04,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.04,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.032,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.029,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.026,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.013,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.013,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.012,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.01,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.007,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.007,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_falcon3-1b-base_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_falcon3-1b-base_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:39:03",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "gemma-2-2b",
   "hf_model_id": "google/gemma-2-2b",
   "hf_model_url": "https://huggingface.co/google/gemma-2-2b",
   "param_count": 2614341888
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.561,
   "TTFT": 0.41,
   "TBT_avg": 0.151,
   "Throughput": 57.036
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 561.046,
     "ttft": 409.834,
     "tbt_avg": 151.212,
     "throughput": 57.036,
     "steps": [
      409.834,
      151.212
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 561.046,
     "min": 561.046,
     "max": 561.046,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 409.834,
     "min": 409.834,
     "max": 409.834,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 151.212,
     "min": 151.212,
     "max": 151.212,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 57.036,
     "min": 57.036,
     "max": 57.036,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 151.212,
     "min": 151.212,
     "max": 151.212,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1466.049,
   "total_nodes": 51,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1466.049,
     "pct": 100.0,
     "nodes": 51,
     "calls": 16816
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 441.768,
     "pct": 30.133,
     "by_device": {
      "CPU": {
       "time_ms": 441.768,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 438.415,
     "pct": 29.905,
     "by_device": {
      "CPU": {
       "time_ms": 438.415,
       "nodes": 1,
       "calls": 368
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 433.857,
     "pct": 29.594,
     "by_device": {
      "CPU": {
       "time_ms": 433.857,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "to",
     "total_ms": 20.544,
     "pct": 1.401,
     "by_device": {
      "CPU": {
       "time_ms": 20.544,
       "nodes": 1,
       "calls": 1567
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 18.523,
     "pct": 1.263,
     "by_device": {
      "CPU": {
       "time_ms": 18.523,
       "nodes": 1,
       "calls": 1282
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 16.804,
     "pct": 1.146,
     "by_device": {
      "CPU": {
       "time_ms": 16.804,
       "nodes": 1,
       "calls": 688
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 11.77,
     "pct": 0.803,
     "by_device": {
      "CPU": {
       "time_ms": 11.77,
       "nodes": 1,
       "calls": 1282
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 9.912,
     "pct": 0.676,
     "by_device": {
      "CPU": {
       "time_ms": 9.912,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "add",
     "total_ms": 8.681,
     "pct": 0.592,
     "by_device": {
      "CPU": {
       "time_ms": 8.681,
       "nodes": 1,
       "calls": 628
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 8.608,
     "pct": 0.587,
     "by_device": {
      "CPU": {
       "time_ms": 8.608,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "type_as",
     "total_ms": 6.394,
     "pct": 0.436,
     "by_device": {
      "CPU": {
       "time_ms": 6.394,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 6.126,
     "pct": 0.418,
     "by_device": {
      "CPU": {
       "time_ms": 6.126,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 5.534,
     "pct": 0.377,
     "by_device": {
      "CPU": {
       "time_ms": 5.534,
       "nodes": 1,
       "calls": 420
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 5.397,
     "pct": 0.368,
     "by_device": {
      "CPU": {
       "time_ms": 5.397,
       "nodes": 1,
       "calls": 218
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 5.03,
     "pct": 0.343,
     "by_device": {
      "CPU": {
       "time_ms": 5.03,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "gelu",
     "total_ms": 4.556,
     "pct": 0.311,
     "by_device": {
      "CPU": {
       "time_ms": 4.556,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 4.514,
     "pct": 0.308,
     "by_device": {
      "CPU": {
       "time_ms": 4.514,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 2.941,
     "pct": 0.201,
     "by_device": {
      "CPU": {
       "time_ms": 2.941,
       "nodes": 1,
       "calls": 888
      }
     }
    },
    {
     "op": "t",
     "total_ms": 2.238,
     "pct": 0.153,
     "by_device": {
      "CPU": {
       "time_ms": 2.238,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 1.909,
     "pct": 0.13,
     "by_device": {
      "CPU": {
       "time_ms": 1.909,
       "nodes": 1,
       "calls": 627
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.819,
     "pct": 0.124,
     "by_device": {
      "CPU": {
       "time_ms": 1.819,
       "nodes": 1,
       "calls": 1334
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 1.517,
     "pct": 0.104,
     "by_device": {
      "CPU": {
       "time_ms": 1.517,
       "nodes": 1,
       "calls": 104
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.45,
     "pct": 0.099,
     "by_device": {
      "CPU": {
       "time_ms": 1.45,
       "nodes": 1,
       "calls": 365
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 1.417,
     "pct": 0.097,
     "by_device": {
      "CPU": {
       "time_ms": 1.417,
       "nodes": 1,
       "calls": 424
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.058,
     "pct": 0.072,
     "by_device": {
      "CPU": {
       "time_ms": 1.058,
       "nodes": 1,
       "calls": 1639
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.926,
     "pct": 0.063,
     "by_device": {
      "CPU": {
       "time_ms": 0.926,
       "nodes": 1,
       "calls": 579
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.642,
     "pct": 0.044,
     "by_device": {
      "CPU": {
       "time_ms": 0.642,
       "nodes": 1,
       "calls": 368
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.5,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.5,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.473,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.473,
       "nodes": 1,
       "calls": 110
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 0.39,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.39,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.316,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.316,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.301,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.301,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.29,
     "pct": 0.02,
     "by_device": {
      "CPU": {
       "time_ms": 0.29,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.284,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.284,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.258,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.258,
       "nodes": 1,
       "calls": 214
      }
     }
    },
    {
     "op": "div",
     "total_ms": 0.21,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.21,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.173,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.173,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.128,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.128,
       "nodes": 1,
       "calls": 736
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.103,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.103,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.073,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.073,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.047,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.047,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.036,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.031,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.031,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.017,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.014,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_gemma-2-2b_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_gemma-2-2b_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:38:55",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "gemma-3-1b-it",
   "hf_model_id": "google/gemma-3-1b-it",
   "hf_model_url": "https://huggingface.co/google/gemma-3-1b-it",
   "param_count": 999885952
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.243,
   "TTFT": 0.179,
   "TBT_avg": 0.064,
   "Throughput": 131.891
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 242.624,
     "ttft": 178.953,
     "tbt_avg": 63.671,
     "throughput": 131.891,
     "steps": [
      178.953,
      63.671
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 242.624,
     "min": 242.624,
     "max": 242.624,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 178.953,
     "min": 178.953,
     "max": 178.953,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 63.671,
     "min": 63.671,
     "max": 63.671,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 131.891,
     "min": 131.891,
     "max": 131.891,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 63.671,
     "min": 63.671,
     "max": 63.671,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 553.337,
   "total_nodes": 49,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 553.337,
     "pct": 100.0,
     "nodes": 49,
     "calls": 20934
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 143.353,
     "pct": 25.907,
     "by_device": {
      "CPU": {
       "time_ms": 143.353,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 140.935,
     "pct": 25.47,
     "by_device": {
      "CPU": {
       "time_ms": 140.935,
       "nodes": 1,
       "calls": 370
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 137.524,
     "pct": 24.854,
     "by_device": {
      "CPU": {
       "time_ms": 137.524,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "to",
     "total_ms": 20.592,
     "pct": 3.722,
     "by_device": {
      "CPU": {
       "time_ms": 20.592,
       "nodes": 1,
       "calls": 2318
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 18.193,
     "pct": 3.288,
     "by_device": {
      "CPU": {
       "time_ms": 18.193,
       "nodes": 1,
       "calls": 1912
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 12.554,
     "pct": 2.269,
     "by_device": {
      "CPU": {
       "time_ms": 12.554,
       "nodes": 1,
       "calls": 898
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 9.985,
     "pct": 1.804,
     "by_device": {
      "CPU": {
       "time_ms": 9.985,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 9.791,
     "pct": 1.769,
     "by_device": {
      "CPU": {
       "time_ms": 9.791,
       "nodes": 1,
       "calls": 1912
      }
     }
    },
    {
     "op": "add",
     "total_ms": 7.995,
     "pct": 1.445,
     "by_device": {
      "CPU": {
       "time_ms": 7.995,
       "nodes": 1,
       "calls": 836
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 6.516,
     "pct": 1.178,
     "by_device": {
      "CPU": {
       "time_ms": 6.516,
       "nodes": 1,
       "calls": 628
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 5.685,
     "pct": 1.027,
     "by_device": {
      "CPU": {
       "time_ms": 5.685,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "type_as",
     "total_ms": 5.595,
     "pct": 1.011,
     "by_device": {
      "CPU": {
       "time_ms": 5.595,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 4.763,
     "pct": 0.861,
     "by_device": {
      "CPU": {
       "time_ms": 4.763,
       "nodes": 1,
       "calls": 322
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 4.497,
     "pct": 0.813,
     "by_device": {
      "CPU": {
       "time_ms": 4.497,
       "nodes": 1,
       "calls": 212
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 3.487,
     "pct": 0.63,
     "by_device": {
      "CPU": {
       "time_ms": 3.487,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 3.22,
     "pct": 0.582,
     "by_device": {
      "CPU": {
       "time_ms": 3.22,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 2.564,
     "pct": 0.463,
     "by_device": {
      "CPU": {
       "time_ms": 2.564,
       "nodes": 1,
       "calls": 1964
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 2.403,
     "pct": 0.434,
     "by_device": {
      "CPU": {
       "time_ms": 2.403,
       "nodes": 1,
       "calls": 890
      }
     }
    },
    {
     "op": "gelu",
     "total_ms": 1.948,
     "pct": 0.352,
     "by_device": {
      "CPU": {
       "time_ms": 1.948,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 1.723,
     "pct": 0.311,
     "by_device": {
      "CPU": {
       "time_ms": 1.723,
       "nodes": 1,
       "calls": 664
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.699,
     "pct": 0.307,
     "by_device": {
      "CPU": {
       "time_ms": 1.699,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.248,
     "pct": 0.225,
     "by_device": {
      "CPU": {
       "time_ms": 1.248,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 1.123,
     "pct": 0.203,
     "by_device": {
      "CPU": {
       "time_ms": 1.123,
       "nodes": 1,
       "calls": 428
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.963,
     "pct": 0.174,
     "by_device": {
      "CPU": {
       "time_ms": 0.963,
       "nodes": 1,
       "calls": 1690
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.814,
     "pct": 0.147,
     "by_device": {
      "CPU": {
       "time_ms": 0.814,
       "nodes": 1,
       "calls": 104
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.742,
     "pct": 0.134,
     "by_device": {
      "CPU": {
       "time_ms": 0.742,
       "nodes": 1,
       "calls": 581
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.56,
     "pct": 0.101,
     "by_device": {
      "CPU": {
       "time_ms": 0.56,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.531,
     "pct": 0.096,
     "by_device": {
      "CPU": {
       "time_ms": 0.531,
       "nodes": 1,
       "calls": 370
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.425,
     "pct": 0.077,
     "by_device": {
      "CPU": {
       "time_ms": 0.425,
       "nodes": 1,
       "calls": 116
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.295,
     "pct": 0.053,
     "by_device": {
      "CPU": {
       "time_ms": 0.295,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.275,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 0.275,
       "nodes": 1,
       "calls": 318
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.266,
     "pct": 0.048,
     "by_device": {
      "CPU": {
       "time_ms": 0.266,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.237,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.237,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.219,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.219,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.107,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.107,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.105,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.105,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.09,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.09,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.088,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.088,
       "nodes": 1,
       "calls": 740
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.065,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.065,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.053,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.053,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.043,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.043,
       "nodes": 1,
       "calls": 12
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.014,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.014,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 5
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.013,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.013,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.01,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_gemma-3-1b-it_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_gemma-3-1b-it_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:38:31",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "gpt2-large",
   "hf_model_id": "openai-community/gpt2-large",
   "hf_model_url": "https://huggingface.co/openai-community/gpt2-large",
   "param_count": 774030080
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.319,
   "TTFT": 0.264,
   "TBT_avg": 0.055,
   "Throughput": 100.177
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 319.436,
     "ttft": 264.338,
     "tbt_avg": 55.098,
     "throughput": 100.177,
     "steps": [
      264.338,
      55.098
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 319.436,
     "min": 319.436,
     "max": 319.436,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 264.338,
     "min": 264.338,
     "max": 264.338,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 55.098,
     "min": 55.098,
     "max": 55.098,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 100.177,
     "min": 100.177,
     "max": 100.177,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 55.098,
     "min": 55.098,
     "max": 55.098,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 320.218,
   "total_nodes": 45,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 320.218,
     "pct": 100.0,
     "nodes": 45,
     "calls": 10060
    }
   ],
   "ops": [
    {
     "op": "addmm",
     "total_ms": 250.428,
     "pct": 78.205,
     "by_device": {
      "CPU": {
       "time_ms": 250.428,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 7.542,
     "pct": 2.355,
     "by_device": {
      "CPU": {
       "time_ms": 7.542,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "add",
     "total_ms": 7.488,
     "pct": 2.338,
     "by_device": {
      "CPU": {
       "time_ms": 7.488,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 6.047,
     "pct": 1.889,
     "by_device": {
      "CPU": {
       "time_ms": 6.047,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 5.668,
     "pct": 1.77,
     "by_device": {
      "CPU": {
       "time_ms": 5.668,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 5.118,
     "pct": 1.598,
     "by_device": {
      "CPU": {
       "time_ms": 5.118,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 4.941,
     "pct": 1.543,
     "by_device": {
      "CPU": {
       "time_ms": 4.941,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 4.814,
     "pct": 1.503,
     "by_device": {
      "CPU": {
       "time_ms": 4.814,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 4.638,
     "pct": 1.448,
     "by_device": {
      "CPU": {
       "time_ms": 4.638,
       "nodes": 1,
       "calls": 580
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 2.194,
     "pct": 0.685,
     "by_device": {
      "CPU": {
       "time_ms": 2.194,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 1.959,
     "pct": 0.612,
     "by_device": {
      "CPU": {
       "time_ms": 1.959,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "to",
     "total_ms": 1.84,
     "pct": 0.574,
     "by_device": {
      "CPU": {
       "time_ms": 1.84,
       "nodes": 1,
       "calls": 438
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.734,
     "pct": 0.542,
     "by_device": {
      "CPU": {
       "time_ms": 1.734,
       "nodes": 1,
       "calls": 722
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.616,
     "pct": 0.505,
     "by_device": {
      "CPU": {
       "time_ms": 1.616,
       "nodes": 1,
       "calls": 432
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 1.54,
     "pct": 0.481,
     "by_device": {
      "CPU": {
       "time_ms": 1.54,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 1.515,
     "pct": 0.473,
     "by_device": {
      "CPU": {
       "time_ms": 1.515,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 1.479,
     "pct": 0.462,
     "by_device": {
      "CPU": {
       "time_ms": 1.479,
       "nodes": 1,
       "calls": 292
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 1.466,
     "pct": 0.458,
     "by_device": {
      "CPU": {
       "time_ms": 1.466,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.416,
     "pct": 0.442,
     "by_device": {
      "CPU": {
       "time_ms": 1.416,
       "nodes": 1,
       "calls": 1171
      }
     }
    },
    {
     "op": "split",
     "total_ms": 1.284,
     "pct": 0.401,
     "by_device": {
      "CPU": {
       "time_ms": 1.284,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 1.029,
     "pct": 0.321,
     "by_device": {
      "CPU": {
       "time_ms": 1.029,
       "nodes": 1,
       "calls": 434
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.752,
     "pct": 0.235,
     "by_device": {
      "CPU": {
       "time_ms": 0.752,
       "nodes": 1,
       "calls": 1454
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.731,
     "pct": 0.228,
     "by_device": {
      "CPU": {
       "time_ms": 0.731,
       "nodes": 1,
       "calls": 730
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.716,
     "pct": 0.224,
     "by_device": {
      "CPU": {
       "time_ms": 0.716,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.47,
     "pct": 0.147,
     "by_device": {
      "CPU": {
       "time_ms": 0.47,
       "nodes": 1,
       "calls": 364
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.38,
     "pct": 0.119,
     "by_device": {
      "CPU": {
       "time_ms": 0.38,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.339,
     "pct": 0.106,
     "by_device": {
      "CPU": {
       "time_ms": 0.339,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.336,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 0.336,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.232,
     "pct": 0.073,
     "by_device": {
      "CPU": {
       "time_ms": 0.232,
       "nodes": 1,
       "calls": 78
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.134,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.134,
       "nodes": 2,
       "calls": 144
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.116,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 0.116,
       "nodes": 1,
       "calls": 868
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.093,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.093,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.062,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.062,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.025,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.017,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.014,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.013,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.013,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.008,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.008,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.007,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.007,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.004,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.004,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_gpt2-large_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_gpt2-large_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:38:47",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "Llama-3.2-1B",
   "hf_model_id": "meta-llama/Llama-3.2-1B",
   "hf_model_url": "https://huggingface.co/meta-llama/Llama-3.2-1B",
   "param_count": 1235814400
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.354,
   "TTFT": 0.266,
   "TBT_avg": 0.089,
   "Throughput": 90.35
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 354.179,
     "ttft": 265.601,
     "tbt_avg": 88.578,
     "throughput": 90.35,
     "steps": [
      265.601,
      88.578
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 354.179,
     "min": 354.179,
     "max": 354.179,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 265.601,
     "min": 265.601,
     "max": 265.601,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 88.578,
     "min": 88.578,
     "max": 88.578,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 90.35,
     "min": 90.35,
     "max": 90.35,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 88.578,
     "min": 88.578,
     "max": 88.578,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 976.15,
   "total_nodes": 48,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 976.15,
     "pct": 100.0,
     "nodes": 48,
     "calls": 7249
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 310.605,
     "pct": 31.819,
     "by_device": {
      "CPU": {
       "time_ms": 310.605,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 309.198,
     "pct": 31.675,
     "by_device": {
      "CPU": {
       "time_ms": 309.198,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 307.0,
     "pct": 31.45,
     "by_device": {
      "CPU": {
       "time_ms": 307.0,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 6.056,
     "pct": 0.62,
     "by_device": {
      "CPU": {
       "time_ms": 6.056,
       "nodes": 1,
       "calls": 296
      }
     }
    },
    {
     "op": "to",
     "total_ms": 4.722,
     "pct": 0.484,
     "by_device": {
      "CPU": {
       "time_ms": 4.722,
       "nodes": 1,
       "calls": 384
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 4.372,
     "pct": 0.448,
     "by_device": {
      "CPU": {
       "time_ms": 4.372,
       "nodes": 1,
       "calls": 278
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 3.72,
     "pct": 0.381,
     "by_device": {
      "CPU": {
       "time_ms": 3.72,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "add",
     "total_ms": 3.485,
     "pct": 0.357,
     "by_device": {
      "CPU": {
       "time_ms": 3.485,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 3.072,
     "pct": 0.315,
     "by_device": {
      "CPU": {
       "time_ms": 3.072,
       "nodes": 1,
       "calls": 278
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.873,
     "pct": 0.294,
     "by_device": {
      "CPU": {
       "time_ms": 2.873,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 2.862,
     "pct": 0.293,
     "by_device": {
      "CPU": {
       "time_ms": 2.862,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 2.667,
     "pct": 0.273,
     "by_device": {
      "CPU": {
       "time_ms": 2.667,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 1.928,
     "pct": 0.197,
     "by_device": {
      "CPU": {
       "time_ms": 1.928,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 1.612,
     "pct": 0.165,
     "by_device": {
      "CPU": {
       "time_ms": 1.612,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 1.595,
     "pct": 0.163,
     "by_device": {
      "CPU": {
       "time_ms": 1.595,
       "nodes": 1,
       "calls": 132
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.449,
     "pct": 0.148,
     "by_device": {
      "CPU": {
       "time_ms": 1.449,
       "nodes": 1,
       "calls": 548
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 1.411,
     "pct": 0.145,
     "by_device": {
      "CPU": {
       "time_ms": 1.411,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.974,
     "pct": 0.1,
     "by_device": {
      "CPU": {
       "time_ms": 0.974,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.908,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 0.908,
       "nodes": 1,
       "calls": 355
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.791,
     "pct": 0.081,
     "by_device": {
      "CPU": {
       "time_ms": 0.791,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.737,
     "pct": 0.075,
     "by_device": {
      "CPU": {
       "time_ms": 0.737,
       "nodes": 1,
       "calls": 225
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.698,
     "pct": 0.071,
     "by_device": {
      "CPU": {
       "time_ms": 0.698,
       "nodes": 1,
       "calls": 264
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.532,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.532,
       "nodes": 1,
       "calls": 985
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.44,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.44,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.42,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.42,
       "nodes": 1,
       "calls": 359
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.328,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.328,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.288,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.288,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.259,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.259,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.244,
     "pct": 0.025,
     "by_device": {
      "CPU": {
       "time_ms": 0.244,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.176,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.176,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.165,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.165,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.151,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.151,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.081,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.081,
       "nodes": 1,
       "calls": 68
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.062,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.062,
       "nodes": 2,
       "calls": 64
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.057,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.057,
       "nodes": 1,
       "calls": 456
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.043,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.043,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.035,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.03,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.03,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.028,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.012,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.01,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.005,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_llama-3.2-1b_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_llama-3.2-1b_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:39:51",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "OLMo-2-0425-1B",
   "hf_model_id": "allenai/OLMo-2-0425-1B",
   "hf_model_url": "https://huggingface.co/allenai/OLMo-2-0425-1B",
   "param_count": 1480000000
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.459,
   "TTFT": 0.376,
   "TBT_avg": 0.084,
   "Throughput": 69.706
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 459.073,
     "ttft": 375.522,
     "tbt_avg": 83.551,
     "throughput": 69.706,
     "steps": [
      375.522,
      83.551
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 459.073,
     "min": 459.073,
     "max": 459.073,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 375.522,
     "min": 375.522,
     "max": 375.522,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 83.551,
     "min": 83.551,
     "max": 83.551,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 69.706,
     "min": 69.706,
     "max": 69.706,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 83.551,
     "min": 83.551,
     "max": 83.551,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1275.484,
   "total_nodes": 48,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1275.484,
     "pct": 100.0,
     "nodes": 48,
     "calls": 10313
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 401.29,
     "pct": 31.462,
     "by_device": {
      "CPU": {
       "time_ms": 401.29,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 399.807,
     "pct": 31.346,
     "by_device": {
      "CPU": {
       "time_ms": 399.807,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 397.448,
     "pct": 31.161,
     "by_device": {
      "CPU": {
       "time_ms": 397.448,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "to",
     "total_ms": 11.48,
     "pct": 0.9,
     "by_device": {
      "CPU": {
       "time_ms": 11.48,
       "nodes": 1,
       "calls": 1022
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 10.911,
     "pct": 0.855,
     "by_device": {
      "CPU": {
       "time_ms": 10.911,
       "nodes": 1,
       "calls": 424
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 10.491,
     "pct": 0.823,
     "by_device": {
      "CPU": {
       "time_ms": 10.491,
       "nodes": 1,
       "calls": 852
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 6.787,
     "pct": 0.532,
     "by_device": {
      "CPU": {
       "time_ms": 6.787,
       "nodes": 1,
       "calls": 852
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 5.17,
     "pct": 0.405,
     "by_device": {
      "CPU": {
       "time_ms": 5.17,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 4.076,
     "pct": 0.32,
     "by_device": {
      "CPU": {
       "time_ms": 4.076,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "add",
     "total_ms": 3.647,
     "pct": 0.286,
     "by_device": {
      "CPU": {
       "time_ms": 3.647,
       "nodes": 1,
       "calls": 258
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 2.901,
     "pct": 0.227,
     "by_device": {
      "CPU": {
       "time_ms": 2.901,
       "nodes": 1,
       "calls": 134
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 2.809,
     "pct": 0.22,
     "by_device": {
      "CPU": {
       "time_ms": 2.809,
       "nodes": 1,
       "calls": 260
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 2.492,
     "pct": 0.195,
     "by_device": {
      "CPU": {
       "time_ms": 2.492,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.434,
     "pct": 0.191,
     "by_device": {
      "CPU": {
       "time_ms": 2.434,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 2.263,
     "pct": 0.177,
     "by_device": {
      "CPU": {
       "time_ms": 2.263,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.491,
     "pct": 0.117,
     "by_device": {
      "CPU": {
       "time_ms": 1.491,
       "nodes": 1,
       "calls": 548
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 1.322,
     "pct": 0.104,
     "by_device": {
      "CPU": {
       "time_ms": 1.322,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.265,
     "pct": 0.099,
     "by_device": {
      "CPU": {
       "time_ms": 1.265,
       "nodes": 1,
       "calls": 884
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.073,
     "pct": 0.084,
     "by_device": {
      "CPU": {
       "time_ms": 1.073,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.963,
     "pct": 0.075,
     "by_device": {
      "CPU": {
       "time_ms": 0.963,
       "nodes": 1,
       "calls": 355
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.8,
     "pct": 0.063,
     "by_device": {
      "CPU": {
       "time_ms": 0.8,
       "nodes": 1,
       "calls": 225
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.787,
     "pct": 0.062,
     "by_device": {
      "CPU": {
       "time_ms": 0.787,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.671,
     "pct": 0.053,
     "by_device": {
      "CPU": {
       "time_ms": 0.671,
       "nodes": 1,
       "calls": 264
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.57,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.57,
       "nodes": 1,
       "calls": 985
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.429,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.429,
       "nodes": 1,
       "calls": 359
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.349,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.349,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.272,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.272,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.249,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.249,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.247,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.247,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.242,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.242,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.148,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.148,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.132,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.132,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.13,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.13,
       "nodes": 1,
       "calls": 132
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.062,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.062,
       "nodes": 2,
       "calls": 64
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.055,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.055,
       "nodes": 1,
       "calls": 456
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.051,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.051,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.047,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.047,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.029,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.027,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.021,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.01,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.004,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_olmo-2-0425-1b_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_olmo-2-0425-1b_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:39:18",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "opt-1.3b",
   "hf_model_id": "facebook/opt-1.3b",
   "hf_model_url": "https://huggingface.co/facebook/opt-1.3b",
   "param_count": 1315758080
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.625,
   "TTFT": 0.521,
   "TBT_avg": 0.104,
   "Throughput": 51.24
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 624.513,
     "ttft": 520.66,
     "tbt_avg": 103.854,
     "throughput": 51.24,
     "steps": [
      520.66,
      103.854
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 624.513,
     "min": 624.513,
     "max": 624.513,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 520.66,
     "min": 520.66,
     "max": 520.66,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 103.854,
     "min": 103.854,
     "max": 103.854,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 51.24,
     "min": 51.24,
     "max": 51.24,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 103.854,
     "min": 103.854,
     "max": 103.854,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1216.014,
   "total_nodes": 43,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1216.014,
     "pct": 100.0,
     "nodes": 43,
     "calls": 7032
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 588.148,
     "pct": 48.367,
     "by_device": {
      "CPU": {
       "time_ms": 588.148,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 583.017,
     "pct": 47.945,
     "by_device": {
      "CPU": {
       "time_ms": 583.017,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 5.885,
     "pct": 0.484,
     "by_device": {
      "CPU": {
       "time_ms": 5.885,
       "nodes": 1,
       "calls": 340
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 3.991,
     "pct": 0.328,
     "by_device": {
      "CPU": {
       "time_ms": 3.991,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 3.867,
     "pct": 0.318,
     "by_device": {
      "CPU": {
       "time_ms": 3.867,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 3.636,
     "pct": 0.299,
     "by_device": {
      "CPU": {
       "time_ms": 3.636,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 3.379,
     "pct": 0.278,
     "by_device": {
      "CPU": {
       "time_ms": 3.379,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 3.125,
     "pct": 0.257,
     "by_device": {
      "CPU": {
       "time_ms": 3.125,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "add",
     "total_ms": 3.01,
     "pct": 0.248,
     "by_device": {
      "CPU": {
       "time_ms": 3.01,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 1.844,
     "pct": 0.152,
     "by_device": {
      "CPU": {
       "time_ms": 1.844,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.823,
     "pct": 0.15,
     "by_device": {
      "CPU": {
       "time_ms": 1.823,
       "nodes": 1,
       "calls": 770
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 1.79,
     "pct": 0.147,
     "by_device": {
      "CPU": {
       "time_ms": 1.79,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "relu",
     "total_ms": 1.714,
     "pct": 0.141,
     "by_device": {
      "CPU": {
       "time_ms": 1.714,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 1.628,
     "pct": 0.134,
     "by_device": {
      "CPU": {
       "time_ms": 1.628,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "clamp_min",
     "total_ms": 1.581,
     "pct": 0.13,
     "by_device": {
      "CPU": {
       "time_ms": 1.581,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.202,
     "pct": 0.099,
     "by_device": {
      "CPU": {
       "time_ms": 1.202,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.994,
     "pct": 0.082,
     "by_device": {
      "CPU": {
       "time_ms": 0.994,
       "nodes": 1,
       "calls": 879
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.621,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 0.621,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.601,
     "pct": 0.049,
     "by_device": {
      "CPU": {
       "time_ms": 0.601,
       "nodes": 1,
       "calls": 1214
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.572,
     "pct": 0.047,
     "by_device": {
      "CPU": {
       "time_ms": 0.572,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.543,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.543,
       "nodes": 1,
       "calls": 490
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.491,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.491,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.399,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.399,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.389,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.389,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.365,
     "pct": 0.03,
     "by_device": {
      "CPU": {
       "time_ms": 0.365,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.345,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.345,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.276,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.276,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.201,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.201,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.194,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.194,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.097,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.097,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.091,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.091,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.08,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.08,
       "nodes": 1,
       "calls": 580
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.053,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.053,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.02,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.007,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.007,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.004,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_opt-1.3b_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_opt-1.3b_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:39:10",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "opt-350m",
   "hf_model_id": "facebook/opt-350m",
   "hf_model_url": "https://huggingface.co/facebook/opt-350m",
   "param_count": 331196416
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.192,
   "TTFT": 0.161,
   "TBT_avg": 0.031,
   "Throughput": 166.972
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 191.649,
     "ttft": 161.143,
     "tbt_avg": 30.506,
     "throughput": 166.972,
     "steps": [
      161.143,
      30.506
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 191.649,
     "min": 191.649,
     "max": 191.649,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 161.143,
     "min": 161.143,
     "max": 161.143,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 30.506,
     "min": 30.506,
     "max": 30.506,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 166.972,
     "min": 166.972,
     "max": 166.972,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 30.506,
     "min": 30.506,
     "max": 30.506,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 341.129,
   "total_nodes": 43,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 341.129,
     "pct": 100.0,
     "nodes": 43,
     "calls": 7062
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 151.857,
     "pct": 44.516,
     "by_device": {
      "CPU": {
       "time_ms": 151.857,
       "nodes": 1,
       "calls": 294
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 146.906,
     "pct": 43.065,
     "by_device": {
      "CPU": {
       "time_ms": 146.906,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 4.247,
     "pct": 1.245,
     "by_device": {
      "CPU": {
       "time_ms": 4.247,
       "nodes": 1,
       "calls": 340
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 3.9,
     "pct": 1.143,
     "by_device": {
      "CPU": {
       "time_ms": 3.9,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 3.861,
     "pct": 1.132,
     "by_device": {
      "CPU": {
       "time_ms": 3.861,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 3.667,
     "pct": 1.075,
     "by_device": {
      "CPU": {
       "time_ms": 3.667,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 3.405,
     "pct": 0.998,
     "by_device": {
      "CPU": {
       "time_ms": 3.405,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 3.099,
     "pct": 0.909,
     "by_device": {
      "CPU": {
       "time_ms": 3.099,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 2.329,
     "pct": 0.683,
     "by_device": {
      "CPU": {
       "time_ms": 2.329,
       "nodes": 1,
       "calls": 774
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.033,
     "pct": 0.596,
     "by_device": {
      "CPU": {
       "time_ms": 2.033,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "relu",
     "total_ms": 1.578,
     "pct": 0.463,
     "by_device": {
      "CPU": {
       "time_ms": 1.578,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 1.527,
     "pct": 0.448,
     "by_device": {
      "CPU": {
       "time_ms": 1.527,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.423,
     "pct": 0.417,
     "by_device": {
      "CPU": {
       "time_ms": 1.423,
       "nodes": 1,
       "calls": 294
      }
     }
    },
    {
     "op": "clamp_min",
     "total_ms": 1.42,
     "pct": 0.416,
     "by_device": {
      "CPU": {
       "time_ms": 1.42,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 1.352,
     "pct": 0.396,
     "by_device": {
      "CPU": {
       "time_ms": 1.352,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 1.243,
     "pct": 0.364,
     "by_device": {
      "CPU": {
       "time_ms": 1.243,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.096,
     "pct": 0.321,
     "by_device": {
      "CPU": {
       "time_ms": 1.096,
       "nodes": 1,
       "calls": 879
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.756,
     "pct": 0.222,
     "by_device": {
      "CPU": {
       "time_ms": 0.756,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.75,
     "pct": 0.22,
     "by_device": {
      "CPU": {
       "time_ms": 0.75,
       "nodes": 1,
       "calls": 1218
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.698,
     "pct": 0.205,
     "by_device": {
      "CPU": {
       "time_ms": 0.698,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.618,
     "pct": 0.181,
     "by_device": {
      "CPU": {
       "time_ms": 0.618,
       "nodes": 1,
       "calls": 484
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.559,
     "pct": 0.164,
     "by_device": {
      "CPU": {
       "time_ms": 0.559,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.438,
     "pct": 0.128,
     "by_device": {
      "CPU": {
       "time_ms": 0.438,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.432,
     "pct": 0.127,
     "by_device": {
      "CPU": {
       "time_ms": 0.432,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.366,
     "pct": 0.107,
     "by_device": {
      "CPU": {
       "time_ms": 0.366,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.363,
     "pct": 0.106,
     "by_device": {
      "CPU": {
       "time_ms": 0.363,
       "nodes": 1,
       "calls": 106
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.313,
     "pct": 0.092,
     "by_device": {
      "CPU": {
       "time_ms": 0.313,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.224,
     "pct": 0.066,
     "by_device": {
      "CPU": {
       "time_ms": 0.224,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.206,
     "pct": 0.061,
     "by_device": {
      "CPU": {
       "time_ms": 0.206,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.119,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.119,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.107,
     "pct": 0.031,
     "by_device": {
      "CPU": {
       "time_ms": 0.107,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.101,
     "pct": 0.03,
     "by_device": {
      "CPU": {
       "time_ms": 0.101,
       "nodes": 1,
       "calls": 588
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.05,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.05,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.023,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.017,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.01,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.01,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.009,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.006,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.005,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_opt-350m_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_opt-350m_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:38:39",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "phi-1_5",
   "hf_model_id": "microsoft/phi-1_5",
   "hf_model_url": "https://huggingface.co/microsoft/phi-1_5",
   "param_count": 1418270720
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.61,
   "TTFT": 0.494,
   "TBT_avg": 0.115,
   "Throughput": 52.48
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 609.756,
     "ttft": 494.287,
     "tbt_avg": 115.469,
     "throughput": 52.48,
     "steps": [
      494.287,
      115.469
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 609.756,
     "min": 609.756,
     "max": 609.756,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 494.287,
     "min": 494.287,
     "max": 494.287,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 115.469,
     "min": 115.469,
     "max": 115.469,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 52.48,
     "min": 52.48,
     "max": 52.48,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 115.469,
     "min": 115.469,
     "max": 115.469,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1154.302,
   "total_nodes": 52,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1154.302,
     "pct": 100.0,
     "nodes": 52,
     "calls": 10547
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 545.909,
     "pct": 47.293,
     "by_device": {
      "CPU": {
       "time_ms": 545.909,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 540.75,
     "pct": 46.847,
     "by_device": {
      "CPU": {
       "time_ms": 540.75,
       "nodes": 1,
       "calls": 289
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 11.042,
     "pct": 0.957,
     "by_device": {
      "CPU": {
       "time_ms": 11.042,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 8.897,
     "pct": 0.771,
     "by_device": {
      "CPU": {
       "time_ms": 8.897,
       "nodes": 1,
       "calls": 388
      }
     }
    },
    {
     "op": "add",
     "total_ms": 6.321,
     "pct": 0.548,
     "by_device": {
      "CPU": {
       "time_ms": 6.321,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 5.623,
     "pct": 0.487,
     "by_device": {
      "CPU": {
       "time_ms": 5.623,
       "nodes": 1,
       "calls": 519
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 3.882,
     "pct": 0.336,
     "by_device": {
      "CPU": {
       "time_ms": 3.882,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 3.637,
     "pct": 0.315,
     "by_device": {
      "CPU": {
       "time_ms": 3.637,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 2.12,
     "pct": 0.184,
     "by_device": {
      "CPU": {
       "time_ms": 2.12,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 2.091,
     "pct": 0.181,
     "by_device": {
      "CPU": {
       "time_ms": 2.091,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 2.022,
     "pct": 0.175,
     "by_device": {
      "CPU": {
       "time_ms": 2.022,
       "nodes": 1,
       "calls": 915
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 1.98,
     "pct": 0.172,
     "by_device": {
      "CPU": {
       "time_ms": 1.98,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 1.95,
     "pct": 0.169,
     "by_device": {
      "CPU": {
       "time_ms": 1.95,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 1.896,
     "pct": 0.164,
     "by_device": {
      "CPU": {
       "time_ms": 1.896,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.828,
     "pct": 0.158,
     "by_device": {
      "CPU": {
       "time_ms": 1.828,
       "nodes": 1,
       "calls": 772
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.581,
     "pct": 0.137,
     "by_device": {
      "CPU": {
       "time_ms": 1.581,
       "nodes": 1,
       "calls": 529
      }
     }
    },
    {
     "op": "to",
     "total_ms": 1.303,
     "pct": 0.113,
     "by_device": {
      "CPU": {
       "time_ms": 1.303,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 1.298,
     "pct": 0.112,
     "by_device": {
      "CPU": {
       "time_ms": 1.298,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.207,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 1.207,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 1.166,
     "pct": 0.101,
     "by_device": {
      "CPU": {
       "time_ms": 1.166,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.075,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 1.075,
       "nodes": 1,
       "calls": 876
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 1.055,
     "pct": 0.091,
     "by_device": {
      "CPU": {
       "time_ms": 1.055,
       "nodes": 1,
       "calls": 206
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.931,
     "pct": 0.081,
     "by_device": {
      "CPU": {
       "time_ms": 0.931,
       "nodes": 1,
       "calls": 2090
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.707,
     "pct": 0.061,
     "by_device": {
      "CPU": {
       "time_ms": 0.707,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.659,
     "pct": 0.057,
     "by_device": {
      "CPU": {
       "time_ms": 0.659,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.605,
     "pct": 0.052,
     "by_device": {
      "CPU": {
       "time_ms": 0.605,
       "nodes": 1,
       "calls": 295
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.416,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 0.416,
       "nodes": 1,
       "calls": 368
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.344,
     "pct": 0.03,
     "by_device": {
      "CPU": {
       "time_ms": 0.344,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.336,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.336,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.322,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.322,
       "nodes": 1,
       "calls": 254
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.311,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.311,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.271,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.271,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.203,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.203,
       "nodes": 1,
       "calls": 55
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.112,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.112,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.095,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.095,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.084,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.084,
       "nodes": 1,
       "calls": 584
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.043,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.043,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.036,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.034,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.033,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "add_",
     "total_ms": 0.033,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.018,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.014,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.012,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.003,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_phi-1_5_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_phi-1_5_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:39:25",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "pythia-1b",
   "hf_model_id": "EleutherAI/pythia-1b",
   "hf_model_url": "https://huggingface.co/EleutherAI/pythia-1b",
   "param_count": 1011781632
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.288,
   "TTFT": 0.202,
   "TBT_avg": 0.086,
   "Throughput": 111.145
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 287.912,
     "ttft": 201.583,
     "tbt_avg": 86.329,
     "throughput": 111.145,
     "steps": [
      201.583,
      86.329
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 287.912,
     "min": 287.912,
     "max": 287.912,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 201.583,
     "min": 201.583,
     "max": 201.583,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 86.329,
     "min": 86.329,
     "max": 86.329,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 111.145,
     "min": 111.145,
     "max": 111.145,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 86.329,
     "min": 86.329,
     "max": 86.329,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 542.064,
   "total_nodes": 51,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 542.064,
     "pct": 100.0,
     "nodes": 51,
     "calls": 5913
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 251.126,
     "pct": 46.328,
     "by_device": {
      "CPU": {
       "time_ms": 251.126,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 247.302,
     "pct": 45.622,
     "by_device": {
      "CPU": {
       "time_ms": 247.302,
       "nodes": 1,
       "calls": 128
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 6.818,
     "pct": 1.258,
     "by_device": {
      "CPU": {
       "time_ms": 6.818,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 3.74,
     "pct": 0.69,
     "by_device": {
      "CPU": {
       "time_ms": 3.74,
       "nodes": 1,
       "calls": 132
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 3.512,
     "pct": 0.648,
     "by_device": {
      "CPU": {
       "time_ms": 3.512,
       "nodes": 1,
       "calls": 158
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.895,
     "pct": 0.534,
     "by_device": {
      "CPU": {
       "time_ms": 2.895,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.749,
     "pct": 0.507,
     "by_device": {
      "CPU": {
       "time_ms": 2.749,
       "nodes": 1,
       "calls": 128
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 2.726,
     "pct": 0.503,
     "by_device": {
      "CPU": {
       "time_ms": 2.726,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 2.521,
     "pct": 0.465,
     "by_device": {
      "CPU": {
       "time_ms": 2.521,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 2.347,
     "pct": 0.433,
     "by_device": {
      "CPU": {
       "time_ms": 2.347,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 2.136,
     "pct": 0.394,
     "by_device": {
      "CPU": {
       "time_ms": 2.136,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 1.993,
     "pct": 0.368,
     "by_device": {
      "CPU": {
       "time_ms": 1.993,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 1.59,
     "pct": 0.293,
     "by_device": {
      "CPU": {
       "time_ms": 1.59,
       "nodes": 1,
       "calls": 707
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.465,
     "pct": 0.27,
     "by_device": {
      "CPU": {
       "time_ms": 1.465,
       "nodes": 1,
       "calls": 449
      }
     }
    },
    {
     "op": "gelu",
     "total_ms": 1.431,
     "pct": 0.264,
     "by_device": {
      "CPU": {
       "time_ms": 1.431,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 0.91,
     "pct": 0.168,
     "by_device": {
      "CPU": {
       "time_ms": 0.91,
       "nodes": 1,
       "calls": 388
      }
     }
    },
    {
     "op": "chunk",
     "total_ms": 0.682,
     "pct": 0.126,
     "by_device": {
      "CPU": {
       "time_ms": 0.682,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.617,
     "pct": 0.114,
     "by_device": {
      "CPU": {
       "time_ms": 0.617,
       "nodes": 1,
       "calls": 1305
      }
     }
    },
    {
     "op": "split",
     "total_ms": 0.598,
     "pct": 0.11,
     "by_device": {
      "CPU": {
       "time_ms": 0.598,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.555,
     "pct": 0.102,
     "by_device": {
      "CPU": {
       "time_ms": 0.555,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.521,
     "pct": 0.096,
     "by_device": {
      "CPU": {
       "time_ms": 0.521,
       "nodes": 1,
       "calls": 459
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.482,
     "pct": 0.089,
     "by_device": {
      "CPU": {
       "time_ms": 0.482,
       "nodes": 1,
       "calls": 16
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.452,
     "pct": 0.083,
     "by_device": {
      "CPU": {
       "time_ms": 0.452,
       "nodes": 1,
       "calls": 16
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.426,
     "pct": 0.079,
     "by_device": {
      "CPU": {
       "time_ms": 0.426,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.358,
     "pct": 0.066,
     "by_device": {
      "CPU": {
       "time_ms": 0.358,
       "nodes": 1,
       "calls": 344
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.292,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.292,
       "nodes": 1,
       "calls": 134
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.282,
     "pct": 0.052,
     "by_device": {
      "CPU": {
       "time_ms": 0.282,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.254,
     "pct": 0.047,
     "by_device": {
      "CPU": {
       "time_ms": 0.254,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.254,
     "pct": 0.047,
     "by_device": {
      "CPU": {
       "time_ms": 0.254,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.198,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.198,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.154,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.154,
       "nodes": 1,
       "calls": 40
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.123,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.123,
       "nodes": 1,
       "calls": 54
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.095,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.095,
       "nodes": 1,
       "calls": 14
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.095,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.095,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.076,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.076,
       "nodes": 1,
       "calls": 46
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.062,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.062,
       "nodes": 2,
       "calls": 64
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.045,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.045,
       "nodes": 1,
       "calls": 264
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.039,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.039,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.03,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.029,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.022,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.011,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.01,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.009,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.008,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.006,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_pythia-1b_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_pythia-1b_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:38:24",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "Qwen2.5-1.5B",
   "hf_model_id": "Qwen/Qwen2.5-1.5B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen2.5-1.5B",
   "param_count": 1543714304
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.39,
   "TTFT": 0.32,
   "TBT_avg": 0.07,
   "Throughput": 82.017
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 390.164,
     "ttft": 320.362,
     "tbt_avg": 69.802,
     "throughput": 82.017,
     "steps": [
      320.362,
      69.802
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 390.164,
     "min": 390.164,
     "max": 390.164,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 320.362,
     "min": 320.362,
     "max": 320.362,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 69.802,
     "min": 69.802,
     "max": 69.802,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 82.017,
     "min": 82.017,
     "max": 82.017,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 69.802,
     "min": 69.802,
     "max": 69.802,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 935.533,
   "total_nodes": 49,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 935.533,
     "pct": 100.0,
     "nodes": 49,
     "calls": 12673
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 320.649,
     "pct": 34.275,
     "by_device": {
      "CPU": {
       "time_ms": 320.649,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 224.406,
     "pct": 23.987,
     "by_device": {
      "CPU": {
       "time_ms": 224.406,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 222.096,
     "pct": 23.74,
     "by_device": {
      "CPU": {
       "time_ms": 222.096,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 92.655,
     "pct": 9.904,
     "by_device": {
      "CPU": {
       "time_ms": 92.655,
       "nodes": 1,
       "calls": 168
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 9.064,
     "pct": 0.969,
     "by_device": {
      "CPU": {
       "time_ms": 9.064,
       "nodes": 1,
       "calls": 512
      }
     }
    },
    {
     "op": "add",
     "total_ms": 8.35,
     "pct": 0.893,
     "by_device": {
      "CPU": {
       "time_ms": 8.35,
       "nodes": 1,
       "calls": 338
      }
     }
    },
    {
     "op": "to",
     "total_ms": 6.463,
     "pct": 0.691,
     "by_device": {
      "CPU": {
       "time_ms": 6.463,
       "nodes": 1,
       "calls": 648
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 5.849,
     "pct": 0.625,
     "by_device": {
      "CPU": {
       "time_ms": 5.849,
       "nodes": 1,
       "calls": 470
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 4.918,
     "pct": 0.526,
     "by_device": {
      "CPU": {
       "time_ms": 4.918,
       "nodes": 1,
       "calls": 638
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 4.68,
     "pct": 0.5,
     "by_device": {
      "CPU": {
       "time_ms": 4.68,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 4.398,
     "pct": 0.47,
     "by_device": {
      "CPU": {
       "time_ms": 4.398,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 4.233,
     "pct": 0.452,
     "by_device": {
      "CPU": {
       "time_ms": 4.233,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 4.107,
     "pct": 0.439,
     "by_device": {
      "CPU": {
       "time_ms": 4.107,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 2.939,
     "pct": 0.314,
     "by_device": {
      "CPU": {
       "time_ms": 2.939,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 2.726,
     "pct": 0.291,
     "by_device": {
      "CPU": {
       "time_ms": 2.726,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 2.335,
     "pct": 0.25,
     "by_device": {
      "CPU": {
       "time_ms": 2.335,
       "nodes": 1,
       "calls": 956
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 2.138,
     "pct": 0.229,
     "by_device": {
      "CPU": {
       "time_ms": 2.138,
       "nodes": 1,
       "calls": 118
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 1.919,
     "pct": 0.205,
     "by_device": {
      "CPU": {
       "time_ms": 1.919,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.712,
     "pct": 0.183,
     "by_device": {
      "CPU": {
       "time_ms": 1.712,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 1.438,
     "pct": 0.154,
     "by_device": {
      "CPU": {
       "time_ms": 1.438,
       "nodes": 1,
       "calls": 619
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.183,
     "pct": 0.126,
     "by_device": {
      "CPU": {
       "time_ms": 1.183,
       "nodes": 1,
       "calls": 393
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.998,
     "pct": 0.107,
     "by_device": {
      "CPU": {
       "time_ms": 0.998,
       "nodes": 1,
       "calls": 791
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.94,
     "pct": 0.1,
     "by_device": {
      "CPU": {
       "time_ms": 0.94,
       "nodes": 1,
       "calls": 112
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.919,
     "pct": 0.098,
     "by_device": {
      "CPU": {
       "time_ms": 0.919,
       "nodes": 1,
       "calls": 1873
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.729,
     "pct": 0.078,
     "by_device": {
      "CPU": {
       "time_ms": 0.729,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.711,
     "pct": 0.076,
     "by_device": {
      "CPU": {
       "time_ms": 0.711,
       "nodes": 1,
       "calls": 526
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.394,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.394,
       "nodes": 1,
       "calls": 174
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.379,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.379,
       "nodes": 1,
       "calls": 118
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.371,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.371,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.284,
     "pct": 0.03,
     "by_device": {
      "CPU": {
       "time_ms": 0.284,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.265,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.265,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.253,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.253,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.23,
     "pct": 0.025,
     "by_device": {
      "CPU": {
       "time_ms": 0.23,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.223,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 0.223,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.146,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.146,
       "nodes": 1,
       "calls": 116
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.103,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.103,
       "nodes": 2,
       "calls": 112
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.102,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.102,
       "nodes": 1,
       "calls": 792
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.06,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.054,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.054,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.034,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.023,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.012,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.01,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.006,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_qwen2.5-1.5b_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_qwen2.5-1.5b_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:38:17",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "Qwen3-0.6B",
   "hf_model_id": "Qwen/Qwen3-0.6B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen3-0.6B",
   "param_count": 596049920
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.151,
   "TTFT": 0.097,
   "TBT_avg": 0.054,
   "Throughput": 211.751
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 151.121,
     "ttft": 96.637,
     "tbt_avg": 54.484,
     "throughput": 211.751,
     "steps": [
      96.637,
      54.484
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 151.121,
     "min": 151.121,
     "max": 151.121,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 96.637,
     "min": 96.637,
     "max": 96.637,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 54.484,
     "min": 54.484,
     "max": 54.484,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 211.751,
     "min": 211.751,
     "max": 211.751,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 54.484,
     "min": 54.484,
     "max": 54.484,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 290.164,
   "total_nodes": 48,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 290.164,
     "pct": 100.0,
     "nodes": 48,
     "calls": 15641
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 65.071,
     "pct": 22.426,
     "by_device": {
      "CPU": {
       "time_ms": 65.071,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 62.846,
     "pct": 21.659,
     "by_device": {
      "CPU": {
       "time_ms": 62.846,
       "nodes": 1,
       "calls": 396
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 59.934,
     "pct": 20.655,
     "by_device": {
      "CPU": {
       "time_ms": 59.934,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 12.435,
     "pct": 4.285,
     "by_device": {
      "CPU": {
       "time_ms": 12.435,
       "nodes": 1,
       "calls": 736
      }
     }
    },
    {
     "op": "to",
     "total_ms": 12.224,
     "pct": 4.213,
     "by_device": {
      "CPU": {
       "time_ms": 12.224,
       "nodes": 1,
       "calls": 1208
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 11.126,
     "pct": 3.835,
     "by_device": {
      "CPU": {
       "time_ms": 11.126,
       "nodes": 1,
       "calls": 918
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 7.734,
     "pct": 2.665,
     "by_device": {
      "CPU": {
       "time_ms": 7.734,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 7.417,
     "pct": 2.556,
     "by_device": {
      "CPU": {
       "time_ms": 7.417,
       "nodes": 1,
       "calls": 918
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 7.083,
     "pct": 2.441,
     "by_device": {
      "CPU": {
       "time_ms": 7.083,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "add",
     "total_ms": 5.62,
     "pct": 1.937,
     "by_device": {
      "CPU": {
       "time_ms": 5.62,
       "nodes": 1,
       "calls": 450
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 5.019,
     "pct": 1.73,
     "by_device": {
      "CPU": {
       "time_ms": 5.019,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 4.67,
     "pct": 1.61,
     "by_device": {
      "CPU": {
       "time_ms": 4.67,
       "nodes": 1,
       "calls": 452
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 4.022,
     "pct": 1.386,
     "by_device": {
      "CPU": {
       "time_ms": 4.022,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 3.981,
     "pct": 1.372,
     "by_device": {
      "CPU": {
       "time_ms": 3.981,
       "nodes": 1,
       "calls": 230
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 3.737,
     "pct": 1.288,
     "by_device": {
      "CPU": {
       "time_ms": 3.737,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 2.965,
     "pct": 1.022,
     "by_device": {
      "CPU": {
       "time_ms": 2.965,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 2.233,
     "pct": 0.77,
     "by_device": {
      "CPU": {
       "time_ms": 2.233,
       "nodes": 1,
       "calls": 956
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.558,
     "pct": 0.537,
     "by_device": {
      "CPU": {
       "time_ms": 1.558,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 1.481,
     "pct": 0.51,
     "by_device": {
      "CPU": {
       "time_ms": 1.481,
       "nodes": 1,
       "calls": 112
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 1.432,
     "pct": 0.494,
     "by_device": {
      "CPU": {
       "time_ms": 1.432,
       "nodes": 1,
       "calls": 619
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.167,
     "pct": 0.402,
     "by_device": {
      "CPU": {
       "time_ms": 1.167,
       "nodes": 1,
       "calls": 974
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.153,
     "pct": 0.397,
     "by_device": {
      "CPU": {
       "time_ms": 1.153,
       "nodes": 1,
       "calls": 393
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.946,
     "pct": 0.326,
     "by_device": {
      "CPU": {
       "time_ms": 0.946,
       "nodes": 1,
       "calls": 456
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.796,
     "pct": 0.274,
     "by_device": {
      "CPU": {
       "time_ms": 0.796,
       "nodes": 1,
       "calls": 1705
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.648,
     "pct": 0.223,
     "by_device": {
      "CPU": {
       "time_ms": 0.648,
       "nodes": 1,
       "calls": 623
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.458,
     "pct": 0.158,
     "by_device": {
      "CPU": {
       "time_ms": 0.458,
       "nodes": 1,
       "calls": 396
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.427,
     "pct": 0.147,
     "by_device": {
      "CPU": {
       "time_ms": 0.427,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.364,
     "pct": 0.125,
     "by_device": {
      "CPU": {
       "time_ms": 0.364,
       "nodes": 1,
       "calls": 118
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.276,
     "pct": 0.095,
     "by_device": {
      "CPU": {
       "time_ms": 0.276,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.243,
     "pct": 0.084,
     "by_device": {
      "CPU": {
       "time_ms": 0.243,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.227,
     "pct": 0.078,
     "by_device": {
      "CPU": {
       "time_ms": 0.227,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.217,
     "pct": 0.075,
     "by_device": {
      "CPU": {
       "time_ms": 0.217,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.204,
     "pct": 0.07,
     "by_device": {
      "CPU": {
       "time_ms": 0.204,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.096,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.096,
       "nodes": 2,
       "calls": 112
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.088,
     "pct": 0.03,
     "by_device": {
      "CPU": {
       "time_ms": 0.088,
       "nodes": 1,
       "calls": 792
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.075,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.075,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.054,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.054,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.034,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.026,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.024,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.01,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.01,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.009,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.008,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_qwen3-0.6b_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_qwen3-0.6b_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:38:10",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "stablelm-2-1_6b",
   "hf_model_id": "stabilityai/stablelm-2-1_6b",
   "hf_model_url": "https://huggingface.co/stabilityai/stablelm-2-1_6b",
   "param_count": 1644417024
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.846,
   "TTFT": 0.784,
   "TBT_avg": 0.061,
   "Throughput": 37.844
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 845.568,
     "ttft": 784.294,
     "tbt_avg": 61.274,
     "throughput": 37.844,
     "steps": [
      784.294,
      61.274
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 845.568,
     "min": 845.568,
     "max": 845.568,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 784.294,
     "min": 784.294,
     "max": 784.294,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 61.274,
     "min": 61.274,
     "max": 61.274,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 37.844,
     "min": 37.844,
     "max": 37.844,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 61.274,
     "min": 61.274,
     "max": 61.274,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2060.662,
   "total_nodes": 49,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 2060.662,
     "pct": 100.0,
     "nodes": 49,
     "calls": 10063
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 779.294,
     "pct": 37.818,
     "by_device": {
      "CPU": {
       "time_ms": 779.294,
       "nodes": 1,
       "calls": 338
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 439.23,
     "pct": 21.315,
     "by_device": {
      "CPU": {
       "time_ms": 439.23,
       "nodes": 1,
       "calls": 196
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 437.12,
     "pct": 21.213,
     "by_device": {
      "CPU": {
       "time_ms": 437.12,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 336.632,
     "pct": 16.336,
     "by_device": {
      "CPU": {
       "time_ms": 336.632,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 13.632,
     "pct": 0.662,
     "by_device": {
      "CPU": {
       "time_ms": 13.632,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "add",
     "total_ms": 6.658,
     "pct": 0.323,
     "by_device": {
      "CPU": {
       "time_ms": 6.658,
       "nodes": 1,
       "calls": 192
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 6.446,
     "pct": 0.313,
     "by_device": {
      "CPU": {
       "time_ms": 6.446,
       "nodes": 1,
       "calls": 244
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 4.658,
     "pct": 0.226,
     "by_device": {
      "CPU": {
       "time_ms": 4.658,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 4.346,
     "pct": 0.211,
     "by_device": {
      "CPU": {
       "time_ms": 4.346,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 4.128,
     "pct": 0.2,
     "by_device": {
      "CPU": {
       "time_ms": 4.128,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 3.864,
     "pct": 0.188,
     "by_device": {
      "CPU": {
       "time_ms": 3.864,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 3.527,
     "pct": 0.171,
     "by_device": {
      "CPU": {
       "time_ms": 3.527,
       "nodes": 1,
       "calls": 182
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 2.567,
     "pct": 0.125,
     "by_device": {
      "CPU": {
       "time_ms": 2.567,
       "nodes": 1,
       "calls": 915
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 2.411,
     "pct": 0.117,
     "by_device": {
      "CPU": {
       "time_ms": 2.411,
       "nodes": 1,
       "calls": 820
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 2.272,
     "pct": 0.11,
     "by_device": {
      "CPU": {
       "time_ms": 2.272,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 2.044,
     "pct": 0.099,
     "by_device": {
      "CPU": {
       "time_ms": 2.044,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.832,
     "pct": 0.089,
     "by_device": {
      "CPU": {
       "time_ms": 1.832,
       "nodes": 1,
       "calls": 529
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.558,
     "pct": 0.076,
     "by_device": {
      "CPU": {
       "time_ms": 1.558,
       "nodes": 1,
       "calls": 338
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.163,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 1.163,
       "nodes": 1,
       "calls": 1993
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.075,
     "pct": 0.052,
     "by_device": {
      "CPU": {
       "time_ms": 1.075,
       "nodes": 1,
       "calls": 875
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.819,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.819,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.761,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.761,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.727,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.727,
       "nodes": 1,
       "calls": 248
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.662,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.662,
       "nodes": 1,
       "calls": 512
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.552,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.552,
       "nodes": 1,
       "calls": 150
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.448,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.448,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.352,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.352,
       "nodes": 1,
       "calls": 196
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.342,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.342,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.336,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.336,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.306,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.306,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.144,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.144,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.124,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.124,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.122,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.122,
       "nodes": 1,
       "calls": 62
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.119,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.119,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.109,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.109,
       "nodes": 1,
       "calls": 14
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.095,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.095,
       "nodes": 1,
       "calls": 680
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.036,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.036,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.029,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.018,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.014,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.013,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.013,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.012,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.009,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_stablelm-2-1_6b_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_stablelm-2-1_6b_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:37:54",
   "framework": "pytorch",
   "device": "cpu"
  },
  "model_info": {
   "model_name": "TinyLlama-1.1B-Chat-v1.0",
   "hf_model_id": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "hf_model_url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "param_count": 1100048384
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.592,
   "TTFT": 0.538,
   "TBT_avg": 0.054,
   "Throughput": 54.069
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 591.839,
     "ttft": 537.534,
     "tbt_avg": 54.306,
     "throughput": 54.069,
     "steps": [
      537.534,
      54.306
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 591.839,
     "min": 591.839,
     "max": 591.839,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 537.534,
     "min": 537.534,
     "max": 537.534,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 54.306,
     "min": 54.306,
     "max": 54.306,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 54.069,
     "min": 54.069,
     "max": 54.069,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 54.306,
     "min": 54.306,
     "max": 54.306,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1652.164,
   "total_nodes": 48,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 1652.164,
     "pct": 100.0,
     "nodes": 48,
     "calls": 9877
    }
   ],
   "ops": [
    {
     "op": "linear",
     "total_ms": 530.693,
     "pct": 32.121,
     "by_device": {
      "CPU": {
       "time_ms": 530.693,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 528.716,
     "pct": 32.001,
     "by_device": {
      "CPU": {
       "time_ms": 528.716,
       "nodes": 1,
       "calls": 312
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 525.838,
     "pct": 31.827,
     "by_device": {
      "CPU": {
       "time_ms": 525.838,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 10.606,
     "pct": 0.642,
     "by_device": {
      "CPU": {
       "time_ms": 10.606,
       "nodes": 1,
       "calls": 404
      }
     }
    },
    {
     "op": "to",
     "total_ms": 6.064,
     "pct": 0.367,
     "by_device": {
      "CPU": {
       "time_ms": 6.064,
       "nodes": 1,
       "calls": 516
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 5.581,
     "pct": 0.338,
     "by_device": {
      "CPU": {
       "time_ms": 5.581,
       "nodes": 1,
       "calls": 374
      }
     }
    },
    {
     "op": "add",
     "total_ms": 4.485,
     "pct": 0.271,
     "by_device": {
      "CPU": {
       "time_ms": 4.485,
       "nodes": 1,
       "calls": 266
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 4.107,
     "pct": 0.249,
     "by_device": {
      "CPU": {
       "time_ms": 4.107,
       "nodes": 1,
       "calls": 178
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 3.841,
     "pct": 0.232,
     "by_device": {
      "CPU": {
       "time_ms": 3.841,
       "nodes": 1,
       "calls": 374
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 3.803,
     "pct": 0.23,
     "by_device": {
      "CPU": {
       "time_ms": 3.803,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 3.585,
     "pct": 0.217,
     "by_device": {
      "CPU": {
       "time_ms": 3.585,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 3.34,
     "pct": 0.202,
     "by_device": {
      "CPU": {
       "time_ms": 3.34,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 3.12,
     "pct": 0.189,
     "by_device": {
      "CPU": {
       "time_ms": 3.12,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 2.863,
     "pct": 0.173,
     "by_device": {
      "CPU": {
       "time_ms": 2.863,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 2.229,
     "pct": 0.135,
     "by_device": {
      "CPU": {
       "time_ms": 2.229,
       "nodes": 1,
       "calls": 94
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 2.012,
     "pct": 0.122,
     "by_device": {
      "CPU": {
       "time_ms": 2.012,
       "nodes": 1,
       "calls": 180
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.903,
     "pct": 0.115,
     "by_device": {
      "CPU": {
       "time_ms": 1.903,
       "nodes": 1,
       "calls": 752
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.349,
     "pct": 0.082,
     "by_device": {
      "CPU": {
       "time_ms": 1.349,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 1.112,
     "pct": 0.067,
     "by_device": {
      "CPU": {
       "time_ms": 1.112,
       "nodes": 1,
       "calls": 487
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 1.032,
     "pct": 0.062,
     "by_device": {
      "CPU": {
       "time_ms": 1.032,
       "nodes": 1,
       "calls": 88
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.929,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 0.929,
       "nodes": 1,
       "calls": 309
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.824,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 0.824,
       "nodes": 1,
       "calls": 360
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.676,
     "pct": 0.041,
     "by_device": {
      "CPU": {
       "time_ms": 0.676,
       "nodes": 1,
       "calls": 1345
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.563,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.563,
       "nodes": 1,
       "calls": 418
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.558,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.558,
       "nodes": 1,
       "calls": 491
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.425,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.425,
       "nodes": 1,
       "calls": 312
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.315,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.315,
       "nodes": 1,
       "calls": 94
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.294,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.294,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.258,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.258,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.184,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.184,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.181,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.181,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.173,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.173,
       "nodes": 1,
       "calls": 178
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.108,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.108,
       "nodes": 1,
       "calls": 92
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.086,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.086,
       "nodes": 2,
       "calls": 88
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.078,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.078,
       "nodes": 1,
       "calls": 624
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.051,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.051,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.037,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.035,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.031,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.031,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.027,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.012,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.008,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.007,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.007,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/cpu/torch_tinyllama-1.1b-chat_b16s8o2_cpu.simplified.json",
   "source_profile": "torch_tinyllama-1.1b-chat_b16s8o2_cpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:30:32",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "bloom-1b7",
   "hf_model_id": "bigscience/bloom-1b7",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-1b7",
   "param_count": 1722408960
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.034,
   "TTFT": 0.019,
   "TBT_avg": 0.015,
   "Throughput": 942.736
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 33.944,
     "ttft": 18.696,
     "tbt_avg": 15.248,
     "throughput": 942.736,
     "steps": [
      18.696,
      15.248
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 33.944,
     "min": 33.944,
     "max": 33.944,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 18.696,
     "min": 18.696,
     "max": 18.696,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 15.248,
     "min": 15.248,
     "max": 15.248,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 942.736,
     "min": 942.736,
     "max": 942.736,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 15.248,
     "min": 15.248,
     "max": 15.248,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 67.085,
   "total_nodes": 98,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 41.14,
     "pct": 61.325,
     "nodes": 56,
     "calls": 5790
    },
    {
     "device": "GPU",
     "time_ms": 25.945,
     "pct": 38.675,
     "nodes": 42,
     "calls": 1354
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 25.914,
     "pct": 38.629,
     "by_device": {
      "GPU": {
       "time_ms": 25.914,
       "nodes": 40,
       "calls": 1326
      }
     }
    },
    {
     "op": "GeLUFunction",
     "total_ms": 5.289,
     "pct": 7.884,
     "by_device": {
      "CPU": {
       "time_ms": 5.289,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 4.998,
     "pct": 7.45,
     "by_device": {
      "CPU": {
       "time_ms": 4.998,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 3.116,
     "pct": 4.644,
     "by_device": {
      "CPU": {
       "time_ms": 3.116,
       "nodes": 1,
       "calls": 192
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 2.414,
     "pct": 3.599,
     "by_device": {
      "CPU": {
       "time_ms": 2.414,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 2.282,
     "pct": 3.401,
     "by_device": {
      "CPU": {
       "time_ms": 2.282,
       "nodes": 1,
       "calls": 292
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 2.116,
     "pct": 3.154,
     "by_device": {
      "CPU": {
       "time_ms": 2.116,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.091,
     "pct": 3.116,
     "by_device": {
      "CPU": {
       "time_ms": 2.091,
       "nodes": 1,
       "calls": 244
      }
     }
    },
    {
     "op": "to",
     "total_ms": 1.864,
     "pct": 2.778,
     "by_device": {
      "CPU": {
       "time_ms": 1.864,
       "nodes": 1,
       "calls": 162
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 1.687,
     "pct": 2.515,
     "by_device": {
      "CPU": {
       "time_ms": 1.687,
       "nodes": 1,
       "calls": 152
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 1.612,
     "pct": 2.402,
     "by_device": {
      "CPU": {
       "time_ms": 1.612,
       "nodes": 1,
       "calls": 252
      }
     }
    },
    {
     "op": "baddbmm",
     "total_ms": 1.431,
     "pct": 2.133,
     "by_device": {
      "CPU": {
       "time_ms": 1.431,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 1.402,
     "pct": 2.09,
     "by_device": {
      "CPU": {
       "time_ms": 1.402,
       "nodes": 1,
       "calls": 202
      }
     }
    },
    {
     "op": "softmax",
     "total_ms": 1.332,
     "pct": 1.986,
     "by_device": {
      "CPU": {
       "time_ms": 1.332,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 0.985,
     "pct": 1.468,
     "by_device": {
      "CPU": {
       "time_ms": 0.985,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.95,
     "pct": 1.416,
     "by_device": {
      "CPU": {
       "time_ms": 0.95,
       "nodes": 1,
       "calls": 422
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 0.91,
     "pct": 1.357,
     "by_device": {
      "CPU": {
       "time_ms": 0.91,
       "nodes": 1,
       "calls": 386
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.754,
     "pct": 1.124,
     "by_device": {
      "CPU": {
       "time_ms": 0.754,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.684,
     "pct": 1.02,
     "by_device": {
      "CPU": {
       "time_ms": 0.684,
       "nodes": 1,
       "calls": 904
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.672,
     "pct": 1.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.672,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.666,
     "pct": 0.993,
     "by_device": {
      "CPU": {
       "time_ms": 0.666,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_softmax",
     "total_ms": 0.44,
     "pct": 0.656,
     "by_device": {
      "CPU": {
       "time_ms": 0.44,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.43,
     "pct": 0.641,
     "by_device": {
      "CPU": {
       "time_ms": 0.43,
       "nodes": 1,
       "calls": 152
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.396,
     "pct": 0.59,
     "by_device": {
      "CPU": {
       "time_ms": 0.396,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 0.394,
     "pct": 0.588,
     "by_device": {
      "CPU": {
       "time_ms": 0.394,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.31,
     "pct": 0.462,
     "by_device": {
      "CPU": {
       "time_ms": 0.31,
       "nodes": 1,
       "calls": 643
      }
     }
    },
    {
     "op": "arange",
     "total_ms": 0.223,
     "pct": 0.332,
     "by_device": {
      "CPU": {
       "time_ms": 0.223,
       "nodes": 1,
       "calls": 20
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.223,
     "pct": 0.333,
     "by_device": {
      "CPU": {
       "time_ms": 0.223,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "permute",
     "total_ms": 0.166,
     "pct": 0.247,
     "by_device": {
      "CPU": {
       "time_ms": 0.166,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "index",
     "total_ms": 0.135,
     "pct": 0.202,
     "by_device": {
      "CPU": {
       "time_ms": 0.135,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "where",
     "total_ms": 0.109,
     "pct": 0.163,
     "by_device": {
      "CPU": {
       "time_ms": 0.109,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.106,
     "pct": 0.158,
     "by_device": {
      "CPU": {
       "time_ms": 0.106,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.104,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 0.104,
       "nodes": 1,
       "calls": 31
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.089,
     "pct": 0.133,
     "by_device": {
      "CPU": {
       "time_ms": 0.089,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.068,
     "pct": 0.101,
     "by_device": {
      "CPU": {
       "time_ms": 0.068,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.066,
     "pct": 0.098,
     "by_device": {
      "CPU": {
       "time_ms": 0.066,
       "nodes": 1,
       "calls": 28
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.06,
     "pct": 0.089,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.059,
     "pct": 0.089,
     "by_device": {
      "CPU": {
       "time_ms": 0.059,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "__and__",
     "total_ms": 0.049,
     "pct": 0.074,
     "by_device": {
      "CPU": {
       "time_ms": 0.049,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "bitwise_and",
     "total_ms": 0.044,
     "pct": 0.066,
     "by_device": {
      "CPU": {
       "time_ms": 0.044,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.04,
     "pct": 0.06,
     "by_device": {
      "CPU": {
       "time_ms": 0.04,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 0.04,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 0.04,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.039,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 0.039,
       "nodes": 1,
       "calls": 29
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.038,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.03,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.03,
     "pct": 0.045,
     "by_device": {
      "GPU": {
       "time_ms": 0.03,
       "nodes": 2,
       "calls": 28
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.029,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "new_ones",
     "total_ms": 0.029,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.028,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 14
      }
     }
    },
    {
     "op": "cumsum",
     "total_ms": 0.027,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "scalar_tensor",
     "total_ms": 0.027,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sub",
     "total_ms": 0.023,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "le",
     "total_ms": 0.022,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.021,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.01,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "new_empty",
     "total_ms": 0.008,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.005,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 52
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_bloom-1b7_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_bloom-1b7_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:29:30",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "bloom-560m",
   "hf_model_id": "bigscience/bloom-560m",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-560m",
   "param_count": 559214592
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.033,
   "TTFT": 0.018,
   "TBT_avg": 0.015,
   "Throughput": 978.603
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 32.7,
     "ttft": 17.927,
     "tbt_avg": 14.773,
     "throughput": 978.603,
     "steps": [
      17.927,
      14.773
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 32.7,
     "min": 32.7,
     "max": 32.7,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 17.927,
     "min": 17.927,
     "max": 17.927,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 14.773,
     "min": 14.773,
     "max": 14.773,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 978.603,
     "min": 978.603,
     "max": 978.603,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 14.773,
     "min": 14.773,
     "max": 14.773,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 49.732,
   "total_nodes": 99,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 39.576,
     "pct": 79.578,
     "nodes": 56,
     "calls": 5790
    },
    {
     "device": "GPU",
     "time_ms": 10.156,
     "pct": 20.422,
     "nodes": 43,
     "calls": 1378
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 10.129,
     "pct": 20.368,
     "by_device": {
      "GPU": {
       "time_ms": 10.129,
       "nodes": 41,
       "calls": 1350
      }
     }
    },
    {
     "op": "GeLUFunction",
     "total_ms": 5.115,
     "pct": 10.284,
     "by_device": {
      "CPU": {
       "time_ms": 5.115,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 4.801,
     "pct": 9.655,
     "by_device": {
      "CPU": {
       "time_ms": 4.801,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 2.993,
     "pct": 6.018,
     "by_device": {
      "CPU": {
       "time_ms": 2.993,
       "nodes": 1,
       "calls": 192
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 2.359,
     "pct": 4.743,
     "by_device": {
      "CPU": {
       "time_ms": 2.359,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 2.184,
     "pct": 4.392,
     "by_device": {
      "CPU": {
       "time_ms": 2.184,
       "nodes": 1,
       "calls": 292
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 2.073,
     "pct": 4.168,
     "by_device": {
      "CPU": {
       "time_ms": 2.073,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.013,
     "pct": 4.047,
     "by_device": {
      "CPU": {
       "time_ms": 2.013,
       "nodes": 1,
       "calls": 244
      }
     }
    },
    {
     "op": "to",
     "total_ms": 1.788,
     "pct": 3.596,
     "by_device": {
      "CPU": {
       "time_ms": 1.788,
       "nodes": 1,
       "calls": 162
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 1.622,
     "pct": 3.262,
     "by_device": {
      "CPU": {
       "time_ms": 1.622,
       "nodes": 1,
       "calls": 152
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 1.52,
     "pct": 3.056,
     "by_device": {
      "CPU": {
       "time_ms": 1.52,
       "nodes": 1,
       "calls": 252
      }
     }
    },
    {
     "op": "baddbmm",
     "total_ms": 1.392,
     "pct": 2.8,
     "by_device": {
      "CPU": {
       "time_ms": 1.392,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 1.347,
     "pct": 2.709,
     "by_device": {
      "CPU": {
       "time_ms": 1.347,
       "nodes": 1,
       "calls": 202
      }
     }
    },
    {
     "op": "softmax",
     "total_ms": 1.278,
     "pct": 2.57,
     "by_device": {
      "CPU": {
       "time_ms": 1.278,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.946,
     "pct": 1.902,
     "by_device": {
      "CPU": {
       "time_ms": 0.946,
       "nodes": 1,
       "calls": 422
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 0.91,
     "pct": 1.83,
     "by_device": {
      "CPU": {
       "time_ms": 0.91,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 0.876,
     "pct": 1.761,
     "by_device": {
      "CPU": {
       "time_ms": 0.876,
       "nodes": 1,
       "calls": 386
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.716,
     "pct": 1.439,
     "by_device": {
      "CPU": {
       "time_ms": 0.716,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.655,
     "pct": 1.317,
     "by_device": {
      "CPU": {
       "time_ms": 0.655,
       "nodes": 1,
       "calls": 904
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.636,
     "pct": 1.28,
     "by_device": {
      "CPU": {
       "time_ms": 0.636,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.633,
     "pct": 1.274,
     "by_device": {
      "CPU": {
       "time_ms": 0.633,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.422,
     "pct": 0.848,
     "by_device": {
      "CPU": {
       "time_ms": 0.422,
       "nodes": 1,
       "calls": 152
      }
     }
    },
    {
     "op": "_softmax",
     "total_ms": 0.414,
     "pct": 0.832,
     "by_device": {
      "CPU": {
       "time_ms": 0.414,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 0.374,
     "pct": 0.752,
     "by_device": {
      "CPU": {
       "time_ms": 0.374,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.365,
     "pct": 0.734,
     "by_device": {
      "CPU": {
       "time_ms": 0.365,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.292,
     "pct": 0.588,
     "by_device": {
      "CPU": {
       "time_ms": 0.292,
       "nodes": 1,
       "calls": 643
      }
     }
    },
    {
     "op": "arange",
     "total_ms": 0.225,
     "pct": 0.453,
     "by_device": {
      "CPU": {
       "time_ms": 0.225,
       "nodes": 1,
       "calls": 20
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.217,
     "pct": 0.436,
     "by_device": {
      "CPU": {
       "time_ms": 0.217,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "permute",
     "total_ms": 0.164,
     "pct": 0.33,
     "by_device": {
      "CPU": {
       "time_ms": 0.164,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "index",
     "total_ms": 0.131,
     "pct": 0.264,
     "by_device": {
      "CPU": {
       "time_ms": 0.131,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "where",
     "total_ms": 0.104,
     "pct": 0.209,
     "by_device": {
      "CPU": {
       "time_ms": 0.104,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.1,
     "pct": 0.201,
     "by_device": {
      "CPU": {
       "time_ms": 0.1,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.095,
     "pct": 0.191,
     "by_device": {
      "CPU": {
       "time_ms": 0.095,
       "nodes": 1,
       "calls": 31
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.089,
     "pct": 0.178,
     "by_device": {
      "CPU": {
       "time_ms": 0.089,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.068,
     "pct": 0.138,
     "by_device": {
      "CPU": {
       "time_ms": 0.068,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.062,
     "pct": 0.124,
     "by_device": {
      "CPU": {
       "time_ms": 0.062,
       "nodes": 1,
       "calls": 28
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.058,
     "pct": 0.117,
     "by_device": {
      "CPU": {
       "time_ms": 0.058,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.056,
     "pct": 0.113,
     "by_device": {
      "CPU": {
       "time_ms": 0.056,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "__and__",
     "total_ms": 0.043,
     "pct": 0.087,
     "by_device": {
      "CPU": {
       "time_ms": 0.043,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.041,
     "pct": 0.082,
     "by_device": {
      "CPU": {
       "time_ms": 0.041,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 0.039,
     "pct": 0.078,
     "by_device": {
      "CPU": {
       "time_ms": 0.039,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bitwise_and",
     "total_ms": 0.038,
     "pct": 0.077,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.036,
     "pct": 0.072,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 29
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.034,
     "pct": 0.069,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "new_ones",
     "total_ms": 0.029,
     "pct": 0.057,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 0.029,
     "pct": 0.058,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.027,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 14
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.027,
     "pct": 0.054,
     "by_device": {
      "GPU": {
       "time_ms": 0.027,
       "nodes": 2,
       "calls": 28
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.026,
     "pct": 0.053,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "scalar_tensor",
     "total_ms": 0.026,
     "pct": 0.052,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cumsum",
     "total_ms": 0.024,
     "pct": 0.049,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "le",
     "total_ms": 0.022,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.021,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "sub",
     "total_ms": 0.021,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.011,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "new_empty",
     "total_ms": 0.008,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.006,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 52
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_bloom-560m_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_bloom-560m_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:19:40",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "Falcon3-1B-Base",
   "hf_model_id": "tiiuae/Falcon3-1B-Base",
   "hf_model_url": "https://huggingface.co/tiiuae/Falcon3-1B-Base",
   "param_count": 1669000000
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.033,
   "TTFT": 0.016,
   "TBT_avg": 0.017,
   "Throughput": 976.458
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 32.772,
     "ttft": 16.19,
     "tbt_avg": 16.582,
     "throughput": 976.458,
     "steps": [
      16.19,
      16.582
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 32.772,
     "min": 32.772,
     "max": 32.772,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 16.19,
     "min": 16.19,
     "max": 16.19,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 16.582,
     "min": 16.582,
     "max": 16.582,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 976.458,
     "min": 976.458,
     "max": 976.458,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 16.582,
     "min": 16.582,
     "max": 16.582,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 64.712,
   "total_nodes": 83,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 43.02,
     "pct": 66.479,
     "nodes": 46,
     "calls": 6236
    },
    {
     "device": "GPU",
     "time_ms": 21.693,
     "pct": 33.521,
     "nodes": 37,
     "calls": 1488
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 21.691,
     "pct": 33.519,
     "by_device": {
      "GPU": {
       "time_ms": 21.691,
       "nodes": 36,
       "calls": 1486
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 6.517,
     "pct": 10.071,
     "by_device": {
      "CPU": {
       "time_ms": 6.517,
       "nodes": 1,
       "calls": 254
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 5.023,
     "pct": 7.762,
     "by_device": {
      "CPU": {
       "time_ms": 5.023,
       "nodes": 1,
       "calls": 256
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 3.228,
     "pct": 4.988,
     "by_device": {
      "CPU": {
       "time_ms": 3.228,
       "nodes": 1,
       "calls": 254
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 2.726,
     "pct": 4.212,
     "by_device": {
      "CPU": {
       "time_ms": 2.726,
       "nodes": 1,
       "calls": 332
      }
     }
    },
    {
     "op": "to",
     "total_ms": 2.493,
     "pct": 3.852,
     "by_device": {
      "CPU": {
       "time_ms": 2.493,
       "nodes": 1,
       "calls": 276
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.387,
     "pct": 3.688,
     "by_device": {
      "CPU": {
       "time_ms": 2.387,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 2.278,
     "pct": 3.52,
     "by_device": {
      "CPU": {
       "time_ms": 2.278,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 2.108,
     "pct": 3.258,
     "by_device": {
      "CPU": {
       "time_ms": 2.108,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "add",
     "total_ms": 1.808,
     "pct": 2.794,
     "by_device": {
      "CPU": {
       "time_ms": 1.808,
       "nodes": 1,
       "calls": 218
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.663,
     "pct": 2.569,
     "by_device": {
      "CPU": {
       "time_ms": 1.663,
       "nodes": 1,
       "calls": 598
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 1.577,
     "pct": 2.437,
     "by_device": {
      "CPU": {
       "time_ms": 1.577,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 1.423,
     "pct": 2.2,
     "by_device": {
      "CPU": {
       "time_ms": 1.423,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.181,
     "pct": 1.824,
     "by_device": {
      "CPU": {
       "time_ms": 1.181,
       "nodes": 1,
       "calls": 254
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 1.081,
     "pct": 1.671,
     "by_device": {
      "CPU": {
       "time_ms": 1.081,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.848,
     "pct": 1.31,
     "by_device": {
      "CPU": {
       "time_ms": 0.848,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 0.781,
     "pct": 1.206,
     "by_device": {
      "CPU": {
       "time_ms": 0.781,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.741,
     "pct": 1.145,
     "by_device": {
      "CPU": {
       "time_ms": 0.741,
       "nodes": 1,
       "calls": 368
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.657,
     "pct": 1.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.657,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.619,
     "pct": 0.957,
     "by_device": {
      "CPU": {
       "time_ms": 0.619,
       "nodes": 1,
       "calls": 230
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.576,
     "pct": 0.89,
     "by_device": {
      "CPU": {
       "time_ms": 0.576,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.502,
     "pct": 0.776,
     "by_device": {
      "CPU": {
       "time_ms": 0.502,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.425,
     "pct": 0.657,
     "by_device": {
      "CPU": {
       "time_ms": 0.425,
       "nodes": 1,
       "calls": 833
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.374,
     "pct": 0.577,
     "by_device": {
      "CPU": {
       "time_ms": 0.374,
       "nodes": 1,
       "calls": 78
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.359,
     "pct": 0.555,
     "by_device": {
      "CPU": {
       "time_ms": 0.359,
       "nodes": 1,
       "calls": 182
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 0.33,
     "pct": 0.51,
     "by_device": {
      "CPU": {
       "time_ms": 0.33,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.315,
     "pct": 0.487,
     "by_device": {
      "CPU": {
       "time_ms": 0.315,
       "nodes": 1,
       "calls": 476
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.175,
     "pct": 0.271,
     "by_device": {
      "CPU": {
       "time_ms": 0.175,
       "nodes": 1,
       "calls": 256
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.172,
     "pct": 0.266,
     "by_device": {
      "CPU": {
       "time_ms": 0.172,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.144,
     "pct": 0.222,
     "by_device": {
      "CPU": {
       "time_ms": 0.144,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.095,
     "pct": 0.146,
     "by_device": {
      "CPU": {
       "time_ms": 0.095,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.073,
     "pct": 0.112,
     "by_device": {
      "CPU": {
       "time_ms": 0.073,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.06,
     "pct": 0.092,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.038,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 2,
       "calls": 72
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.035,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.032,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.029,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.024,
     "pct": 0.038,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.024,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.022,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.022,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.021,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.02,
     "pct": 0.031,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.006,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.005,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.003,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.003,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_falcon3-1b-base_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_falcon3-1b-base_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:37:34",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "gemma-2-2b",
   "hf_model_id": "google/gemma-2-2b",
   "hf_model_url": "https://huggingface.co/google/gemma-2-2b",
   "param_count": 2614341888
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.058,
   "TTFT": 0.03,
   "TBT_avg": 0.027,
   "Throughput": 555.833
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 57.571,
     "ttft": 30.086,
     "tbt_avg": 27.485,
     "throughput": 555.833,
     "steps": [
      30.086,
      27.485
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 57.571,
     "min": 57.571,
     "max": 57.571,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 30.086,
     "min": 30.086,
     "max": 30.086,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 27.485,
     "min": 27.485,
     "max": 27.485,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 555.833,
     "min": 555.833,
     "max": 555.833,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 27.485,
     "min": 27.485,
     "max": 27.485,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 116.218,
   "total_nodes": 87,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 75.122,
     "pct": 64.639,
     "nodes": 49,
     "calls": 12044
    },
    {
     "device": "GPU",
     "time_ms": 41.096,
     "pct": 35.361,
     "nodes": 38,
     "calls": 3405
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 41.089,
     "pct": 35.355,
     "by_device": {
      "GPU": {
       "time_ms": 41.089,
       "nodes": 36,
       "calls": 3388
      }
     }
    },
    {
     "op": "to",
     "total_ms": 8.296,
     "pct": 7.138,
     "by_device": {
      "CPU": {
       "time_ms": 8.296,
       "nodes": 1,
       "calls": 929
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 7.969,
     "pct": 6.857,
     "by_device": {
      "CPU": {
       "time_ms": 7.969,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 7.491,
     "pct": 6.446,
     "by_device": {
      "CPU": {
       "time_ms": 7.491,
       "nodes": 1,
       "calls": 709
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 6.468,
     "pct": 5.565,
     "by_device": {
      "CPU": {
       "time_ms": 6.468,
       "nodes": 1,
       "calls": 368
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 5.043,
     "pct": 4.339,
     "by_device": {
      "CPU": {
       "time_ms": 5.043,
       "nodes": 1,
       "calls": 688
      }
     }
    },
    {
     "op": "add",
     "total_ms": 4.467,
     "pct": 3.844,
     "by_device": {
      "CPU": {
       "time_ms": 4.467,
       "nodes": 1,
       "calls": 628
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 4.389,
     "pct": 3.777,
     "by_device": {
      "CPU": {
       "time_ms": 4.389,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 3.961,
     "pct": 3.408,
     "by_device": {
      "CPU": {
       "time_ms": 3.961,
       "nodes": 1,
       "calls": 709
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.838,
     "pct": 2.442,
     "by_device": {
      "CPU": {
       "time_ms": 2.838,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "type_as",
     "total_ms": 2.601,
     "pct": 2.238,
     "by_device": {
      "CPU": {
       "time_ms": 2.601,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 2.473,
     "pct": 2.128,
     "by_device": {
      "CPU": {
       "time_ms": 2.473,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 2.106,
     "pct": 1.812,
     "by_device": {
      "CPU": {
       "time_ms": 2.106,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 1.89,
     "pct": 1.626,
     "by_device": {
      "CPU": {
       "time_ms": 1.89,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 1.873,
     "pct": 1.611,
     "by_device": {
      "CPU": {
       "time_ms": 1.873,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.859,
     "pct": 1.599,
     "by_device": {
      "CPU": {
       "time_ms": 1.859,
       "nodes": 1,
       "calls": 761
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 1.816,
     "pct": 1.563,
     "by_device": {
      "CPU": {
       "time_ms": 1.816,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.44,
     "pct": 1.239,
     "by_device": {
      "CPU": {
       "time_ms": 1.44,
       "nodes": 1,
       "calls": 862
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 1.405,
     "pct": 1.209,
     "by_device": {
      "CPU": {
       "time_ms": 1.405,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.998,
     "pct": 0.859,
     "by_device": {
      "CPU": {
       "time_ms": 0.998,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.831,
     "pct": 0.715,
     "by_device": {
      "CPU": {
       "time_ms": 0.831,
       "nodes": 1,
       "calls": 104
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.804,
     "pct": 0.691,
     "by_device": {
      "CPU": {
       "time_ms": 0.804,
       "nodes": 1,
       "calls": 528
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.65,
     "pct": 0.559,
     "by_device": {
      "CPU": {
       "time_ms": 0.65,
       "nodes": 1,
       "calls": 262
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.49,
     "pct": 0.422,
     "by_device": {
      "CPU": {
       "time_ms": 0.49,
       "nodes": 1,
       "calls": 262
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.466,
     "pct": 0.401,
     "by_device": {
      "CPU": {
       "time_ms": 0.466,
       "nodes": 1,
       "calls": 1247
      }
     }
    },
    {
     "op": "gelu",
     "total_ms": 0.466,
     "pct": 0.401,
     "by_device": {
      "CPU": {
       "time_ms": 0.466,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.462,
     "pct": 0.397,
     "by_device": {
      "CPU": {
       "time_ms": 0.462,
       "nodes": 1,
       "calls": 684
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.258,
     "pct": 0.222,
     "by_device": {
      "CPU": {
       "time_ms": 0.258,
       "nodes": 1,
       "calls": 110
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.227,
     "pct": 0.196,
     "by_device": {
      "CPU": {
       "time_ms": 0.227,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.214,
     "pct": 0.184,
     "by_device": {
      "CPU": {
       "time_ms": 0.214,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.207,
     "pct": 0.178,
     "by_device": {
      "CPU": {
       "time_ms": 0.207,
       "nodes": 1,
       "calls": 368
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.091,
     "pct": 0.079,
     "by_device": {
      "CPU": {
       "time_ms": 0.091,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.069,
     "pct": 0.06,
     "by_device": {
      "CPU": {
       "time_ms": 0.069,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.058,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 0.058,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.058,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 0.058,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.057,
     "pct": 0.049,
     "by_device": {
      "CPU": {
       "time_ms": 0.057,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.054,
     "pct": 0.046,
     "by_device": {
      "CPU": {
       "time_ms": 0.054,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.049,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.049,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.048,
     "pct": 0.041,
     "by_device": {
      "CPU": {
       "time_ms": 0.048,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.04,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.04,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "div",
     "total_ms": 0.025,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.021,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.021,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.019,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 0.019,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.017,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.008,
     "pct": 0.006,
     "by_device": {
      "GPU": {
       "time_ms": 0.008,
       "nodes": 2,
       "calls": 17
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.006,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.006,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.005,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 52
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_gemma-2-2b_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_gemma-2-2b_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:33:42",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "gemma-3-1b-it",
   "hf_model_id": "google/gemma-3-1b-it",
   "hf_model_url": "https://huggingface.co/google/gemma-3-1b-it",
   "param_count": 999885952
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.074,
   "TTFT": 0.039,
   "TBT_avg": 0.035,
   "Throughput": 432.615
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 73.969,
     "ttft": 38.549,
     "tbt_avg": 35.42,
     "throughput": 432.615,
     "steps": [
      38.549,
      35.42
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 73.969,
     "min": 73.969,
     "max": 73.969,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 38.549,
     "min": 38.549,
     "max": 38.549,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 35.42,
     "min": 35.42,
     "max": 35.42,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 432.615,
     "min": 432.615,
     "max": 432.615,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 35.42,
     "min": 35.42,
     "max": 35.42,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 120.408,
   "total_nodes": 85,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 98.445,
     "pct": 81.76,
     "nodes": 47,
     "calls": 14518
    },
    {
     "device": "GPU",
     "time_ms": 21.963,
     "pct": 18.24,
     "nodes": 38,
     "calls": 4628
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 21.953,
     "pct": 18.232,
     "by_device": {
      "GPU": {
       "time_ms": 21.953,
       "nodes": 36,
       "calls": 4602
      }
     }
    },
    {
     "op": "to",
     "total_ms": 12.595,
     "pct": 10.461,
     "by_device": {
      "CPU": {
       "time_ms": 12.595,
       "nodes": 1,
       "calls": 1368
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 11.49,
     "pct": 9.542,
     "by_device": {
      "CPU": {
       "time_ms": 11.49,
       "nodes": 1,
       "calls": 1036
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 9.011,
     "pct": 7.484,
     "by_device": {
      "CPU": {
       "time_ms": 9.011,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 7.605,
     "pct": 6.316,
     "by_device": {
      "CPU": {
       "time_ms": 7.605,
       "nodes": 1,
       "calls": 370
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 6.547,
     "pct": 5.437,
     "by_device": {
      "CPU": {
       "time_ms": 6.547,
       "nodes": 1,
       "calls": 898
      }
     }
    },
    {
     "op": "add",
     "total_ms": 6.007,
     "pct": 4.989,
     "by_device": {
      "CPU": {
       "time_ms": 6.007,
       "nodes": 1,
       "calls": 836
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 5.899,
     "pct": 4.899,
     "by_device": {
      "CPU": {
       "time_ms": 5.899,
       "nodes": 1,
       "calls": 1036
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 5.334,
     "pct": 4.43,
     "by_device": {
      "CPU": {
       "time_ms": 5.334,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "type_as",
     "total_ms": 4.016,
     "pct": 3.335,
     "by_device": {
      "CPU": {
       "time_ms": 4.016,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 3.224,
     "pct": 2.677,
     "by_device": {
      "CPU": {
       "time_ms": 3.224,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 3.171,
     "pct": 2.633,
     "by_device": {
      "CPU": {
       "time_ms": 3.171,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 2.956,
     "pct": 2.455,
     "by_device": {
      "CPU": {
       "time_ms": 2.956,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 2.793,
     "pct": 2.32,
     "by_device": {
      "CPU": {
       "time_ms": 2.793,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 2.757,
     "pct": 2.29,
     "by_device": {
      "CPU": {
       "time_ms": 2.757,
       "nodes": 1,
       "calls": 1088
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 2.15,
     "pct": 1.785,
     "by_device": {
      "CPU": {
       "time_ms": 2.15,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 2.032,
     "pct": 1.688,
     "by_device": {
      "CPU": {
       "time_ms": 2.032,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 1.872,
     "pct": 1.555,
     "by_device": {
      "CPU": {
       "time_ms": 1.872,
       "nodes": 1,
       "calls": 212
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.555,
     "pct": 1.291,
     "by_device": {
      "CPU": {
       "time_ms": 1.555,
       "nodes": 1,
       "calls": 864
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.037,
     "pct": 0.861,
     "by_device": {
      "CPU": {
       "time_ms": 1.037,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.937,
     "pct": 0.778,
     "by_device": {
      "CPU": {
       "time_ms": 0.937,
       "nodes": 1,
       "calls": 532
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.853,
     "pct": 0.709,
     "by_device": {
      "CPU": {
       "time_ms": 0.853,
       "nodes": 1,
       "calls": 104
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.799,
     "pct": 0.664,
     "by_device": {
      "CPU": {
       "time_ms": 0.799,
       "nodes": 1,
       "calls": 298
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.541,
     "pct": 0.449,
     "by_device": {
      "CPU": {
       "time_ms": 0.541,
       "nodes": 1,
       "calls": 1297
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.532,
     "pct": 0.442,
     "by_device": {
      "CPU": {
       "time_ms": 0.532,
       "nodes": 1,
       "calls": 686
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.513,
     "pct": 0.426,
     "by_device": {
      "CPU": {
       "time_ms": 0.513,
       "nodes": 1,
       "calls": 262
      }
     }
    },
    {
     "op": "gelu",
     "total_ms": 0.513,
     "pct": 0.426,
     "by_device": {
      "CPU": {
       "time_ms": 0.513,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.317,
     "pct": 0.263,
     "by_device": {
      "CPU": {
       "time_ms": 0.317,
       "nodes": 1,
       "calls": 116
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.231,
     "pct": 0.192,
     "by_device": {
      "CPU": {
       "time_ms": 0.231,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.224,
     "pct": 0.186,
     "by_device": {
      "CPU": {
       "time_ms": 0.224,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.213,
     "pct": 0.177,
     "by_device": {
      "CPU": {
       "time_ms": 0.213,
       "nodes": 1,
       "calls": 370
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.091,
     "pct": 0.076,
     "by_device": {
      "CPU": {
       "time_ms": 0.091,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.09,
     "pct": 0.075,
     "by_device": {
      "CPU": {
       "time_ms": 0.09,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.068,
     "pct": 0.057,
     "by_device": {
      "CPU": {
       "time_ms": 0.068,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.065,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.065,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.064,
     "pct": 0.053,
     "by_device": {
      "CPU": {
       "time_ms": 0.064,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.06,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.055,
     "pct": 0.046,
     "by_device": {
      "CPU": {
       "time_ms": 0.055,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.05,
     "pct": 0.041,
     "by_device": {
      "CPU": {
       "time_ms": 0.05,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.041,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.041,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.036,
     "pct": 0.03,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.032,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 1,
       "calls": 13
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.032,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.021,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.01,
     "pct": 0.009,
     "by_device": {
      "GPU": {
       "time_ms": 0.01,
       "nodes": 2,
       "calls": 26
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.008,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 5
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.006,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.005,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 52
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_gemma-3-1b-it_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_gemma-3-1b-it_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:22:05",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "gpt2-large",
   "hf_model_id": "openai-community/gpt2-large",
   "hf_model_url": "https://huggingface.co/openai-community/gpt2-large",
   "param_count": 774030080
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.04,
   "TTFT": 0.022,
   "TBT_avg": 0.018,
   "Throughput": 798.541
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 40.073,
     "ttft": 21.984,
     "tbt_avg": 18.089,
     "throughput": 798.541,
     "steps": [
      21.984,
      18.089
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 40.073,
     "min": 40.073,
     "max": 40.073,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 21.984,
     "min": 21.984,
     "max": 21.984,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 18.089,
     "min": 18.089,
     "max": 18.089,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 798.541,
     "min": 798.541,
     "max": 798.541,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 18.089,
     "min": 18.089,
     "max": 18.089,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 53.964,
   "total_nodes": 69,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 36.343,
     "pct": 67.348,
     "nodes": 45,
     "calls": 6598
    },
    {
     "device": "GPU",
     "time_ms": 17.62,
     "pct": 32.652,
     "nodes": 24,
     "calls": 1388
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 17.619,
     "pct": 32.649,
     "by_device": {
      "GPU": {
       "time_ms": 17.619,
       "nodes": 23,
       "calls": 1386
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 4.664,
     "pct": 8.643,
     "by_device": {
      "CPU": {
       "time_ms": 4.664,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 3.457,
     "pct": 6.406,
     "by_device": {
      "CPU": {
       "time_ms": 3.457,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 3.198,
     "pct": 5.925,
     "by_device": {
      "CPU": {
       "time_ms": 3.198,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 2.998,
     "pct": 5.556,
     "by_device": {
      "CPU": {
       "time_ms": 2.998,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 2.901,
     "pct": 5.375,
     "by_device": {
      "CPU": {
       "time_ms": 2.901,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.461,
     "pct": 4.561,
     "by_device": {
      "CPU": {
       "time_ms": 2.461,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 2.421,
     "pct": 4.486,
     "by_device": {
      "CPU": {
       "time_ms": 2.421,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 2.018,
     "pct": 3.739,
     "by_device": {
      "CPU": {
       "time_ms": 2.018,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 1.685,
     "pct": 3.122,
     "by_device": {
      "CPU": {
       "time_ms": 1.685,
       "nodes": 1,
       "calls": 802
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 1.443,
     "pct": 2.675,
     "by_device": {
      "CPU": {
       "time_ms": 1.443,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.119,
     "pct": 2.073,
     "by_device": {
      "CPU": {
       "time_ms": 1.119,
       "nodes": 1,
       "calls": 578
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.928,
     "pct": 1.72,
     "by_device": {
      "CPU": {
       "time_ms": 0.928,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.91,
     "pct": 1.687,
     "by_device": {
      "CPU": {
       "time_ms": 0.91,
       "nodes": 1,
       "calls": 1173
      }
     }
    },
    {
     "op": "split",
     "total_ms": 0.909,
     "pct": 1.685,
     "by_device": {
      "CPU": {
       "time_ms": 0.909,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.681,
     "pct": 1.262,
     "by_device": {
      "CPU": {
       "time_ms": 0.681,
       "nodes": 1,
       "calls": 216
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.651,
     "pct": 1.206,
     "by_device": {
      "CPU": {
       "time_ms": 0.651,
       "nodes": 1,
       "calls": 150
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 0.58,
     "pct": 1.074,
     "by_device": {
      "CPU": {
       "time_ms": 0.58,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.511,
     "pct": 0.947,
     "by_device": {
      "CPU": {
       "time_ms": 0.511,
       "nodes": 1,
       "calls": 76
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.47,
     "pct": 0.87,
     "by_device": {
      "CPU": {
       "time_ms": 0.47,
       "nodes": 1,
       "calls": 218
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.394,
     "pct": 0.729,
     "by_device": {
      "CPU": {
       "time_ms": 0.394,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.378,
     "pct": 0.701,
     "by_device": {
      "CPU": {
       "time_ms": 0.378,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.304,
     "pct": 0.564,
     "by_device": {
      "CPU": {
       "time_ms": 0.304,
       "nodes": 1,
       "calls": 804
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.174,
     "pct": 0.322,
     "by_device": {
      "CPU": {
       "time_ms": 0.174,
       "nodes": 1,
       "calls": 78
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.141,
     "pct": 0.261,
     "by_device": {
      "CPU": {
       "time_ms": 0.141,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.135,
     "pct": 0.25,
     "by_device": {
      "CPU": {
       "time_ms": 0.135,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.109,
     "pct": 0.202,
     "by_device": {
      "CPU": {
       "time_ms": 0.109,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 0.106,
     "pct": 0.197,
     "by_device": {
      "CPU": {
       "time_ms": 0.106,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 0.103,
     "pct": 0.191,
     "by_device": {
      "CPU": {
       "time_ms": 0.103,
       "nodes": 1,
       "calls": 76
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.089,
     "pct": 0.165,
     "by_device": {
      "CPU": {
       "time_ms": 0.089,
       "nodes": 2,
       "calls": 144
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.088,
     "pct": 0.164,
     "by_device": {
      "CPU": {
       "time_ms": 0.088,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 0.05,
     "pct": 0.094,
     "by_device": {
      "CPU": {
       "time_ms": 0.05,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.043,
     "pct": 0.079,
     "by_device": {
      "CPU": {
       "time_ms": 0.043,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.04,
     "pct": 0.074,
     "by_device": {
      "CPU": {
       "time_ms": 0.04,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.037,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.036,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.034,
     "pct": 0.063,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.024,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.016,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.009,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.009,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.008,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.006,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.004,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.003,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.003,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_gpt2-large_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_gpt2-large_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:33:01",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "Llama-3.2-1B",
   "hf_model_id": "meta-llama/Llama-3.2-1B",
   "hf_model_url": "https://huggingface.co/meta-llama/Llama-3.2-1B",
   "param_count": 1235814400
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.029,
   "TTFT": 0.015,
   "TBT_avg": 0.014,
   "Throughput": 1095.287
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 29.216,
     "ttft": 15.446,
     "tbt_avg": 13.77,
     "throughput": 1095.287,
     "steps": [
      15.446,
      13.77
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 29.216,
     "min": 29.216,
     "max": 29.216,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 15.446,
     "min": 15.446,
     "max": 15.446,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 13.77,
     "min": 13.77,
     "max": 13.77,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 1095.287,
     "min": 1095.287,
     "max": 1095.287,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 13.77,
     "min": 13.77,
     "max": 13.77,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 55.778,
   "total_nodes": 85,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 37.458,
     "pct": 67.156,
     "nodes": 46,
     "calls": 5564
    },
    {
     "device": "GPU",
     "time_ms": 18.32,
     "pct": 32.844,
     "nodes": 39,
     "calls": 1388
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 18.296,
     "pct": 32.801,
     "by_device": {
      "GPU": {
       "time_ms": 18.296,
       "nodes": 37,
       "calls": 1354
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 5.798,
     "pct": 10.394,
     "by_device": {
      "CPU": {
       "time_ms": 5.798,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 4.678,
     "pct": 8.387,
     "by_device": {
      "CPU": {
       "time_ms": 4.678,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 3.101,
     "pct": 5.56,
     "by_device": {
      "CPU": {
       "time_ms": 3.101,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 2.335,
     "pct": 4.187,
     "by_device": {
      "CPU": {
       "time_ms": 2.335,
       "nodes": 1,
       "calls": 296
      }
     }
    },
    {
     "op": "to",
     "total_ms": 2.222,
     "pct": 3.984,
     "by_device": {
      "CPU": {
       "time_ms": 2.222,
       "nodes": 1,
       "calls": 248
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 2.008,
     "pct": 3.601,
     "by_device": {
      "CPU": {
       "time_ms": 2.008,
       "nodes": 1,
       "calls": 174
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 1.942,
     "pct": 3.481,
     "by_device": {
      "CPU": {
       "time_ms": 1.942,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 1.705,
     "pct": 3.056,
     "by_device": {
      "CPU": {
       "time_ms": 1.705,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "add",
     "total_ms": 1.542,
     "pct": 2.765,
     "by_device": {
      "CPU": {
       "time_ms": 1.542,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.34,
     "pct": 2.402,
     "by_device": {
      "CPU": {
       "time_ms": 1.34,
       "nodes": 1,
       "calls": 532
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 1.244,
     "pct": 2.231,
     "by_device": {
      "CPU": {
       "time_ms": 1.244,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 1.191,
     "pct": 2.135,
     "by_device": {
      "CPU": {
       "time_ms": 1.191,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 0.954,
     "pct": 1.71,
     "by_device": {
      "CPU": {
       "time_ms": 0.954,
       "nodes": 1,
       "calls": 174
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.848,
     "pct": 1.521,
     "by_device": {
      "CPU": {
       "time_ms": 0.848,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.782,
     "pct": 1.403,
     "by_device": {
      "CPU": {
       "time_ms": 0.782,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 0.685,
     "pct": 1.228,
     "by_device": {
      "CPU": {
       "time_ms": 0.685,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.59,
     "pct": 1.058,
     "by_device": {
      "CPU": {
       "time_ms": 0.59,
       "nodes": 1,
       "calls": 328
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.558,
     "pct": 1.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.558,
       "nodes": 1,
       "calls": 206
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.549,
     "pct": 0.984,
     "by_device": {
      "CPU": {
       "time_ms": 0.549,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.486,
     "pct": 0.871,
     "by_device": {
      "CPU": {
       "time_ms": 0.486,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.438,
     "pct": 0.785,
     "by_device": {
      "CPU": {
       "time_ms": 0.438,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.369,
     "pct": 0.661,
     "by_device": {
      "CPU": {
       "time_ms": 0.369,
       "nodes": 1,
       "calls": 743
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.335,
     "pct": 0.601,
     "by_device": {
      "CPU": {
       "time_ms": 0.335,
       "nodes": 1,
       "calls": 162
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 0.312,
     "pct": 0.56,
     "by_device": {
      "CPU": {
       "time_ms": 0.312,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.304,
     "pct": 0.544,
     "by_device": {
      "CPU": {
       "time_ms": 0.304,
       "nodes": 1,
       "calls": 424
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.2,
     "pct": 0.359,
     "by_device": {
      "CPU": {
       "time_ms": 0.2,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.154,
     "pct": 0.277,
     "by_device": {
      "CPU": {
       "time_ms": 0.154,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.146,
     "pct": 0.262,
     "by_device": {
      "CPU": {
       "time_ms": 0.146,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.136,
     "pct": 0.244,
     "by_device": {
      "CPU": {
       "time_ms": 0.136,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.095,
     "pct": 0.17,
     "by_device": {
      "CPU": {
       "time_ms": 0.095,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.069,
     "pct": 0.124,
     "by_device": {
      "CPU": {
       "time_ms": 0.069,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.06,
     "pct": 0.107,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.038,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 2,
       "calls": 64
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.036,
     "pct": 0.065,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.033,
     "pct": 0.06,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.03,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.025,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.024,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "gpu_memset",
     "total_ms": 0.023,
     "pct": 0.04,
     "by_device": {
      "GPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.022,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.021,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.019,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.018,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.005,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.005,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.004,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.003,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_llama-3.2-1b_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_llama-3.2-1b_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:31:57",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "OLMo-2-0425-1B",
   "hf_model_id": "allenai/OLMo-2-0425-1B",
   "hf_model_url": "https://huggingface.co/allenai/OLMo-2-0425-1B",
   "param_count": 1480000000
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.034,
   "TTFT": 0.018,
   "TBT_avg": 0.016,
   "Throughput": 942.69
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 33.945,
     "ttft": 17.929,
     "tbt_avg": 16.016,
     "throughput": 942.69,
     "steps": [
      17.929,
      16.016
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 33.945,
     "min": 33.945,
     "max": 33.945,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 17.929,
     "min": 17.929,
     "max": 17.929,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 16.016,
     "min": 16.016,
     "max": 16.016,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 942.69,
     "min": 942.69,
     "max": 942.69,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 16.016,
     "min": 16.016,
     "max": 16.016,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 63.498,
   "total_nodes": 82,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 41.915,
     "pct": 66.01,
     "nodes": 46,
     "calls": 6604
    },
    {
     "device": "GPU",
     "time_ms": 21.583,
     "pct": 33.99,
     "nodes": 36,
     "calls": 1900
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 21.581,
     "pct": 33.987,
     "by_device": {
      "GPU": {
       "time_ms": 21.581,
       "nodes": 35,
       "calls": 1898
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 5.18,
     "pct": 8.157,
     "by_device": {
      "CPU": {
       "time_ms": 5.18,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "to",
     "total_ms": 4.257,
     "pct": 6.704,
     "by_device": {
      "CPU": {
       "time_ms": 4.257,
       "nodes": 1,
       "calls": 500
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 4.16,
     "pct": 6.552,
     "by_device": {
      "CPU": {
       "time_ms": 4.16,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 3.861,
     "pct": 6.08,
     "by_device": {
      "CPU": {
       "time_ms": 3.861,
       "nodes": 1,
       "calls": 362
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 3.137,
     "pct": 4.94,
     "by_device": {
      "CPU": {
       "time_ms": 3.137,
       "nodes": 1,
       "calls": 424
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 2.744,
     "pct": 4.321,
     "by_device": {
      "CPU": {
       "time_ms": 2.744,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 1.928,
     "pct": 3.036,
     "by_device": {
      "CPU": {
       "time_ms": 1.928,
       "nodes": 1,
       "calls": 362
      }
     }
    },
    {
     "op": "add",
     "total_ms": 1.907,
     "pct": 3.003,
     "by_device": {
      "CPU": {
       "time_ms": 1.907,
       "nodes": 1,
       "calls": 258
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 1.621,
     "pct": 2.552,
     "by_device": {
      "CPU": {
       "time_ms": 1.621,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 1.395,
     "pct": 2.196,
     "by_device": {
      "CPU": {
       "time_ms": 1.395,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 1.297,
     "pct": 2.042,
     "by_device": {
      "CPU": {
       "time_ms": 1.297,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 1.172,
     "pct": 1.846,
     "by_device": {
      "CPU": {
       "time_ms": 1.172,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 1.145,
     "pct": 1.804,
     "by_device": {
      "CPU": {
       "time_ms": 1.145,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.03,
     "pct": 1.622,
     "by_device": {
      "CPU": {
       "time_ms": 1.03,
       "nodes": 1,
       "calls": 484
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.974,
     "pct": 1.535,
     "by_device": {
      "CPU": {
       "time_ms": 0.974,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 0.926,
     "pct": 1.458,
     "by_device": {
      "CPU": {
       "time_ms": 0.926,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.873,
     "pct": 1.375,
     "by_device": {
      "CPU": {
       "time_ms": 0.873,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.763,
     "pct": 1.201,
     "by_device": {
      "CPU": {
       "time_ms": 0.763,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.517,
     "pct": 0.814,
     "by_device": {
      "CPU": {
       "time_ms": 0.517,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.459,
     "pct": 0.722,
     "by_device": {
      "CPU": {
       "time_ms": 0.459,
       "nodes": 1,
       "calls": 264
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.335,
     "pct": 0.528,
     "by_device": {
      "CPU": {
       "time_ms": 0.335,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.297,
     "pct": 0.468,
     "by_device": {
      "CPU": {
       "time_ms": 0.297,
       "nodes": 1,
       "calls": 162
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.294,
     "pct": 0.464,
     "by_device": {
      "CPU": {
       "time_ms": 0.294,
       "nodes": 1,
       "calls": 695
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 0.287,
     "pct": 0.453,
     "by_device": {
      "CPU": {
       "time_ms": 0.287,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.274,
     "pct": 0.432,
     "by_device": {
      "CPU": {
       "time_ms": 0.274,
       "nodes": 1,
       "calls": 360
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.178,
     "pct": 0.28,
     "by_device": {
      "CPU": {
       "time_ms": 0.178,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.135,
     "pct": 0.213,
     "by_device": {
      "CPU": {
       "time_ms": 0.135,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.13,
     "pct": 0.204,
     "by_device": {
      "CPU": {
       "time_ms": 0.13,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.129,
     "pct": 0.203,
     "by_device": {
      "CPU": {
       "time_ms": 0.129,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.09,
     "pct": 0.141,
     "by_device": {
      "CPU": {
       "time_ms": 0.09,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.068,
     "pct": 0.107,
     "by_device": {
      "CPU": {
       "time_ms": 0.068,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.059,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 0.059,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.041,
     "pct": 0.064,
     "by_device": {
      "CPU": {
       "time_ms": 0.041,
       "nodes": 2,
       "calls": 64
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.035,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.033,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.03,
     "pct": 0.047,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.029,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.027,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.025,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.023,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.022,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.018,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.006,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.005,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.003,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.003,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_olmo-2-0425-1b_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_olmo-2-0425-1b_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:27:01",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "opt-1.3b",
   "hf_model_id": "facebook/opt-1.3b",
   "hf_model_url": "https://huggingface.co/facebook/opt-1.3b",
   "param_count": 1315758080
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.023,
   "TTFT": 0.013,
   "TBT_avg": 0.01,
   "Throughput": 1381.188
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 23.168,
     "ttft": 13.002,
     "tbt_avg": 10.167,
     "throughput": 1381.188,
     "steps": [
      13.002,
      10.167
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 23.168,
     "min": 23.168,
     "max": 23.168,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 13.002,
     "min": 13.002,
     "max": 13.002,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 10.167,
     "min": 10.167,
     "max": 10.167,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 1381.188,
     "min": 1381.188,
     "max": 1381.188,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 10.167,
     "min": 10.167,
     "max": 10.167,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 48.831,
   "total_nodes": 63,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 28.051,
     "pct": 57.445,
     "nodes": 42,
     "calls": 5010
    },
    {
     "device": "GPU",
     "time_ms": 20.78,
     "pct": 42.555,
     "nodes": 21,
     "calls": 742
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 20.778,
     "pct": 42.552,
     "by_device": {
      "GPU": {
       "time_ms": 20.778,
       "nodes": 20,
       "calls": 740
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 6.018,
     "pct": 12.324,
     "by_device": {
      "CPU": {
       "time_ms": 6.018,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 4.023,
     "pct": 8.239,
     "by_device": {
      "CPU": {
       "time_ms": 4.023,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 2.072,
     "pct": 4.242,
     "by_device": {
      "CPU": {
       "time_ms": 2.072,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.007,
     "pct": 4.11,
     "by_device": {
      "CPU": {
       "time_ms": 2.007,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 1.886,
     "pct": 3.862,
     "by_device": {
      "CPU": {
       "time_ms": 1.886,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 1.749,
     "pct": 3.582,
     "by_device": {
      "CPU": {
       "time_ms": 1.749,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 1.22,
     "pct": 2.499,
     "by_device": {
      "CPU": {
       "time_ms": 1.22,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.189,
     "pct": 2.435,
     "by_device": {
      "CPU": {
       "time_ms": 1.189,
       "nodes": 1,
       "calls": 674
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 1.092,
     "pct": 2.235,
     "by_device": {
      "CPU": {
       "time_ms": 1.092,
       "nodes": 1,
       "calls": 538
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 0.845,
     "pct": 1.731,
     "by_device": {
      "CPU": {
       "time_ms": 0.845,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "add",
     "total_ms": 0.831,
     "pct": 1.702,
     "by_device": {
      "CPU": {
       "time_ms": 0.831,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.69,
     "pct": 1.413,
     "by_device": {
      "CPU": {
       "time_ms": 0.69,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.586,
     "pct": 1.2,
     "by_device": {
      "CPU": {
       "time_ms": 0.586,
       "nodes": 1,
       "calls": 881
      }
     }
    },
    {
     "op": "relu",
     "total_ms": 0.485,
     "pct": 0.992,
     "by_device": {
      "CPU": {
       "time_ms": 0.485,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 0.461,
     "pct": 0.943,
     "by_device": {
      "CPU": {
       "time_ms": 0.461,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.414,
     "pct": 0.848,
     "by_device": {
      "CPU": {
       "time_ms": 0.414,
       "nodes": 1,
       "calls": 54
      }
     }
    },
    {
     "op": "clamp_min",
     "total_ms": 0.385,
     "pct": 0.789,
     "by_device": {
      "CPU": {
       "time_ms": 0.385,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.37,
     "pct": 0.757,
     "by_device": {
      "CPU": {
       "time_ms": 0.37,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.248,
     "pct": 0.508,
     "by_device": {
      "CPU": {
       "time_ms": 0.248,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.235,
     "pct": 0.482,
     "by_device": {
      "CPU": {
       "time_ms": 0.235,
       "nodes": 1,
       "calls": 684
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.213,
     "pct": 0.436,
     "by_device": {
      "CPU": {
       "time_ms": 0.213,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.178,
     "pct": 0.365,
     "by_device": {
      "CPU": {
       "time_ms": 0.178,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.151,
     "pct": 0.309,
     "by_device": {
      "CPU": {
       "time_ms": 0.151,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.133,
     "pct": 0.272,
     "by_device": {
      "CPU": {
       "time_ms": 0.133,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.112,
     "pct": 0.229,
     "by_device": {
      "CPU": {
       "time_ms": 0.112,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 0.079,
     "pct": 0.162,
     "by_device": {
      "CPU": {
       "time_ms": 0.079,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.078,
     "pct": 0.16,
     "by_device": {
      "CPU": {
       "time_ms": 0.078,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.046,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 0.046,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 0.042,
     "pct": 0.087,
     "by_device": {
      "CPU": {
       "time_ms": 0.042,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.035,
     "pct": 0.071,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.034,
     "pct": 0.07,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.032,
     "pct": 0.066,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.029,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.023,
     "pct": 0.047,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.023,
     "pct": 0.047,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.01,
     "pct": 0.02,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.01,
     "pct": 0.02,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.006,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.004,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.003,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.003,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.003,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_opt-1.3b_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_opt-1.3b_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:22:52",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "opt-350m",
   "hf_model_id": "facebook/opt-350m",
   "hf_model_url": "https://huggingface.co/facebook/opt-350m",
   "param_count": 331196416
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.025,
   "TTFT": 0.014,
   "TBT_avg": 0.011,
   "Throughput": 1303.269
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 24.554,
     "ttft": 13.62,
     "tbt_avg": 10.934,
     "throughput": 1303.269,
     "steps": [
      13.62,
      10.934
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 24.554,
     "min": 24.554,
     "max": 24.554,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 13.62,
     "min": 13.62,
     "max": 13.62,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 10.934,
     "min": 10.934,
     "max": 10.934,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 1303.269,
     "min": 1303.269,
     "max": 1303.269,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 10.934,
     "min": 10.934,
     "max": 10.934,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 37.174,
   "total_nodes": 66,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 30.001,
     "pct": 80.705,
     "nodes": 42,
     "calls": 5032
    },
    {
     "device": "GPU",
     "time_ms": 7.173,
     "pct": 19.295,
     "nodes": 24,
     "calls": 768
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 7.171,
     "pct": 19.29,
     "by_device": {
      "GPU": {
       "time_ms": 7.171,
       "nodes": 23,
       "calls": 766
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 6.542,
     "pct": 17.599,
     "by_device": {
      "CPU": {
       "time_ms": 6.542,
       "nodes": 1,
       "calls": 294
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 4.136,
     "pct": 11.125,
     "by_device": {
      "CPU": {
       "time_ms": 4.136,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.117,
     "pct": 5.694,
     "by_device": {
      "CPU": {
       "time_ms": 2.117,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 2.052,
     "pct": 5.52,
     "by_device": {
      "CPU": {
       "time_ms": 2.052,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 1.84,
     "pct": 4.951,
     "by_device": {
      "CPU": {
       "time_ms": 1.84,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 1.815,
     "pct": 4.882,
     "by_device": {
      "CPU": {
       "time_ms": 1.815,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 1.314,
     "pct": 3.535,
     "by_device": {
      "CPU": {
       "time_ms": 1.314,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.175,
     "pct": 3.162,
     "by_device": {
      "CPU": {
       "time_ms": 1.175,
       "nodes": 1,
       "calls": 678
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 1.085,
     "pct": 2.92,
     "by_device": {
      "CPU": {
       "time_ms": 1.085,
       "nodes": 1,
       "calls": 532
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 0.943,
     "pct": 2.535,
     "by_device": {
      "CPU": {
       "time_ms": 0.943,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "add",
     "total_ms": 0.915,
     "pct": 2.462,
     "by_device": {
      "CPU": {
       "time_ms": 0.915,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.903,
     "pct": 2.43,
     "by_device": {
      "CPU": {
       "time_ms": 0.903,
       "nodes": 1,
       "calls": 294
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.582,
     "pct": 1.566,
     "by_device": {
      "CPU": {
       "time_ms": 0.582,
       "nodes": 1,
       "calls": 881
      }
     }
    },
    {
     "op": "relu",
     "total_ms": 0.536,
     "pct": 1.443,
     "by_device": {
      "CPU": {
       "time_ms": 0.536,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 0.489,
     "pct": 1.315,
     "by_device": {
      "CPU": {
       "time_ms": 0.489,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "clamp_min",
     "total_ms": 0.429,
     "pct": 1.155,
     "by_device": {
      "CPU": {
       "time_ms": 0.429,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.411,
     "pct": 1.107,
     "by_device": {
      "CPU": {
       "time_ms": 0.411,
       "nodes": 1,
       "calls": 54
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.37,
     "pct": 0.996,
     "by_device": {
      "CPU": {
       "time_ms": 0.37,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.311,
     "pct": 0.836,
     "by_device": {
      "CPU": {
       "time_ms": 0.311,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 0.308,
     "pct": 0.828,
     "by_device": {
      "CPU": {
       "time_ms": 0.308,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.258,
     "pct": 0.695,
     "by_device": {
      "CPU": {
       "time_ms": 0.258,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.245,
     "pct": 0.659,
     "by_device": {
      "CPU": {
       "time_ms": 0.245,
       "nodes": 1,
       "calls": 688
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.223,
     "pct": 0.601,
     "by_device": {
      "CPU": {
       "time_ms": 0.223,
       "nodes": 1,
       "calls": 106
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.194,
     "pct": 0.523,
     "by_device": {
      "CPU": {
       "time_ms": 0.194,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.164,
     "pct": 0.441,
     "by_device": {
      "CPU": {
       "time_ms": 0.164,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.15,
     "pct": 0.405,
     "by_device": {
      "CPU": {
       "time_ms": 0.15,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.126,
     "pct": 0.339,
     "by_device": {
      "CPU": {
       "time_ms": 0.126,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 0.084,
     "pct": 0.226,
     "by_device": {
      "CPU": {
       "time_ms": 0.084,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.047,
     "pct": 0.126,
     "by_device": {
      "CPU": {
       "time_ms": 0.047,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.039,
     "pct": 0.106,
     "by_device": {
      "CPU": {
       "time_ms": 0.039,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.036,
     "pct": 0.098,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.034,
     "pct": 0.091,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.031,
     "pct": 0.082,
     "by_device": {
      "CPU": {
       "time_ms": 0.031,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.026,
     "pct": 0.07,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.026,
     "pct": 0.07,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.011,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.01,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.008,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.006,
     "pct": 0.017,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.004,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.003,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.005,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_opt-350m_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_opt-350m_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:22:12",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "phi-1_5",
   "hf_model_id": "microsoft/phi-1_5",
   "hf_model_url": "https://huggingface.co/microsoft/phi-1_5",
   "param_count": 1418270720
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.037,
   "TTFT": 0.02,
   "TBT_avg": 0.017,
   "Throughput": 874.66
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 36.586,
     "ttft": 19.974,
     "tbt_avg": 16.611,
     "throughput": 874.66,
     "steps": [
      19.974,
      16.611
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 36.586,
     "min": 36.586,
     "max": 36.586,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 19.974,
     "min": 19.974,
     "max": 19.974,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 16.611,
     "min": 16.611,
     "max": 16.611,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 874.66,
     "min": 874.66,
     "max": 874.66,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 16.611,
     "min": 16.611,
     "max": 16.611,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 61.147,
   "total_nodes": 87,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 37.723,
     "pct": 61.692,
     "nodes": 51,
     "calls": 6771
    },
    {
     "device": "GPU",
     "time_ms": 23.424,
     "pct": 38.308,
     "nodes": 36,
     "calls": 1594
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 23.423,
     "pct": 38.305,
     "by_device": {
      "GPU": {
       "time_ms": 23.423,
       "nodes": 35,
       "calls": 1592
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 6.79,
     "pct": 11.104,
     "by_device": {
      "CPU": {
       "time_ms": 6.79,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 4.382,
     "pct": 7.167,
     "by_device": {
      "CPU": {
       "time_ms": 4.382,
       "nodes": 1,
       "calls": 289
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 3.254,
     "pct": 5.322,
     "by_device": {
      "CPU": {
       "time_ms": 3.254,
       "nodes": 1,
       "calls": 388
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 2.541,
     "pct": 4.155,
     "by_device": {
      "CPU": {
       "time_ms": 2.541,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.382,
     "pct": 3.896,
     "by_device": {
      "CPU": {
       "time_ms": 2.382,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.256,
     "pct": 3.69,
     "by_device": {
      "CPU": {
       "time_ms": 2.256,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 2.072,
     "pct": 3.388,
     "by_device": {
      "CPU": {
       "time_ms": 2.072,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 1.364,
     "pct": 2.231,
     "by_device": {
      "CPU": {
       "time_ms": 1.364,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.247,
     "pct": 2.04,
     "by_device": {
      "CPU": {
       "time_ms": 1.247,
       "nodes": 1,
       "calls": 676
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 1.188,
     "pct": 1.943,
     "by_device": {
      "CPU": {
       "time_ms": 1.188,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 1.069,
     "pct": 1.749,
     "by_device": {
      "CPU": {
       "time_ms": 1.069,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.866,
     "pct": 1.417,
     "by_device": {
      "CPU": {
       "time_ms": 0.866,
       "nodes": 1,
       "calls": 416
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.841,
     "pct": 1.376,
     "by_device": {
      "CPU": {
       "time_ms": 0.841,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.828,
     "pct": 1.354,
     "by_device": {
      "CPU": {
       "time_ms": 0.828,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.821,
     "pct": 1.343,
     "by_device": {
      "CPU": {
       "time_ms": 0.821,
       "nodes": 1,
       "calls": 386
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.623,
     "pct": 1.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.623,
       "nodes": 1,
       "calls": 877
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.59,
     "pct": 0.965,
     "by_device": {
      "CPU": {
       "time_ms": 0.59,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.475,
     "pct": 0.777,
     "by_device": {
      "CPU": {
       "time_ms": 0.475,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.458,
     "pct": 0.748,
     "by_device": {
      "CPU": {
       "time_ms": 0.458,
       "nodes": 1,
       "calls": 1175
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.407,
     "pct": 0.666,
     "by_device": {
      "CPU": {
       "time_ms": 0.407,
       "nodes": 1,
       "calls": 58
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.383,
     "pct": 0.626,
     "by_device": {
      "CPU": {
       "time_ms": 0.383,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 0.378,
     "pct": 0.619,
     "by_device": {
      "CPU": {
       "time_ms": 0.378,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.349,
     "pct": 0.571,
     "by_device": {
      "CPU": {
       "time_ms": 0.349,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 0.327,
     "pct": 0.535,
     "by_device": {
      "CPU": {
       "time_ms": 0.327,
       "nodes": 1,
       "calls": 82
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.27,
     "pct": 0.442,
     "by_device": {
      "CPU": {
       "time_ms": 0.27,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.264,
     "pct": 0.432,
     "by_device": {
      "CPU": {
       "time_ms": 0.264,
       "nodes": 1,
       "calls": 106
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.252,
     "pct": 0.412,
     "by_device": {
      "CPU": {
       "time_ms": 0.252,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.155,
     "pct": 0.253,
     "by_device": {
      "CPU": {
       "time_ms": 0.155,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.139,
     "pct": 0.227,
     "by_device": {
      "CPU": {
       "time_ms": 0.139,
       "nodes": 1,
       "calls": 55
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.136,
     "pct": 0.222,
     "by_device": {
      "CPU": {
       "time_ms": 0.136,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.099,
     "pct": 0.162,
     "by_device": {
      "CPU": {
       "time_ms": 0.099,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.07,
     "pct": 0.114,
     "by_device": {
      "CPU": {
       "time_ms": 0.07,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.058,
     "pct": 0.095,
     "by_device": {
      "CPU": {
       "time_ms": 0.058,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.056,
     "pct": 0.092,
     "by_device": {
      "CPU": {
       "time_ms": 0.056,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.047,
     "pct": 0.077,
     "by_device": {
      "CPU": {
       "time_ms": 0.047,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.041,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 0.041,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.033,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 0.029,
     "pct": 0.048,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.026,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.023,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.023,
     "pct": 0.038,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.023,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.021,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.02,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.012,
     "pct": 0.02,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "add_",
     "total_ms": 0.012,
     "pct": 0.02,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.006,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.006,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.005,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.005,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.003,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_phi-1_5_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_phi-1_5_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:28:37",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "pythia-1b",
   "hf_model_id": "EleutherAI/pythia-1b",
   "hf_model_url": "https://huggingface.co/EleutherAI/pythia-1b",
   "param_count": 1011781632
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.023,
   "TTFT": 0.013,
   "TBT_avg": 0.01,
   "Throughput": 1410.893
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 22.681,
     "ttft": 12.589,
     "tbt_avg": 10.091,
     "throughput": 1410.893,
     "steps": [
      12.589,
      10.091
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 22.681,
     "min": 22.681,
     "max": 22.681,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 12.589,
     "min": 12.589,
     "max": 12.589,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 10.091,
     "min": 10.091,
     "max": 10.091,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 1410.893,
     "min": 1410.893,
     "max": 1410.893,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 10.091,
     "min": 10.091,
     "max": 10.091,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 39.308,
   "total_nodes": 84,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 24.915,
     "pct": 63.385,
     "nodes": 51,
     "calls": 4220
    },
    {
     "device": "GPU",
     "time_ms": 14.393,
     "pct": 36.615,
     "nodes": 33,
     "calls": 818
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 14.391,
     "pct": 36.611,
     "by_device": {
      "GPU": {
       "time_ms": 14.391,
       "nodes": 32,
       "calls": 816
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 3.415,
     "pct": 8.687,
     "by_device": {
      "CPU": {
       "time_ms": 3.415,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 2.212,
     "pct": 5.626,
     "by_device": {
      "CPU": {
       "time_ms": 2.212,
       "nodes": 1,
       "calls": 128
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 1.693,
     "pct": 4.307,
     "by_device": {
      "CPU": {
       "time_ms": 1.693,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 1.649,
     "pct": 4.195,
     "by_device": {
      "CPU": {
       "time_ms": 1.649,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 1.523,
     "pct": 3.876,
     "by_device": {
      "CPU": {
       "time_ms": 1.523,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 1.428,
     "pct": 3.634,
     "by_device": {
      "CPU": {
       "time_ms": 1.428,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 1.384,
     "pct": 3.52,
     "by_device": {
      "CPU": {
       "time_ms": 1.384,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 1.166,
     "pct": 2.967,
     "by_device": {
      "CPU": {
       "time_ms": 1.166,
       "nodes": 1,
       "calls": 132
      }
     }
    },
    {
     "op": "add",
     "total_ms": 1.064,
     "pct": 2.708,
     "by_device": {
      "CPU": {
       "time_ms": 1.064,
       "nodes": 1,
       "calls": 128
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 1.018,
     "pct": 2.59,
     "by_device": {
      "CPU": {
       "time_ms": 1.018,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.852,
     "pct": 2.166,
     "by_device": {
      "CPU": {
       "time_ms": 0.852,
       "nodes": 1,
       "calls": 376
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.734,
     "pct": 1.869,
     "by_device": {
      "CPU": {
       "time_ms": 0.734,
       "nodes": 1,
       "calls": 354
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 0.588,
     "pct": 1.496,
     "by_device": {
      "CPU": {
       "time_ms": 0.588,
       "nodes": 1,
       "calls": 324
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.568,
     "pct": 1.445,
     "by_device": {
      "CPU": {
       "time_ms": 0.568,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "chunk",
     "total_ms": 0.551,
     "pct": 1.402,
     "by_device": {
      "CPU": {
       "time_ms": 0.551,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "split",
     "total_ms": 0.442,
     "pct": 1.125,
     "by_device": {
      "CPU": {
       "time_ms": 0.442,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.377,
     "pct": 0.959,
     "by_device": {
      "CPU": {
       "time_ms": 0.377,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.343,
     "pct": 0.873,
     "by_device": {
      "CPU": {
       "time_ms": 0.343,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.33,
     "pct": 0.84,
     "by_device": {
      "CPU": {
       "time_ms": 0.33,
       "nodes": 1,
       "calls": 42
      }
     }
    },
    {
     "op": "gelu",
     "total_ms": 0.318,
     "pct": 0.808,
     "by_device": {
      "CPU": {
       "time_ms": 0.318,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.311,
     "pct": 0.792,
     "by_device": {
      "CPU": {
       "time_ms": 0.311,
       "nodes": 1,
       "calls": 460
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.307,
     "pct": 0.782,
     "by_device": {
      "CPU": {
       "time_ms": 0.307,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.289,
     "pct": 0.736,
     "by_device": {
      "CPU": {
       "time_ms": 0.289,
       "nodes": 1,
       "calls": 759
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.279,
     "pct": 0.71,
     "by_device": {
      "CPU": {
       "time_ms": 0.279,
       "nodes": 1,
       "calls": 16
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.258,
     "pct": 0.657,
     "by_device": {
      "CPU": {
       "time_ms": 0.258,
       "nodes": 1,
       "calls": 16
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 0.255,
     "pct": 0.648,
     "by_device": {
      "CPU": {
       "time_ms": 0.255,
       "nodes": 1,
       "calls": 58
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.202,
     "pct": 0.514,
     "by_device": {
      "CPU": {
       "time_ms": 0.202,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.197,
     "pct": 0.501,
     "by_device": {
      "CPU": {
       "time_ms": 0.197,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.192,
     "pct": 0.487,
     "by_device": {
      "CPU": {
       "time_ms": 0.192,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.175,
     "pct": 0.446,
     "by_device": {
      "CPU": {
       "time_ms": 0.175,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.139,
     "pct": 0.353,
     "by_device": {
      "CPU": {
       "time_ms": 0.139,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.111,
     "pct": 0.282,
     "by_device": {
      "CPU": {
       "time_ms": 0.111,
       "nodes": 1,
       "calls": 40
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.09,
     "pct": 0.229,
     "by_device": {
      "CPU": {
       "time_ms": 0.09,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.068,
     "pct": 0.172,
     "by_device": {
      "CPU": {
       "time_ms": 0.068,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.061,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 0.061,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 0.044,
     "pct": 0.111,
     "by_device": {
      "CPU": {
       "time_ms": 0.044,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.041,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 0.041,
       "nodes": 2,
       "calls": 64
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.036,
     "pct": 0.091,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.033,
     "pct": 0.084,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.03,
     "pct": 0.076,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.025,
     "pct": 0.063,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.022,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.021,
     "pct": 0.053,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.019,
     "pct": 0.048,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.019,
     "pct": 0.048,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.018,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.005,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.005,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.005,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.003,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.004,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_pythia-1b_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_pythia-1b_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:21:58",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "Qwen2.5-1.5B",
   "hf_model_id": "Qwen/Qwen2.5-1.5B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen2.5-1.5B",
   "param_count": 1543714304
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.047,
   "TTFT": 0.024,
   "TBT_avg": 0.023,
   "Throughput": 678.191
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 47.184,
     "ttft": 24.307,
     "tbt_avg": 22.877,
     "throughput": 678.191,
     "steps": [
      24.307,
      22.877
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 47.184,
     "min": 47.184,
     "max": 47.184,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 24.307,
     "min": 24.307,
     "max": 24.307,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 22.877,
     "min": 22.877,
     "max": 22.877,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 678.191,
     "min": 678.191,
     "max": 678.191,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 22.877,
     "min": 22.877,
     "max": 22.877,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 82.51,
   "total_nodes": 87,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 56.986,
     "pct": 69.066,
     "nodes": 47,
     "calls": 9260
    },
    {
     "device": "GPU",
     "time_ms": 25.523,
     "pct": 30.934,
     "nodes": 40,
     "calls": 2456
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 25.512,
     "pct": 30.92,
     "by_device": {
      "GPU": {
       "time_ms": 25.512,
       "nodes": 38,
       "calls": 2426
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 9.743,
     "pct": 11.809,
     "by_device": {
      "CPU": {
       "time_ms": 9.743,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 4.303,
     "pct": 5.215,
     "by_device": {
      "CPU": {
       "time_ms": 4.303,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 3.93,
     "pct": 4.763,
     "by_device": {
      "CPU": {
       "time_ms": 3.93,
       "nodes": 1,
       "calls": 512
      }
     }
    },
    {
     "op": "to",
     "total_ms": 3.504,
     "pct": 4.247,
     "by_device": {
      "CPU": {
       "time_ms": 3.504,
       "nodes": 1,
       "calls": 416
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 3.185,
     "pct": 3.86,
     "by_device": {
      "CPU": {
       "time_ms": 3.185,
       "nodes": 1,
       "calls": 294
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 3.166,
     "pct": 3.838,
     "by_device": {
      "CPU": {
       "time_ms": 3.166,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 3.014,
     "pct": 3.653,
     "by_device": {
      "CPU": {
       "time_ms": 3.014,
       "nodes": 1,
       "calls": 168
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 2.969,
     "pct": 3.598,
     "by_device": {
      "CPU": {
       "time_ms": 2.969,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 2.762,
     "pct": 3.347,
     "by_device": {
      "CPU": {
       "time_ms": 2.762,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.618,
     "pct": 3.173,
     "by_device": {
      "CPU": {
       "time_ms": 2.618,
       "nodes": 1,
       "calls": 338
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 2.069,
     "pct": 2.508,
     "by_device": {
      "CPU": {
       "time_ms": 2.069,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 1.915,
     "pct": 2.321,
     "by_device": {
      "CPU": {
       "time_ms": 1.915,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.58,
     "pct": 1.915,
     "by_device": {
      "CPU": {
       "time_ms": 1.58,
       "nodes": 1,
       "calls": 928
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 1.541,
     "pct": 1.868,
     "by_device": {
      "CPU": {
       "time_ms": 1.541,
       "nodes": 1,
       "calls": 294
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 1.201,
     "pct": 1.455,
     "by_device": {
      "CPU": {
       "time_ms": 1.201,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 1.162,
     "pct": 1.408,
     "by_device": {
      "CPU": {
       "time_ms": 1.162,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.1,
     "pct": 1.333,
     "by_device": {
      "CPU": {
       "time_ms": 1.1,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.908,
     "pct": 1.1,
     "by_device": {
      "CPU": {
       "time_ms": 0.908,
       "nodes": 1,
       "calls": 112
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.903,
     "pct": 1.094,
     "by_device": {
      "CPU": {
       "time_ms": 0.903,
       "nodes": 1,
       "calls": 350
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.823,
     "pct": 0.998,
     "by_device": {
      "CPU": {
       "time_ms": 0.823,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.616,
     "pct": 0.746,
     "by_device": {
      "CPU": {
       "time_ms": 0.616,
       "nodes": 1,
       "calls": 904
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.609,
     "pct": 0.738,
     "by_device": {
      "CPU": {
       "time_ms": 0.609,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.564,
     "pct": 0.684,
     "by_device": {
      "CPU": {
       "time_ms": 0.564,
       "nodes": 1,
       "calls": 400
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.548,
     "pct": 0.664,
     "by_device": {
      "CPU": {
       "time_ms": 0.548,
       "nodes": 1,
       "calls": 282
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 0.51,
     "pct": 0.618,
     "by_device": {
      "CPU": {
       "time_ms": 0.51,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.469,
     "pct": 0.568,
     "by_device": {
      "CPU": {
       "time_ms": 0.469,
       "nodes": 1,
       "calls": 1283
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.271,
     "pct": 0.328,
     "by_device": {
      "CPU": {
       "time_ms": 0.271,
       "nodes": 1,
       "calls": 118
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.23,
     "pct": 0.279,
     "by_device": {
      "CPU": {
       "time_ms": 0.23,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.136,
     "pct": 0.165,
     "by_device": {
      "CPU": {
       "time_ms": 0.136,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.128,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 0.128,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.085,
     "pct": 0.103,
     "by_device": {
      "CPU": {
       "time_ms": 0.085,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.064,
     "pct": 0.077,
     "by_device": {
      "CPU": {
       "time_ms": 0.064,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.064,
     "pct": 0.077,
     "by_device": {
      "CPU": {
       "time_ms": 0.064,
       "nodes": 2,
       "calls": 112
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.056,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 0.056,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.033,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.031,
     "pct": 0.038,
     "by_device": {
      "CPU": {
       "time_ms": 0.031,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.03,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.028,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.023,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.022,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.02,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.019,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.019,
     "pct": 0.023,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gpu_memset",
     "total_ms": 0.01,
     "pct": 0.012,
     "by_device": {
      "GPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 28
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.005,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.005,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.004,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.002,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_qwen2.5-1.5b_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_qwen2.5-1.5b_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:21:50",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "Qwen3-0.6B",
   "hf_model_id": "Qwen/Qwen3-0.6B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen3-0.6B",
   "param_count": 596049920
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.058,
   "TTFT": 0.03,
   "TBT_avg": 0.028,
   "Throughput": 549.416
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 58.244,
     "ttft": 30.45,
     "tbt_avg": 27.793,
     "throughput": 549.416,
     "steps": [
      30.45,
      27.793
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 58.244,
     "min": 58.244,
     "max": 58.244,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 30.45,
     "min": 30.45,
     "max": 30.45,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 27.793,
     "min": 27.793,
     "max": 27.793,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 549.416,
     "min": 549.416,
     "max": 549.416,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 27.793,
     "min": 27.793,
     "max": 27.793,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 87.838,
   "total_nodes": 85,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 72.724,
     "pct": 82.793,
     "nodes": 46,
     "calls": 11388
    },
    {
     "device": "GPU",
     "time_ms": 15.115,
     "pct": 17.207,
     "nodes": 39,
     "calls": 3268
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 15.104,
     "pct": 17.195,
     "by_device": {
      "GPU": {
       "time_ms": 15.104,
       "nodes": 37,
       "calls": 3238
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 9.316,
     "pct": 10.606,
     "by_device": {
      "CPU": {
       "time_ms": 9.316,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 7.595,
     "pct": 8.646,
     "by_device": {
      "CPU": {
       "time_ms": 7.595,
       "nodes": 1,
       "calls": 396
      }
     }
    },
    {
     "op": "to",
     "total_ms": 6.233,
     "pct": 7.096,
     "by_device": {
      "CPU": {
       "time_ms": 6.233,
       "nodes": 1,
       "calls": 752
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 5.703,
     "pct": 6.492,
     "by_device": {
      "CPU": {
       "time_ms": 5.703,
       "nodes": 1,
       "calls": 736
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 5.682,
     "pct": 6.468,
     "by_device": {
      "CPU": {
       "time_ms": 5.682,
       "nodes": 1,
       "calls": 518
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 5.392,
     "pct": 6.139,
     "by_device": {
      "CPU": {
       "time_ms": 5.392,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "add",
     "total_ms": 3.629,
     "pct": 4.132,
     "by_device": {
      "CPU": {
       "time_ms": 3.629,
       "nodes": 1,
       "calls": 450
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 3.097,
     "pct": 3.526,
     "by_device": {
      "CPU": {
       "time_ms": 3.097,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 2.998,
     "pct": 3.413,
     "by_device": {
      "CPU": {
       "time_ms": 2.998,
       "nodes": 1,
       "calls": 518
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 2.698,
     "pct": 3.072,
     "by_device": {
      "CPU": {
       "time_ms": 2.698,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 2.427,
     "pct": 2.763,
     "by_device": {
      "CPU": {
       "time_ms": 2.427,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 2.216,
     "pct": 2.522,
     "by_device": {
      "CPU": {
       "time_ms": 2.216,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 2.067,
     "pct": 2.353,
     "by_device": {
      "CPU": {
       "time_ms": 2.067,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 2.056,
     "pct": 2.34,
     "by_device": {
      "CPU": {
       "time_ms": 2.056,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.672,
     "pct": 1.903,
     "by_device": {
      "CPU": {
       "time_ms": 1.672,
       "nodes": 1,
       "calls": 928
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 1.664,
     "pct": 1.894,
     "by_device": {
      "CPU": {
       "time_ms": 1.664,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.401,
     "pct": 1.595,
     "by_device": {
      "CPU": {
       "time_ms": 1.401,
       "nodes": 1,
       "calls": 574
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.142,
     "pct": 1.301,
     "by_device": {
      "CPU": {
       "time_ms": 1.142,
       "nodes": 1,
       "calls": 394
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.929,
     "pct": 1.057,
     "by_device": {
      "CPU": {
       "time_ms": 0.929,
       "nodes": 1,
       "calls": 112
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.873,
     "pct": 0.994,
     "by_device": {
      "CPU": {
       "time_ms": 0.873,
       "nodes": 1,
       "calls": 568
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.524,
     "pct": 0.596,
     "by_device": {
      "CPU": {
       "time_ms": 0.524,
       "nodes": 1,
       "calls": 282
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.519,
     "pct": 0.591,
     "by_device": {
      "CPU": {
       "time_ms": 0.519,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 0.513,
     "pct": 0.584,
     "by_device": {
      "CPU": {
       "time_ms": 0.513,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.505,
     "pct": 0.575,
     "by_device": {
      "CPU": {
       "time_ms": 0.505,
       "nodes": 1,
       "calls": 736
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.478,
     "pct": 0.544,
     "by_device": {
      "CPU": {
       "time_ms": 0.478,
       "nodes": 1,
       "calls": 1283
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.271,
     "pct": 0.309,
     "by_device": {
      "CPU": {
       "time_ms": 0.271,
       "nodes": 1,
       "calls": 118
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.224,
     "pct": 0.255,
     "by_device": {
      "CPU": {
       "time_ms": 0.224,
       "nodes": 1,
       "calls": 396
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.216,
     "pct": 0.246,
     "by_device": {
      "CPU": {
       "time_ms": 0.216,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.13,
     "pct": 0.148,
     "by_device": {
      "CPU": {
       "time_ms": 0.13,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.094,
     "pct": 0.107,
     "by_device": {
      "CPU": {
       "time_ms": 0.094,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.07,
     "pct": 0.08,
     "by_device": {
      "CPU": {
       "time_ms": 0.07,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.063,
     "pct": 0.072,
     "by_device": {
      "CPU": {
       "time_ms": 0.063,
       "nodes": 2,
       "calls": 112
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.06,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 0.06,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.055,
     "pct": 0.062,
     "by_device": {
      "CPU": {
       "time_ms": 0.055,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.035,
     "pct": 0.039,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.032,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.029,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.023,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.023,
     "pct": 0.026,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.021,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.019,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.019,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gpu_memset",
     "total_ms": 0.009,
     "pct": 0.01,
     "by_device": {
      "GPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 28
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.006,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.005,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.005,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.002,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_qwen3-0.6b_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_qwen3-0.6b_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:21:43",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "stablelm-2-1_6b",
   "hf_model_id": "stabilityai/stablelm-2-1_6b",
   "hf_model_url": "https://huggingface.co/stabilityai/stablelm-2-1_6b",
   "param_count": 1644417024
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.035,
   "TTFT": 0.019,
   "TBT_avg": 0.016,
   "Throughput": 909.093
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 35.2,
     "ttft": 19.324,
     "tbt_avg": 15.876,
     "throughput": 909.093,
     "steps": [
      19.324,
      15.876
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 35.2,
     "min": 35.2,
     "max": 35.2,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 19.324,
     "min": 19.324,
     "max": 19.324,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 15.876,
     "min": 15.876,
     "max": 15.876,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 909.093,
     "min": 909.093,
     "max": 909.093,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 15.876,
     "min": 15.876,
     "max": 15.876,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 67.113,
   "total_nodes": 83,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 42.657,
     "pct": 63.559,
     "nodes": 48,
     "calls": 7394
    },
    {
     "device": "GPU",
     "time_ms": 24.456,
     "pct": 36.441,
     "nodes": 35,
     "calls": 1426
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 24.455,
     "pct": 36.438,
     "by_device": {
      "GPU": {
       "time_ms": 24.455,
       "nodes": 34,
       "calls": 1424
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 7.754,
     "pct": 11.553,
     "by_device": {
      "CPU": {
       "time_ms": 7.754,
       "nodes": 1,
       "calls": 338
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 3.762,
     "pct": 5.606,
     "by_device": {
      "CPU": {
       "time_ms": 3.762,
       "nodes": 1,
       "calls": 196
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 2.686,
     "pct": 4.002,
     "by_device": {
      "CPU": {
       "time_ms": 2.686,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 2.543,
     "pct": 3.789,
     "by_device": {
      "CPU": {
       "time_ms": 2.543,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.242,
     "pct": 3.341,
     "by_device": {
      "CPU": {
       "time_ms": 2.242,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 2.175,
     "pct": 3.241,
     "by_device": {
      "CPU": {
       "time_ms": 2.175,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "addmm",
     "total_ms": 2.162,
     "pct": 3.222,
     "by_device": {
      "CPU": {
       "time_ms": 2.162,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 2.081,
     "pct": 3.101,
     "by_device": {
      "CPU": {
       "time_ms": 2.081,
       "nodes": 1,
       "calls": 244
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 1.981,
     "pct": 2.952,
     "by_device": {
      "CPU": {
       "time_ms": 1.981,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 1.92,
     "pct": 2.861,
     "by_device": {
      "CPU": {
       "time_ms": 1.92,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "add",
     "total_ms": 1.583,
     "pct": 2.359,
     "by_device": {
      "CPU": {
       "time_ms": 1.583,
       "nodes": 1,
       "calls": 192
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 1.362,
     "pct": 2.03,
     "by_device": {
      "CPU": {
       "time_ms": 1.362,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 1.185,
     "pct": 1.765,
     "by_device": {
      "CPU": {
       "time_ms": 1.185,
       "nodes": 1,
       "calls": 560
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.177,
     "pct": 1.754,
     "by_device": {
      "CPU": {
       "time_ms": 1.177,
       "nodes": 1,
       "calls": 724
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.881,
     "pct": 1.313,
     "by_device": {
      "CPU": {
       "time_ms": 0.881,
       "nodes": 1,
       "calls": 338
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.871,
     "pct": 1.298,
     "by_device": {
      "CPU": {
       "time_ms": 0.871,
       "nodes": 1,
       "calls": 386
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.815,
     "pct": 1.214,
     "by_device": {
      "CPU": {
       "time_ms": 0.815,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.529,
     "pct": 0.788,
     "by_device": {
      "CPU": {
       "time_ms": 0.529,
       "nodes": 1,
       "calls": 876
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.521,
     "pct": 0.776,
     "by_device": {
      "CPU": {
       "time_ms": 0.521,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.465,
     "pct": 0.693,
     "by_device": {
      "CPU": {
       "time_ms": 0.465,
       "nodes": 1,
       "calls": 58
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 0.464,
     "pct": 0.692,
     "by_device": {
      "CPU": {
       "time_ms": 0.464,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.439,
     "pct": 0.654,
     "by_device": {
      "CPU": {
       "time_ms": 0.439,
       "nodes": 1,
       "calls": 1223
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.397,
     "pct": 0.592,
     "by_device": {
      "CPU": {
       "time_ms": 0.397,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.392,
     "pct": 0.584,
     "by_device": {
      "CPU": {
       "time_ms": 0.392,
       "nodes": 1,
       "calls": 248
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.366,
     "pct": 0.546,
     "by_device": {
      "CPU": {
       "time_ms": 0.366,
       "nodes": 1,
       "calls": 24
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 0.337,
     "pct": 0.502,
     "by_device": {
      "CPU": {
       "time_ms": 0.337,
       "nodes": 1,
       "calls": 82
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.28,
     "pct": 0.417,
     "by_device": {
      "CPU": {
       "time_ms": 0.28,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.259,
     "pct": 0.386,
     "by_device": {
      "CPU": {
       "time_ms": 0.259,
       "nodes": 1,
       "calls": 106
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.243,
     "pct": 0.362,
     "by_device": {
      "CPU": {
       "time_ms": 0.243,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.135,
     "pct": 0.201,
     "by_device": {
      "CPU": {
       "time_ms": 0.135,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.116,
     "pct": 0.173,
     "by_device": {
      "CPU": {
       "time_ms": 0.116,
       "nodes": 1,
       "calls": 196
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.093,
     "pct": 0.139,
     "by_device": {
      "CPU": {
       "time_ms": 0.093,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.07,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 0.07,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.063,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 0.063,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.052,
     "pct": 0.077,
     "by_device": {
      "CPU": {
       "time_ms": 0.052,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.041,
     "pct": 0.061,
     "by_device": {
      "CPU": {
       "time_ms": 0.041,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.038,
     "pct": 0.057,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.035,
     "pct": 0.052,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.024,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.023,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.021,
     "pct": 0.031,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.021,
     "pct": 0.031,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.02,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.016,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.005,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.005,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.005,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.002,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_stablelm-2-1_6b_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_stablelm-2-1_6b_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:19:10",
   "framework": "pytorch",
   "device": "cpu+gpu"
  },
  "model_info": {
   "model_name": "TinyLlama-1.1B-Chat-v1.0",
   "hf_model_id": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "hf_model_url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "param_count": 1100048384
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 0.041,
   "TTFT": 0.021,
   "TBT_avg": 0.02,
   "Throughput": 773.107
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 41.391,
     "ttft": 21.411,
     "tbt_avg": 19.98,
     "throughput": 773.107,
     "steps": [
      21.411,
      19.98
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 41.391,
     "min": 41.391,
     "max": 41.391,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 21.411,
     "min": 21.411,
     "max": 21.411,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 19.98,
     "min": 19.98,
     "max": 19.98,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 773.107,
     "min": 773.107,
     "max": 773.107,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 19.98,
     "min": 19.98,
     "max": 19.98,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 72.424,
   "total_nodes": 85,
   "devices": [
    {
     "device": "CPU",
     "time_ms": 53.366,
     "pct": 73.686,
     "nodes": 46,
     "calls": 7580
    },
    {
     "device": "GPU",
     "time_ms": 19.058,
     "pct": 26.314,
     "nodes": 39,
     "calls": 1918
    }
   ],
   "ops": [
    {
     "op": "kernel",
     "total_ms": 19.056,
     "pct": 26.312,
     "by_device": {
      "GPU": {
       "time_ms": 19.056,
       "nodes": 38,
       "calls": 1916
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 8.485,
     "pct": 11.715,
     "by_device": {
      "CPU": {
       "time_ms": 8.485,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 6.828,
     "pct": 9.428,
     "by_device": {
      "CPU": {
       "time_ms": 6.828,
       "nodes": 1,
       "calls": 312
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 4.86,
     "pct": 6.711,
     "by_device": {
      "CPU": {
       "time_ms": 4.86,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 3.424,
     "pct": 4.727,
     "by_device": {
      "CPU": {
       "time_ms": 3.424,
       "nodes": 1,
       "calls": 404
      }
     }
    },
    {
     "op": "to",
     "total_ms": 3.059,
     "pct": 4.224,
     "by_device": {
      "CPU": {
       "time_ms": 3.059,
       "nodes": 1,
       "calls": 332
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 2.792,
     "pct": 3.854,
     "by_device": {
      "CPU": {
       "time_ms": 2.792,
       "nodes": 1,
       "calls": 234
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.686,
     "pct": 3.708,
     "by_device": {
      "CPU": {
       "time_ms": 2.686,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.343,
     "pct": 3.235,
     "by_device": {
      "CPU": {
       "time_ms": 2.343,
       "nodes": 1,
       "calls": 266
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention",
     "total_ms": 2.335,
     "pct": 3.224,
     "by_device": {
      "CPU": {
       "time_ms": 2.335,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "_flash_attention_forward",
     "total_ms": 1.748,
     "pct": 2.414,
     "by_device": {
      "CPU": {
       "time_ms": 1.748,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 1.734,
     "pct": 2.394,
     "by_device": {
      "CPU": {
       "time_ms": 1.734,
       "nodes": 1,
       "calls": 178
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.594,
     "pct": 2.201,
     "by_device": {
      "CPU": {
       "time_ms": 1.594,
       "nodes": 1,
       "calls": 730
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 1.406,
     "pct": 1.942,
     "by_device": {
      "CPU": {
       "time_ms": 1.406,
       "nodes": 1,
       "calls": 234
      }
     }
    },
    {
     "op": "t",
     "total_ms": 1.14,
     "pct": 1.575,
     "by_device": {
      "CPU": {
       "time_ms": 1.14,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 1.09,
     "pct": 1.505,
     "by_device": {
      "CPU": {
       "time_ms": 1.09,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 1.01,
     "pct": 1.395,
     "by_device": {
      "CPU": {
       "time_ms": 1.01,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 0.819,
     "pct": 1.131,
     "by_device": {
      "CPU": {
       "time_ms": 0.819,
       "nodes": 1,
       "calls": 448
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.803,
     "pct": 1.108,
     "by_device": {
      "CPU": {
       "time_ms": 0.803,
       "nodes": 1,
       "calls": 278
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.79,
     "pct": 1.09,
     "by_device": {
      "CPU": {
       "time_ms": 0.79,
       "nodes": 1,
       "calls": 88
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.762,
     "pct": 1.052,
     "by_device": {
      "CPU": {
       "time_ms": 0.762,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 0.486,
     "pct": 0.672,
     "by_device": {
      "CPU": {
       "time_ms": 0.486,
       "nodes": 1,
       "calls": 178
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 0.471,
     "pct": 0.651,
     "by_device": {
      "CPU": {
       "time_ms": 0.471,
       "nodes": 1,
       "calls": 222
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.446,
     "pct": 0.616,
     "by_device": {
      "CPU": {
       "time_ms": 0.446,
       "nodes": 1,
       "calls": 1013
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.442,
     "pct": 0.61,
     "by_device": {
      "CPU": {
       "time_ms": 0.442,
       "nodes": 1,
       "calls": 580
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 0.442,
     "pct": 0.61,
     "by_device": {
      "CPU": {
       "time_ms": 0.442,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.315,
     "pct": 0.436,
     "by_device": {
      "CPU": {
       "time_ms": 0.315,
       "nodes": 1,
       "calls": 94
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.203,
     "pct": 0.28,
     "by_device": {
      "CPU": {
       "time_ms": 0.203,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.179,
     "pct": 0.248,
     "by_device": {
      "CPU": {
       "time_ms": 0.179,
       "nodes": 1,
       "calls": 312
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.14,
     "pct": 0.194,
     "by_device": {
      "CPU": {
       "time_ms": 0.14,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.087,
     "pct": 0.12,
     "by_device": {
      "CPU": {
       "time_ms": 0.087,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.067,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 0.067,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.064,
     "pct": 0.088,
     "by_device": {
      "CPU": {
       "time_ms": 0.064,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.057,
     "pct": 0.078,
     "by_device": {
      "CPU": {
       "time_ms": 0.057,
       "nodes": 2,
       "calls": 88
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.036,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.033,
     "pct": 0.046,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.033,
     "pct": 0.046,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.03,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.027,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.025,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "gather",
     "total_ms": 0.022,
     "pct": 0.03,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.02,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.02,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.019,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 7
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.005,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.005,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 3
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.005,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "gpu_memcpy",
     "total_ms": 0.002,
     "pct": 0.002,
     "by_device": {
      "GPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "gpu": {
    "name": "NVIDIA RTX 4000 Ada Generation",
    "count": 1,
    "memory_gb": 20,
    "architecture": "Ada Lovelace",
    "cuda_version": "13.0"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/gpu/torch_tinyllama-1.1b-chat_b16s8o2_gpu.simplified.json",
   "source_profile": "torch_tinyllama-1.1b-chat_b16s8o2_gpu.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)",
    "environment_spec.gpu  (kineto deviceProperties)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:02:05",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "bloom-1b7",
   "hf_model_id": "bigscience/bloom-1b7",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-1b7",
   "param_count": 1722408960
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.764,
   "TTFT": 2.206,
   "TBT_avg": 0.558,
   "Throughput": 11.577
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2764.088,
     "ttft": 2206.407,
     "tbt_avg": 557.681,
     "throughput": 11.577,
     "steps": [
      2206.407,
      557.681
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2764.088,
     "min": 2764.088,
     "max": 2764.088,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 2206.407,
     "min": 2206.407,
     "max": 2206.407,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 557.681,
     "min": 557.681,
     "max": 557.681,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 11.577,
     "min": 11.577,
     "max": 11.577,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 557.681,
     "min": 557.681,
     "max": 557.681,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2986.53,
   "total_nodes": 61,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 2650.075,
     "pct": 88.734,
     "nodes": 3,
     "calls": 672
    },
    {
     "device": "CPU",
     "time_ms": 336.454,
     "pct": 11.266,
     "nodes": 58,
     "calls": 12275
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 2335.11,
     "pct": 78.188,
     "by_device": {
      "PIM": {
       "time_ms": 2335.11,
       "nodes": 1,
       "calls": 192
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 271.212,
     "pct": 9.081,
     "by_device": {
      "PIM": {
       "time_ms": 271.212,
       "nodes": 1,
       "calls": 384
      }
     }
    },
    {
     "op": "GeLUFunction",
     "total_ms": 133.545,
     "pct": 4.472,
     "by_device": {
      "CPU": {
       "time_ms": 133.545,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 43.753,
     "pct": 1.465,
     "by_device": {
      "PIM": {
       "time_ms": 43.753,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 20.247,
     "pct": 0.678,
     "by_device": {
      "CPU": {
       "time_ms": 20.247,
       "nodes": 1,
       "calls": 292
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 19.255,
     "pct": 0.645,
     "by_device": {
      "CPU": {
       "time_ms": 19.255,
       "nodes": 1,
       "calls": 482
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 15.439,
     "pct": 0.517,
     "by_device": {
      "CPU": {
       "time_ms": 15.439,
       "nodes": 1,
       "calls": 482
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 14.702,
     "pct": 0.492,
     "by_device": {
      "CPU": {
       "time_ms": 14.702,
       "nodes": 1,
       "calls": 484
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 10.728,
     "pct": 0.359,
     "by_device": {
      "CPU": {
       "time_ms": 10.728,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 10.721,
     "pct": 0.359,
     "by_device": {
      "CPU": {
       "time_ms": 10.721,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 10.29,
     "pct": 0.345,
     "by_device": {
      "CPU": {
       "time_ms": 10.29,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 9.227,
     "pct": 0.309,
     "by_device": {
      "CPU": {
       "time_ms": 9.227,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 9.175,
     "pct": 0.307,
     "by_device": {
      "CPU": {
       "time_ms": 9.175,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 9.135,
     "pct": 0.306,
     "by_device": {
      "CPU": {
       "time_ms": 9.135,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "baddbmm",
     "total_ms": 7.336,
     "pct": 0.246,
     "by_device": {
      "CPU": {
       "time_ms": 7.336,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 6.486,
     "pct": 0.217,
     "by_device": {
      "CPU": {
       "time_ms": 6.486,
       "nodes": 1,
       "calls": 588
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 5.273,
     "pct": 0.177,
     "by_device": {
      "CPU": {
       "time_ms": 5.273,
       "nodes": 1,
       "calls": 168
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 5.057,
     "pct": 0.169,
     "by_device": {
      "CPU": {
       "time_ms": 5.057,
       "nodes": 1,
       "calls": 560
      }
     }
    },
    {
     "op": "add",
     "total_ms": 4.982,
     "pct": 0.167,
     "by_device": {
      "CPU": {
       "time_ms": 4.982,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "to",
     "total_ms": 4.806,
     "pct": 0.161,
     "by_device": {
      "CPU": {
       "time_ms": 4.806,
       "nodes": 1,
       "calls": 600
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 4.122,
     "pct": 0.138,
     "by_device": {
      "CPU": {
       "time_ms": 4.122,
       "nodes": 1,
       "calls": 2313
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 4.116,
     "pct": 0.138,
     "by_device": {
      "CPU": {
       "time_ms": 4.116,
       "nodes": 1,
       "calls": 344
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 3.984,
     "pct": 0.133,
     "by_device": {
      "CPU": {
       "time_ms": 3.984,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 3.703,
     "pct": 0.124,
     "by_device": {
      "CPU": {
       "time_ms": 3.703,
       "nodes": 1,
       "calls": 1210
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 3.549,
     "pct": 0.119,
     "by_device": {
      "CPU": {
       "time_ms": 3.549,
       "nodes": 1,
       "calls": 482
      }
     }
    },
    {
     "op": "softmax",
     "total_ms": 3.017,
     "pct": 0.101,
     "by_device": {
      "CPU": {
       "time_ms": 3.017,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 3.008,
     "pct": 0.101,
     "by_device": {
      "CPU": {
       "time_ms": 3.008,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 2.447,
     "pct": 0.082,
     "by_device": {
      "CPU": {
       "time_ms": 2.447,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.678,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 1.678,
       "nodes": 1,
       "calls": 761
      }
     }
    },
    {
     "op": "_softmax",
     "total_ms": 1.533,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 1.533,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.336,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 1.336,
       "nodes": 1,
       "calls": 932
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.242,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 1.242,
       "nodes": 1,
       "calls": 344
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.112,
     "pct": 0.037,
     "by_device": {
      "CPU": {
       "time_ms": 1.112,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.954,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.954,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.912,
     "pct": 0.031,
     "by_device": {
      "CPU": {
       "time_ms": 0.912,
       "nodes": 1,
       "calls": 168
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.838,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.838,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.375,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.375,
       "nodes": 1,
       "calls": 149
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.289,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.289,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "permute",
     "total_ms": 0.288,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.288,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.251,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.251,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.226,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.226,
       "nodes": 1,
       "calls": 30
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.172,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.172,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.14,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.14,
       "nodes": 1,
       "calls": 28
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.136,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.136,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "arange",
     "total_ms": 0.133,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.133,
       "nodes": 1,
       "calls": 20
      }
     }
    },
    {
     "op": "index",
     "total_ms": 0.114,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.114,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "where",
     "total_ms": 0.076,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.076,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.061,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.061,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.036,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 196
      }
     }
    },
    {
     "op": "__and__",
     "total_ms": 0.034,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.029,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.029,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bitwise_and",
     "total_ms": 0.026,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "cumsum",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "new_ones",
     "total_ms": 0.021,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sub",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "le",
     "total_ms": 0.013,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.013,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.012,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 10
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "new_empty",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "scalar_tensor",
     "total_ms": 0.004,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_bloom-1b7_b16s8o2_pim.simplified.json",
   "source_profile": "torch_bloom-1b7_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:01:43",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "bloom-560m",
   "hf_model_id": "bigscience/bloom-560m",
   "hf_model_url": "https://huggingface.co/bigscience/bloom-560m",
   "param_count": 559214592
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.139,
   "TTFT": 0.778,
   "TBT_avg": 0.361,
   "Throughput": 28.101
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1138.745,
     "ttft": 777.945,
     "tbt_avg": 360.8,
     "throughput": 28.101,
     "steps": [
      777.945,
      360.8
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1138.745,
     "min": 1138.745,
     "max": 1138.745,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 777.945,
     "min": 777.945,
     "max": 777.945,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 360.8,
     "min": 360.8,
     "max": 360.8,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 28.101,
     "min": 28.101,
     "max": 28.101,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 360.8,
     "min": 360.8,
     "max": 360.8,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1384.603,
   "total_nodes": 62,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 1044.574,
     "pct": 75.442,
     "nodes": 4,
     "calls": 720
    },
    {
     "device": "CPU",
     "time_ms": 340.029,
     "pct": 24.558,
     "nodes": 58,
     "calls": 12922
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 709.273,
     "pct": 51.226,
     "by_device": {
      "PIM": {
       "time_ms": 709.273,
       "nodes": 1,
       "calls": 192
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 244.914,
     "pct": 17.688,
     "by_device": {
      "PIM": {
       "time_ms": 244.914,
       "nodes": 1,
       "calls": 384
      }
     }
    },
    {
     "op": "GeLUFunction",
     "total_ms": 177.55,
     "pct": 12.823,
     "by_device": {
      "CPU": {
       "time_ms": 177.55,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "pim_mul",
     "total_ms": 59.711,
     "pct": 4.312,
     "by_device": {
      "PIM": {
       "time_ms": 59.711,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 30.676,
     "pct": 2.215,
     "by_device": {
      "PIM": {
       "time_ms": 30.676,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 18.536,
     "pct": 1.339,
     "by_device": {
      "CPU": {
       "time_ms": 18.536,
       "nodes": 1,
       "calls": 578
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 14.213,
     "pct": 1.027,
     "by_device": {
      "CPU": {
       "time_ms": 14.213,
       "nodes": 1,
       "calls": 578
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 12.891,
     "pct": 0.931,
     "by_device": {
      "CPU": {
       "time_ms": 12.891,
       "nodes": 1,
       "calls": 411
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 11.613,
     "pct": 0.839,
     "by_device": {
      "CPU": {
       "time_ms": 11.613,
       "nodes": 1,
       "calls": 244
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 8.802,
     "pct": 0.636,
     "by_device": {
      "CPU": {
       "time_ms": 8.802,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 6.884,
     "pct": 0.497,
     "by_device": {
      "CPU": {
       "time_ms": 6.884,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 6.879,
     "pct": 0.497,
     "by_device": {
      "CPU": {
       "time_ms": 6.879,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 6.843,
     "pct": 0.494,
     "by_device": {
      "CPU": {
       "time_ms": 6.843,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 6.8,
     "pct": 0.491,
     "by_device": {
      "CPU": {
       "time_ms": 6.8,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 6.507,
     "pct": 0.47,
     "by_device": {
      "CPU": {
       "time_ms": 6.507,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 5.708,
     "pct": 0.412,
     "by_device": {
      "CPU": {
       "time_ms": 5.708,
       "nodes": 1,
       "calls": 684
      }
     }
    },
    {
     "op": "baddbmm",
     "total_ms": 5.512,
     "pct": 0.398,
     "by_device": {
      "CPU": {
       "time_ms": 5.512,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 4.652,
     "pct": 0.336,
     "by_device": {
      "CPU": {
       "time_ms": 4.652,
       "nodes": 1,
       "calls": 168
      }
     }
    },
    {
     "op": "add",
     "total_ms": 4.566,
     "pct": 0.33,
     "by_device": {
      "CPU": {
       "time_ms": 4.566,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 4.469,
     "pct": 0.323,
     "by_device": {
      "CPU": {
       "time_ms": 4.469,
       "nodes": 1,
       "calls": 560
      }
     }
    },
    {
     "op": "to",
     "total_ms": 4.122,
     "pct": 0.298,
     "by_device": {
      "CPU": {
       "time_ms": 4.122,
       "nodes": 1,
       "calls": 648
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 4.066,
     "pct": 0.294,
     "by_device": {
      "CPU": {
       "time_ms": 4.066,
       "nodes": 1,
       "calls": 1354
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 3.414,
     "pct": 0.247,
     "by_device": {
      "CPU": {
       "time_ms": 3.414,
       "nodes": 1,
       "calls": 530
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 3.38,
     "pct": 0.244,
     "by_device": {
      "CPU": {
       "time_ms": 3.38,
       "nodes": 1,
       "calls": 344
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 3.365,
     "pct": 0.243,
     "by_device": {
      "CPU": {
       "time_ms": 3.365,
       "nodes": 1,
       "calls": 2409
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 2.839,
     "pct": 0.205,
     "by_device": {
      "CPU": {
       "time_ms": 2.839,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 2.761,
     "pct": 0.199,
     "by_device": {
      "CPU": {
       "time_ms": 2.761,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "softmax",
     "total_ms": 2.363,
     "pct": 0.171,
     "by_device": {
      "CPU": {
       "time_ms": 2.363,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.607,
     "pct": 0.116,
     "by_device": {
      "CPU": {
       "time_ms": 1.607,
       "nodes": 1,
       "calls": 857
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 1.453,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 1.453,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_softmax",
     "total_ms": 1.154,
     "pct": 0.083,
     "by_device": {
      "CPU": {
       "time_ms": 1.154,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.934,
     "pct": 0.067,
     "by_device": {
      "CPU": {
       "time_ms": 0.934,
       "nodes": 1,
       "calls": 980
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 0.893,
     "pct": 0.065,
     "by_device": {
      "CPU": {
       "time_ms": 0.893,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.874,
     "pct": 0.063,
     "by_device": {
      "CPU": {
       "time_ms": 0.874,
       "nodes": 1,
       "calls": 168
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.725,
     "pct": 0.052,
     "by_device": {
      "CPU": {
       "time_ms": 0.725,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.716,
     "pct": 0.052,
     "by_device": {
      "CPU": {
       "time_ms": 0.716,
       "nodes": 1,
       "calls": 344
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.618,
     "pct": 0.045,
     "by_device": {
      "CPU": {
       "time_ms": 0.618,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.35,
     "pct": 0.025,
     "by_device": {
      "CPU": {
       "time_ms": 0.35,
       "nodes": 1,
       "calls": 149
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.335,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 0.335,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.305,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.305,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "permute",
     "total_ms": 0.267,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.267,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "arange",
     "total_ms": 0.143,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.143,
       "nodes": 1,
       "calls": 20
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.141,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.141,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.14,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.14,
       "nodes": 1,
       "calls": 30
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.114,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.114,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "index",
     "total_ms": 0.103,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.103,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.082,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.082,
       "nodes": 1,
       "calls": 28
      }
     }
    },
    {
     "op": "where",
     "total_ms": 0.064,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.064,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 0.054,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.054,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.038,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 1,
       "calls": 196
      }
     }
    },
    {
     "op": "__and__",
     "total_ms": 0.032,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "cumsum",
     "total_ms": 0.027,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.027,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "bitwise_and",
     "total_ms": 0.024,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.023,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resize_",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 10
      }
     }
    },
    {
     "op": "sub",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "new_ones",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "le",
     "total_ms": 0.015,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.015,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.009,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "new_empty",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "scalar_tensor",
     "total_ms": 0.003,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_bloom-560m_b16s8o2_pim.simplified.json",
   "source_profile": "torch_bloom-560m_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:56:56",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "Falcon3-1B-Base",
   "hf_model_id": "tiiuae/Falcon3-1B-Base",
   "hf_model_url": "https://huggingface.co/tiiuae/Falcon3-1B-Base",
   "param_count": 1669000000
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.737,
   "TTFT": 2.132,
   "TBT_avg": 0.605,
   "Throughput": 11.693
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2736.647,
     "ttft": 2131.973,
     "tbt_avg": 604.675,
     "throughput": 11.693,
     "steps": [
      2131.973,
      604.675
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2736.647,
     "min": 2736.647,
     "max": 2736.647,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 2131.973,
     "min": 2131.973,
     "max": 2131.973,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 604.675,
     "min": 604.675,
     "max": 604.675,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 11.693,
     "min": 11.693,
     "max": 11.693,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 604.675,
     "min": 604.675,
     "max": 604.675,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2826.591,
   "total_nodes": 53,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 2633.841,
     "pct": 93.181,
     "nodes": 3,
     "calls": 650
    },
    {
     "device": "CPU",
     "time_ms": 192.751,
     "pct": 6.819,
     "nodes": 50,
     "calls": 12338
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 2337.555,
     "pct": 82.699,
     "by_device": {
      "PIM": {
       "time_ms": 2337.555,
       "nodes": 1,
       "calls": 254
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 266.303,
     "pct": 9.421,
     "by_device": {
      "PIM": {
       "time_ms": 266.303,
       "nodes": 1,
       "calls": 324
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 31.974,
     "pct": 1.131,
     "by_device": {
      "CPU": {
       "time_ms": 31.974,
       "nodes": 1,
       "calls": 582
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 29.983,
     "pct": 1.061,
     "by_device": {
      "PIM": {
       "time_ms": 29.983,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 27.791,
     "pct": 0.983,
     "by_device": {
      "CPU": {
       "time_ms": 27.791,
       "nodes": 1,
       "calls": 582
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 26.848,
     "pct": 0.95,
     "by_device": {
      "CPU": {
       "time_ms": 26.848,
       "nodes": 1,
       "calls": 622
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 17.703,
     "pct": 0.626,
     "by_device": {
      "CPU": {
       "time_ms": 17.703,
       "nodes": 1,
       "calls": 332
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 9.427,
     "pct": 0.334,
     "by_device": {
      "CPU": {
       "time_ms": 9.427,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 8.633,
     "pct": 0.305,
     "by_device": {
      "CPU": {
       "time_ms": 8.633,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "to",
     "total_ms": 7.23,
     "pct": 0.256,
     "by_device": {
      "CPU": {
       "time_ms": 7.23,
       "nodes": 1,
       "calls": 684
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 6.945,
     "pct": 0.246,
     "by_device": {
      "CPU": {
       "time_ms": 6.945,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 6.559,
     "pct": 0.232,
     "by_device": {
      "CPU": {
       "time_ms": 6.559,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 4.582,
     "pct": 0.162,
     "by_device": {
      "CPU": {
       "time_ms": 4.582,
       "nodes": 1,
       "calls": 831
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 4.316,
     "pct": 0.153,
     "by_device": {
      "CPU": {
       "time_ms": 4.316,
       "nodes": 1,
       "calls": 78
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 4.049,
     "pct": 0.143,
     "by_device": {
      "CPU": {
       "time_ms": 4.049,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 3.891,
     "pct": 0.138,
     "by_device": {
      "CPU": {
       "time_ms": 3.891,
       "nodes": 1,
       "calls": 2394
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 3.8,
     "pct": 0.134,
     "by_device": {
      "CPU": {
       "time_ms": 3.8,
       "nodes": 1,
       "calls": 311
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 3.786,
     "pct": 0.134,
     "by_device": {
      "CPU": {
       "time_ms": 3.786,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 3.562,
     "pct": 0.126,
     "by_device": {
      "CPU": {
       "time_ms": 3.562,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.321,
     "pct": 0.117,
     "by_device": {
      "CPU": {
       "time_ms": 3.321,
       "nodes": 1,
       "calls": 550
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 2.916,
     "pct": 0.103,
     "by_device": {
      "CPU": {
       "time_ms": 2.916,
       "nodes": 1,
       "calls": 981
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.5,
     "pct": 0.088,
     "by_device": {
      "CPU": {
       "time_ms": 2.5,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 2.056,
     "pct": 0.073,
     "by_device": {
      "CPU": {
       "time_ms": 2.056,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.524,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 1.524,
       "nodes": 1,
       "calls": 477
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.448,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 1.448,
       "nodes": 1,
       "calls": 362
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.434,
     "pct": 0.051,
     "by_device": {
      "CPU": {
       "time_ms": 1.434,
       "nodes": 1,
       "calls": 1283
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.256,
     "pct": 0.044,
     "by_device": {
      "CPU": {
       "time_ms": 1.256,
       "nodes": 1,
       "calls": 346
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 1.232,
     "pct": 0.044,
     "by_device": {
      "CPU": {
       "time_ms": 1.232,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.148,
     "pct": 0.041,
     "by_device": {
      "CPU": {
       "time_ms": 1.148,
       "nodes": 1,
       "calls": 253
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.391,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.391,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.38,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.38,
       "nodes": 1,
       "calls": 78
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.376,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.376,
       "nodes": 1,
       "calls": 183
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.326,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.326,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.301,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.301,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.231,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.231,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.22,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.22,
       "nodes": 1,
       "calls": 37
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.17,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.17,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.089,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.089,
       "nodes": 2,
       "calls": 72
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.084,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.084,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.058,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.058,
       "nodes": 1,
       "calls": 74
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.033,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.033,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.026,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.022,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.021,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.021,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.015,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.015,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.015,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.015,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 36
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.008,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_falcon3-1b-base_b16s8o2_pim.simplified.json",
   "source_profile": "torch_falcon3-1b-base_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:00:43",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "gemma-2-2b",
   "hf_model_id": "google/gemma-2-2b",
   "hf_model_url": "https://huggingface.co/google/gemma-2-2b",
   "param_count": 2614341888
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 4.933,
   "TTFT": 3.982,
   "TBT_avg": 0.951,
   "Throughput": 6.487
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 4933.102,
     "ttft": 3982.417,
     "tbt_avg": 950.686,
     "throughput": 6.487,
     "steps": [
      3982.417,
      950.686
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 4933.102,
     "min": 4933.102,
     "max": 4933.102,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 3982.417,
     "min": 3982.417,
     "max": 3982.417,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 950.686,
     "min": 950.686,
     "max": 950.686,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 6.487,
     "min": 6.487,
     "max": 6.487,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 950.686,
     "min": 950.686,
     "max": 950.686,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 5180.803,
   "total_nodes": 56,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 4698.262,
     "pct": 90.686,
     "nodes": 2,
     "calls": 832
    },
    {
     "device": "CPU",
     "time_ms": 482.541,
     "pct": 9.314,
     "nodes": 54,
     "calls": 23645
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 4260.954,
     "pct": 82.245,
     "by_device": {
      "PIM": {
       "time_ms": 4260.954,
       "nodes": 1,
       "calls": 364
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 437.308,
     "pct": 8.441,
     "by_device": {
      "PIM": {
       "time_ms": 437.308,
       "nodes": 1,
       "calls": 468
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 62.339,
     "pct": 1.203,
     "by_device": {
      "CPU": {
       "time_ms": 62.339,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 62.205,
     "pct": 1.201,
     "by_device": {
      "CPU": {
       "time_ms": 62.205,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 62.086,
     "pct": 1.198,
     "by_device": {
      "CPU": {
       "time_ms": 62.086,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "to",
     "total_ms": 32.963,
     "pct": 0.636,
     "by_device": {
      "CPU": {
       "time_ms": 32.963,
       "nodes": 1,
       "calls": 1931
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 31.86,
     "pct": 0.615,
     "by_device": {
      "CPU": {
       "time_ms": 31.86,
       "nodes": 1,
       "calls": 728
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 30.947,
     "pct": 0.597,
     "by_device": {
      "CPU": {
       "time_ms": 30.947,
       "nodes": 1,
       "calls": 1282
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 27.125,
     "pct": 0.524,
     "by_device": {
      "CPU": {
       "time_ms": 27.125,
       "nodes": 1,
       "calls": 728
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 26.115,
     "pct": 0.504,
     "by_device": {
      "CPU": {
       "time_ms": 26.115,
       "nodes": 1,
       "calls": 890
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 24.774,
     "pct": 0.478,
     "by_device": {
      "CPU": {
       "time_ms": 24.774,
       "nodes": 1,
       "calls": 688
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 24.364,
     "pct": 0.47,
     "by_device": {
      "CPU": {
       "time_ms": 24.364,
       "nodes": 1,
       "calls": 1282
      }
     }
    },
    {
     "op": "gelu",
     "total_ms": 14.689,
     "pct": 0.284,
     "by_device": {
      "CPU": {
       "time_ms": 14.689,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 8.656,
     "pct": 0.167,
     "by_device": {
      "CPU": {
       "time_ms": 8.656,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "add",
     "total_ms": 8.015,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 8.015,
       "nodes": 1,
       "calls": 628
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 6.861,
     "pct": 0.132,
     "by_device": {
      "CPU": {
       "time_ms": 6.861,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 5.709,
     "pct": 0.11,
     "by_device": {
      "CPU": {
       "time_ms": 5.709,
       "nodes": 1,
       "calls": 1355
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 5.462,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 5.462,
       "nodes": 1,
       "calls": 420
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 5.009,
     "pct": 0.097,
     "by_device": {
      "CPU": {
       "time_ms": 5.009,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 4.931,
     "pct": 0.095,
     "by_device": {
      "CPU": {
       "time_ms": 4.931,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "type_as",
     "total_ms": 4.786,
     "pct": 0.092,
     "by_device": {
      "CPU": {
       "time_ms": 4.786,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 4.628,
     "pct": 0.089,
     "by_device": {
      "CPU": {
       "time_ms": 4.628,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 4.612,
     "pct": 0.089,
     "by_device": {
      "CPU": {
       "time_ms": 4.612,
       "nodes": 1,
       "calls": 4333
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 4.342,
     "pct": 0.084,
     "by_device": {
      "CPU": {
       "time_ms": 4.342,
       "nodes": 1,
       "calls": 218
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.917,
     "pct": 0.076,
     "by_device": {
      "CPU": {
       "time_ms": 3.917,
       "nodes": 1,
       "calls": 788
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 2.916,
     "pct": 0.056,
     "by_device": {
      "CPU": {
       "time_ms": 2.916,
       "nodes": 1,
       "calls": 1302
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 1.956,
     "pct": 0.038,
     "by_device": {
      "CPU": {
       "time_ms": 1.956,
       "nodes": 1,
       "calls": 104
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.891,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 1.891,
       "nodes": 1,
       "calls": 1334
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.699,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 1.699,
       "nodes": 1,
       "calls": 524
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.378,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 1.378,
       "nodes": 1,
       "calls": 365
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.267,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 1.267,
       "nodes": 1,
       "calls": 579
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.246,
     "pct": 0.024,
     "by_device": {
      "CPU": {
       "time_ms": 1.246,
       "nodes": 1,
       "calls": 2003
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.636,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.636,
       "nodes": 1,
       "calls": 367
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.558,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.558,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.406,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.406,
       "nodes": 1,
       "calls": 110
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 0.402,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.402,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.377,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.377,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.339,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.339,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.258,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.258,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "div",
     "total_ms": 0.226,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.226,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.13,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.13,
       "nodes": 1,
       "calls": 210
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.121,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.121,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.08,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.08,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.068,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.068,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.046,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.046,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.031,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.031,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.022,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.019,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.018,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.017,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.012,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.004,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_gemma-2-2b_b16s8o2_pim.simplified.json",
   "source_profile": "torch_gemma-2-2b_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:59:41",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "gemma-3-1b-it",
   "hf_model_id": "google/gemma-3-1b-it",
   "hf_model_url": "https://huggingface.co/google/gemma-3-1b-it",
   "param_count": 999885952
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.34,
   "TTFT": 1.762,
   "TBT_avg": 0.578,
   "Throughput": 13.675
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2339.972,
     "ttft": 1762.001,
     "tbt_avg": 577.972,
     "throughput": 13.675,
     "steps": [
      1762.001,
      577.972
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2339.972,
     "min": 2339.972,
     "max": 2339.972,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 1762.001,
     "min": 1762.001,
     "max": 1762.001,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 577.972,
     "min": 577.972,
     "max": 577.972,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 13.675,
     "min": 13.675,
     "max": 13.675,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 577.972,
     "min": 577.972,
     "max": 577.972,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2401.997,
   "total_nodes": 54,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 2149.976,
     "pct": 89.508,
     "nodes": 2,
     "calls": 832
    },
    {
     "device": "CPU",
     "time_ms": 252.021,
     "pct": 10.492,
     "nodes": 52,
     "calls": 28391
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 1785.122,
     "pct": 74.318,
     "by_device": {
      "PIM": {
       "time_ms": 1785.122,
       "nodes": 1,
       "calls": 364
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 364.854,
     "pct": 15.19,
     "by_device": {
      "PIM": {
       "time_ms": 364.854,
       "nodes": 1,
       "calls": 468
      }
     }
    },
    {
     "op": "to",
     "total_ms": 27.584,
     "pct": 1.148,
     "by_device": {
      "CPU": {
       "time_ms": 27.584,
       "nodes": 1,
       "calls": 2682
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 24.896,
     "pct": 1.036,
     "by_device": {
      "CPU": {
       "time_ms": 24.896,
       "nodes": 1,
       "calls": 1912
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 21.578,
     "pct": 0.898,
     "by_device": {
      "CPU": {
       "time_ms": 21.578,
       "nodes": 1,
       "calls": 898
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 15.395,
     "pct": 0.641,
     "by_device": {
      "CPU": {
       "time_ms": 15.395,
       "nodes": 1,
       "calls": 1912
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 14.832,
     "pct": 0.617,
     "by_device": {
      "CPU": {
       "time_ms": 14.832,
       "nodes": 1,
       "calls": 728
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 10.925,
     "pct": 0.455,
     "by_device": {
      "CPU": {
       "time_ms": 10.925,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 10.294,
     "pct": 0.429,
     "by_device": {
      "CPU": {
       "time_ms": 10.294,
       "nodes": 1,
       "calls": 728
      }
     }
    },
    {
     "op": "add",
     "total_ms": 8.898,
     "pct": 0.37,
     "by_device": {
      "CPU": {
       "time_ms": 8.898,
       "nodes": 1,
       "calls": 836
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 8.826,
     "pct": 0.367,
     "by_device": {
      "CPU": {
       "time_ms": 8.826,
       "nodes": 1,
       "calls": 760
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 8.694,
     "pct": 0.362,
     "by_device": {
      "CPU": {
       "time_ms": 8.694,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 8.411,
     "pct": 0.35,
     "by_device": {
      "CPU": {
       "time_ms": 8.411,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "gelu",
     "total_ms": 7.826,
     "pct": 0.326,
     "by_device": {
      "CPU": {
       "time_ms": 7.826,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 7.444,
     "pct": 0.31,
     "by_device": {
      "CPU": {
       "time_ms": 7.444,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 7.415,
     "pct": 0.309,
     "by_device": {
      "CPU": {
       "time_ms": 7.415,
       "nodes": 1,
       "calls": 628
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 7.241,
     "pct": 0.301,
     "by_device": {
      "CPU": {
       "time_ms": 7.241,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 7.166,
     "pct": 0.298,
     "by_device": {
      "CPU": {
       "time_ms": 7.166,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 7.13,
     "pct": 0.297,
     "by_device": {
      "CPU": {
       "time_ms": 7.13,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "type_as",
     "total_ms": 6.416,
     "pct": 0.267,
     "by_device": {
      "CPU": {
       "time_ms": 6.416,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 5.612,
     "pct": 0.234,
     "by_device": {
      "CPU": {
       "time_ms": 5.612,
       "nodes": 1,
       "calls": 212
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 5.612,
     "pct": 0.234,
     "by_device": {
      "CPU": {
       "time_ms": 5.612,
       "nodes": 1,
       "calls": 1392
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 5.108,
     "pct": 0.213,
     "by_device": {
      "CPU": {
       "time_ms": 5.108,
       "nodes": 1,
       "calls": 322
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 5.068,
     "pct": 0.211,
     "by_device": {
      "CPU": {
       "time_ms": 5.068,
       "nodes": 1,
       "calls": 5195
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.799,
     "pct": 0.158,
     "by_device": {
      "CPU": {
       "time_ms": 3.799,
       "nodes": 1,
       "calls": 792
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 3.228,
     "pct": 0.134,
     "by_device": {
      "CPU": {
       "time_ms": 3.228,
       "nodes": 1,
       "calls": 1302
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 2.919,
     "pct": 0.122,
     "by_device": {
      "CPU": {
       "time_ms": 2.919,
       "nodes": 1,
       "calls": 1964
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.574,
     "pct": 0.066,
     "by_device": {
      "CPU": {
       "time_ms": 1.574,
       "nodes": 1,
       "calls": 526
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.369,
     "pct": 0.057,
     "by_device": {
      "CPU": {
       "time_ms": 1.369,
       "nodes": 1,
       "calls": 2054
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.196,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 1.196,
       "nodes": 1,
       "calls": 581
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.163,
     "pct": 0.048,
     "by_device": {
      "CPU": {
       "time_ms": 1.163,
       "nodes": 1,
       "calls": 366
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 1.04,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 1.04,
       "nodes": 1,
       "calls": 104
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.747,
     "pct": 0.031,
     "by_device": {
      "CPU": {
       "time_ms": 0.747,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.666,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.666,
       "nodes": 1,
       "calls": 369
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.421,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.421,
       "nodes": 1,
       "calls": 116
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.32,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.32,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.287,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.287,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.219,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.219,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.157,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.157,
       "nodes": 1,
       "calls": 314
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.133,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.133,
       "nodes": 2,
       "calls": 104
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.101,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.101,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.087,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.087,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.076,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.076,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.049,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.049,
       "nodes": 1,
       "calls": 12
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.017,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.017,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.015,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.015,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.015,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.015,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.013,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.013,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.009,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.003,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 12
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_gemma-3-1b-it_b16s8o2_pim.simplified.json",
   "source_profile": "torch_gemma-3-1b-it_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:24:23",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "gpt2-large",
   "hf_model_id": "openai-community/gpt2-large",
   "hf_model_url": "https://huggingface.co/openai-community/gpt2-large",
   "param_count": 774030080
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.195,
   "TTFT": 1.624,
   "TBT_avg": 0.571,
   "Throughput": 14.581
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2194.675,
     "ttft": 1624.032,
     "tbt_avg": 570.642,
     "throughput": 14.581,
     "steps": [
      1624.032,
      570.642
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2194.675,
     "min": 2194.675,
     "max": 2194.675,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 1624.032,
     "min": 1624.032,
     "max": 1624.032,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 570.642,
     "min": 570.642,
     "max": 570.642,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 14.581,
     "min": 14.581,
     "max": 14.581,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 570.642,
     "min": 570.642,
     "max": 570.642,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2192.624,
   "total_nodes": 48,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 1980.657,
     "pct": 90.333,
     "nodes": 2,
     "calls": 576
    },
    {
     "device": "CPU",
     "time_ms": 211.967,
     "pct": 9.667,
     "nodes": 46,
     "calls": 16223
    }
   ],
   "ops": [
    {
     "op": "pim_addmm",
     "total_ms": 1678.773,
     "pct": 76.565,
     "by_device": {
      "PIM": {
       "time_ms": 1678.773,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 301.883,
     "pct": 13.768,
     "by_device": {
      "PIM": {
       "time_ms": 301.883,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 27.413,
     "pct": 1.25,
     "by_device": {
      "CPU": {
       "time_ms": 27.413,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "add",
     "total_ms": 23.031,
     "pct": 1.05,
     "by_device": {
      "CPU": {
       "time_ms": 23.031,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 21.276,
     "pct": 0.97,
     "by_device": {
      "CPU": {
       "time_ms": 21.276,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 18.959,
     "pct": 0.865,
     "by_device": {
      "CPU": {
       "time_ms": 18.959,
       "nodes": 1,
       "calls": 576
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 14.423,
     "pct": 0.658,
     "by_device": {
      "CPU": {
       "time_ms": 14.423,
       "nodes": 1,
       "calls": 576
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 12.682,
     "pct": 0.578,
     "by_device": {
      "CPU": {
       "time_ms": 12.682,
       "nodes": 1,
       "calls": 398
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 8.506,
     "pct": 0.388,
     "by_device": {
      "CPU": {
       "time_ms": 8.506,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 8.471,
     "pct": 0.386,
     "by_device": {
      "CPU": {
       "time_ms": 8.471,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 8.41,
     "pct": 0.384,
     "by_device": {
      "CPU": {
       "time_ms": 8.41,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 7.802,
     "pct": 0.356,
     "by_device": {
      "CPU": {
       "time_ms": 7.802,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 7.383,
     "pct": 0.337,
     "by_device": {
      "CPU": {
       "time_ms": 7.383,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 6.355,
     "pct": 0.29,
     "by_device": {
      "CPU": {
       "time_ms": 6.355,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 5.778,
     "pct": 0.264,
     "by_device": {
      "CPU": {
       "time_ms": 5.778,
       "nodes": 1,
       "calls": 146
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 4.984,
     "pct": 0.227,
     "by_device": {
      "CPU": {
       "time_ms": 4.984,
       "nodes": 1,
       "calls": 1010
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 4.398,
     "pct": 0.201,
     "by_device": {
      "CPU": {
       "time_ms": 4.398,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 4.139,
     "pct": 0.189,
     "by_device": {
      "CPU": {
       "time_ms": 4.139,
       "nodes": 1,
       "calls": 3320
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 4.106,
     "pct": 0.187,
     "by_device": {
      "CPU": {
       "time_ms": 4.106,
       "nodes": 1,
       "calls": 1594
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 3.837,
     "pct": 0.175,
     "by_device": {
      "CPU": {
       "time_ms": 3.837,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.153,
     "pct": 0.144,
     "by_device": {
      "CPU": {
       "time_ms": 3.153,
       "nodes": 1,
       "calls": 654
      }
     }
    },
    {
     "op": "to",
     "total_ms": 2.764,
     "pct": 0.126,
     "by_device": {
      "CPU": {
       "time_ms": 2.764,
       "nodes": 1,
       "calls": 726
      }
     }
    },
    {
     "op": "view",
     "total_ms": 2.184,
     "pct": 0.1,
     "by_device": {
      "CPU": {
       "time_ms": 2.184,
       "nodes": 1,
       "calls": 1459
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 2.141,
     "pct": 0.098,
     "by_device": {
      "CPU": {
       "time_ms": 2.141,
       "nodes": 1,
       "calls": 292
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 2.046,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 2.046,
       "nodes": 1,
       "calls": 722
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.775,
     "pct": 0.081,
     "by_device": {
      "CPU": {
       "time_ms": 1.775,
       "nodes": 1,
       "calls": 432
      }
     }
    },
    {
     "op": "split",
     "total_ms": 1.339,
     "pct": 0.061,
     "by_device": {
      "CPU": {
       "time_ms": 1.339,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.216,
     "pct": 0.055,
     "by_device": {
      "CPU": {
       "time_ms": 1.216,
       "nodes": 1,
       "calls": 1742
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.645,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.645,
       "nodes": 1,
       "calls": 364
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 0.588,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.588,
       "nodes": 1,
       "calls": 292
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.585,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.585,
       "nodes": 1,
       "calls": 289
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.467,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.467,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.392,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.392,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.339,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.339,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.146,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.146,
       "nodes": 2,
       "calls": 144
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.104,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.104,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.037,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.037,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.022,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.019,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.012,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 72
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.009,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 1
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_gpt2-large_b16s8o2_pim.simplified.json",
   "source_profile": "torch_gpt2-large_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:59:23",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "Llama-3.2-1B",
   "hf_model_id": "meta-llama/Llama-3.2-1B",
   "hf_model_url": "https://huggingface.co/meta-llama/Llama-3.2-1B",
   "param_count": 1235814400
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.233,
   "TTFT": 1.784,
   "TBT_avg": 0.449,
   "Throughput": 14.333
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2232.587,
     "ttft": 1784.047,
     "tbt_avg": 448.54,
     "throughput": 14.333,
     "steps": [
      1784.047,
      448.54
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2232.587,
     "min": 2232.587,
     "max": 2232.587,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 1784.047,
     "min": 1784.047,
     "max": 1784.047,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 448.54,
     "min": 448.54,
     "max": 448.54,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 14.333,
     "min": 14.333,
     "max": 14.333,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 448.54,
     "min": 448.54,
     "max": 448.54,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2344.203,
   "total_nodes": 54,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 2137.384,
     "pct": 91.177,
     "nodes": 3,
     "calls": 576
    },
    {
     "device": "CPU",
     "time_ms": 206.819,
     "pct": 8.823,
     "nodes": 51,
     "calls": 10988
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 1881.817,
     "pct": 80.275,
     "by_device": {
      "PIM": {
       "time_ms": 1881.817,
       "nodes": 1,
       "calls": 224
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 221.872,
     "pct": 9.465,
     "by_device": {
      "PIM": {
       "time_ms": 221.872,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 33.696,
     "pct": 1.437,
     "by_device": {
      "PIM": {
       "time_ms": 33.696,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 29.6,
     "pct": 1.263,
     "by_device": {
      "CPU": {
       "time_ms": 29.6,
       "nodes": 1,
       "calls": 514
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 25.966,
     "pct": 1.108,
     "by_device": {
      "CPU": {
       "time_ms": 25.966,
       "nodes": 1,
       "calls": 514
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 25.201,
     "pct": 1.075,
     "by_device": {
      "CPU": {
       "time_ms": 25.201,
       "nodes": 1,
       "calls": 550
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 13.045,
     "pct": 0.556,
     "by_device": {
      "CPU": {
       "time_ms": 13.045,
       "nodes": 1,
       "calls": 296
      }
     }
    },
    {
     "op": "to",
     "total_ms": 8.675,
     "pct": 0.37,
     "by_device": {
      "CPU": {
       "time_ms": 8.675,
       "nodes": 1,
       "calls": 610
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 8.576,
     "pct": 0.366,
     "by_device": {
      "CPU": {
       "time_ms": 8.576,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 8.437,
     "pct": 0.36,
     "by_device": {
      "CPU": {
       "time_ms": 8.437,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 8.36,
     "pct": 0.357,
     "by_device": {
      "CPU": {
       "time_ms": 8.36,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 8.146,
     "pct": 0.347,
     "by_device": {
      "CPU": {
       "time_ms": 8.146,
       "nodes": 1,
       "calls": 278
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 7.37,
     "pct": 0.314,
     "by_device": {
      "CPU": {
       "time_ms": 7.37,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 6.789,
     "pct": 0.29,
     "by_device": {
      "CPU": {
       "time_ms": 6.789,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 6.067,
     "pct": 0.259,
     "by_device": {
      "CPU": {
       "time_ms": 6.067,
       "nodes": 1,
       "calls": 278
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 5.365,
     "pct": 0.229,
     "by_device": {
      "CPU": {
       "time_ms": 5.365,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "add",
     "total_ms": 4.945,
     "pct": 0.211,
     "by_device": {
      "CPU": {
       "time_ms": 4.945,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 4.537,
     "pct": 0.194,
     "by_device": {
      "CPU": {
       "time_ms": 4.537,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 4.343,
     "pct": 0.185,
     "by_device": {
      "CPU": {
       "time_ms": 4.343,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 4.162,
     "pct": 0.178,
     "by_device": {
      "CPU": {
       "time_ms": 4.162,
       "nodes": 1,
       "calls": 739
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 3.912,
     "pct": 0.167,
     "by_device": {
      "CPU": {
       "time_ms": 3.912,
       "nodes": 1,
       "calls": 2125
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 3.61,
     "pct": 0.154,
     "by_device": {
      "CPU": {
       "time_ms": 3.61,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.043,
     "pct": 0.13,
     "by_device": {
      "CPU": {
       "time_ms": 3.043,
       "nodes": 1,
       "calls": 490
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 2.391,
     "pct": 0.102,
     "by_device": {
      "CPU": {
       "time_ms": 2.391,
       "nodes": 1,
       "calls": 868
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 2.272,
     "pct": 0.097,
     "by_device": {
      "CPU": {
       "time_ms": 2.272,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 2.233,
     "pct": 0.095,
     "by_device": {
      "CPU": {
       "time_ms": 2.233,
       "nodes": 1,
       "calls": 132
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 1.549,
     "pct": 0.066,
     "by_device": {
      "CPU": {
       "time_ms": 1.549,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.387,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 1.387,
       "nodes": 1,
       "calls": 1145
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.378,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 1.378,
       "nodes": 1,
       "calls": 425
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.27,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 1.27,
       "nodes": 1,
       "calls": 324
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.025,
     "pct": 0.044,
     "by_device": {
      "CPU": {
       "time_ms": 1.025,
       "nodes": 1,
       "calls": 225
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.847,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 0.847,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.357,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.357,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.347,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.347,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.34,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.34,
       "nodes": 1,
       "calls": 163
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.321,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.321,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.318,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.318,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.217,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.217,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.07,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.07,
       "nodes": 2,
       "calls": 64
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.066,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.066,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.056,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.056,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.048,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.048,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.045,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.045,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.043,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.043,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.018,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.008,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.007,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.007,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.003,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_llama-3.2-1b_b16s8o2_pim.simplified.json",
   "source_profile": "torch_llama-3.2-1b_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:02:25",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "OLMo-2-0425-1B",
   "hf_model_id": "allenai/OLMo-2-0425-1B",
   "hf_model_url": "https://huggingface.co/allenai/OLMo-2-0425-1B",
   "param_count": 1480000000
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.492,
   "TTFT": 1.948,
   "TBT_avg": 0.544,
   "Throughput": 12.841
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2491.935,
     "ttft": 1947.92,
     "tbt_avg": 544.014,
     "throughput": 12.841,
     "steps": [
      1947.92,
      544.014
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2491.935,
     "min": 2491.935,
     "max": 2491.935,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 1947.92,
     "min": 1947.92,
     "max": 1947.92,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 544.014,
     "min": 544.014,
     "max": 544.014,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 12.841,
     "min": 12.841,
     "max": 12.841,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 544.014,
     "min": 544.014,
     "max": 544.014,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2569.891,
   "total_nodes": 52,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 2369.628,
     "pct": 92.207,
     "nodes": 2,
     "calls": 578
    },
    {
     "device": "CPU",
     "time_ms": 200.263,
     "pct": 7.793,
     "nodes": 50,
     "calls": 14469
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 2133.486,
     "pct": 83.019,
     "by_device": {
      "PIM": {
       "time_ms": 2133.486,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 236.141,
     "pct": 9.189,
     "by_device": {
      "PIM": {
       "time_ms": 236.141,
       "nodes": 1,
       "calls": 352
      }
     }
    },
    {
     "op": "to",
     "total_ms": 27.772,
     "pct": 1.081,
     "by_device": {
      "CPU": {
       "time_ms": 27.772,
       "nodes": 1,
       "calls": 1248
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 26.485,
     "pct": 1.031,
     "by_device": {
      "CPU": {
       "time_ms": 26.485,
       "nodes": 1,
       "calls": 852
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 21.716,
     "pct": 0.845,
     "by_device": {
      "CPU": {
       "time_ms": 21.716,
       "nodes": 1,
       "calls": 424
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 20.838,
     "pct": 0.811,
     "by_device": {
      "CPU": {
       "time_ms": 20.838,
       "nodes": 1,
       "calls": 853
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 14.173,
     "pct": 0.551,
     "by_device": {
      "CPU": {
       "time_ms": 14.173,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 10.605,
     "pct": 0.413,
     "by_device": {
      "CPU": {
       "time_ms": 10.605,
       "nodes": 1,
       "calls": 134
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 8.698,
     "pct": 0.338,
     "by_device": {
      "CPU": {
       "time_ms": 8.698,
       "nodes": 1,
       "calls": 452
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 8.583,
     "pct": 0.334,
     "by_device": {
      "CPU": {
       "time_ms": 8.583,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 5.897,
     "pct": 0.229,
     "by_device": {
      "CPU": {
       "time_ms": 5.897,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 5.781,
     "pct": 0.225,
     "by_device": {
      "CPU": {
       "time_ms": 5.781,
       "nodes": 1,
       "calls": 452
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 5.276,
     "pct": 0.205,
     "by_device": {
      "CPU": {
       "time_ms": 5.276,
       "nodes": 1,
       "calls": 584
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 4.751,
     "pct": 0.185,
     "by_device": {
      "CPU": {
       "time_ms": 4.751,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "add",
     "total_ms": 4.733,
     "pct": 0.184,
     "by_device": {
      "CPU": {
       "time_ms": 4.733,
       "nodes": 1,
       "calls": 258
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 4.562,
     "pct": 0.178,
     "by_device": {
      "CPU": {
       "time_ms": 4.562,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 4.535,
     "pct": 0.176,
     "by_device": {
      "CPU": {
       "time_ms": 4.535,
       "nodes": 1,
       "calls": 260
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 3.976,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 3.976,
       "nodes": 1,
       "calls": 739
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 3.48,
     "pct": 0.135,
     "by_device": {
      "CPU": {
       "time_ms": 3.48,
       "nodes": 1,
       "calls": 2711
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 3.478,
     "pct": 0.135,
     "by_device": {
      "CPU": {
       "time_ms": 3.478,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 2.953,
     "pct": 0.115,
     "by_device": {
      "CPU": {
       "time_ms": 2.953,
       "nodes": 1,
       "calls": 488
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 2.112,
     "pct": 0.082,
     "by_device": {
      "CPU": {
       "time_ms": 2.112,
       "nodes": 1,
       "calls": 884
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 1.941,
     "pct": 0.076,
     "by_device": {
      "CPU": {
       "time_ms": 1.941,
       "nodes": 1,
       "calls": 807
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.222,
     "pct": 0.048,
     "by_device": {
      "CPU": {
       "time_ms": 1.222,
       "nodes": 1,
       "calls": 322
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.071,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 1.071,
       "nodes": 1,
       "calls": 225
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.062,
     "pct": 0.041,
     "by_device": {
      "CPU": {
       "time_ms": 1.062,
       "nodes": 1,
       "calls": 1143
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.04,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 1.04,
       "nodes": 1,
       "calls": 423
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.824,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.824,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.472,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.472,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.46,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.46,
       "nodes": 1,
       "calls": 163
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.338,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.338,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.311,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.311,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.297,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.297,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.202,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.202,
       "nodes": 1,
       "calls": 33
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.144,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.144,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.083,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.083,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.071,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.071,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.063,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.063,
       "nodes": 2,
       "calls": 64
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.062,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.062,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.038,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.036,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.034,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.034,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.033,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.033,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.014,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_olmo-2-0425-1b_b16s8o2_pim.simplified.json",
   "source_profile": "torch_olmo-2-0425-1b_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:01:16",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "opt-1.3b",
   "hf_model_id": "facebook/opt-1.3b",
   "hf_model_url": "https://huggingface.co/facebook/opt-1.3b",
   "param_count": 1315758080
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.864,
   "TTFT": 2.246,
   "TBT_avg": 0.618,
   "Throughput": 11.173
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2864.086,
     "ttft": 2245.638,
     "tbt_avg": 618.449,
     "throughput": 11.173,
     "steps": [
      2245.638,
      618.449
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2864.086,
     "min": 2864.086,
     "max": 2864.086,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 2245.638,
     "min": 2245.638,
     "max": 2245.638,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 618.449,
     "min": 618.449,
     "max": 618.449,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 11.173,
     "min": 11.173,
     "max": 11.173,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 618.449,
     "min": 618.449,
     "max": 618.449,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 3071.184,
   "total_nodes": 48,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 2760.016,
     "pct": 89.868,
     "nodes": 3,
     "calls": 768
    },
    {
     "device": "CPU",
     "time_ms": 311.168,
     "pct": 10.132,
     "nodes": 45,
     "calls": 12736
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 2346.626,
     "pct": 76.408,
     "by_device": {
      "PIM": {
       "time_ms": 2346.626,
       "nodes": 1,
       "calls": 288
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 280.03,
     "pct": 9.118,
     "by_device": {
      "PIM": {
       "time_ms": 280.03,
       "nodes": 1,
       "calls": 384
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 133.36,
     "pct": 4.342,
     "by_device": {
      "PIM": {
       "time_ms": 133.36,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 38.271,
     "pct": 1.246,
     "by_device": {
      "CPU": {
       "time_ms": 38.271,
       "nodes": 1,
       "calls": 768
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 33.233,
     "pct": 1.082,
     "by_device": {
      "CPU": {
       "time_ms": 33.233,
       "nodes": 1,
       "calls": 768
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 31.99,
     "pct": 1.042,
     "by_device": {
      "CPU": {
       "time_ms": 31.99,
       "nodes": 1,
       "calls": 770
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 17.971,
     "pct": 0.585,
     "by_device": {
      "CPU": {
       "time_ms": 17.971,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 17.94,
     "pct": 0.584,
     "by_device": {
      "CPU": {
       "time_ms": 17.94,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 17.88,
     "pct": 0.582,
     "by_device": {
      "CPU": {
       "time_ms": 17.88,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "relu",
     "total_ms": 15.792,
     "pct": 0.514,
     "by_device": {
      "CPU": {
       "time_ms": 15.792,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "clamp_min",
     "total_ms": 15.617,
     "pct": 0.509,
     "by_device": {
      "CPU": {
       "time_ms": 15.617,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 15.35,
     "pct": 0.5,
     "by_device": {
      "CPU": {
       "time_ms": 15.35,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 13.263,
     "pct": 0.432,
     "by_device": {
      "CPU": {
       "time_ms": 13.263,
       "nodes": 1,
       "calls": 190
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 13.141,
     "pct": 0.428,
     "by_device": {
      "CPU": {
       "time_ms": 13.141,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 12.909,
     "pct": 0.42,
     "by_device": {
      "CPU": {
       "time_ms": 12.909,
       "nodes": 1,
       "calls": 190
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 12.776,
     "pct": 0.416,
     "by_device": {
      "CPU": {
       "time_ms": 12.776,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 11.498,
     "pct": 0.374,
     "by_device": {
      "CPU": {
       "time_ms": 11.498,
       "nodes": 1,
       "calls": 242
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 7.131,
     "pct": 0.232,
     "by_device": {
      "CPU": {
       "time_ms": 7.131,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 7.089,
     "pct": 0.231,
     "by_device": {
      "CPU": {
       "time_ms": 7.089,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 6.789,
     "pct": 0.221,
     "by_device": {
      "CPU": {
       "time_ms": 6.789,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 3.997,
     "pct": 0.13,
     "by_device": {
      "CPU": {
       "time_ms": 3.997,
       "nodes": 1,
       "calls": 1736
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.747,
     "pct": 0.122,
     "by_device": {
      "CPU": {
       "time_ms": 3.747,
       "nodes": 1,
       "calls": 774
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 3.225,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 3.225,
       "nodes": 1,
       "calls": 674
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 2.645,
     "pct": 0.086,
     "by_device": {
      "CPU": {
       "time_ms": 2.645,
       "nodes": 1,
       "calls": 2066
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.536,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 1.536,
       "nodes": 1,
       "calls": 482
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.276,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 1.276,
       "nodes": 1,
       "calls": 785
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 1.051,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 1.051,
       "nodes": 1,
       "calls": 238
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.873,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.873,
       "nodes": 1,
       "calls": 1166
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.854,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.854,
       "nodes": 1,
       "calls": 486
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.676,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.676,
       "nodes": 1,
       "calls": 383
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.583,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.583,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.571,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.571,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.445,
     "pct": 0.014,
     "by_device": {
      "CPU": {
       "time_ms": 0.445,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.384,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.384,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.257,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.257,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.134,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.134,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.123,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.123,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "add",
     "total_ms": 0.061,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.061,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.024,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.016,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.009,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_opt-1.3b_b16s8o2_pim.simplified.json",
   "source_profile": "torch_opt-1.3b_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:00:56",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "opt-350m",
   "hf_model_id": "facebook/opt-350m",
   "hf_model_url": "https://huggingface.co/facebook/opt-350m",
   "param_count": 331196416
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.215,
   "TTFT": 0.823,
   "TBT_avg": 0.392,
   "Throughput": 26.338
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1214.985,
     "ttft": 822.542,
     "tbt_avg": 392.443,
     "throughput": 26.338,
     "steps": [
      822.542,
      392.443
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1214.985,
     "min": 1214.985,
     "max": 1214.985,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 822.542,
     "min": 822.542,
     "max": 822.542,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 392.443,
     "min": 392.443,
     "max": 392.443,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 26.338,
     "min": 26.338,
     "max": 26.338,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 392.443,
     "min": 392.443,
     "max": 392.443,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1274.018,
   "total_nodes": 48,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 1151.22,
     "pct": 90.361,
     "nodes": 3,
     "calls": 684
    },
    {
     "device": "CPU",
     "time_ms": 122.798,
     "pct": 9.639,
     "nodes": 45,
     "calls": 11871
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 802.637,
     "pct": 63.0,
     "by_device": {
      "PIM": {
       "time_ms": 802.637,
       "nodes": 1,
       "calls": 292
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 227.435,
     "pct": 17.852,
     "by_device": {
      "PIM": {
       "time_ms": 227.435,
       "nodes": 1,
       "calls": 294
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 121.148,
     "pct": 9.509,
     "by_device": {
      "PIM": {
       "time_ms": 121.148,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 18.259,
     "pct": 1.433,
     "by_device": {
      "CPU": {
       "time_ms": 18.259,
       "nodes": 1,
       "calls": 772
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 13.197,
     "pct": 1.036,
     "by_device": {
      "CPU": {
       "time_ms": 13.197,
       "nodes": 1,
       "calls": 772
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 11.68,
     "pct": 0.917,
     "by_device": {
      "CPU": {
       "time_ms": 11.68,
       "nodes": 1,
       "calls": 436
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 9.403,
     "pct": 0.738,
     "by_device": {
      "CPU": {
       "time_ms": 9.403,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 6.756,
     "pct": 0.53,
     "by_device": {
      "CPU": {
       "time_ms": 6.756,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 6.381,
     "pct": 0.501,
     "by_device": {
      "CPU": {
       "time_ms": 6.381,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "relu",
     "total_ms": 6.032,
     "pct": 0.474,
     "by_device": {
      "CPU": {
       "time_ms": 6.032,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "clamp_min",
     "total_ms": 5.86,
     "pct": 0.46,
     "by_device": {
      "CPU": {
       "time_ms": 5.86,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 4.583,
     "pct": 0.36,
     "by_device": {
      "CPU": {
       "time_ms": 4.583,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 4.303,
     "pct": 0.338,
     "by_device": {
      "CPU": {
       "time_ms": 4.303,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.751,
     "pct": 0.294,
     "by_device": {
      "CPU": {
       "time_ms": 3.751,
       "nodes": 1,
       "calls": 778
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 3.578,
     "pct": 0.281,
     "by_device": {
      "CPU": {
       "time_ms": 3.578,
       "nodes": 1,
       "calls": 686
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 3.535,
     "pct": 0.278,
     "by_device": {
      "CPU": {
       "time_ms": 3.535,
       "nodes": 1,
       "calls": 1646
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 3.117,
     "pct": 0.245,
     "by_device": {
      "CPU": {
       "time_ms": 3.117,
       "nodes": 1,
       "calls": 1973
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 2.619,
     "pct": 0.206,
     "by_device": {
      "CPU": {
       "time_ms": 2.619,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 2.431,
     "pct": 0.191,
     "by_device": {
      "CPU": {
       "time_ms": 2.431,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 1.908,
     "pct": 0.15,
     "by_device": {
      "CPU": {
       "time_ms": 1.908,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 1.783,
     "pct": 0.14,
     "by_device": {
      "CPU": {
       "time_ms": 1.783,
       "nodes": 1,
       "calls": 148
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.73,
     "pct": 0.136,
     "by_device": {
      "CPU": {
       "time_ms": 1.73,
       "nodes": 1,
       "calls": 482
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.644,
     "pct": 0.129,
     "by_device": {
      "CPU": {
       "time_ms": 1.644,
       "nodes": 1,
       "calls": 920
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 1.473,
     "pct": 0.116,
     "by_device": {
      "CPU": {
       "time_ms": 1.473,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 1.442,
     "pct": 0.113,
     "by_device": {
      "CPU": {
       "time_ms": 1.442,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 1.393,
     "pct": 0.109,
     "by_device": {
      "CPU": {
       "time_ms": 1.393,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.045,
     "pct": 0.082,
     "by_device": {
      "CPU": {
       "time_ms": 1.045,
       "nodes": 1,
       "calls": 1178
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.925,
     "pct": 0.073,
     "by_device": {
      "CPU": {
       "time_ms": 0.925,
       "nodes": 1,
       "calls": 484
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.731,
     "pct": 0.057,
     "by_device": {
      "CPU": {
       "time_ms": 0.731,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.688,
     "pct": 0.054,
     "by_device": {
      "CPU": {
       "time_ms": 0.688,
       "nodes": 1,
       "calls": 144
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.622,
     "pct": 0.049,
     "by_device": {
      "CPU": {
       "time_ms": 0.622,
       "nodes": 1,
       "calls": 52
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.538,
     "pct": 0.042,
     "by_device": {
      "CPU": {
       "time_ms": 0.538,
       "nodes": 1,
       "calls": 248
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.416,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.416,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.358,
     "pct": 0.028,
     "by_device": {
      "CPU": {
       "time_ms": 0.358,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.27,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.27,
       "nodes": 1,
       "calls": 100
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.121,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.121,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.112,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.112,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.031,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.031,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.018,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "add",
     "total_ms": 0.01,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.01,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.006,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_opt-350m_b16s8o2_pim.simplified.json",
   "source_profile": "torch_opt-350m_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:58:48",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "phi-1_5",
   "hf_model_id": "microsoft/phi-1_5",
   "hf_model_url": "https://huggingface.co/microsoft/phi-1_5",
   "param_count": 1418270720
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.913,
   "TTFT": 2.3,
   "TBT_avg": 0.613,
   "Throughput": 10.985
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2913.037,
     "ttft": 2299.976,
     "tbt_avg": 613.062,
     "throughput": 10.985,
     "steps": [
      2299.976,
      613.062
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2913.037,
     "min": 2913.037,
     "max": 2913.037,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 2299.976,
     "min": 2299.976,
     "max": 2299.976,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 613.062,
     "min": 613.062,
     "max": 613.062,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 10.985,
     "min": 10.985,
     "max": 10.985,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 613.062,
     "min": 613.062,
     "max": 613.062,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 3081.483,
   "total_nodes": 53,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 2749.425,
     "pct": 89.224,
     "nodes": 3,
     "calls": 722
    },
    {
     "device": "CPU",
     "time_ms": 332.058,
     "pct": 10.776,
     "nodes": 50,
     "calls": 15205
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 2443.039,
     "pct": 79.281,
     "by_device": {
      "PIM": {
       "time_ms": 2443.039,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 269.371,
     "pct": 8.742,
     "by_device": {
      "PIM": {
       "time_ms": 269.371,
       "nodes": 1,
       "calls": 336
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 51.569,
     "pct": 1.674,
     "by_device": {
      "CPU": {
       "time_ms": 51.569,
       "nodes": 1,
       "calls": 678
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 46.209,
     "pct": 1.5,
     "by_device": {
      "CPU": {
       "time_ms": 46.209,
       "nodes": 1,
       "calls": 678
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 44.956,
     "pct": 1.459,
     "by_device": {
      "CPU": {
       "time_ms": 44.956,
       "nodes": 1,
       "calls": 680
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 40.845,
     "pct": 1.325,
     "by_device": {
      "CPU": {
       "time_ms": 40.845,
       "nodes": 1,
       "calls": 388
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 37.014,
     "pct": 1.201,
     "by_device": {
      "PIM": {
       "time_ms": 37.014,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 23.669,
     "pct": 0.768,
     "by_device": {
      "CPU": {
       "time_ms": 23.669,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "add",
     "total_ms": 16.745,
     "pct": 0.543,
     "by_device": {
      "CPU": {
       "time_ms": 16.745,
       "nodes": 1,
       "calls": 192
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 11.938,
     "pct": 0.387,
     "by_device": {
      "CPU": {
       "time_ms": 11.938,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 11.486,
     "pct": 0.373,
     "by_device": {
      "CPU": {
       "time_ms": 11.486,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 11.275,
     "pct": 0.366,
     "by_device": {
      "CPU": {
       "time_ms": 11.275,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 11.042,
     "pct": 0.358,
     "by_device": {
      "CPU": {
       "time_ms": 11.042,
       "nodes": 1,
       "calls": 50
      }
     }
    },
    {
     "op": "tanh",
     "total_ms": 6.844,
     "pct": 0.222,
     "by_device": {
      "CPU": {
       "time_ms": 6.844,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 6.027,
     "pct": 0.196,
     "by_device": {
      "CPU": {
       "time_ms": 6.027,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 5.987,
     "pct": 0.194,
     "by_device": {
      "CPU": {
       "time_ms": 5.987,
       "nodes": 1,
       "calls": 73
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 5.811,
     "pct": 0.189,
     "by_device": {
      "CPU": {
       "time_ms": 5.811,
       "nodes": 1,
       "calls": 73
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 5.501,
     "pct": 0.179,
     "by_device": {
      "CPU": {
       "time_ms": 5.501,
       "nodes": 1,
       "calls": 279
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 5.023,
     "pct": 0.163,
     "by_device": {
      "CPU": {
       "time_ms": 5.023,
       "nodes": 1,
       "calls": 1347
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 4.136,
     "pct": 0.134,
     "by_device": {
      "CPU": {
       "time_ms": 4.136,
       "nodes": 1,
       "calls": 1335
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 3.71,
     "pct": 0.12,
     "by_device": {
      "CPU": {
       "time_ms": 3.71,
       "nodes": 1,
       "calls": 2914
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.09,
     "pct": 0.1,
     "by_device": {
      "CPU": {
       "time_ms": 3.09,
       "nodes": 1,
       "calls": 586
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 2.496,
     "pct": 0.081,
     "by_device": {
      "CPU": {
       "time_ms": 2.496,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "to",
     "total_ms": 2.167,
     "pct": 0.07,
     "by_device": {
      "CPU": {
       "time_ms": 2.167,
       "nodes": 1,
       "calls": 602
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.993,
     "pct": 0.065,
     "by_device": {
      "CPU": {
       "time_ms": 1.993,
       "nodes": 1,
       "calls": 529
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.814,
     "pct": 0.059,
     "by_device": {
      "CPU": {
       "time_ms": 1.814,
       "nodes": 1,
       "calls": 482
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 1.683,
     "pct": 0.055,
     "by_device": {
      "CPU": {
       "time_ms": 1.683,
       "nodes": 1,
       "calls": 206
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.433,
     "pct": 0.046,
     "by_device": {
      "CPU": {
       "time_ms": 1.433,
       "nodes": 1,
       "calls": 637
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.246,
     "pct": 0.04,
     "by_device": {
      "CPU": {
       "time_ms": 1.246,
       "nodes": 1,
       "calls": 1943
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.759,
     "pct": 0.025,
     "by_device": {
      "CPU": {
       "time_ms": 0.759,
       "nodes": 1,
       "calls": 121
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.504,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.504,
       "nodes": 1,
       "calls": 254
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.414,
     "pct": 0.013,
     "by_device": {
      "CPU": {
       "time_ms": 0.414,
       "nodes": 1,
       "calls": 195
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.38,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.38,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.342,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.342,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.306,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.306,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.166,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.166,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.122,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.122,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.121,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.121,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.049,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.049,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.036,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.036,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.035,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.035,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.03,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.023,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.018,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.014,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.011,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_phi-1_5_b16s8o2_pim.simplified.json",
   "source_profile": "torch_phi-1_5_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 17:01:31",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "pythia-1b",
   "hf_model_id": "EleutherAI/pythia-1b",
   "hf_model_url": "https://huggingface.co/EleutherAI/pythia-1b",
   "param_count": 1011781632
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.806,
   "TTFT": 1.465,
   "TBT_avg": 0.341,
   "Throughput": 17.717
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1806.168,
     "ttft": 1464.812,
     "tbt_avg": 341.356,
     "throughput": 17.717,
     "steps": [
      1464.812,
      341.356
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1806.168,
     "min": 1806.168,
     "max": 1806.168,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 1464.812,
     "min": 1464.812,
     "max": 1464.812,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 341.356,
     "min": 341.356,
     "max": 341.356,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 17.717,
     "min": 17.717,
     "max": 17.717,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 341.356,
     "min": 341.356,
     "max": 341.356,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1871.528,
   "total_nodes": 53,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 1746.785,
     "pct": 93.335,
     "nodes": 3,
     "calls": 320
    },
    {
     "device": "CPU",
     "time_ms": 124.742,
     "pct": 6.665,
     "nodes": 50,
     "calls": 8380
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 1590.481,
     "pct": 84.983,
     "by_device": {
      "PIM": {
       "time_ms": 1590.481,
       "nodes": 1,
       "calls": 130
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 131.421,
     "pct": 7.022,
     "by_device": {
      "PIM": {
       "time_ms": 131.421,
       "nodes": 1,
       "calls": 126
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 24.884,
     "pct": 1.33,
     "by_device": {
      "PIM": {
       "time_ms": 24.884,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 18.838,
     "pct": 1.007,
     "by_device": {
      "CPU": {
       "time_ms": 18.838,
       "nodes": 1,
       "calls": 326
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 17.098,
     "pct": 0.914,
     "by_device": {
      "CPU": {
       "time_ms": 17.098,
       "nodes": 1,
       "calls": 326
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 16.676,
     "pct": 0.891,
     "by_device": {
      "CPU": {
       "time_ms": 16.676,
       "nodes": 1,
       "calls": 328
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 9.644,
     "pct": 0.515,
     "by_device": {
      "CPU": {
       "time_ms": 9.644,
       "nodes": 1,
       "calls": 194
      }
     }
    },
    {
     "op": "gelu",
     "total_ms": 8.22,
     "pct": 0.439,
     "by_device": {
      "CPU": {
       "time_ms": 8.22,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 7.631,
     "pct": 0.408,
     "by_device": {
      "CPU": {
       "time_ms": 7.631,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 7.439,
     "pct": 0.397,
     "by_device": {
      "CPU": {
       "time_ms": 7.439,
       "nodes": 1,
       "calls": 66
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 7.18,
     "pct": 0.384,
     "by_device": {
      "CPU": {
       "time_ms": 7.18,
       "nodes": 1,
       "calls": 132
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 4.161,
     "pct": 0.222,
     "by_device": {
      "CPU": {
       "time_ms": 4.161,
       "nodes": 1,
       "calls": 79
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 4.034,
     "pct": 0.216,
     "by_device": {
      "CPU": {
       "time_ms": 4.034,
       "nodes": 1,
       "calls": 79
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 3.563,
     "pct": 0.19,
     "by_device": {
      "CPU": {
       "time_ms": 3.563,
       "nodes": 1,
       "calls": 93
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 2.685,
     "pct": 0.143,
     "by_device": {
      "CPU": {
       "time_ms": 2.685,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 2.533,
     "pct": 0.135,
     "by_device": {
      "CPU": {
       "time_ms": 2.533,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 2.475,
     "pct": 0.132,
     "by_device": {
      "CPU": {
       "time_ms": 2.475,
       "nodes": 1,
       "calls": 867
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 1.476,
     "pct": 0.079,
     "by_device": {
      "CPU": {
       "time_ms": 1.476,
       "nodes": 1,
       "calls": 1531
      }
     }
    },
    {
     "op": "add",
     "total_ms": 1.446,
     "pct": 0.077,
     "by_device": {
      "CPU": {
       "time_ms": 1.446,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 1.411,
     "pct": 0.075,
     "by_device": {
      "CPU": {
       "time_ms": 1.411,
       "nodes": 1,
       "calls": 829
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.345,
     "pct": 0.072,
     "by_device": {
      "CPU": {
       "time_ms": 1.345,
       "nodes": 1,
       "calls": 449
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 1.073,
     "pct": 0.057,
     "by_device": {
      "CPU": {
       "time_ms": 1.073,
       "nodes": 1,
       "calls": 266
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.725,
     "pct": 0.039,
     "by_device": {
      "CPU": {
       "time_ms": 0.725,
       "nodes": 1,
       "calls": 64
      }
     }
    },
    {
     "op": "chunk",
     "total_ms": 0.625,
     "pct": 0.033,
     "by_device": {
      "CPU": {
       "time_ms": 0.625,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 0.6,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 0.6,
       "nodes": 1,
       "calls": 258
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.567,
     "pct": 0.03,
     "by_device": {
      "CPU": {
       "time_ms": 0.567,
       "nodes": 1,
       "calls": 1207
      }
     }
    },
    {
     "op": "split",
     "total_ms": 0.542,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.542,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.514,
     "pct": 0.027,
     "by_device": {
      "CPU": {
       "time_ms": 0.514,
       "nodes": 1,
       "calls": 365
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.38,
     "pct": 0.02,
     "by_device": {
      "CPU": {
       "time_ms": 0.38,
       "nodes": 1,
       "calls": 111
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.362,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.362,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.332,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.332,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.218,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.218,
       "nodes": 1,
       "calls": 70
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.187,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.187,
       "nodes": 1,
       "calls": 186
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.125,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.125,
       "nodes": 1,
       "calls": 67
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.113,
     "pct": 0.006,
     "by_device": {
      "CPU": {
       "time_ms": 0.113,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.097,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.097,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.094,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.094,
       "nodes": 1,
       "calls": 14
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.082,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.082,
       "nodes": 1,
       "calls": 46
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.063,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.063,
       "nodes": 2,
       "calls": 64
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.049,
     "pct": 0.003,
     "by_device": {
      "CPU": {
       "time_ms": 0.049,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.04,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.04,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.026,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.026,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.024,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.024,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.01,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.009,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 32
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "dropout",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 1
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_pythia-1b_b16s8o2_pim.simplified.json",
   "source_profile": "torch_pythia-1b_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:58:19",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "Qwen2.5-1.5B",
   "hf_model_id": "Qwen/Qwen2.5-1.5B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen2.5-1.5B",
   "param_count": 1543714304
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 3.725,
   "TTFT": 2.915,
   "TBT_avg": 0.809,
   "Throughput": 8.592
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 3724.6,
     "ttft": 2915.306,
     "tbt_avg": 809.294,
     "throughput": 8.592,
     "steps": [
      2915.306,
      809.294
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 3724.6,
     "min": 3724.6,
     "max": 3724.6,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 2915.306,
     "min": 2915.306,
     "max": 2915.306,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 809.294,
     "min": 809.294,
     "max": 809.294,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 8.592,
     "min": 8.592,
     "max": 8.592,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 809.294,
     "min": 809.294,
     "max": 809.294,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 3802.832,
   "total_nodes": 54,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 3578.304,
     "pct": 94.096,
     "nodes": 3,
     "calls": 1008
    },
    {
     "device": "CPU",
     "time_ms": 224.528,
     "pct": 5.904,
     "nodes": 51,
     "calls": 18943
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 3034.447,
     "pct": 79.794,
     "by_device": {
      "PIM": {
       "time_ms": 3034.447,
       "nodes": 1,
       "calls": 392
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 497.168,
     "pct": 13.074,
     "by_device": {
      "PIM": {
       "time_ms": 497.168,
       "nodes": 1,
       "calls": 504
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 46.689,
     "pct": 1.228,
     "by_device": {
      "PIM": {
       "time_ms": 46.689,
       "nodes": 1,
       "calls": 112
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 28.26,
     "pct": 0.743,
     "by_device": {
      "CPU": {
       "time_ms": 28.26,
       "nodes": 1,
       "calls": 898
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 23.096,
     "pct": 0.607,
     "by_device": {
      "CPU": {
       "time_ms": 23.096,
       "nodes": 1,
       "calls": 898
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 21.59,
     "pct": 0.568,
     "by_device": {
      "CPU": {
       "time_ms": 21.59,
       "nodes": 1,
       "calls": 649
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 21.517,
     "pct": 0.566,
     "by_device": {
      "CPU": {
       "time_ms": 21.517,
       "nodes": 1,
       "calls": 512
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 13.841,
     "pct": 0.364,
     "by_device": {
      "CPU": {
       "time_ms": 13.841,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 10.72,
     "pct": 0.282,
     "by_device": {
      "CPU": {
       "time_ms": 10.72,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 10.288,
     "pct": 0.271,
     "by_device": {
      "CPU": {
       "time_ms": 10.288,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 8.313,
     "pct": 0.219,
     "by_device": {
      "CPU": {
       "time_ms": 8.313,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "to",
     "total_ms": 8.254,
     "pct": 0.217,
     "by_device": {
      "CPU": {
       "time_ms": 8.254,
       "nodes": 1,
       "calls": 1042
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 8.213,
     "pct": 0.216,
     "by_device": {
      "CPU": {
       "time_ms": 8.213,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 8.117,
     "pct": 0.213,
     "by_device": {
      "CPU": {
       "time_ms": 8.117,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 7.338,
     "pct": 0.193,
     "by_device": {
      "CPU": {
       "time_ms": 7.338,
       "nodes": 1,
       "calls": 470
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 5.459,
     "pct": 0.144,
     "by_device": {
      "CPU": {
       "time_ms": 5.459,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 5.277,
     "pct": 0.139,
     "by_device": {
      "CPU": {
       "time_ms": 5.277,
       "nodes": 1,
       "calls": 1403
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 5.21,
     "pct": 0.137,
     "by_device": {
      "CPU": {
       "time_ms": 5.21,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 4.584,
     "pct": 0.121,
     "by_device": {
      "CPU": {
       "time_ms": 4.584,
       "nodes": 1,
       "calls": 470
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 4.078,
     "pct": 0.107,
     "by_device": {
      "CPU": {
       "time_ms": 4.078,
       "nodes": 1,
       "calls": 850
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 3.857,
     "pct": 0.101,
     "by_device": {
      "CPU": {
       "time_ms": 3.857,
       "nodes": 1,
       "calls": 3709
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 3.538,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 3.538,
       "nodes": 1,
       "calls": 228
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 3.526,
     "pct": 0.093,
     "by_device": {
      "CPU": {
       "time_ms": 3.526,
       "nodes": 1,
       "calls": 1516
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 3.221,
     "pct": 0.085,
     "by_device": {
      "CPU": {
       "time_ms": 3.221,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "add",
     "total_ms": 3.092,
     "pct": 0.081,
     "by_device": {
      "CPU": {
       "time_ms": 3.092,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 2.714,
     "pct": 0.071,
     "by_device": {
      "CPU": {
       "time_ms": 2.714,
       "nodes": 1,
       "calls": 118
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.637,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 1.637,
       "nodes": 1,
       "calls": 564
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 1.309,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 1.309,
       "nodes": 1,
       "calls": 112
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.292,
     "pct": 0.034,
     "by_device": {
      "CPU": {
       "time_ms": 1.292,
       "nodes": 1,
       "calls": 625
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.222,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 1.222,
       "nodes": 1,
       "calls": 2097
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.214,
     "pct": 0.032,
     "by_device": {
      "CPU": {
       "time_ms": 1.214,
       "nodes": 1,
       "calls": 393
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.846,
     "pct": 0.022,
     "by_device": {
      "CPU": {
       "time_ms": 0.846,
       "nodes": 1,
       "calls": 526
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.706,
     "pct": 0.019,
     "by_device": {
      "CPU": {
       "time_ms": 0.706,
       "nodes": 1,
       "calls": 395
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.423,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.423,
       "nodes": 1,
       "calls": 118
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.375,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.375,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.319,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.319,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.29,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.29,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.285,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.285,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.161,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.161,
       "nodes": 2,
       "calls": 112
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.094,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.094,
       "nodes": 1,
       "calls": 114
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.066,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.066,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.046,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.046,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.028,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.028,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.016,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.016,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.016,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.014,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.009,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.009,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.008,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.003,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.003,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_qwen2.5-1.5b_b16s8o2_pim.simplified.json",
   "source_profile": "torch_qwen2.5-1.5b_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:57:54",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "Qwen3-0.6B",
   "hf_model_id": "Qwen/Qwen3-0.6B",
   "hf_model_url": "https://huggingface.co/Qwen/Qwen3-0.6B",
   "param_count": 596049920
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 1.774,
   "TTFT": 1.182,
   "TBT_avg": 0.592,
   "Throughput": 18.042
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 1773.628,
     "ttft": 1181.629,
     "tbt_avg": 591.999,
     "throughput": 18.042,
     "steps": [
      1181.629,
      591.999
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 1773.628,
     "min": 1773.628,
     "max": 1773.628,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 1181.629,
     "min": 1181.629,
     "max": 1181.629,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 591.999,
     "min": 591.999,
     "max": 591.999,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 18.042,
     "min": 18.042,
     "max": 18.042,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 591.999,
     "min": 591.999,
     "max": 591.999,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 1852.342,
   "total_nodes": 55,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 1585.097,
     "pct": 85.573,
     "nodes": 4,
     "calls": 1008
    },
    {
     "device": "CPU",
     "time_ms": 267.245,
     "pct": 14.427,
     "nodes": 51,
     "calls": 23199
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 1111.309,
     "pct": 59.995,
     "by_device": {
      "PIM": {
       "time_ms": 1111.309,
       "nodes": 1,
       "calls": 392
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 375.045,
     "pct": 20.247,
     "by_device": {
      "PIM": {
       "time_ms": 375.045,
       "nodes": 1,
       "calls": 448
      }
     }
    },
    {
     "op": "pim_mul",
     "total_ms": 67.444,
     "pct": 3.641,
     "by_device": {
      "PIM": {
       "time_ms": 67.444,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 31.298,
     "pct": 1.69,
     "by_device": {
      "PIM": {
       "time_ms": 31.298,
       "nodes": 1,
       "calls": 112
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 30.683,
     "pct": 1.656,
     "by_device": {
      "CPU": {
       "time_ms": 30.683,
       "nodes": 1,
       "calls": 954
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 24.428,
     "pct": 1.319,
     "by_device": {
      "CPU": {
       "time_ms": 24.428,
       "nodes": 1,
       "calls": 954
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 23.072,
     "pct": 1.246,
     "by_device": {
      "CPU": {
       "time_ms": 23.072,
       "nodes": 1,
       "calls": 873
      }
     }
    },
    {
     "op": "to",
     "total_ms": 19.095,
     "pct": 1.031,
     "by_device": {
      "CPU": {
       "time_ms": 19.095,
       "nodes": 1,
       "calls": 1602
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 17.369,
     "pct": 0.938,
     "by_device": {
      "CPU": {
       "time_ms": 17.369,
       "nodes": 1,
       "calls": 918
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 15.764,
     "pct": 0.851,
     "by_device": {
      "CPU": {
       "time_ms": 15.764,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 12.291,
     "pct": 0.664,
     "by_device": {
      "CPU": {
       "time_ms": 12.291,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 12.093,
     "pct": 0.653,
     "by_device": {
      "CPU": {
       "time_ms": 12.093,
       "nodes": 1,
       "calls": 680
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 11.273,
     "pct": 0.609,
     "by_device": {
      "CPU": {
       "time_ms": 11.273,
       "nodes": 1,
       "calls": 918
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 9.974,
     "pct": 0.538,
     "by_device": {
      "CPU": {
       "time_ms": 9.974,
       "nodes": 1,
       "calls": 230
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 9.529,
     "pct": 0.514,
     "by_device": {
      "CPU": {
       "time_ms": 9.529,
       "nodes": 1,
       "calls": 1572
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 7.351,
     "pct": 0.397,
     "by_device": {
      "CPU": {
       "time_ms": 7.351,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 7.314,
     "pct": 0.395,
     "by_device": {
      "CPU": {
       "time_ms": 7.314,
       "nodes": 1,
       "calls": 452
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 6.594,
     "pct": 0.356,
     "by_device": {
      "CPU": {
       "time_ms": 6.594,
       "nodes": 1,
       "calls": 1403
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 5.63,
     "pct": 0.304,
     "by_device": {
      "CPU": {
       "time_ms": 5.63,
       "nodes": 1,
       "calls": 4605
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 5.48,
     "pct": 0.296,
     "by_device": {
      "CPU": {
       "time_ms": 5.48,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 5.422,
     "pct": 0.293,
     "by_device": {
      "CPU": {
       "time_ms": 5.422,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 5.075,
     "pct": 0.274,
     "by_device": {
      "CPU": {
       "time_ms": 5.075,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 5.059,
     "pct": 0.273,
     "by_device": {
      "CPU": {
       "time_ms": 5.059,
       "nodes": 1,
       "calls": 850
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 5.059,
     "pct": 0.273,
     "by_device": {
      "CPU": {
       "time_ms": 5.059,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "linear",
     "total_ms": 4.949,
     "pct": 0.267,
     "by_device": {
      "CPU": {
       "time_ms": 4.949,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "add",
     "total_ms": 4.904,
     "pct": 0.265,
     "by_device": {
      "CPU": {
       "time_ms": 4.904,
       "nodes": 1,
       "calls": 338
      }
     }
    },
    {
     "op": "mm",
     "total_ms": 4.862,
     "pct": 0.262,
     "by_device": {
      "CPU": {
       "time_ms": 4.862,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 2.096,
     "pct": 0.113,
     "by_device": {
      "CPU": {
       "time_ms": 2.096,
       "nodes": 1,
       "calls": 564
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 1.971,
     "pct": 0.106,
     "by_device": {
      "CPU": {
       "time_ms": 1.971,
       "nodes": 1,
       "calls": 974
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.695,
     "pct": 0.092,
     "by_device": {
      "CPU": {
       "time_ms": 1.695,
       "nodes": 1,
       "calls": 625
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 1.611,
     "pct": 0.087,
     "by_device": {
      "CPU": {
       "time_ms": 1.611,
       "nodes": 1,
       "calls": 2097
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 1.487,
     "pct": 0.08,
     "by_device": {
      "CPU": {
       "time_ms": 1.487,
       "nodes": 1,
       "calls": 393
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 1.439,
     "pct": 0.078,
     "by_device": {
      "CPU": {
       "time_ms": 1.439,
       "nodes": 1,
       "calls": 112
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.932,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 0.932,
       "nodes": 1,
       "calls": 395
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.804,
     "pct": 0.043,
     "by_device": {
      "CPU": {
       "time_ms": 0.804,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.459,
     "pct": 0.025,
     "by_device": {
      "CPU": {
       "time_ms": 0.459,
       "nodes": 1,
       "calls": 118
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.325,
     "pct": 0.018,
     "by_device": {
      "CPU": {
       "time_ms": 0.325,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.297,
     "pct": 0.016,
     "by_device": {
      "CPU": {
       "time_ms": 0.297,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.287,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.287,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.172,
     "pct": 0.009,
     "by_device": {
      "CPU": {
       "time_ms": 0.172,
       "nodes": 1,
       "calls": 226
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.134,
     "pct": 0.007,
     "by_device": {
      "CPU": {
       "time_ms": 0.134,
       "nodes": 2,
       "calls": 112
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.072,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.072,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.039,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.039,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.032,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.032,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.028,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.028,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.019,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "t",
     "total_ms": 0.018,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.018,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.014,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 56
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.011,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.011,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.004,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.004,
       "nodes": 1,
       "calls": 8
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_qwen3-0.6b_b16s8o2_pim.simplified.json",
   "source_profile": "torch_qwen3-0.6b_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:57:40",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "stablelm-2-1_6b",
   "hf_model_id": "stabilityai/stablelm-2-1_6b",
   "hf_model_url": "https://huggingface.co/stabilityai/stablelm-2-1_6b",
   "param_count": 1644417024
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 3.086,
   "TTFT": 2.371,
   "TBT_avg": 0.715,
   "Throughput": 10.369
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 3086.047,
     "ttft": 2371.436,
     "tbt_avg": 714.611,
     "throughput": 10.369,
     "steps": [
      2371.436,
      714.611
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 3086.047,
     "min": 3086.047,
     "max": 3086.047,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 2371.436,
     "min": 2371.436,
     "max": 2371.436,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 714.611,
     "min": 714.611,
     "max": 714.611,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 10.369,
     "min": 10.369,
     "max": 10.369,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 714.611,
     "min": 714.611,
     "max": 714.611,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 3222.582,
   "total_nodes": 50,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 2968.069,
     "pct": 92.102,
     "nodes": 3,
     "calls": 770
    },
    {
     "device": "CPU",
     "time_ms": 254.513,
     "pct": 7.898,
     "nodes": 47,
     "calls": 15829
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 2590.862,
     "pct": 80.397,
     "by_device": {
      "PIM": {
       "time_ms": 2590.862,
       "nodes": 1,
       "calls": 338
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 336.787,
     "pct": 10.451,
     "by_device": {
      "PIM": {
       "time_ms": 336.787,
       "nodes": 1,
       "calls": 336
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 44.225,
     "pct": 1.372,
     "by_device": {
      "CPU": {
       "time_ms": 44.225,
       "nodes": 1,
       "calls": 774
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 40.42,
     "pct": 1.254,
     "by_device": {
      "PIM": {
       "time_ms": 40.42,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 37.73,
     "pct": 1.171,
     "by_device": {
      "CPU": {
       "time_ms": 37.73,
       "nodes": 1,
       "calls": 774
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 36.293,
     "pct": 1.126,
     "by_device": {
      "CPU": {
       "time_ms": 36.293,
       "nodes": 1,
       "calls": 776
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 20.765,
     "pct": 0.644,
     "by_device": {
      "CPU": {
       "time_ms": 20.765,
       "nodes": 1,
       "calls": 290
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 20.67,
     "pct": 0.641,
     "by_device": {
      "CPU": {
       "time_ms": 20.67,
       "nodes": 1,
       "calls": 244
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 12.252,
     "pct": 0.38,
     "by_device": {
      "CPU": {
       "time_ms": 12.252,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "layer_norm",
     "total_ms": 10.211,
     "pct": 0.317,
     "by_device": {
      "CPU": {
       "time_ms": 10.211,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "native_layer_norm",
     "total_ms": 9.828,
     "pct": 0.305,
     "by_device": {
      "CPU": {
       "time_ms": 9.828,
       "nodes": 1,
       "calls": 98
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 7.009,
     "pct": 0.218,
     "by_device": {
      "CPU": {
       "time_ms": 7.009,
       "nodes": 1,
       "calls": 1587
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 6.073,
     "pct": 0.188,
     "by_device": {
      "CPU": {
       "time_ms": 6.073,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 5.751,
     "pct": 0.178,
     "by_device": {
      "CPU": {
       "time_ms": 5.751,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 4.998,
     "pct": 0.155,
     "by_device": {
      "CPU": {
       "time_ms": 4.998,
       "nodes": 1,
       "calls": 1719
      }
     }
    },
    {
     "op": "add",
     "total_ms": 4.61,
     "pct": 0.143,
     "by_device": {
      "CPU": {
       "time_ms": 4.61,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 4.536,
     "pct": 0.141,
     "by_device": {
      "CPU": {
       "time_ms": 4.536,
       "nodes": 1,
       "calls": 2819
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 4.433,
     "pct": 0.138,
     "by_device": {
      "CPU": {
       "time_ms": 4.433,
       "nodes": 1,
       "calls": 730
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 3.888,
     "pct": 0.121,
     "by_device": {
      "CPU": {
       "time_ms": 3.888,
       "nodes": 1,
       "calls": 121
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 3.614,
     "pct": 0.112,
     "by_device": {
      "CPU": {
       "time_ms": 3.614,
       "nodes": 1,
       "calls": 121
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 3.051,
     "pct": 0.095,
     "by_device": {
      "CPU": {
       "time_ms": 3.051,
       "nodes": 1,
       "calls": 96
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 2.448,
     "pct": 0.076,
     "by_device": {
      "CPU": {
       "time_ms": 2.448,
       "nodes": 1,
       "calls": 135
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 2.3,
     "pct": 0.071,
     "by_device": {
      "CPU": {
       "time_ms": 2.3,
       "nodes": 1,
       "calls": 529
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 2.187,
     "pct": 0.068,
     "by_device": {
      "CPU": {
       "time_ms": 2.187,
       "nodes": 1,
       "calls": 2183
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.834,
     "pct": 0.057,
     "by_device": {
      "CPU": {
       "time_ms": 1.834,
       "nodes": 1,
       "calls": 482
      }
     }
    },
    {
     "op": "view",
     "total_ms": 1.689,
     "pct": 0.052,
     "by_device": {
      "CPU": {
       "time_ms": 1.689,
       "nodes": 1,
       "calls": 733
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.933,
     "pct": 0.029,
     "by_device": {
      "CPU": {
       "time_ms": 0.933,
       "nodes": 1,
       "calls": 169
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.64,
     "pct": 0.02,
     "by_device": {
      "CPU": {
       "time_ms": 0.64,
       "nodes": 1,
       "calls": 339
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.492,
     "pct": 0.015,
     "by_device": {
      "CPU": {
       "time_ms": 0.492,
       "nodes": 1,
       "calls": 102
      }
     }
    },
    {
     "op": "to",
     "total_ms": 0.396,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.396,
       "nodes": 1,
       "calls": 410
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.378,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.378,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.34,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.34,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.168,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.168,
       "nodes": 1,
       "calls": 62
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 0.153,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.153,
       "nodes": 1,
       "calls": 14
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 0.145,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.145,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.136,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.136,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.13,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.13,
       "nodes": 2,
       "calls": 96
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.046,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.046,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.041,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.041,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.03,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.03,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.019,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.019,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.014,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.014,
       "nodes": 1,
       "calls": 48
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.013,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.013,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.012,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.012,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.012,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.007,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.007,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_stablelm-2-1_6b_b16s8o2_pim.simplified.json",
   "source_profile": "torch_stablelm-2-1_6b_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 },
 {
  "metadata": {
   "timestamp": "2026-07-27 16:56:15",
   "framework": "pytorch",
   "device": "cpu+pim"
  },
  "model_info": {
   "model_name": "TinyLlama-1.1B-Chat-v1.0",
   "hf_model_id": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "hf_model_url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0",
   "param_count": 1100048384
  },
  "test_conditions": {
   "dtype": "bfloat16",
   "input_tokens": 8,
   "batch_size": 16,
   "output_tokens": 2,
   "iterations": 1
  },
  "performance_metrics": {
   "E2E": 2.446,
   "TTFT": 1.897,
   "TBT_avg": 0.549,
   "Throughput": 13.082
  },
  "run_breakdown": {
   "unit": "ms",
   "steps_per_iteration": 2,
   "prefill_steps": 1,
   "iterations": [
    {
     "e2e": 2446.133,
     "ttft": 1897.154,
     "tbt_avg": 548.979,
     "throughput": 13.082,
     "steps": [
      1897.154,
      548.979
     ]
    }
   ],
   "summary": {
    "e2e": {
     "mean": 2446.133,
     "min": 2446.133,
     "max": 2446.133,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "ttft": {
     "mean": 1897.154,
     "min": 1897.154,
     "max": 1897.154,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "tbt": {
     "mean": 548.979,
     "min": 548.979,
     "max": 548.979,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "throughput": {
     "mean": 13.082,
     "min": 13.082,
     "max": 13.082,
     "std": 0.0,
     "cv_pct": 0.0
    },
    "decode_step": {
     "mean": 548.979,
     "min": 548.979,
     "max": 548.979,
     "std": 0.0,
     "cv_pct": 0.0
    }
   }
  },
  "op_breakdown": {
   "unit": "ms",
   "total_ms": 2513.58,
   "total_nodes": 53,
   "devices": [
    {
     "device": "PIM",
     "time_ms": 2350.239,
     "pct": 93.502,
     "nodes": 3,
     "calls": 794
    },
    {
     "device": "CPU",
     "time_ms": 163.341,
     "pct": 6.498,
     "nodes": 50,
     "calls": 15194
    }
   ],
   "ops": [
    {
     "op": "pim_linear",
     "total_ms": 2032.555,
     "pct": 80.863,
     "by_device": {
      "PIM": {
       "time_ms": 2032.555,
       "nodes": 1,
       "calls": 310
      }
     }
    },
    {
     "op": "pim_dispatch",
     "total_ms": 282.451,
     "pct": 11.237,
     "by_device": {
      "PIM": {
       "time_ms": 282.451,
       "nodes": 1,
       "calls": 396
      }
     }
    },
    {
     "op": "pim_add",
     "total_ms": 35.234,
     "pct": 1.402,
     "by_device": {
      "PIM": {
       "time_ms": 35.234,
       "nodes": 1,
       "calls": 88
      }
     }
    },
    {
     "op": "zeros",
     "total_ms": 26.832,
     "pct": 1.067,
     "by_device": {
      "CPU": {
       "time_ms": 26.832,
       "nodes": 1,
       "calls": 710
      }
     }
    },
    {
     "op": "zero_",
     "total_ms": 22.908,
     "pct": 0.911,
     "by_device": {
      "CPU": {
       "time_ms": 22.908,
       "nodes": 1,
       "calls": 710
      }
     }
    },
    {
     "op": "fill_",
     "total_ms": 22.053,
     "pct": 0.877,
     "by_device": {
      "CPU": {
       "time_ms": 22.053,
       "nodes": 1,
       "calls": 758
      }
     }
    },
    {
     "op": "mul",
     "total_ms": 14.432,
     "pct": 0.574,
     "by_device": {
      "CPU": {
       "time_ms": 14.432,
       "nodes": 1,
       "calls": 404
      }
     }
    },
    {
     "op": "scaled_dot_product_attention",
     "total_ms": 7.839,
     "pct": 0.312,
     "by_device": {
      "CPU": {
       "time_ms": 7.839,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "_scaled_dot_product_flash_attention_for_cpu",
     "total_ms": 7.614,
     "pct": 0.303,
     "by_device": {
      "CPU": {
       "time_ms": 7.614,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "silu",
     "total_ms": 7.343,
     "pct": 0.292,
     "by_device": {
      "CPU": {
       "time_ms": 7.343,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "to",
     "total_ms": 6.289,
     "pct": 0.25,
     "by_device": {
      "CPU": {
       "time_ms": 6.289,
       "nodes": 1,
       "calls": 828
      }
     }
    },
    {
     "op": "_to_copy",
     "total_ms": 5.651,
     "pct": 0.225,
     "by_device": {
      "CPU": {
       "time_ms": 5.651,
       "nodes": 1,
       "calls": 374
      }
     }
    },
    {
     "op": "mean",
     "total_ms": 4.341,
     "pct": 0.173,
     "by_device": {
      "CPU": {
       "time_ms": 4.341,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "cat",
     "total_ms": 4.135,
     "pct": 0.164,
     "by_device": {
      "CPU": {
       "time_ms": 4.135,
       "nodes": 1,
       "calls": 178
      }
     }
    },
    {
     "op": "slice",
     "total_ms": 3.981,
     "pct": 0.158,
     "by_device": {
      "CPU": {
       "time_ms": 3.981,
       "nodes": 1,
       "calls": 1103
      }
     }
    },
    {
     "op": "copy_",
     "total_ms": 3.667,
     "pct": 0.146,
     "by_device": {
      "CPU": {
       "time_ms": 3.667,
       "nodes": 1,
       "calls": 375
      }
     }
    },
    {
     "op": "reshape",
     "total_ms": 3.051,
     "pct": 0.121,
     "by_device": {
      "CPU": {
       "time_ms": 3.051,
       "nodes": 1,
       "calls": 670
      }
     }
    },
    {
     "op": "alias",
     "total_ms": 2.762,
     "pct": 0.11,
     "by_device": {
      "CPU": {
       "time_ms": 2.762,
       "nodes": 1,
       "calls": 2922
      }
     }
    },
    {
     "op": "add",
     "total_ms": 2.627,
     "pct": 0.105,
     "by_device": {
      "CPU": {
       "time_ms": 2.627,
       "nodes": 1,
       "calls": 178
      }
     }
    },
    {
     "op": "pow",
     "total_ms": 2.571,
     "pct": 0.102,
     "by_device": {
      "CPU": {
       "time_ms": 2.571,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "div_",
     "total_ms": 2.45,
     "pct": 0.097,
     "by_device": {
      "CPU": {
       "time_ms": 2.45,
       "nodes": 1,
       "calls": 180
      }
     }
    },
    {
     "op": "empty",
     "total_ms": 2.44,
     "pct": 0.097,
     "by_device": {
      "CPU": {
       "time_ms": 2.44,
       "nodes": 1,
       "calls": 1197
      }
     }
    },
    {
     "op": "sum",
     "total_ms": 2.418,
     "pct": 0.096,
     "by_device": {
      "CPU": {
       "time_ms": 2.418,
       "nodes": 1,
       "calls": 94
      }
     }
    },
    {
     "op": "transpose",
     "total_ms": 1.249,
     "pct": 0.05,
     "by_device": {
      "CPU": {
       "time_ms": 1.249,
       "nodes": 1,
       "calls": 442
      }
     }
    },
    {
     "op": "neg",
     "total_ms": 0.988,
     "pct": 0.039,
     "by_device": {
      "CPU": {
       "time_ms": 0.988,
       "nodes": 1,
       "calls": 88
      }
     }
    },
    {
     "op": "view",
     "total_ms": 0.959,
     "pct": 0.038,
     "by_device": {
      "CPU": {
       "time_ms": 0.959,
       "nodes": 1,
       "calls": 493
      }
     }
    },
    {
     "op": "narrow",
     "total_ms": 0.904,
     "pct": 0.036,
     "by_device": {
      "CPU": {
       "time_ms": 0.904,
       "nodes": 1,
       "calls": 309
      }
     }
    },
    {
     "op": "as_strided",
     "total_ms": 0.891,
     "pct": 0.035,
     "by_device": {
      "CPU": {
       "time_ms": 0.891,
       "nodes": 1,
       "calls": 1651
      }
     }
    },
    {
     "op": "empty_strided",
     "total_ms": 0.633,
     "pct": 0.025,
     "by_device": {
      "CPU": {
       "time_ms": 0.633,
       "nodes": 1,
       "calls": 418
      }
     }
    },
    {
     "op": "_reshape_alias",
     "total_ms": 0.522,
     "pct": 0.021,
     "by_device": {
      "CPU": {
       "time_ms": 0.522,
       "nodes": 1,
       "calls": 311
      }
     }
    },
    {
     "op": "unsqueeze",
     "total_ms": 0.301,
     "pct": 0.012,
     "by_device": {
      "CPU": {
       "time_ms": 0.301,
       "nodes": 1,
       "calls": 94
      }
     }
    },
    {
     "op": "embedding",
     "total_ms": 0.282,
     "pct": 0.011,
     "by_device": {
      "CPU": {
       "time_ms": 0.282,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "index_select",
     "total_ms": 0.255,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.255,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "rsqrt",
     "total_ms": 0.253,
     "pct": 0.01,
     "by_device": {
      "CPU": {
       "time_ms": 0.253,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "empty_like",
     "total_ms": 0.189,
     "pct": 0.008,
     "by_device": {
      "CPU": {
       "time_ms": 0.189,
       "nodes": 1,
       "calls": 45
      }
     }
    },
    {
     "op": "matmul",
     "total_ms": 0.119,
     "pct": 0.005,
     "by_device": {
      "CPU": {
       "time_ms": 0.119,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "detach_",
     "total_ms": 0.096,
     "pct": 0.004,
     "by_device": {
      "CPU": {
       "time_ms": 0.096,
       "nodes": 2,
       "calls": 88
      }
     }
    },
    {
     "op": "result_type",
     "total_ms": 0.058,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.058,
       "nodes": 1,
       "calls": 90
      }
     }
    },
    {
     "op": "bmm",
     "total_ms": 0.056,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.056,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "cos",
     "total_ms": 0.038,
     "pct": 0.002,
     "by_device": {
      "CPU": {
       "time_ms": 0.038,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "expand",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 6
      }
     }
    },
    {
     "op": "contiguous",
     "total_ms": 0.025,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.025,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "sin",
     "total_ms": 0.023,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.023,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "clone",
     "total_ms": 0.022,
     "pct": 0.001,
     "by_device": {
      "CPU": {
       "time_ms": 0.022,
       "nodes": 1,
       "calls": 1
      }
     }
    },
    {
     "op": "select",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "lift_fresh",
     "total_ms": 0.01,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.01,
       "nodes": 1,
       "calls": 44
      }
     }
    },
    {
     "op": "eq",
     "total_ms": 0.009,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.009,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "is_nonzero",
     "total_ms": 0.008,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.008,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "_unsafe_view",
     "total_ms": 0.006,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.006,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "item",
     "total_ms": 0.005,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.005,
       "nodes": 1,
       "calls": 2
      }
     }
    },
    {
     "op": "resolve_conj",
     "total_ms": 0.002,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.002,
       "nodes": 1,
       "calls": 4
      }
     }
    },
    {
     "op": "_local_scalar_dense",
     "total_ms": 0.001,
     "pct": 0.0,
     "by_device": {
      "CPU": {
       "time_ms": 0.001,
       "nodes": 1,
       "calls": 2
      }
     }
    }
   ]
  },
  "environment_spec": {
   "pim": {
    "device": "AiM-PIM"
   }
  },
  "_provenance": {
   "note": "임시 데이터 — 기존 trace에서 추출. 재측정 아님.",
   "source_simplified": "torch/pim/torch_tinyllama-1.1b-chat_b16s8o2_pim.simplified.json",
   "source_profile": "torch_tinyllama-1.1b-chat_b16s8o2_pim.json",
   "measured": [
    "performance_metrics.*  (role=model phase 소요시간에서 산출)",
    "run_breakdown.*        (동일)",
    "op_breakdown.*         (측정구간 leaf 이벤트 집계)",
    "test_conditions.*      (sidecar)",
    "model_info.hf_model_id (sidecar)",
    "metadata.device        (leaf 이벤트에서 관측된 디바이스)"
   ],
   "derived": [
    "model_info.param_count (HF 모델카드 공개값, 근사 — trace에 없음)",
    "model_info.model_name / hf_model_url (model_id 에서 생성)",
    "metadata.timestamp     (원본 파일 mtime)"
   ],
   "missing": [
    "io_details.*           (입출력 텍스트·토큰수 — trace에 없음)",
    "environment_spec.cpu   (CPU 모델명·코어수·governor — trace에 없음)",
    "pim_features.pim_devices (실제 사용 PIM 대수 — trace에 없음)",
    "environment_spec.pim 의 count/memory_gb/clock_mhz (device 이름만 확보)"
   ],
   "caveats": [
    "iterations=1 이라 iteration 간 통계(std/cv)는 0이다.",
    "Throughput = batch_size × output_tokens / E2E.",
    "GPU 커널은 CPU와 비동기로 겹치므로 op_breakdown 시간 합이 벽시계보다 클 수 있다.",
    "timestamp 는 진짜 측정 시각이 아니다 — kineto baseTimeNanoseconds 가 48개 파일 전부 같은 상수라 쓸 수 없어 원본 파일 mtime 으로 대체했다."
   ]
  }
 }
]