{
  "article_key": "llm-cpu-offload",
  "cases": [
    {
      "case_wall_seconds": 185.7723383929697,
      "completed_response_resources": {
        "gpu_energy": {
          "available": true,
          "interpolation": "linear_at_bracketed_boundaries",
          "joules": 43316.62505863769,
          "reason": null,
          "response_union_seconds": 144.44773938611615,
          "scope": "GPU board only; not whole-server AC energy",
          "time_weighted_mean_watts": 299.8774867833006,
          "trapezoid_pieces": 155,
          "watt_hours": 12.032395849621581
        },
        "gpu_sample_count": 145,
        "host_sample_count": 142,
        "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
        "metrics": {
          "cgroup_memory_current_bytes": {
            "maximum": 783441920,
            "mean": 744136242.4788733,
            "median": 753983488.0,
            "minimum": 708935680,
            "sample_count": 142
          },
          "gpu_board_watts": {
            "maximum": 302.93,
            "mean": 300.00172413793103,
            "median": 300.0,
            "minimum": 297.31,
            "sample_count": 145
          },
          "gpu_memory_used_bytes": {
            "maximum": 47104131072.0,
            "mean": 47104131072.0,
            "median": 47104131072.0,
            "minimum": 47104131072.0,
            "sample_count": 145
          },
          "gpu_sm_clock_mhz": {
            "maximum": 1635.0,
            "mean": 1481.2551724137932,
            "median": 1485.0,
            "minimum": 1312.0,
            "sample_count": 145
          },
          "gpu_utilization_percent": {
            "maximum": 100.0,
            "mean": 99.84827586206896,
            "median": 100.0,
            "minimum": 83.0,
            "sample_count": 145
          },
          "host_mem_available_bytes": {
            "maximum": 96093003776,
            "mean": 96040983393.35211,
            "median": 96037416960.0,
            "minimum": 96006701056,
            "sample_count": 142
          },
          "process_pss_bytes": {
            "maximum": 1257277440,
            "mean": 1236598834.4788733,
            "median": 1246618624.0,
            "minimum": 1205324800,
            "sample_count": 142
          },
          "process_rss_bytes": {
            "maximum": 1260240896,
            "mean": 1239593392.6760564,
            "median": 1249583104.0,
            "minimum": 1209503744,
            "sample_count": 142
          }
        },
        "process_sample_count": 142,
        "response_union_seconds": 144.44773938611615,
        "sampled_counter_deltas": {
          "deltas": {
            "major_faults": 0,
            "read_syscall_bytes": 0,
            "storage_read_bytes": 0,
            "storage_write_bytes": 102400
          },
          "observed_endpoint_seconds": 135.37246441596653,
          "observed_intervals": 10,
          "requested_intervals": 10
        }
      },
      "groups": [
        {
          "aggregate_output_tokens_per_second": 22.20832513756411,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 2048,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 3.1211448738400773,
            "interval_count": 5,
            "interval_seconds": 57.6511207500007,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 99.87663596288247,
            "process_cpu_seconds": 57.58
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 17282.92564468072,
              "reason": null,
              "response_union_seconds": 57.63604378409218,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 299.86315003547986,
              "trapezoid_pieces": 63,
              "watt_hours": 4.8008126790779775
            },
            "gpu_sample_count": 58,
            "host_sample_count": 57,
            "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
            "metrics": {
              "cgroup_memory_current_bytes": {
                "maximum": 748466176,
                "mean": 716765866.6666666,
                "median": 711774208,
                "minimum": 708935680,
                "sample_count": 57
              },
              "gpu_board_watts": {
                "maximum": 300.95,
                "mean": 300.01241379310346,
                "median": 299.99,
                "minimum": 299.66,
                "sample_count": 58
              },
              "gpu_memory_used_bytes": {
                "maximum": 47104131072.0,
                "mean": 47104131072.0,
                "median": 47104131072.0,
                "minimum": 47104131072.0,
                "sample_count": 58
              },
              "gpu_sm_clock_mhz": {
                "maximum": 1627.0,
                "mean": 1446.4827586206898,
                "median": 1451.0,
                "minimum": 1312.0,
                "sample_count": 58
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 100.0,
                "median": 100.0,
                "minimum": 100.0,
                "sample_count": 58
              },
              "host_mem_available_bytes": {
                "maximum": 96093003776,
                "mean": 96060373171.64912,
                "median": 96058638336,
                "minimum": 96031748096,
                "sample_count": 57
              },
              "process_pss_bytes": {
                "maximum": 1225331712,
                "mean": 1211821738.6666667,
                "median": 1207718912,
                "minimum": 1205324800,
                "sample_count": 57
              },
              "process_rss_bytes": {
                "maximum": 1228296192,
                "mean": 1214836376.7017543,
                "median": 1210683392,
                "minimum": 1209503744,
                "sample_count": 57
              }
            },
            "process_sample_count": 57,
            "response_union_seconds": 57.63604378409218,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 40960
              },
              "observed_endpoint_seconds": 53.343125255953055,
              "observed_intervals": 5,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 57.6594875259907,
          "response_seconds": {
            "maximum": 11.647003977035638,
            "mean": 11.527208756818435,
            "median": 11.566605183004867,
            "minimum": 11.37437688402133,
            "sample_count": 5
          },
          "response_window_union_seconds": 57.63604378409218,
          "server_decode_tokens_per_second": {
            "maximum": 26.211279772416194,
            "mean": 25.85845974942399,
            "median": 25.778169650885744,
            "minimum": 25.605412361610277,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 1.6881801380077377,
            "mean": 1.665065498405602,
            "median": 1.6596780359977856,
            "minimum": 1.645603878016118,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        },
        {
          "aggregate_output_tokens_per_second": 14.7445570683008,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 8192,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 3.1216834193859877,
            "interval_count": 5,
            "interval_seconds": 86.83215546992142,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 99.8938694203516,
            "process_cpu_seconds": 86.74000000000001
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 26033.69941395698,
              "reason": null,
              "response_union_seconds": 86.81169560202397,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 299.88700524068577,
              "trapezoid_pieces": 92,
              "watt_hours": 7.2315831705436056
            },
            "gpu_sample_count": 87,
            "host_sample_count": 85,
            "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
            "metrics": {
              "cgroup_memory_current_bytes": {
                "maximum": 783441920,
                "mean": 762490494.4941176,
                "median": 764178432,
                "minimum": 749699072,
                "sample_count": 85
              },
              "gpu_board_watts": {
                "maximum": 302.93,
                "mean": 299.9945977011494,
                "median": 300.0,
                "minimum": 297.31,
                "sample_count": 87
              },
              "gpu_memory_used_bytes": {
                "maximum": 47104131072.0,
                "mean": 47104131072.0,
                "median": 47104131072.0,
                "minimum": 47104131072.0,
                "sample_count": 87
              },
              "gpu_sm_clock_mhz": {
                "maximum": 1635.0,
                "mean": 1504.4367816091954,
                "median": 1485.0,
                "minimum": 1462.0,
                "sample_count": 87
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 99.74712643678161,
                "median": 100.0,
                "minimum": 83.0,
                "sample_count": 87
              },
              "host_mem_available_bytes": {
                "maximum": 96051191808,
                "mean": 96027980836.14117,
                "median": 96027217920,
                "minimum": 96006701056,
                "sample_count": 85
              },
              "process_pss_bytes": {
                "maximum": 1257277440,
                "mean": 1253214063.4352942,
                "median": 1255912448,
                "minimum": 1243612160,
                "sample_count": 85
              },
              "process_rss_bytes": {
                "maximum": 1260240896,
                "mean": 1256195156.3294117,
                "median": 1258876928,
                "minimum": 1246576640,
                "sample_count": 85
              }
            },
            "process_sample_count": 85,
            "response_union_seconds": 86.81169560202397,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 61440
              },
              "observed_endpoint_seconds": 82.02933916001348,
              "observed_intervals": 5,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 86.84236659400631,
          "response_seconds": {
            "maximum": 17.491869097051676,
            "mean": 17.362339120404794,
            "median": 17.35889967199182,
            "minimum": 17.22252958599711,
            "sample_count": 5
          },
          "response_window_union_seconds": 86.81169560202397,
          "server_decode_tokens_per_second": {
            "maximum": 25.05988575640317,
            "mean": 24.836448251688132,
            "median": 24.82518929206835,
            "minimum": 24.64976311577646,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 7.146799207024742,
            "mean": 7.094700342812575,
            "median": 7.08692858600989,
            "minimum": 7.046769339009188,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        }
      ],
      "id": "offload-72b-full",
      "model": "qwen25-72b-q4",
      "module": "dense-offload",
      "placement": {
        "buffers": [
          {
            "device": "CPU_Mapped",
            "reported_mib": 668.25
          },
          {
            "device": "CUDA0",
            "reported_mib": 41297.69
          }
        ],
        "gpu_layers": 81,
        "meaning": "Runtime buffer allocation is not measured resident process RAM.",
        "requested_cpu_expert_layers": 0,
        "total_layers": 81
      },
      "record_sha256": "f8b20b521b99946cae8150a28b03c57e74bb9a19fc93a5809019ef9e4d3899c0",
      "resources_by_phase": {
        "baseline": {
          "gpu_energy": {
            "available": false,
            "joules": null,
            "reason": "unbracketed_boundary"
          },
          "gpu_sample_count": 2,
          "host_sample_count": 2,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 358690816,
              "mean": 349544448,
              "median": 349544448.0,
              "minimum": 340398080,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 15.22,
              "mean": 14.965,
              "median": 14.965,
              "minimum": 14.71,
              "sample_count": 2
            },
            "gpu_memory_used_bytes": {
              "maximum": 2097152.0,
              "mean": 2097152.0,
              "median": 2097152.0,
              "minimum": 2097152.0,
              "sample_count": 2
            },
            "gpu_sm_clock_mhz": {
              "maximum": 180.0,
              "mean": 180.0,
              "median": 180.0,
              "minimum": 180.0,
              "sample_count": 2
            },
            "gpu_utilization_percent": {
              "maximum": 0.0,
              "mean": 0.0,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 2
            },
            "host_mem_available_bytes": {
              "maximum": 96567312384,
              "mean": 96558721024,
              "median": 96558721024.0,
              "minimum": 96550129664,
              "sample_count": 2
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1008167749969289,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 1
          }
        },
        "cleanup": {
          "gpu_energy": {
            "available": false,
            "joules": null,
            "reason": "unbracketed_boundary"
          },
          "gpu_sample_count": 1,
          "host_sample_count": 1,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 352747520,
              "mean": 352747520,
              "median": 352747520,
              "minimum": 352747520,
              "sample_count": 1
            },
            "gpu_board_watts": {
              "maximum": 103.77,
              "mean": 103.77,
              "median": 103.77,
              "minimum": 103.77,
              "sample_count": 1
            },
            "gpu_memory_used_bytes": {
              "maximum": 2097152.0,
              "mean": 2097152.0,
              "median": 2097152.0,
              "minimum": 2097152.0,
              "sample_count": 1
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2340.0,
              "mean": 2340.0,
              "median": 2340.0,
              "minimum": 2340.0,
              "sample_count": 1
            },
            "gpu_utilization_percent": {
              "maximum": 0.0,
              "mean": 0.0,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 1
            },
            "host_mem_available_bytes": {
              "maximum": 96511545344,
              "mean": 96511545344,
              "median": 96511545344,
              "minimum": 96511545344,
              "sample_count": 1
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1564114179927856,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 1
          }
        },
        "load": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 478.07168217720226,
            "reason": null,
            "response_union_seconds": 7.210961889009923,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 66.29790720511522,
            "trapezoid_pieces": 8,
            "watt_hours": 0.1327976894936673
          },
          "gpu_sample_count": 7,
          "host_sample_count": 4,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 656150528,
              "mean": 638523392,
              "median": 638926848.0,
              "minimum": 620089344,
              "sample_count": 4
            },
            "gpu_board_watts": {
              "maximum": 74.94,
              "mean": 69.43428571428572,
              "median": 74.31,
              "minimum": 42.68,
              "sample_count": 7
            },
            "gpu_memory_used_bytes": {
              "maximum": 43891294208.0,
              "mean": 37704995401.14286,
              "median": 43891294208.0,
              "minimum": 587202560.0,
              "sample_count": 7
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2295.0,
              "mean": 2291.714285714286,
              "median": 2295.0,
              "minimum": 2272.0,
              "sample_count": 7
            },
            "gpu_utilization_percent": {
              "maximum": 83.0,
              "mean": 58.857142857142854,
              "median": 82.0,
              "minimum": 0.0,
              "sample_count": 7
            },
            "host_mem_available_bytes": {
              "maximum": 96185810944,
              "mean": 96155684864,
              "median": 96151988224.0,
              "minimum": 96132952064,
              "sample_count": 4
            },
            "process_pss_bytes": {
              "maximum": 44348953600,
              "mean": 42197760000,
              "median": 44348720128.0,
              "minimum": 35744646144,
              "sample_count": 4
            },
            "process_rss_bytes": {
              "maximum": 44351918080,
              "mean": 42200724480,
              "median": 44351684608.0,
              "minimum": 35747610624,
              "sample_count": 4
            }
          },
          "process_sample_count": 4,
          "response_union_seconds": 7.210961889009923,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 5,
              "storage_read_bytes": 0,
              "storage_write_bytes": 0
            },
            "observed_endpoint_seconds": 5.209772021975368,
            "observed_intervals": 1,
            "requested_intervals": 1
          }
        },
        "ready": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 551.952217391011,
            "reason": null,
            "response_union_seconds": 3.3186683430103585,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 166.3173780391493,
            "trapezoid_pieces": 6,
            "watt_hours": 0.15332006038639195
          },
          "gpu_sample_count": 3,
          "host_sample_count": 3,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 769622016,
              "mean": 725244586.6666666,
              "median": 737775616,
              "minimum": 668336128,
              "sample_count": 3
            },
            "gpu_board_watts": {
              "maximum": 205.83,
              "mean": 158.66,
              "median": 175.53,
              "minimum": 94.62,
              "sample_count": 3
            },
            "gpu_memory_used_bytes": {
              "maximum": 47104131072.0,
              "mean": 47093645312.0,
              "median": 47104131072.0,
              "minimum": 47072673792.0,
              "sample_count": 3
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2340.0,
              "mean": 2314.6666666666665,
              "median": 2332.0,
              "minimum": 2272.0,
              "sample_count": 3
            },
            "gpu_utilization_percent": {
              "maximum": 0.0,
              "mean": 0.0,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 3
            },
            "host_mem_available_bytes": {
              "maximum": 96088260608,
              "mean": 96051089408,
              "median": 96037539840,
              "minimum": 96027467776,
              "sample_count": 3
            },
            "process_pss_bytes": {
              "maximum": 1257276416,
              "mean": 1216586752,
              "median": 1231762432,
              "minimum": 1160721408,
              "sample_count": 3
            },
            "process_rss_bytes": {
              "maximum": 1260240896,
              "mean": 1219551232,
              "median": 1234726912,
              "minimum": 1163685888,
              "sample_count": 3
            }
          },
          "process_sample_count": 3,
          "response_union_seconds": 3.3186683430103585,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 3
          }
        },
        "responses": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 43336.282681437755,
            "reason": null,
            "response_union_seconds": 144.51526583201485,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 299.87339006670777,
            "trapezoid_pieces": 147,
            "watt_hours": 12.037856300399376
          },
          "gpu_sample_count": 145,
          "host_sample_count": 142,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 783441920,
              "mean": 744136242.4788733,
              "median": 753983488.0,
              "minimum": 708935680,
              "sample_count": 142
            },
            "gpu_board_watts": {
              "maximum": 302.93,
              "mean": 300.00172413793103,
              "median": 300.0,
              "minimum": 297.31,
              "sample_count": 145
            },
            "gpu_memory_used_bytes": {
              "maximum": 47104131072.0,
              "mean": 47104131072.0,
              "median": 47104131072.0,
              "minimum": 47104131072.0,
              "sample_count": 145
            },
            "gpu_sm_clock_mhz": {
              "maximum": 1635.0,
              "mean": 1481.2551724137932,
              "median": 1485.0,
              "minimum": 1312.0,
              "sample_count": 145
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 99.84827586206896,
              "median": 100.0,
              "minimum": 83.0,
              "sample_count": 145
            },
            "host_mem_available_bytes": {
              "maximum": 96093003776,
              "mean": 96040983393.35211,
              "median": 96037416960.0,
              "minimum": 96006701056,
              "sample_count": 142
            },
            "process_pss_bytes": {
              "maximum": 1257277440,
              "mean": 1236598834.4788733,
              "median": 1246618624.0,
              "minimum": 1205324800,
              "sample_count": 142
            },
            "process_rss_bytes": {
              "maximum": 1260240896,
              "mean": 1239593392.6760564,
              "median": 1249583104.0,
              "minimum": 1209503744,
              "sample_count": 142
            }
          },
          "process_sample_count": 142,
          "response_union_seconds": 144.51526583201485,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 122880
            },
            "observed_endpoint_seconds": 143.58244017901598,
            "observed_intervals": 2,
            "requested_intervals": 2
          }
        },
        "warmup": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 8285.867276863784,
            "reason": null,
            "response_union_seconds": 28.43168306199368,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 291.4307696381153,
            "trapezoid_pieces": 30,
            "watt_hours": 2.301629799128829
          },
          "gpu_sample_count": 28,
          "host_sample_count": 27,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 748781568,
              "mean": 728878952.2962962,
              "median": 738770944,
              "minimum": 700723200,
              "sample_count": 27
            },
            "gpu_board_watts": {
              "maximum": 300.16,
              "mean": 292.4089285714286,
              "median": 300.0,
              "minimum": 186.04,
              "sample_count": 28
            },
            "gpu_memory_used_bytes": {
              "maximum": 47104131072.0,
              "mean": 47104131072.0,
              "median": 47104131072.0,
              "minimum": 47104131072.0,
              "sample_count": 28
            },
            "gpu_sm_clock_mhz": {
              "maximum": 1642.0,
              "mean": 1531.3214285714287,
              "median": 1511.0,
              "minimum": 1492.0,
              "sample_count": 28
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 100.0,
              "median": 100.0,
              "minimum": 100.0,
              "sample_count": 28
            },
            "host_mem_available_bytes": {
              "maximum": 96063094784,
              "mean": 96039131060.14815,
              "median": 96037965824,
              "minimum": 96024268800,
              "sample_count": 27
            },
            "process_pss_bytes": {
              "maximum": 1242903552,
              "mean": 1224316662.5185184,
              "median": 1232471040,
              "minimum": 1196975104,
              "sample_count": 27
            },
            "process_rss_bytes": {
              "maximum": 1245868032,
              "mean": 1227281142.5185184,
              "median": 1235435520,
              "minimum": 1199939584,
              "sample_count": 27
            }
          },
          "process_sample_count": 27,
          "response_union_seconds": 28.43168306199368,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 12288
            },
            "observed_endpoint_seconds": 25.652496571012307,
            "observed_intervals": 2,
            "requested_intervals": 2
          }
        }
      },
      "reviewed_export_sha256": "4b17dd2a9c39d7cab209a102bfdbd49f8324b54625bc01cbe7d9aeaf6b87dd5a",
      "settings": {
        "batch": 2048,
        "cache_prompt": false,
        "cache_ram_bytes": 0,
        "concurrency": 1,
        "context_shift": false,
        "contexts": [
          2048,
          8192
        ],
        "endpoint": "/completion",
        "fit": "off",
        "flash_attention": true,
        "ignore_eos": true,
        "kv_cache": "f16",
        "lazy_mode": "off",
        "load_mode": "mmap",
        "measured_rounds": 5,
        "mlock": false,
        "network_scope": "loopback",
        "output_tokens": 256,
        "p95_minimum_requests": 200,
        "placement": {
          "cpu_expert_layers": 0,
          "gpu_layers": "all"
        },
        "reasoning": "not_applicable_raw_completion",
        "seed": 20260927,
        "server_slots": 1,
        "synthetic_prompt_sha256": "c099b5abc034050fac1133e0214990d1f72a124b64f36e8a80ba3209bc07d187",
        "temperature": 0,
        "threads": 16,
        "tokenization": "model_tokenizer_add_special_false_repeat_then_truncate",
        "total_context_tokens": 8704,
        "ubatch": 512,
        "warmup_rounds": 1
      },
      "status": "completed",
      "telemetry": {
        "gpu_samples": 186,
        "host_samples": 179,
        "maximum_gpu_gap_seconds": 1.0065537570044398,
        "maximum_host_gap_seconds": 1.7706437300075777,
        "meaning": "Sampled observations; no numeric temperatures published and no warranty between samples.",
        "thermal_slowdown_fully_available": true,
        "thermal_slowdown_observed_samples": 0,
        "thermal_slowdown_states": {
          "hardware": {
            "active": 0,
            "not_active": 186,
            "unavailable": 0
          },
          "software": {
            "active": 0,
            "not_active": 186,
            "unavailable": 0
          }
        }
      }
    },
    {
      "case_wall_seconds": 3542.812662684999,
      "completed_response_resources": {
        "gpu_energy": {
          "available": true,
          "interpolation": "linear_at_bracketed_boundaries",
          "joules": 236530.70309813673,
          "reason": null,
          "response_union_seconds": 2749.5454220350366,
          "scope": "GPU board only; not whole-server AC energy",
          "time_weighted_mean_watts": 86.02538485182468,
          "trapezoid_pieces": 2760,
          "watt_hours": 65.70297308281576
        },
        "gpu_sample_count": 2750,
        "host_sample_count": 1620,
        "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
        "metrics": {
          "cgroup_memory_current_bytes": {
            "maximum": 2323677184,
            "mean": 2286991319.545679,
            "median": 2292633600.0,
            "minimum": 2259800064,
            "sample_count": 1620
          },
          "gpu_board_watts": {
            "maximum": 152.49,
            "mean": 86.01490909090909,
            "median": 81.61,
            "minimum": 79.5,
            "sample_count": 2750
          },
          "gpu_memory_used_bytes": {
            "maximum": 24192745472.0,
            "mean": 24186094068.456726,
            "median": 24192745472.0,
            "minimum": 24178065408.0,
            "sample_count": 2750
          },
          "gpu_sm_clock_mhz": {
            "maximum": 2332.0,
            "mean": 2263.0723636363637,
            "median": 2272.0,
            "minimum": 1785.0,
            "sample_count": 2750
          },
          "gpu_utilization_percent": {
            "maximum": 100.0,
            "mean": 11.226545454545455,
            "median": 0.0,
            "minimum": 0.0,
            "sample_count": 2750
          },
          "host_mem_available_bytes": {
            "maximum": 94687850496,
            "mean": 94599706644.22716,
            "median": 94597986304.0,
            "minimum": 94527299584,
            "sample_count": 1620
          },
          "process_pss_bytes": {
            "maximum": 24359494656,
            "mean": 24348639060.701233,
            "median": 24354731008.0,
            "minimum": 24331427840,
            "sample_count": 1620
          },
          "process_rss_bytes": {
            "maximum": 24362508288,
            "mean": 24351724731.101234,
            "median": 24357744640.0,
            "minimum": 24335863808,
            "sample_count": 1620
          }
        },
        "process_sample_count": 1620,
        "response_union_seconds": 2749.5454220350366,
        "sampled_counter_deltas": {
          "deltas": {
            "major_faults": 0,
            "read_syscall_bytes": 0,
            "storage_read_bytes": 0,
            "storage_write_bytes": 1228800
          },
          "observed_endpoint_seconds": 2732.816400779062,
          "observed_intervals": 10,
          "requested_intervals": 10
        }
      },
      "groups": [
        {
          "aggregate_output_tokens_per_second": 1.0272049883875025,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 2048,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 46.961556060353814,
            "interval_count": 5,
            "interval_seconds": 1246.1183393240208,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 1502.769793931322,
            "process_cpu_seconds": 18726.29
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 103758.67294673451,
              "reason": null,
              "response_union_seconds": 1246.099867573008,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 83.2667394057446,
              "trapezoid_pieces": 1251,
              "watt_hours": 28.821853596315144
            },
            "gpu_sample_count": 1246,
            "host_sample_count": 696,
            "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
            "metrics": {
              "cgroup_memory_current_bytes": {
                "maximum": 2293600256,
                "mean": 2270765685.7011495,
                "median": 2271793152.0,
                "minimum": 2259800064,
                "sample_count": 696
              },
              "gpu_board_watts": {
                "maximum": 146.64,
                "mean": 83.25452648475121,
                "median": 80.98,
                "minimum": 79.5,
                "sample_count": 1246
              },
              "gpu_memory_used_bytes": {
                "maximum": 24178065408.0,
                "mean": 24178065408.0,
                "median": 24178065408.0,
                "minimum": 24178065408.0,
                "sample_count": 1246
              },
              "gpu_sm_clock_mhz": {
                "maximum": 2325.0,
                "mean": 2263.345104333868,
                "median": 2272.0,
                "minimum": 1852.0,
                "sample_count": 1246
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 5.826645264847512,
                "median": 0.0,
                "minimum": 0.0,
                "sample_count": 1246
              },
              "host_mem_available_bytes": {
                "maximum": 94687850496,
                "mean": 94607496833.47127,
                "median": 94607792128.0,
                "minimum": 94549704704,
                "sample_count": 696
              },
              "process_pss_bytes": {
                "maximum": 24341984256,
                "mean": 24338712246.436783,
                "median": 24339825664.0,
                "minimum": 24331427840,
                "sample_count": 696
              },
              "process_rss_bytes": {
                "maximum": 24344997888,
                "mean": 24341802213.517242,
                "median": 24342839296.0,
                "minimum": 24335863808,
                "sample_count": 696
              }
            },
            "process_sample_count": 696,
            "response_union_seconds": 1246.099867573008,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 565248
              },
              "observed_endpoint_seconds": 1236.6610553499777,
              "observed_intervals": 5,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 1246.124862863042,
          "response_seconds": {
            "maximum": 249.32341260602698,
            "mean": 249.21997351460158,
            "median": 249.17530603701016,
            "minimum": 249.15829364600359,
            "sample_count": 5
          },
          "response_window_union_seconds": 1246.099867573008,
          "server_decode_tokens_per_second": {
            "maximum": 1.0771389966646243,
            "mean": 1.0769035802653517,
            "median": 1.0769342620955185,
            "minimum": 1.0766185877515206,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 12.426381879020482,
            "mean": 12.409445786592551,
            "median": 12.411477557965554,
            "minimum": 12.391928925993852,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        },
        {
          "aggregate_output_tokens_per_second": 0.8513776878724529,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 8192,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 41.45806174570441,
            "interval_count": 5,
            "interval_seconds": 1503.4704017839977,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 1326.6579758625412,
            "process_cpu_seconds": 19945.91
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 132772.0301514023,
              "reason": null,
              "response_union_seconds": 1503.4455544620287,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 88.31183128471288,
              "trapezoid_pieces": 1509,
              "watt_hours": 36.88111948650064
            },
            "gpu_sample_count": 1504,
            "host_sample_count": 924,
            "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
            "metrics": {
              "cgroup_memory_current_bytes": {
                "maximum": 2323677184,
                "mean": 2299213225.5584416,
                "median": 2299183104.0,
                "minimum": 2290978816,
                "sample_count": 924
              },
              "gpu_board_watts": {
                "maximum": 152.49,
                "mean": 88.30176861702128,
                "median": 82.01,
                "minimum": 80.39,
                "sample_count": 1504
              },
              "gpu_memory_used_bytes": {
                "maximum": 24192745472.0,
                "mean": 24192745472.0,
                "median": 24192745472.0,
                "minimum": 24192745472.0,
                "sample_count": 1504
              },
              "gpu_sm_clock_mhz": {
                "maximum": 2332.0,
                "mean": 2262.846409574468,
                "median": 2272.0,
                "minimum": 1785.0,
                "sample_count": 1504
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 15.700132978723405,
                "median": 0.0,
                "minimum": 0.0,
                "sample_count": 1504
              },
              "host_mem_available_bytes": {
                "maximum": 94663417856,
                "mean": 94593838709.47186,
                "median": 94591438848.0,
                "minimum": 94527299584,
                "sample_count": 924
              },
              "process_pss_bytes": {
                "maximum": 24359494656,
                "mean": 24356116401.316017,
                "median": 24356086784.0,
                "minimum": 24353292288,
                "sample_count": 924
              },
              "process_rss_bytes": {
                "maximum": 24362508288,
                "mean": 24359198835.255413,
                "median": 24359100416.0,
                "minimum": 24357724160,
                "sample_count": 924
              }
            },
            "process_sample_count": 924,
            "response_union_seconds": 1503.4455544620287,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 663552
              },
              "observed_endpoint_seconds": 1496.1553454290843,
              "observed_intervals": 5,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 1503.4815187479835,
          "response_seconds": {
            "maximum": 300.788018139021,
            "mean": 300.6891108924057,
            "median": 300.6939707050333,
            "minimum": 300.5792841439834,
            "sample_count": 5
          },
          "response_window_union_seconds": 1503.4455544620287,
          "server_decode_tokens_per_second": {
            "maximum": 1.02139510723465,
            "mean": 1.0210610239530291,
            "median": 1.0211173638020445,
            "minimum": 1.0206264196713375,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 50.99202599900309,
            "mean": 50.948651273804714,
            "median": 50.94126555201365,
            "minimum": 50.9204786689952,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        }
      ],
      "id": "offload-72b-40",
      "model": "qwen25-72b-q4",
      "module": "dense-offload",
      "placement": {
        "buffers": [
          {
            "device": "CPU_Mapped",
            "reported_mib": 3796.51
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3698.21
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3780.0
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3758.87
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3780.06
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 2489.76
          },
          {
            "device": "CUDA0",
            "reported_mib": 20662.52
          }
        ],
        "gpu_layers": 40,
        "meaning": "Runtime buffer allocation is not measured resident process RAM.",
        "requested_cpu_expert_layers": 0,
        "total_layers": 81
      },
      "record_sha256": "6e94d76fd5f84834595a846d9cdb805969e602d7e4891c723a3d7f43a132755d",
      "resources_by_phase": {
        "baseline": {
          "gpu_energy": {
            "available": false,
            "joules": null,
            "reason": "unbracketed_boundary"
          },
          "gpu_sample_count": 2,
          "host_sample_count": 2,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 376336384,
              "mean": 368064512,
              "median": 368064512.0,
              "minimum": 359792640,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 15.95,
              "mean": 15.95,
              "median": 15.95,
              "minimum": 15.95,
              "sample_count": 2
            },
            "gpu_memory_used_bytes": {
              "maximum": 2097152.0,
              "mean": 2097152.0,
              "median": 2097152.0,
              "minimum": 2097152.0,
              "sample_count": 2
            },
            "gpu_sm_clock_mhz": {
              "maximum": 180.0,
              "mean": 180.0,
              "median": 180.0,
              "minimum": 180.0,
              "sample_count": 2
            },
            "gpu_utilization_percent": {
              "maximum": 0.0,
              "mean": 0.0,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 2
            },
            "host_mem_available_bytes": {
              "maximum": 96496439296,
              "mean": 96496254976,
              "median": 96496254976.0,
              "minimum": 96496070656,
              "sample_count": 2
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1007298770127818,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 1
          }
        },
        "cleanup": {
          "gpu_energy": {
            "available": false,
            "joules": null,
            "reason": "unbracketed_boundary"
          },
          "gpu_sample_count": 1,
          "host_sample_count": 1,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 603852800,
              "mean": 603852800,
              "median": 603852800,
              "minimum": 603852800,
              "sample_count": 1
            },
            "gpu_board_watts": {
              "maximum": 79.95,
              "mean": 79.95,
              "median": 79.95,
              "minimum": 79.95,
              "sample_count": 1
            },
            "gpu_memory_used_bytes": {
              "maximum": 2097152.0,
              "mean": 2097152.0,
              "median": 2097152.0,
              "minimum": 2097152.0,
              "sample_count": 1
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2272.0,
              "mean": 2272.0,
              "median": 2272.0,
              "minimum": 2272.0,
              "sample_count": 1
            },
            "gpu_utilization_percent": {
              "maximum": 0.0,
              "mean": 0.0,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 1
            },
            "host_mem_available_bytes": {
              "maximum": 96244224000,
              "mean": 96244224000,
              "median": 96244224000,
              "minimum": 96244224000,
              "sample_count": 1
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1574398530065082,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 1
          }
        },
        "load": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 481.14320595099116,
            "reason": null,
            "response_union_seconds": 6.915469526022207,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 69.57491521587917,
            "trapezoid_pieces": 7,
            "watt_hours": 0.13365089054194199
          },
          "gpu_sample_count": 6,
          "host_sample_count": 5,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 2173902848,
              "mean": 1243803648,
              "median": 657190912,
              "minimum": 605564928,
              "sample_count": 5
            },
            "gpu_board_watts": {
              "maximum": 79.13,
              "mean": 72.73333333333333,
              "median": 76.555,
              "minimum": 51.58,
              "sample_count": 6
            },
            "gpu_memory_used_bytes": {
              "maximum": 24106762240.0,
              "mean": 19248010581.333332,
              "median": 22254977024.0,
              "minimum": 587202560.0,
              "sample_count": 6
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2310.0,
              "mean": 2297.3333333333335,
              "median": 2306.0,
              "minimum": 2265.0,
              "sample_count": 6
            },
            "gpu_utilization_percent": {
              "maximum": 81.0,
              "mean": 40.333333333333336,
              "median": 40.0,
              "minimum": 0.0,
              "sample_count": 6
            },
            "host_mem_available_bytes": {
              "maximum": 96189173760,
              "mean": 95540204339.2,
              "median": 96136634368,
              "minimum": 94583787520,
              "sample_count": 5
            },
            "process_pss_bytes": {
              "maximum": 44348638208,
              "mean": 31194135961.6,
              "median": 24240718848,
              "minimum": 18852537344,
              "sample_count": 5
            },
            "process_rss_bytes": {
              "maximum": 44351598592,
              "mean": 31197117644.8,
              "median": 24243732480,
              "minimum": 18855497728,
              "sample_count": 5
            }
          },
          "process_sample_count": 5,
          "response_union_seconds": 6.915469526022207,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 13,
              "storage_read_bytes": 0,
              "storage_write_bytes": 8192
            },
            "observed_endpoint_seconds": 5.654299405985512,
            "observed_intervals": 1,
            "requested_intervals": 1
          }
        },
        "ready": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 272.823505245464,
            "reason": null,
            "response_union_seconds": 3.400849792989902,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 80.22215677029581,
            "trapezoid_pieces": 8,
            "watt_hours": 0.0757843070126289
          },
          "gpu_sample_count": 5,
          "host_sample_count": 2,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 2307264512,
              "mean": 2245849088,
              "median": 2245849088.0,
              "minimum": 2184433664,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 85.62,
              "mean": 80.17,
              "median": 80.06,
              "minimum": 77.05,
              "sample_count": 5
            },
            "gpu_memory_used_bytes": {
              "maximum": 24192745472.0,
              "mean": 24154157875.2,
              "median": 24178065408.0,
              "minimum": 24110956544.0,
              "sample_count": 5
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2272.0,
              "mean": 2264.6,
              "median": 2272.0,
              "minimum": 2235.0,
              "sample_count": 5
            },
            "gpu_utilization_percent": {
              "maximum": 1.0,
              "mean": 0.2,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 5
            },
            "host_mem_available_bytes": {
              "maximum": 94570885120,
              "mean": 94556760064,
              "median": 94556760064.0,
              "minimum": 94542635008,
              "sample_count": 2
            },
            "process_pss_bytes": {
              "maximum": 24359494656,
              "mean": 24305228800,
              "median": 24305228800.0,
              "minimum": 24250962944,
              "sample_count": 2
            },
            "process_rss_bytes": {
              "maximum": 24362508288,
              "mean": 24308242432,
              "median": 24308242432.0,
              "minimum": 24253976576,
              "sample_count": 2
            }
          },
          "process_sample_count": 2,
          "response_union_seconds": 3.400849792989902,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 3
          }
        },
        "responses": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 236537.1246580265,
            "reason": null,
            "response_union_seconds": 2749.6215623870376,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 86.02533813878037,
            "trapezoid_pieces": 2752,
            "watt_hours": 65.70475684945181
          },
          "gpu_sample_count": 2750,
          "host_sample_count": 1622,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 2323677184,
              "mean": 2286969816.8581996,
              "median": 2292627456.0,
              "minimum": 2259800064,
              "sample_count": 1622
            },
            "gpu_board_watts": {
              "maximum": 152.49,
              "mean": 86.01490909090909,
              "median": 81.61,
              "minimum": 79.5,
              "sample_count": 2750
            },
            "gpu_memory_used_bytes": {
              "maximum": 24192745472.0,
              "mean": 24186094068.456726,
              "median": 24192745472.0,
              "minimum": 24178065408.0,
              "sample_count": 2750
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2332.0,
              "mean": 2263.0723636363637,
              "median": 2272.0,
              "minimum": 1785.0,
              "sample_count": 2750
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 11.226545454545455,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 2750
            },
            "host_mem_available_bytes": {
              "maximum": 94687850496,
              "mean": 94599713823.56596,
              "median": 94597986304.0,
              "minimum": 94527299584,
              "sample_count": 1622
            },
            "process_pss_bytes": {
              "maximum": 24359494656,
              "mean": 24348625269.10974,
              "median": 24354731008.0,
              "minimum": 24331427840,
              "sample_count": 1622
            },
            "process_rss_bytes": {
              "maximum": 24362508288,
              "mean": 24351710850.683105,
              "median": 24357744640.0,
              "minimum": 24335863808,
              "sample_count": 1622
            }
          },
          "process_sample_count": 1622,
          "response_union_seconds": 2749.6215623870376,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 1261568
            },
            "observed_endpoint_seconds": 2746.6284927970264,
            "observed_intervals": 2,
            "requested_intervals": 2
          }
        },
        "warmup": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 46701.3050646631,
            "reason": null,
            "response_union_seconds": 550.41680468095,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 84.84716430802578,
            "trapezoid_pieces": 551,
            "watt_hours": 12.972584740184194
          },
          "gpu_sample_count": 549,
          "host_sample_count": 327,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 2307465216,
              "mean": 2272184376.3669724,
              "median": 2278948864,
              "minimum": 2212904960,
              "sample_count": 327
            },
            "gpu_board_watts": {
              "maximum": 152.14,
              "mean": 84.84987249544626,
              "median": 80.91,
              "minimum": 76.67,
              "sample_count": 549
            },
            "gpu_memory_used_bytes": {
              "maximum": 24192745472.0,
              "mean": 24185243092.167576,
              "median": 24192745472.0,
              "minimum": 24142413824.0,
              "sample_count": 549
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2332.0,
              "mean": 2263.2768670309656,
              "median": 2272.0,
              "minimum": 1792.0,
              "sample_count": 549
            },
            "gpu_utilization_percent": {
              "maximum": 96.0,
              "mean": 11.145719489981785,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 549
            },
            "host_mem_available_bytes": {
              "maximum": 94694584320,
              "mean": 94598962323.18044,
              "median": 94595268608,
              "minimum": 94545797120,
              "sample_count": 327
            },
            "process_pss_bytes": {
              "maximum": 24354034688,
              "mean": 24339138647.681957,
              "median": 24342766592,
              "minimum": 24285017088,
              "sample_count": 327
            },
            "process_rss_bytes": {
              "maximum": 24357048320,
              "mean": 24342234703.85321,
              "median": 24345780224,
              "minimum": 24288030720,
              "sample_count": 327
            }
          },
          "process_sample_count": 327,
          "response_union_seconds": 550.41680468095,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 249856
            },
            "observed_endpoint_seconds": 548.5689892289811,
            "observed_intervals": 2,
            "requested_intervals": 2
          }
        }
      },
      "reviewed_export_sha256": "4b17dd2a9c39d7cab209a102bfdbd49f8324b54625bc01cbe7d9aeaf6b87dd5a",
      "settings": {
        "batch": 2048,
        "cache_prompt": false,
        "cache_ram_bytes": 0,
        "concurrency": 1,
        "context_shift": false,
        "contexts": [
          2048,
          8192
        ],
        "endpoint": "/completion",
        "fit": "off",
        "flash_attention": true,
        "ignore_eos": true,
        "kv_cache": "f16",
        "lazy_mode": "off",
        "load_mode": "mmap",
        "measured_rounds": 5,
        "mlock": false,
        "network_scope": "loopback",
        "output_tokens": 256,
        "p95_minimum_requests": 200,
        "placement": {
          "cpu_expert_layers": 0,
          "gpu_layers": 40
        },
        "reasoning": "not_applicable_raw_completion",
        "seed": 20260927,
        "server_slots": 1,
        "synthetic_prompt_sha256": "c099b5abc034050fac1133e0214990d1f72a124b64f36e8a80ba3209bc07d187",
        "temperature": 0,
        "threads": 16,
        "tokenization": "model_tokenizer_add_special_false_repeat_then_truncate",
        "total_context_tokens": 8704,
        "ubatch": 512,
        "warmup_rounds": 1
      },
      "status": "completed",
      "telemetry": {
        "gpu_samples": 3313,
        "host_samples": 1959,
        "maximum_gpu_gap_seconds": 1.0563414199859835,
        "maximum_host_gap_seconds": 1.9819231039728038,
        "meaning": "Sampled observations; no numeric temperatures published and no warranty between samples.",
        "thermal_slowdown_fully_available": true,
        "thermal_slowdown_observed_samples": 0,
        "thermal_slowdown_states": {
          "hardware": {
            "active": 0,
            "not_active": 3313,
            "unavailable": 0
          },
          "software": {
            "active": 0,
            "not_active": 3313,
            "unavailable": 0
          }
        }
      }
    },
    {
      "case_wall_seconds": 4961.106233133993,
      "completed_response_resources": {
        "gpu_energy": {
          "available": true,
          "interpolation": "linear_at_bracketed_boundaries",
          "joules": 326983.819602327,
          "reason": null,
          "response_union_seconds": 4008.546728375077,
          "scope": "GPU board only; not whole-server AC energy",
          "time_weighted_mean_watts": 81.57166219062006,
          "trapezoid_pieces": 4019,
          "watt_hours": 90.82883877842417
        },
        "gpu_sample_count": 4009,
        "host_sample_count": 2035,
        "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
        "metrics": {
          "cgroup_memory_current_bytes": {
            "maximum": 3084062720,
            "mean": 3045302493.4054055,
            "median": 3055714304,
            "minimum": 3018514432,
            "sample_count": 2035
          },
          "gpu_board_watts": {
            "maximum": 129.75,
            "mean": 81.56763532052881,
            "median": 78.23,
            "minimum": 74.81,
            "sample_count": 4009
          },
          "gpu_memory_used_bytes": {
            "maximum": 13128171520.0,
            "mean": 13116759330.418558,
            "median": 13128171520.0,
            "minimum": 13103005696.0,
            "sample_count": 4009
          },
          "gpu_sm_clock_mhz": {
            "maximum": 2332.0,
            "mean": 2271.545522574208,
            "median": 2272.0,
            "minimum": 2100.0,
            "sample_count": 4009
          },
          "gpu_utilization_percent": {
            "maximum": 100.0,
            "mean": 11.124220503866301,
            "median": 0.0,
            "minimum": 0.0,
            "sample_count": 4009
          },
          "host_mem_available_bytes": {
            "maximum": 93943697408,
            "mean": 93853951377.0457,
            "median": 93853831168,
            "minimum": 93781438464,
            "sample_count": 2035
          },
          "process_pss_bytes": {
            "maximum": 35444739072,
            "mean": 35430992087.36707,
            "median": 35441887232,
            "minimum": 35412776960,
            "sample_count": 2035
          },
          "process_rss_bytes": {
            "maximum": 35447758848,
            "mean": 35434078916.4973,
            "median": 35444908032,
            "minimum": 35416961024,
            "sample_count": 2035
          }
        },
        "process_sample_count": 2035,
        "response_union_seconds": 4008.546728375077,
        "sampled_counter_deltas": {
          "deltas": {
            "major_faults": 0,
            "read_syscall_bytes": 0,
            "storage_read_bytes": 0,
            "storage_write_bytes": 1728512
          },
          "observed_endpoint_seconds": 3988.5196464579785,
          "observed_intervals": 10,
          "requested_intervals": 10
        }
      },
      "groups": [
        {
          "aggregate_output_tokens_per_second": 0.7043362587375134,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 2048,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 47.48170472432676,
            "interval_count": 5,
            "interval_seconds": 1817.331549088005,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 1519.4145511784563,
            "process_cpu_seconds": 27612.8
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 144503.54057542377,
              "reason": null,
              "response_union_seconds": 1817.31379596208,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 79.5149087056394,
              "trapezoid_pieces": 1823,
              "watt_hours": 40.13987238206216
            },
            "gpu_sample_count": 1818,
            "host_sample_count": 890,
            "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
            "metrics": {
              "cgroup_memory_current_bytes": {
                "maximum": 3044786176,
                "mean": 3024263241.635955,
                "median": 3023706112.0,
                "minimum": 3018514432,
                "sample_count": 890
              },
              "gpu_board_watts": {
                "maximum": 125.38,
                "mean": 79.50995599559955,
                "median": 78.14500000000001,
                "minimum": 75.15,
                "sample_count": 1818
              },
              "gpu_memory_used_bytes": {
                "maximum": 13103005696.0,
                "mean": 13103005696.0,
                "median": 13103005696.0,
                "minimum": 13103005696.0,
                "sample_count": 1818
              },
              "gpu_sm_clock_mhz": {
                "maximum": 2325.0,
                "mean": 2270.7398239823983,
                "median": 2272.0,
                "minimum": 2122.0,
                "sample_count": 1818
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 5.860836083608361,
                "median": 0.0,
                "minimum": 0.0,
                "sample_count": 1818
              },
              "host_mem_available_bytes": {
                "maximum": 93943697408,
                "mean": 93876701759.2809,
                "median": 93871968256.0,
                "minimum": 93820104704,
                "sample_count": 890
              },
              "process_pss_bytes": {
                "maximum": 35415967744,
                "mean": 35415209832.12584,
                "median": 35415275520.0,
                "minimum": 35412776960,
                "sample_count": 890
              },
              "process_rss_bytes": {
                "maximum": 35418988544,
                "mean": 35418297640.84494,
                "median": 35418296320.0,
                "minimum": 35416961024,
                "sample_count": 890
              }
            },
            "process_sample_count": 890,
            "response_union_seconds": 1817.31379596208,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 806912
              },
              "observed_endpoint_seconds": 1805.0209762880113,
              "observed_intervals": 5,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 1817.338583304023,
          "response_seconds": {
            "maximum": 363.6143847160274,
            "mean": 363.462759192416,
            "median": 363.4720513680368,
            "minimum": 363.25934934196994,
            "sample_count": 5
          },
          "response_window_union_seconds": 1817.31379596208,
          "server_decode_tokens_per_second": {
            "maximum": 0.7378754690175205,
            "mean": 0.7375854372949258,
            "median": 0.7376701920484424,
            "minimum": 0.7372008255111242,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 17.720808499027044,
            "mean": 17.69595089161303,
            "median": 17.700019541021902,
            "minimum": 17.672509245981928,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        },
        {
          "aggregate_output_tokens_per_second": 0.5841460216602611,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 8192,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 42.039219255208145,
            "interval_count": 5,
            "interval_seconds": 2191.255163203052,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 1345.2550161666607,
            "process_cpu_seconds": 29477.97
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 182480.27902690295,
              "reason": null,
              "response_union_seconds": 2191.232932412997,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 83.27744455079667,
              "trapezoid_pieces": 2196,
              "watt_hours": 50.688966396361934
            },
            "gpu_sample_count": 2191,
            "host_sample_count": 1145,
            "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
            "metrics": {
              "cgroup_memory_current_bytes": {
                "maximum": 3084062720,
                "mean": 3061656147.619214,
                "median": 3061407744,
                "minimum": 3053445120,
                "sample_count": 1145
              },
              "gpu_board_watts": {
                "maximum": 129.75,
                "mean": 83.27501141031493,
                "median": 78.35,
                "minimum": 74.81,
                "sample_count": 2191
              },
              "gpu_memory_used_bytes": {
                "maximum": 13128171520.0,
                "mean": 13128171520.0,
                "median": 13128171520.0,
                "minimum": 13128171520.0,
                "sample_count": 2191
              },
              "gpu_sm_clock_mhz": {
                "maximum": 2332.0,
                "mean": 2272.214057507987,
                "median": 2272.0,
                "minimum": 2100.0,
                "sample_count": 2191
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 15.491556366955727,
                "median": 0.0,
                "minimum": 0.0,
                "sample_count": 2191
              },
              "host_mem_available_bytes": {
                "maximum": 93916975104,
                "mean": 93836267673.82358,
                "median": 93834838016,
                "minimum": 93781438464,
                "sample_count": 1145
              },
              "process_pss_bytes": {
                "maximum": 35444739072,
                "mean": 35443259517.20524,
                "median": 35443341312,
                "minimum": 35440470016,
                "sample_count": 1145
              },
              "process_rss_bytes": {
                "maximum": 35447758848,
                "mean": 35446345584.908295,
                "median": 35446362112,
                "minimum": 35444199424,
                "sample_count": 1145
              }
            },
            "process_sample_count": 1145,
            "response_union_seconds": 2191.232932412997,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 921600
              },
              "observed_endpoint_seconds": 2183.498670169967,
              "observed_intervals": 5,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 2191.2666058779578,
          "response_seconds": {
            "maximum": 438.39019691804424,
            "mean": 438.24658648259935,
            "median": 438.2723452189821,
            "minimum": 438.1088442449691,
            "sample_count": 5
          },
          "response_window_union_seconds": 2191.232932412997,
          "server_decode_tokens_per_second": {
            "maximum": 0.6975875589334964,
            "mean": 0.6973100693905205,
            "median": 0.6972552794747908,
            "minimum": 0.6971042864759415,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 72.59099566203076,
            "mean": 72.55541585101746,
            "median": 72.55246841500048,
            "minimum": 72.52375948702684,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        }
      ],
      "id": "offload-72b-20",
      "model": "qwen25-72b-q4",
      "module": "dense-offload",
      "placement": {
        "buffers": [
          {
            "device": "CPU_Mapped",
            "reported_mib": 3796.51
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3698.21
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3780.0
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3758.87
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3780.06
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3758.87
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3649.56
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 3799.34
          },
          {
            "device": "CPU_Mapped",
            "reported_mib": 1180.14
          },
          {
            "device": "CUDA0",
            "reported_mib": 10764.37
          }
        ],
        "gpu_layers": 20,
        "meaning": "Runtime buffer allocation is not measured resident process RAM.",
        "requested_cpu_expert_layers": 0,
        "total_layers": 81
      },
      "record_sha256": "79ea9153d287fb1c63a4773c4e576d5dead6f1aca8f6f5fadd95c4ed32e82ff3",
      "resources_by_phase": {
        "baseline": {
          "gpu_energy": {
            "available": false,
            "joules": null,
            "reason": "unbracketed_boundary"
          },
          "gpu_sample_count": 2,
          "host_sample_count": 2,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 411516928,
              "mean": 402950144,
              "median": 402950144.0,
              "minimum": 394383360,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 16.41,
              "mean": 16.395,
              "median": 16.395,
              "minimum": 16.38,
              "sample_count": 2
            },
            "gpu_memory_used_bytes": {
              "maximum": 2097152.0,
              "mean": 2097152.0,
              "median": 2097152.0,
              "minimum": 2097152.0,
              "sample_count": 2
            },
            "gpu_sm_clock_mhz": {
              "maximum": 180.0,
              "mean": 176.0,
              "median": 176.0,
              "minimum": 172.0,
              "sample_count": 2
            },
            "gpu_utilization_percent": {
              "maximum": 0.0,
              "mean": 0.0,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 2
            },
            "host_mem_available_bytes": {
              "maximum": 96501207040,
              "mean": 96496873472,
              "median": 96496873472.0,
              "minimum": 96492539904,
              "sample_count": 2
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1008169630076736,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 1
          }
        },
        "cleanup": {
          "gpu_energy": {
            "available": false,
            "joules": null,
            "reason": "unbracketed_boundary"
          },
          "gpu_sample_count": 1,
          "host_sample_count": 1,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 650592256,
              "mean": 650592256,
              "median": 650592256,
              "minimum": 650592256,
              "sample_count": 1
            },
            "gpu_board_watts": {
              "maximum": 77.62,
              "mean": 77.62,
              "median": 77.62,
              "minimum": 77.62,
              "sample_count": 1
            },
            "gpu_memory_used_bytes": {
              "maximum": 2097152.0,
              "mean": 2097152.0,
              "median": 2097152.0,
              "minimum": 2097152.0,
              "sample_count": 1
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2272.0,
              "mean": 2272.0,
              "median": 2272.0,
              "minimum": 2272.0,
              "sample_count": 1
            },
            "gpu_utilization_percent": {
              "maximum": 25.0,
              "mean": 25.0,
              "median": 25.0,
              "minimum": 25.0,
              "sample_count": 1
            },
            "host_mem_available_bytes": {
              "maximum": 96264818688,
              "mean": 96264818688,
              "median": 96264818688,
              "minimum": 96264818688,
              "sample_count": 1
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1522748679853976,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 1
          }
        },
        "load": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 447.90588874900686,
            "reason": null,
            "response_union_seconds": 6.576283341972157,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 68.10927471605636,
            "trapezoid_pieces": 7,
            "watt_hours": 0.12441830243027968
          },
          "gpu_sample_count": 6,
          "host_sample_count": 4,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 2943393792,
              "mean": 1787288576,
              "median": 1792483328.0,
              "minimum": 620793856,
              "sample_count": 4
            },
            "gpu_board_watts": {
              "maximum": 78.7,
              "mean": 71.50666666666666,
              "median": 74.795,
              "minimum": 52.23,
              "sample_count": 6
            },
            "gpu_memory_used_bytes": {
              "maximum": 13014925312.0,
              "mean": 10551121237.333334,
              "median": 12406751232.0,
              "minimum": 587202560.0,
              "sample_count": 6
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2310.0,
              "mean": 2289.6666666666665,
              "median": 2294.5,
              "minimum": 2265.0,
              "sample_count": 6
            },
            "gpu_utilization_percent": {
              "maximum": 83.0,
              "mean": 27.666666666666668,
              "median": 0.5,
              "minimum": 0.0,
              "sample_count": 6
            },
            "host_mem_available_bytes": {
              "maximum": 96198549504,
              "mean": 95033121792,
              "median": 95025416192.0,
              "minimum": 93883105280,
              "sample_count": 4
            },
            "process_pss_bytes": {
              "maximum": 44347988992,
              "mean": 31105929216,
              "median": 35304272896.0,
              "minimum": 9467182080,
              "sample_count": 4
            },
            "process_rss_bytes": {
              "maximum": 44350955520,
              "mean": 31108922368,
              "median": 35307292672.0,
              "minimum": 9470148608,
              "sample_count": 4
            }
          },
          "process_sample_count": 4,
          "response_union_seconds": 6.576283341972157,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 13,
              "storage_read_bytes": 0,
              "storage_write_bytes": 4096
            },
            "observed_endpoint_seconds": 5.295854870986659,
            "observed_intervals": 1,
            "requested_intervals": 1
          }
        },
        "ready": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 264.97328211031623,
            "reason": null,
            "response_union_seconds": 3.4458548309630714,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 76.8962405872042,
            "trapezoid_pieces": 6,
            "watt_hours": 0.07360368947508784
          },
          "gpu_sample_count": 3,
          "host_sample_count": 2,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 3067142144,
              "mean": 3007987712,
              "median": 3007987712.0,
              "minimum": 2948833280,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 78.24,
              "mean": 77.15,
              "median": 78.06,
              "minimum": 75.15,
              "sample_count": 3
            },
            "gpu_memory_used_bytes": {
              "maximum": 13128171520.0,
              "mean": 13082034176.0,
              "median": 13103005696.0,
              "minimum": 13014925312.0,
              "sample_count": 3
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2272.0,
              "mean": 2272.0,
              "median": 2272.0,
              "minimum": 2272.0,
              "sample_count": 3
            },
            "gpu_utilization_percent": {
              "maximum": 0.0,
              "mean": 0.0,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 3
            },
            "host_mem_available_bytes": {
              "maximum": 93878493184,
              "mean": 93867741184,
              "median": 93867741184.0,
              "minimum": 93856989184,
              "sample_count": 2
            },
            "process_pss_bytes": {
              "maximum": 35444738048,
              "mean": 35390990848,
              "median": 35390990848.0,
              "minimum": 35337243648,
              "sample_count": 2
            },
            "process_rss_bytes": {
              "maximum": 35447758848,
              "mean": 35394011136,
              "median": 35394011136.0,
              "minimum": 35340263424,
              "sample_count": 2
            }
          },
          "process_sample_count": 2,
          "response_union_seconds": 3.4458548309630714,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 3
          }
        },
        "responses": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 326989.6537589074,
            "reason": null,
            "response_union_seconds": 4008.619653561036,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 81.57163363414332,
            "trapezoid_pieces": 4011,
            "watt_hours": 90.83045937747428
          },
          "gpu_sample_count": 4009,
          "host_sample_count": 2038,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 3084062720,
              "mean": 3045271785.138371,
              "median": 3055710208.0,
              "minimum": 3018514432,
              "sample_count": 2038
            },
            "gpu_board_watts": {
              "maximum": 129.75,
              "mean": 81.56763532052881,
              "median": 78.23,
              "minimum": 74.81,
              "sample_count": 4009
            },
            "gpu_memory_used_bytes": {
              "maximum": 13128171520.0,
              "mean": 13116759330.418558,
              "median": 13128171520.0,
              "minimum": 13103005696.0,
              "sample_count": 4009
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2332.0,
              "mean": 2271.545522574208,
              "median": 2272.0,
              "minimum": 2100.0,
              "sample_count": 4009
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 11.124220503866301,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 4009
            },
            "host_mem_available_bytes": {
              "maximum": 93943697408,
              "mean": 93854005643.93327,
              "median": 93853870080.0,
              "minimum": 93781438464,
              "sample_count": 2038
            },
            "process_pss_bytes": {
              "maximum": 35444739072,
              "mean": 35430969175.17566,
              "median": 35441887232.0,
              "minimum": 35412776960,
              "sample_count": 2038
            },
            "process_rss_bytes": {
              "maximum": 35447758848,
              "mean": 35434055907.10893,
              "median": 35444908032.0,
              "minimum": 35416961024,
              "sample_count": 2038
            }
          },
          "process_sample_count": 2038,
          "response_union_seconds": 4008.619653561036,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 1781760
            },
            "observed_endpoint_seconds": 4006.5016962169902,
            "observed_intervals": 2,
            "requested_intervals": 2
          }
        },
        "warmup": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 65077.663397590804,
            "reason": null,
            "response_union_seconds": 802.8343216970097,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 81.05989198373007,
            "trapezoid_pieces": 805,
            "watt_hours": 18.077128721553002
          },
          "gpu_sample_count": 803,
          "host_sample_count": 409,
          "meaning": "Samples inside declared interval union. Means are arithmetic sample means; peaks need not coincide. Counter deltas exclude unsampled edges; not minimum required RAM.",
          "metrics": {
            "cgroup_memory_current_bytes": {
              "maximum": 3071012864,
              "mean": 3032417838.317848,
              "median": 3029204992,
              "minimum": 2971230208,
              "sample_count": 409
            },
            "gpu_board_watts": {
              "maximum": 129.58,
              "mean": 81.05648816936488,
              "median": 77.86,
              "minimum": 72.65,
              "sample_count": 803
            },
            "gpu_memory_used_bytes": {
              "maximum": 13128171520.0,
              "mean": 13115147239.770859,
              "median": 13126074368.0,
              "minimum": 13044285440.0,
              "sample_count": 803
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2325.0,
              "mean": 2270.7372353673723,
              "median": 2272.0,
              "minimum": 2107.0,
              "sample_count": 803
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 11.643835616438356,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 803
            },
            "host_mem_available_bytes": {
              "maximum": 93974298624,
              "mean": 93874269837.45721,
              "median": 93873455104,
              "minimum": 93823176704,
              "sample_count": 409
            },
            "process_pss_bytes": {
              "maximum": 35441178624,
              "mean": 35423200719.17848,
              "median": 35416747008,
              "minimum": 35365141504,
              "sample_count": 409
            },
            "process_rss_bytes": {
              "maximum": 35444199424,
              "mean": 35426312392.293396,
              "median": 35419983872,
              "minimum": 35368161280,
              "sample_count": 409
            }
          },
          "process_sample_count": 409,
          "response_union_seconds": 802.8343216970097,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 339968
            },
            "observed_endpoint_seconds": 799.775153718947,
            "observed_intervals": 2,
            "requested_intervals": 2
          }
        }
      },
      "reviewed_export_sha256": "4b17dd2a9c39d7cab209a102bfdbd49f8324b54625bc01cbe7d9aeaf6b87dd5a",
      "settings": {
        "batch": 2048,
        "cache_prompt": false,
        "cache_ram_bytes": 0,
        "concurrency": 1,
        "context_shift": false,
        "contexts": [
          2048,
          8192
        ],
        "endpoint": "/completion",
        "fit": "off",
        "flash_attention": true,
        "ignore_eos": true,
        "kv_cache": "f16",
        "lazy_mode": "off",
        "load_mode": "mmap",
        "measured_rounds": 5,
        "mlock": false,
        "network_scope": "loopback",
        "output_tokens": 256,
        "p95_minimum_requests": 200,
        "placement": {
          "cpu_expert_layers": 0,
          "gpu_layers": 20
        },
        "reasoning": "not_applicable_raw_completion",
        "seed": 20260927,
        "server_slots": 1,
        "synthetic_prompt_sha256": "c099b5abc034050fac1133e0214990d1f72a124b64f36e8a80ba3209bc07d187",
        "temperature": 0,
        "threads": 16,
        "tokenization": "model_tokenizer_add_special_false_repeat_then_truncate",
        "total_context_tokens": 8704,
        "ubatch": 512,
        "warmup_rounds": 1
      },
      "status": "completed",
      "telemetry": {
        "gpu_samples": 4824,
        "host_samples": 2456,
        "maximum_gpu_gap_seconds": 1.0568395520094782,
        "maximum_host_gap_seconds": 2.3296370519674383,
        "meaning": "Sampled observations; no numeric temperatures published and no warranty between samples.",
        "thermal_slowdown_fully_available": true,
        "thermal_slowdown_observed_samples": 0,
        "thermal_slowdown_states": {
          "hardware": {
            "active": 0,
            "not_active": 4824,
            "unavailable": 0
          },
          "software": {
            "active": 0,
            "not_active": 4824,
            "unavailable": 0
          }
        }
      }
    }
  ],
  "comparisons": [
    {
      "axis": "gpu_layers",
      "cases": [
        "offload-72b-full",
        "offload-72b-40",
        "offload-72b-20"
      ]
    }
  ],
  "environment": {
    "kernel": "6.12.0-211.56.1.el10_2.0.1.x86_64",
    "nvidia_driver": "595.91.07",
    "os": {
      "id": "rocky",
      "name": "Rocky Linux",
      "version_id": "10.2"
    }
  },
  "hardware": {
    "cpu": "AMD EPYC 4565P 16-Core Processor",
    "fan_profile": "GPU300WAuto",
    "gpu": "NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition",
    "gpu_nominal_vram_gb": 96,
    "gpu_vram_bytes": 102641958912,
    "installed_ram_bytes": 103079215104,
    "kernel_mem_total_bytes": 100271353856,
    "logical_cpu_count": 32,
    "physical_cpu_count": 16
  },
  "historical_results_pooled": false,
  "limitations": [
    "Externally pinned capture integrity is not remote hardware attestation.",
    "Five repeated rounds share a process and are not five independent installations.",
    "Synthetic token latency is not practical first-visible answer latency.",
    "GPU board power is not whole-server AC power; energy is unavailable without bracketing samples.",
    "Sampled memory peaks may not coincide and do not establish minimum required system RAM.",
    "The practical corpus contains 24 authored tasks, not a broad model-quality benchmark.",
    "Engine telemetry includes process loading and internal warmup; exact per-sample resource windows are unavailable.",
    "No historical result is relabeled or pooled into this versioned series."
  ],
  "models": [
    {
      "architecture": "dense",
      "chat_template_sha256": "d5495a1e5db0611132a97e46a65dbb64a642a499421228b9c8b93229097fa9a4",
      "effective_chat_template_rule": "llama.cpp-7fe450e19-chat-source-v1",
      "effective_chat_template_sha256": "4e9918361c284a93880606d182d64da6a9fe97cdc1f5c5a78c1c8840444246fc",
      "files": [
        {
          "bytes": 3986869120,
          "sha256": "3531a90691b5f4edb4bdf488246d72ce0b5d40cacfab1489718a662cec386e43"
        },
        {
          "bytes": 3877864544,
          "sha256": "a4cfb079461591e2c23ef73953aeba4b2520a9b93afaa841f31adf4e480647ab"
        },
        {
          "bytes": 3963622720,
          "sha256": "c67c3641d018b7ca646e7c37b50152687d4b52ce6ab6bf170b006a756b52778d"
        },
        {
          "bytes": 3941463040,
          "sha256": "92f5149d685ad5c5ab0fc30f5e0725d95a746d9a8e54d8397707b97dc435cd76"
        },
        {
          "bytes": 3963688384,
          "sha256": "f5abb6ecd7f3ec4438a4e89ae704acd43405182a0477776d5e8aa35a71a5ecdc"
        },
        {
          "bytes": 3941463040,
          "sha256": "ead67a4ac2648b0f7b71fea755f910ed208f3f8b881b55c208fc3fdca67359ee"
        },
        {
          "bytes": 3826849152,
          "sha256": "f7b168a5fdf78695736a6441dc846b19251339750ea78bf993057b7bf499671d"
        },
        {
          "bytes": 3983901792,
          "sha256": "4ab253ada85c14097597546ac13cfb1426cb175d1d8da520b7a2904d0c323e32"
        },
        {
          "bytes": 3921183968,
          "sha256": "51661b80af773b488e807a03fc7ea31bb64630c50e3ee77fbb5e37bdd9ff9979"
        },
        {
          "bytes": 3995005888,
          "sha256": "fcc180433606a81e2137ba1f088defe3ad51ec9353f5d129a7209614597e16f9"
        },
        {
          "bytes": 3586683584,
          "sha256": "57b9437649f12062bdbed6c623c3f6adc6924a3d2f4e7aabdd3ab32883f5485d"
        },
        {
          "bytes": 1021870240,
          "sha256": "60ae16428f9edb62b60e6c8657ba9344c6fae6ef44c497845dc6a6fd002e4d5b"
        }
      ],
      "id": "qwen25-72b-q4",
      "name": "Qwen2.5-72B-Instruct",
      "quantization": "Q4_K_M",
      "repository": "Qwen/Qwen2.5-72B-Instruct-GGUF",
      "revision": "7ca3bc388f97b264c4283bc9bf1055e2abc38441",
      "tokenizer_sha256": "67b3d77a05acfb80f96742cec2740cc6ee8fffadef98e655e855baf58a3946b1"
    }
  ],
  "protocol": "gsh-llm-core/1.1",
  "runtime": {
    "archives_sha256": [
      "c2ab9e19838513ff69d1af8d999ad717dd3c7ee4714ac04c7ed5ab9077c50e4e",
      "1466daea60aad1144819e151b2bae19d54556cf1da6c129c4f55a5ded2637c25"
    ],
    "binary_sha256": "376626f953efd7bbf16dee112747196d2b1de26717bcd1d37e203340dc5c043c",
    "commit": "7fe450e19305b828c199d602c23a8337aaa1f03b",
    "name": "llama.cpp",
    "release": "b11146"
  },
  "schema": "gsh-llm-core-article/1",
  "sources": [
    {
      "exporter_sha256": "4f4046b91d99ba35175c3c6c04473a0021f54c6baa4892625cb66435f4c7e2ec",
      "finished_at": "2026-09-27T20:35:44.649240+00:00",
      "plan_id": "core-offload-20260927t180810-offloade234fdb2",
      "plan_sha256": "337114b4eef68828eabe560dad626826411d46979051d71e89bd51d5941dca27",
      "protocol_sha256": "3ce8ce9f2b17e66a655c07fd7e7aa3f1762790ad9be38f87d29e5e1279dec3ac",
      "record_sha256": "711f4d34a08d22e6937d72802601ed59fe398c64f9fb191925cb9eebb2cd6b2d",
      "reviewed_export_sha256": "4b17dd2a9c39d7cab209a102bfdbd49f8324b54625bc01cbe7d9aeaf6b87dd5a",
      "runner_sha256": "1bac2436c5aa2eca4f4b42cc9b8f0942b5c66256eb1a41a2169b88c12b044f4e",
      "started_at": "2026-09-27T18:08:57.430966+00:00",
      "task_corpus_sha256": "0642db889565ce25b980e24afa5048eb83e4d15c36754457474a65b2c1f71c21"
    }
  ],
  "status": "completed",
  "test_date_timezone": "Europe/Warsaw",
  "tested_at": "2026-09-27"
}
