{
  "article_key": "llm-gpu-power",
  "cases": [
    {
      "case_wall_seconds": 14450.58151647501,
      "completed_response_resources": {
        "gpu_energy": {
          "available": true,
          "interpolation": "linear_at_bracketed_boundaries",
          "joules": 4322526.512667371,
          "reason": null,
          "response_union_seconds": 14409.160561117635,
          "scope": "GPU board only; not whole-server AC energy",
          "time_weighted_mean_watts": 299.9846170311602,
          "trapezoid_pieces": 14984,
          "watt_hours": 1200.7018090742697
        },
        "gpu_sample_count": 14408,
        "host_sample_count": 14044,
        "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": 7715704832,
            "mean": 6801911819.374537,
            "median": 6799910912.0,
            "minimum": 5833928704,
            "sample_count": 14044
          },
          "gpu_board_watts": {
            "maximum": 305.13,
            "mean": 299.9847897001666,
            "median": 299.99,
            "minimum": 283.16,
            "sample_count": 14408
          },
          "gpu_memory_used_bytes": {
            "maximum": 29349642240.0,
            "mean": 29349642240.0,
            "median": 29349642240.0,
            "minimum": 29349642240.0,
            "sample_count": 14408
          },
          "gpu_sm_clock_mhz": {
            "maximum": 1792.0,
            "mean": 1531.5502498611881,
            "median": 1545.0,
            "minimum": 1477.0,
            "sample_count": 14408
          },
          "gpu_utilization_percent": {
            "maximum": 100.0,
            "mean": 99.21002220988339,
            "median": 100.0,
            "minimum": 50.0,
            "sample_count": 14408
          },
          "host_mem_available_bytes": {
            "maximum": 96192446464,
            "mean": 95707652121.37396,
            "median": 95703994368.0,
            "minimum": 95185977344,
            "sample_count": 14044
          },
          "process_pss_bytes": {
            "maximum": 1000995840,
            "mean": 994586522.5041299,
            "median": 999561216.0,
            "minimum": 939293696,
            "sample_count": 14044
          },
          "process_rss_bytes": {
            "maximum": 1003978752,
            "mean": 997597475.3631444,
            "median": 1002545152.0,
            "minimum": 942276608,
            "sample_count": 14044
          }
        },
        "process_sample_count": 14044,
        "response_union_seconds": 14409.160561117635,
        "sampled_counter_deltas": {
          "deltas": {
            "major_faults": 0,
            "read_syscall_bytes": 0,
            "storage_read_bytes": 0,
            "storage_write_bytes": 14311424
          },
          "observed_endpoint_seconds": 13820.278715579421,
          "observed_intervals": 576,
          "requested_intervals": 576
        }
      },
      "groups": [
        {
          "aggregate_output_tokens_per_second": 40.933959858259136,
          "completed_output_tokens": 589824,
          "completed_request_count": 2304,
          "context_tokens": 8192,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 3.1251006357900364,
            "interval_count": 576,
            "interval_seconds": 14411.525899105764,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 100.00322034528116,
            "process_cpu_seconds": 14411.99
          },
          "failed_request_count": 0,
          "measured_rounds": 576,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 4322526.512667371,
              "reason": null,
              "response_union_seconds": 14409.160561117635,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 299.9846170311602,
              "trapezoid_pieces": 14984,
              "watt_hours": 1200.7018090742697
            },
            "gpu_sample_count": 14408,
            "host_sample_count": 14044,
            "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": 7715704832,
                "mean": 6801911819.374537,
                "median": 6799910912.0,
                "minimum": 5833928704,
                "sample_count": 14044
              },
              "gpu_board_watts": {
                "maximum": 305.13,
                "mean": 299.9847897001666,
                "median": 299.99,
                "minimum": 283.16,
                "sample_count": 14408
              },
              "gpu_memory_used_bytes": {
                "maximum": 29349642240.0,
                "mean": 29349642240.0,
                "median": 29349642240.0,
                "minimum": 29349642240.0,
                "sample_count": 14408
              },
              "gpu_sm_clock_mhz": {
                "maximum": 1792.0,
                "mean": 1531.5502498611881,
                "median": 1545.0,
                "minimum": 1477.0,
                "sample_count": 14408
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 99.21002220988339,
                "median": 100.0,
                "minimum": 50.0,
                "sample_count": 14408
              },
              "host_mem_available_bytes": {
                "maximum": 96192446464,
                "mean": 95707652121.37396,
                "median": 95703994368.0,
                "minimum": 95185977344,
                "sample_count": 14044
              },
              "process_pss_bytes": {
                "maximum": 1000995840,
                "mean": 994586522.5041299,
                "median": 999561216.0,
                "minimum": 939293696,
                "sample_count": 14044
              },
              "process_rss_bytes": {
                "maximum": 1003978752,
                "mean": 997597475.3631444,
                "median": 1002545152.0,
                "minimum": 942276608,
                "sample_count": 14044
              }
            },
            "process_sample_count": 14044,
            "response_union_seconds": 14409.160561117635,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 14311424
              },
              "observed_endpoint_seconds": 13820.278715579421,
              "observed_intervals": 576,
              "requested_intervals": 576
            }
          },
          "response_outer_span_seconds": 14418.665552712977,
          "response_seconds": {
            "maximum": 25.164631637977436,
            "mean": 24.864257757205955,
            "median": 24.896307882503606,
            "minimum": 23.494174277991988,
            "p95": 25.100366784119977,
            "sample_count": 2304
          },
          "response_window_union_seconds": 14409.160561117635,
          "server_decode_tokens_per_second": {
            "maximum": 25.484744831743715,
            "mean": 18.02991749691153,
            "median": 17.848019291580243,
            "minimum": 11.97676620079551,
            "sample_count": 2304
          },
          "synthetic_first_token_seconds": {
            "maximum": 14.590205368993338,
            "mean": 9.797891346179211,
            "median": 10.10864435750409,
            "minimum": 3.4045858300523832,
            "p95": 14.525741860616836,
            "sample_count": 2304
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 4
        }
      ],
      "id": "sustained-32b",
      "model": "qwen3-32b-q4",
      "module": "sustained",
      "placement": {
        "buffers": [
          {
            "device": "CPU_Mapped",
            "reported_mib": 417.3
          },
          {
            "device": "CUDA0",
            "reported_mib": 18423.65
          }
        ],
        "gpu_layers": 65,
        "meaning": "Runtime buffer allocation is not measured resident process RAM.",
        "requested_cpu_expert_layers": 0,
        "total_layers": 65
      },
      "record_sha256": "79abc8b59d526e04286b44ba1fdecc454ad81d540397d9dad2c2634dda632816",
      "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": 5514604544,
              "mean": 5505738752,
              "median": 5505738752.0,
              "minimum": 5496872960,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 15.58,
              "mean": 15.555,
              "median": 15.555,
              "minimum": 15.53,
              "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": 96590258176,
              "mean": 96524697600,
              "median": 96524697600.0,
              "minimum": 96459137024,
              "sample_count": 2
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1006686040200293,
          "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": 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": 7321759744,
              "mean": 7300308992,
              "median": 7300308992.0,
              "minimum": 7278858240,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 108.08,
              "mean": 108.08,
              "median": 108.08,
              "minimum": 108.08,
              "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": 95783763968,
              "mean": 95729772544,
              "median": 95729772544.0,
              "minimum": 95675781120,
              "sample_count": 2
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1560256039956585,
          "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": 213.13813387022867,
            "reason": null,
            "response_union_seconds": 3.6325854180031456,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 58.6739496376088,
            "trapezoid_pieces": 4,
            "watt_hours": 0.05920503718617463
          },
          "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": 5706874880,
              "mean": 5706758144,
              "median": 5706758144.0,
              "minimum": 5706641408,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 77.09,
              "mean": 62.513333333333335,
              "median": 76.77,
              "minimum": 33.68,
              "sample_count": 3
            },
            "gpu_memory_used_bytes": {
              "maximum": 19906166784.0,
              "mean": 19906166784.0,
              "median": 19906166784.0,
              "minimum": 19906166784.0,
              "sample_count": 3
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2302.0,
              "mean": 2292.0,
              "median": 2302.0,
              "minimum": 2272.0,
              "sample_count": 3
            },
            "gpu_utilization_percent": {
              "maximum": 82.0,
              "mean": 79.33333333333333,
              "median": 82.0,
              "minimum": 74.0,
              "sample_count": 3
            },
            "host_mem_available_bytes": {
              "maximum": 96227672064,
              "mean": 96217997312,
              "median": 96217997312.0,
              "minimum": 96208322560,
              "sample_count": 2
            },
            "process_pss_bytes": {
              "maximum": 20099197952,
              "mean": 20099126272,
              "median": 20099126272.0,
              "minimum": 20099054592,
              "sample_count": 2
            },
            "process_rss_bytes": {
              "maximum": 20102180864,
              "mean": 20102109184,
              "median": 20102109184.0,
              "minimum": 20102037504,
              "sample_count": 2
            }
          },
          "process_sample_count": 2,
          "response_union_seconds": 3.6325854180031456,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 5,
              "storage_read_bytes": 0,
              "storage_write_bytes": 0
            },
            "observed_endpoint_seconds": 1.3338036570348777,
            "observed_intervals": 1,
            "requested_intervals": 1
          }
        },
        "ready": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 327.03305081688416,
            "reason": null,
            "response_union_seconds": 2.223970894992817,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 147.0491594800931,
            "trapezoid_pieces": 4,
            "watt_hours": 0.09084251411580116
          },
          "gpu_sample_count": 2,
          "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": 7702061056,
              "mean": 6413938688,
              "median": 5770027008,
              "minimum": 5769728000,
              "sample_count": 3
            },
            "gpu_board_watts": {
              "maximum": 183.77,
              "mean": 136.445,
              "median": 136.445,
              "minimum": 89.12,
              "sample_count": 2
            },
            "gpu_memory_used_bytes": {
              "maximum": 29349642240.0,
              "mean": 29336010752.0,
              "median": 29336010752.0,
              "minimum": 29322379264.0,
              "sample_count": 2
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2340.0,
              "mean": 2321.0,
              "median": 2321.0,
              "minimum": 2302.0,
              "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": 96134074368,
              "mean": 95850448213.33333,
              "median": 96131477504,
              "minimum": 95285792768,
              "sample_count": 3
            },
            "process_pss_bytes": {
              "maximum": 1000994816,
              "mean": 919594325.3333334,
              "median": 878894080,
              "minimum": 878894080,
              "sample_count": 3
            },
            "process_rss_bytes": {
              "maximum": 1003978752,
              "mean": 922577578.6666666,
              "median": 881876992,
              "minimum": 881876992,
              "sample_count": 3
            }
          },
          "process_sample_count": 3,
          "response_union_seconds": 2.223970894992817,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 0
            },
            "observed_endpoint_seconds": 1.021025077032391,
            "observed_intervals": 1,
            "requested_intervals": 2
          }
        },
        "responses": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 4325377.306417703,
            "reason": null,
            "response_union_seconds": 14418.681973422004,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 299.98423672778705,
            "trapezoid_pieces": 14420,
            "watt_hours": 1201.4936962271397
          },
          "gpu_sample_count": 14419,
          "host_sample_count": 14050,
          "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": 7715704832,
              "mean": 6801967864.01936,
              "median": 6799910912.0,
              "minimum": 5833928704,
              "sample_count": 14050
            },
            "gpu_board_watts": {
              "maximum": 305.13,
              "mean": 299.9844940703239,
              "median": 299.99,
              "minimum": 283.16,
              "sample_count": 14419
            },
            "gpu_memory_used_bytes": {
              "maximum": 29349642240.0,
              "mean": 29349642240.0,
              "median": 29349642240.0,
              "minimum": 29349642240.0,
              "sample_count": 14419
            },
            "gpu_sm_clock_mhz": {
              "maximum": 1792.0,
              "mean": 1531.5317289687218,
              "median": 1545.0,
              "minimum": 1477.0,
              "sample_count": 14419
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 99.20694916429711,
              "median": 100.0,
              "minimum": 50.0,
              "sample_count": 14419
            },
            "host_mem_available_bytes": {
              "maximum": 96192446464,
              "mean": 95707636376.68896,
              "median": 95703994368.0,
              "minimum": 95185977344,
              "sample_count": 14050
            },
            "process_pss_bytes": {
              "maximum": 1000995840,
              "mean": 994587084.7635587,
              "median": 999561216.0,
              "minimum": 939293696,
              "sample_count": 14050
            },
            "process_rss_bytes": {
              "maximum": 1003978752,
              "mean": 997598025.9393594,
              "median": 1002545152.0,
              "minimum": 942276608,
              "sample_count": 14050
            }
          },
          "process_sample_count": 14050,
          "response_union_seconds": 14418.681973422004,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 17612800
            },
            "observed_endpoint_seconds": 14416.98345157504,
            "observed_intervals": 1,
            "requested_intervals": 1
          }
        },
        "warmup": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 6987.65928224834,
            "reason": null,
            "response_union_seconds": 23.74817975901533,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 294.23978398157743,
            "trapezoid_pieces": 25,
            "watt_hours": 1.9410164672912056
          },
          "gpu_sample_count": 24,
          "host_sample_count": 23,
          "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": 5842784256,
              "mean": 5822150210.782609,
              "median": 5826940928,
              "minimum": 5801627648,
              "sample_count": 23
            },
            "gpu_board_watts": {
              "maximum": 306.76,
              "mean": 291.29333333333335,
              "median": 300.005,
              "minimum": 85.69,
              "sample_count": 24
            },
            "gpu_memory_used_bytes": {
              "maximum": 29349642240.0,
              "mean": 29348943189.333332,
              "median": 29349642240.0,
              "minimum": 29347545088.0,
              "sample_count": 24
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2302.0,
              "mean": 1641.0416666666667,
              "median": 1635.0,
              "minimum": 1552.0,
              "sample_count": 24
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 98.875,
              "median": 99.0,
              "minimum": 90.0,
              "sample_count": 24
            },
            "host_mem_available_bytes": {
              "maximum": 96146788352,
              "mean": 96126758912,
              "median": 96126504960,
              "minimum": 96112836608,
              "sample_count": 23
            },
            "process_pss_bytes": {
              "maximum": 936446976,
              "mean": 929384670.6086956,
              "median": 933833728,
              "minimum": 910089216,
              "sample_count": 23
            },
            "process_rss_bytes": {
              "maximum": 939429888,
              "mean": 932429289.7391304,
              "median": 936816640,
              "minimum": 913072128,
              "sample_count": 23
            }
          },
          "process_sample_count": 23,
          "response_union_seconds": 23.74817975901533,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 8,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 1441792,
              "storage_write_bytes": 20480
            },
            "observed_endpoint_seconds": 22.51987132197246,
            "observed_intervals": 1,
            "requested_intervals": 1
          }
        }
      },
      "reviewed_export_sha256": "7e9bf3881a4bd446b234cda9c158b651eb9d123d6af6090c39eed1e26ad459ac",
      "settings": {
        "batch": 2048,
        "cache_prompt": false,
        "cache_ram_bytes": 0,
        "concurrency": 4,
        "context_shift": false,
        "contexts": [
          8192
        ],
        "endpoint": "/completion",
        "fit": "off",
        "flash_attention": true,
        "ignore_eos": true,
        "kv_cache": "f16",
        "lazy_mode": "off",
        "load_mode": "mmap",
        "measured_rounds": null,
        "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": 4,
        "synthetic_prompt_sha256": "c099b5abc034050fac1133e0214990d1f72a124b64f36e8a80ba3209bc07d187",
        "temperature": 0,
        "threads": 16,
        "tokenization": "model_tokenizer_add_special_false_repeat_then_truncate",
        "total_context_tokens": 34816,
        "ubatch": 512,
        "warmup_rounds": 1
      },
      "status": "completed",
      "telemetry": {
        "gpu_samples": 14451,
        "host_samples": 14082,
        "maximum_gpu_gap_seconds": 1.0334704880369827,
        "maximum_host_gap_seconds": 1.5620193640352227,
        "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": 14451,
            "unavailable": 0
          },
          "software": {
            "active": 0,
            "not_active": 14451,
            "unavailable": 0
          }
        }
      }
    },
    {
      "case_wall_seconds": 29.91674370097462,
      "completed_response_resources": {
        "gpu_energy": {
          "available": true,
          "interpolation": "linear_at_bracketed_boundaries",
          "joules": 5704.656739645703,
          "reason": null,
          "response_union_seconds": 19.113167581963353,
          "scope": "GPU board only; not whole-server AC energy",
          "time_weighted_mean_watts": 298.4673636738817,
          "trapezoid_pieces": 30,
          "watt_hours": 1.5846268721238064
        },
        "gpu_sample_count": 20,
        "host_sample_count": 19,
        "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": 6720995328,
            "mean": 6712309221.052631,
            "median": 6713368576,
            "minimum": 6703079424,
            "sample_count": 19
          },
          "gpu_board_watts": {
            "maximum": 301.0,
            "mean": 299.713,
            "median": 299.95,
            "minimum": 297.7,
            "sample_count": 20
          },
          "gpu_memory_used_bytes": {
            "maximum": 6752829440.0,
            "mean": 6752829440.0,
            "median": 6752829440.0,
            "minimum": 6752829440.0,
            "sample_count": 20
          },
          "gpu_sm_clock_mhz": {
            "maximum": 1845.0,
            "mean": 1793.7,
            "median": 1800.0,
            "minimum": 1710.0,
            "sample_count": 20
          },
          "gpu_utilization_percent": {
            "maximum": 100.0,
            "mean": 95.9,
            "median": 97.0,
            "minimum": 85.0,
            "sample_count": 20
          },
          "host_mem_available_bytes": {
            "maximum": 96199720960,
            "mean": 96188027095.57895,
            "median": 96187265024,
            "minimum": 96176087040,
            "sample_count": 19
          },
          "process_pss_bytes": {
            "maximum": 819189760,
            "mean": 815475550.3157895,
            "median": 816355328,
            "minimum": 811747328,
            "sample_count": 19
          },
          "process_rss_bytes": {
            "maximum": 822177792,
            "mean": 818463582.3157895,
            "median": 819343360,
            "minimum": 814735360,
            "sample_count": 19
          }
        },
        "process_sample_count": 19,
        "response_union_seconds": 19.113167581963353,
        "sampled_counter_deltas": {
          "deltas": {
            "major_faults": 0,
            "read_syscall_bytes": 0,
            "storage_read_bytes": 0,
            "storage_write_bytes": 0
          },
          "observed_endpoint_seconds": 9.175436379155144,
          "observed_intervals": 8,
          "requested_intervals": 10
        }
      },
      "groups": [
        {
          "aggregate_output_tokens_per_second": 172.7690055501098,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 2048,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 3.119277758589535,
            "interval_count": 5,
            "interval_seconds": 7.423593470070045,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 99.81688827486512,
            "process_cpu_seconds": 7.41
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 2207.7645157894385,
              "reason": null,
              "response_union_seconds": 7.408736283017788,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 297.99474990762576,
              "trapezoid_pieces": 13,
              "watt_hours": 0.6132679210526218
            },
            "gpu_sample_count": 8,
            "host_sample_count": 8,
            "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": 6708359168,
                "mean": 6705936384,
                "median": 6706384896.0,
                "minimum": 6703079424,
                "sample_count": 8
              },
              "gpu_board_watts": {
                "maximum": 300.43,
                "mean": 299.49625,
                "median": 299.95,
                "minimum": 297.7,
                "sample_count": 8
              },
              "gpu_memory_used_bytes": {
                "maximum": 6752829440.0,
                "mean": 6752829440.0,
                "median": 6752829440.0,
                "minimum": 6752829440.0,
                "sample_count": 8
              },
              "gpu_sm_clock_mhz": {
                "maximum": 1845.0,
                "mean": 1832.5,
                "median": 1833.5,
                "minimum": 1822.0,
                "sample_count": 8
              },
              "gpu_utilization_percent": {
                "maximum": 97.0,
                "mean": 94.25,
                "median": 97.0,
                "minimum": 86.0,
                "sample_count": 8
              },
              "host_mem_available_bytes": {
                "maximum": 96199720960,
                "mean": 96194545152,
                "median": 96195299328.0,
                "minimum": 96187265024,
                "sample_count": 8
              },
              "process_pss_bytes": {
                "maximum": 813074432,
                "mean": 812405760,
                "median": 812410880.0,
                "minimum": 811747328,
                "sample_count": 8
              },
              "process_rss_bytes": {
                "maximum": 816062464,
                "mean": 815393792,
                "median": 815398912.0,
                "minimum": 814735360,
                "sample_count": 8
              }
            },
            "process_sample_count": 8,
            "response_union_seconds": 7.408736283017788,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 0
              },
              "observed_endpoint_seconds": 3.05030764703406,
              "observed_intervals": 3,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 7.430938940960914,
          "response_seconds": {
            "maximum": 1.484641135961283,
            "mean": 1.4817472566035577,
            "median": 1.481124191021081,
            "minimum": 1.479376372997649,
            "sample_count": 5
          },
          "response_window_union_seconds": 7.408736283017788,
          "server_decode_tokens_per_second": {
            "maximum": 197.7636422013653,
            "mean": 197.5064926865939,
            "median": 197.53690841569204,
            "minimum": 197.21684494811652,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 0.19157884595915675,
            "mean": 0.1905686531914398,
            "median": 0.19060329400235787,
            "minimum": 0.1899017799878493,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        },
        {
          "aggregate_output_tokens_per_second": 109.36028990279206,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 8192,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 3.118724755561472,
            "interval_count": 5,
            "interval_seconds": 11.72354178893147,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 99.7991921779671,
            "process_cpu_seconds": 11.7
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 3496.8922238562645,
              "reason": null,
              "response_union_seconds": 11.704431298945565,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 298.76652137479715,
              "trapezoid_pieces": 17,
              "watt_hours": 0.9713589510711846
            },
            "gpu_sample_count": 12,
            "host_sample_count": 11,
            "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": 6720995328,
                "mean": 6716944011.636364,
                "median": 6717145088,
                "minimum": 6713245696,
                "sample_count": 11
              },
              "gpu_board_watts": {
                "maximum": 301.0,
                "mean": 299.8575,
                "median": 299.995,
                "minimum": 298.01,
                "sample_count": 12
              },
              "gpu_memory_used_bytes": {
                "maximum": 6752829440.0,
                "mean": 6752829440.0,
                "median": 6752829440.0,
                "minimum": 6752829440.0,
                "sample_count": 12
              },
              "gpu_sm_clock_mhz": {
                "maximum": 1822.0,
                "mean": 1767.8333333333333,
                "median": 1773.5,
                "minimum": 1710.0,
                "sample_count": 12
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 97.0,
                "median": 97.5,
                "minimum": 85.0,
                "sample_count": 12
              },
              "host_mem_available_bytes": {
                "maximum": 96188547072,
                "mean": 96183286690.90909,
                "median": 96184393728,
                "minimum": 96176087040,
                "sample_count": 11
              },
              "process_pss_bytes": {
                "maximum": 819189760,
                "mean": 817708125.0909091,
                "median": 817772544,
                "minimum": 816355328,
                "sample_count": 11
              },
              "process_rss_bytes": {
                "maximum": 822177792,
                "mean": 820696157.0909091,
                "median": 820760576,
                "minimum": 819343360,
                "sample_count": 11
              }
            },
            "process_sample_count": 11,
            "response_union_seconds": 11.704431298945565,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 0
              },
              "observed_endpoint_seconds": 6.125128732121084,
              "observed_intervals": 5,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 11.733725362981204,
          "response_seconds": {
            "maximum": 2.3454450629651546,
            "mean": 2.340886259789113,
            "median": 2.3407816879916936,
            "minimum": 2.3358659429941326,
            "sample_count": 5
          },
          "response_window_union_seconds": 11.704431298945565,
          "server_decode_tokens_per_second": {
            "maximum": 169.3727562261093,
            "mean": 169.06737906840763,
            "median": 169.01162004597117,
            "minimum": 168.80352064566353,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 0.834626693977043,
            "mean": 0.8324093010160141,
            "median": 0.8318316700169817,
            "minimum": 0.8301292260293849,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        }
      ],
      "id": "baseline-8b",
      "model": "qwen3-8b-q4",
      "module": "baseline",
      "placement": {
        "buffers": [
          {
            "device": "CPU_Mapped",
            "reported_mib": 333.84
          },
          {
            "device": "CUDA0",
            "reported_mib": 4455.34
          }
        ],
        "gpu_layers": 37,
        "meaning": "Runtime buffer allocation is not measured resident process RAM.",
        "requested_cpu_expert_layers": 0,
        "total_layers": 37
      },
      "record_sha256": "2db66c2c90e7ccf91ace8aaf7f6b8e1865a1de2453d6547e769e2dda7e176b21",
      "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": 6373740544,
              "mean": 6365132800,
              "median": 6365132800.0,
              "minimum": 6356525056,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 15.03,
              "mean": 15.024999999999999,
              "median": 15.024999999999999,
              "minimum": 15.02,
              "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": 96634912768,
              "mean": 96623800320,
              "median": 96623800320.0,
              "minimum": 96612687872,
              "sample_count": 2
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1007707070093602,
          "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": 6392254464,
              "mean": 6392254464,
              "median": 6392254464,
              "minimum": 6392254464,
              "sample_count": 1
            },
            "gpu_board_watts": {
              "maximum": 78.38,
              "mean": 78.38,
              "median": 78.38,
              "minimum": 78.38,
              "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": 2325.0,
              "mean": 2325.0,
              "median": 2325.0,
              "minimum": 2325.0,
              "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": 96599707648,
              "mean": 96599707648,
              "median": 96599707648,
              "minimum": 96599707648,
              "sample_count": 1
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.155548368988093,
          "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": 51.318043537193375,
            "reason": null,
            "response_union_seconds": 1.2843944630003534,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 39.95504886973283,
            "trapezoid_pieces": 2,
            "watt_hours": 0.014255012093664827
          },
          "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": 6579744768,
              "mean": 6579744768,
              "median": 6579744768,
              "minimum": 6579744768,
              "sample_count": 1
            },
            "gpu_board_watts": {
              "maximum": 50.74,
              "mean": 50.74,
              "median": 50.74,
              "minimum": 50.74,
              "sample_count": 1
            },
            "gpu_memory_used_bytes": {
              "maximum": 5259657216.0,
              "mean": 5259657216.0,
              "median": 5259657216.0,
              "minimum": 5259657216.0,
              "sample_count": 1
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2280.0,
              "mean": 2280.0,
              "median": 2280.0,
              "minimum": 2280.0,
              "sample_count": 1
            },
            "gpu_utilization_percent": {
              "maximum": 76.0,
              "mean": 76.0,
              "median": 76.0,
              "minimum": 76.0,
              "sample_count": 1
            },
            "host_mem_available_bytes": {
              "maximum": 96307732480,
              "mean": 96307732480,
              "median": 96307732480,
              "minimum": 96307732480,
              "sample_count": 1
            },
            "process_pss_bytes": {
              "maximum": 5364942848,
              "mean": 5364942848,
              "median": 5364942848,
              "minimum": 5364942848,
              "sample_count": 1
            },
            "process_rss_bytes": {
              "maximum": 5367930880,
              "mean": 5367930880,
              "median": 5367930880,
              "minimum": 5367930880,
              "sample_count": 1
            }
          },
          "process_sample_count": 1,
          "response_union_seconds": 1.2843944630003534,
          "sampled_counter_deltas": {
            "deltas": null,
            "observed_endpoint_seconds": 0.0,
            "observed_intervals": 0,
            "requested_intervals": 1
          }
        },
        "ready": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 521.8941446628455,
            "reason": null,
            "response_union_seconds": 3.3171223630197346,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 157.33340152931243,
            "trapezoid_pieces": 6,
            "watt_hours": 0.1449705957396793
          },
          "gpu_sample_count": 3,
          "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": 6721740800,
              "mean": 6704672768,
              "median": 6714998784.0,
              "minimum": 6666952704,
              "sample_count": 4
            },
            "gpu_board_watts": {
              "maximum": 226.38,
              "mean": 158.82666666666665,
              "median": 176.54,
              "minimum": 73.56,
              "sample_count": 3
            },
            "gpu_memory_used_bytes": {
              "maximum": 6752829440.0,
              "mean": 6746537984.0,
              "median": 6752829440.0,
              "minimum": 6733955072.0,
              "sample_count": 3
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2325.0,
              "mean": 2310.0,
              "median": 2310.0,
              "minimum": 2295.0,
              "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": 96215072768,
              "mean": 96190546944,
              "median": 96184281088.0,
              "minimum": 96178552832,
              "sample_count": 4
            },
            "process_pss_bytes": {
              "maximum": 819189760,
              "mean": 807820288,
              "median": 816168960.0,
              "minimum": 779753472,
              "sample_count": 4
            },
            "process_rss_bytes": {
              "maximum": 822177792,
              "mean": 810808320,
              "median": 819156992.0,
              "minimum": 782741504,
              "sample_count": 4
            }
          },
          "process_sample_count": 4,
          "response_union_seconds": 3.3171223630197346,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 0
            },
            "observed_endpoint_seconds": 1.0194021269562654,
            "observed_intervals": 1,
            "requested_intervals": 3
          }
        },
        "responses": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 5723.3732039281185,
            "reason": null,
            "response_union_seconds": 19.17764565697871,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 298.4398244862446,
            "trapezoid_pieces": 22,
            "watt_hours": 1.5898258899800328
          },
          "gpu_sample_count": 20,
          "host_sample_count": 19,
          "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": 6720995328,
              "mean": 6712309221.052631,
              "median": 6713368576,
              "minimum": 6703079424,
              "sample_count": 19
            },
            "gpu_board_watts": {
              "maximum": 301.0,
              "mean": 299.713,
              "median": 299.95,
              "minimum": 297.7,
              "sample_count": 20
            },
            "gpu_memory_used_bytes": {
              "maximum": 6752829440.0,
              "mean": 6752829440.0,
              "median": 6752829440.0,
              "minimum": 6752829440.0,
              "sample_count": 20
            },
            "gpu_sm_clock_mhz": {
              "maximum": 1845.0,
              "mean": 1793.7,
              "median": 1800.0,
              "minimum": 1710.0,
              "sample_count": 20
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 95.9,
              "median": 97.0,
              "minimum": 85.0,
              "sample_count": 20
            },
            "host_mem_available_bytes": {
              "maximum": 96199720960,
              "mean": 96188027095.57895,
              "median": 96187265024,
              "minimum": 96176087040,
              "sample_count": 19
            },
            "process_pss_bytes": {
              "maximum": 819189760,
              "mean": 815475550.3157895,
              "median": 816355328,
              "minimum": 811747328,
              "sample_count": 19
            },
            "process_rss_bytes": {
              "maximum": 822177792,
              "mean": 818463582.3157895,
              "median": 819343360,
              "minimum": 814735360,
              "sample_count": 19
            }
          },
          "process_sample_count": 19,
          "response_union_seconds": 19.17764565697871,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 16384
            },
            "observed_endpoint_seconds": 17.353597005014308,
            "observed_intervals": 2,
            "requested_intervals": 2
          }
        },
        "warmup": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 888.6925389093352,
            "reason": null,
            "response_union_seconds": 3.8429829700035043,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 231.25070963000516,
            "trapezoid_pieces": 5,
            "watt_hours": 0.24685903858592645
          },
          "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": 6711808000,
              "mean": 6708021930.666667,
              "median": 6709710848,
              "minimum": 6702546944,
              "sample_count": 3
            },
            "gpu_board_watts": {
              "maximum": 300.38,
              "mean": 216.23666666666668,
              "median": 180.59,
              "minimum": 167.74,
              "sample_count": 3
            },
            "gpu_memory_used_bytes": {
              "maximum": 6752829440.0,
              "mean": 6752829440.0,
              "median": 6752829440.0,
              "minimum": 6752829440.0,
              "sample_count": 3
            },
            "gpu_sm_clock_mhz": {
              "maximum": 1905.0,
              "mean": 1807.3333333333333,
              "median": 1777.0,
              "minimum": 1740.0,
              "sample_count": 3
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 98.0,
              "median": 97.0,
              "minimum": 97.0,
              "sample_count": 3
            },
            "host_mem_available_bytes": {
              "maximum": 96199966720,
              "mean": 96187517610.66667,
              "median": 96182734848,
              "minimum": 96179851264,
              "sample_count": 3
            },
            "process_pss_bytes": {
              "maximum": 815646720,
              "mean": 813635584,
              "median": 813856768,
              "minimum": 811403264,
              "sample_count": 3
            },
            "process_rss_bytes": {
              "maximum": 818634752,
              "mean": 816623616,
              "median": 816844800,
              "minimum": 814391296,
              "sample_count": 3
            }
          },
          "process_sample_count": 3,
          "response_union_seconds": 3.8429829700035043,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 0
            },
            "observed_endpoint_seconds": 1.046531479980331,
            "observed_intervals": 1,
            "requested_intervals": 2
          }
        }
      },
      "reviewed_export_sha256": "2d1a82644c4ba9ba504ddf8fd2c7fe3c253e15cc9811363c52186bd4b7510a07",
      "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": 30,
        "host_samples": 30,
        "maximum_gpu_gap_seconds": 1.0048957320395857,
        "maximum_host_gap_seconds": 1.0701882099965587,
        "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": 30,
            "unavailable": 0
          },
          "software": {
            "active": 0,
            "not_active": 30,
            "unavailable": 0
          }
        }
      }
    },
    {
      "case_wall_seconds": 268.15063832100714,
      "completed_response_resources": {
        "gpu_energy": {
          "available": true,
          "interpolation": "linear_at_bracketed_boundaries",
          "joules": 20773.514431872452,
          "reason": null,
          "response_union_seconds": 69.25191831210395,
          "scope": "GPU board only; not whole-server AC energy",
          "time_weighted_mean_watts": 299.9702381997644,
          "trapezoid_pieces": 78,
          "watt_hours": 5.770420675520126
        },
        "gpu_sample_count": 68,
        "host_sample_count": 68,
        "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": 17240788992,
            "mean": 17231835497.411766,
            "median": 17233764352.0,
            "minimum": 17221709824,
            "sample_count": 68
          },
          "gpu_board_watts": {
            "maximum": 301.79,
            "mean": 299.9689705882353,
            "median": 300.0,
            "minimum": 297.03,
            "sample_count": 68
          },
          "gpu_memory_used_bytes": {
            "maximum": 22502440960.0,
            "mean": 22502440960.0,
            "median": 22502440960.0,
            "minimum": 22502440960.0,
            "sample_count": 68
          },
          "gpu_sm_clock_mhz": {
            "maximum": 1687.0,
            "mean": 1608.3088235294117,
            "median": 1605.0,
            "minimum": 1575.0,
            "sample_count": 68
          },
          "gpu_utilization_percent": {
            "maximum": 100.0,
            "mean": 99.02941176470588,
            "median": 99.0,
            "minimum": 84.0,
            "sample_count": 68
          },
          "host_mem_available_bytes": {
            "maximum": 96136097792,
            "mean": 96108860235.29411,
            "median": 96106113024.0,
            "minimum": 96084488192,
            "sample_count": 68
          },
          "process_pss_bytes": {
            "maximum": 925034496,
            "mean": 919076412.2352941,
            "median": 921339904.0,
            "minimum": 911089664,
            "sample_count": 68
          },
          "process_rss_bytes": {
            "maximum": 928030720,
            "mean": 922093507.7647059,
            "median": 924336128.0,
            "minimum": 915148800,
            "sample_count": 68
          }
        },
        "process_sample_count": 68,
        "response_union_seconds": 69.25191831210395,
        "sampled_counter_deltas": {
          "deltas": {
            "major_faults": 0,
            "read_syscall_bytes": 0,
            "storage_read_bytes": 0,
            "storage_write_bytes": 45056
          },
          "observed_endpoint_seconds": 59.35470441589132,
          "observed_intervals": 10,
          "requested_intervals": 10
        }
      },
      "groups": [
        {
          "aggregate_output_tokens_per_second": 46.274918602713925,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 2048,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 3.1198924303190445,
            "interval_count": 5,
            "interval_seconds": 27.675233017944265,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 99.83655777020942,
            "process_cpu_seconds": 27.63
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 8298.985135398982,
              "reason": null,
              "response_union_seconds": 27.660772588045802,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 300.02723564509427,
              "trapezoid_pieces": 32,
              "watt_hours": 2.3052736487219394
            },
            "gpu_sample_count": 27,
            "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": 17240031232,
                "mean": 17223894964.148148,
                "median": 17223282688,
                "minimum": 17221709824,
                "sample_count": 27
              },
              "gpu_board_watts": {
                "maximum": 301.79,
                "mean": 300.0274074074074,
                "median": 299.99,
                "minimum": 299.24,
                "sample_count": 27
              },
              "gpu_memory_used_bytes": {
                "maximum": 22502440960.0,
                "mean": 22502440960.0,
                "median": 22502440960.0,
                "minimum": 22502440960.0,
                "sample_count": 27
              },
              "gpu_sm_clock_mhz": {
                "maximum": 1687.0,
                "mean": 1607.5925925925926,
                "median": 1605.0,
                "minimum": 1575.0,
                "sample_count": 27
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 99.0,
                "median": 99.0,
                "minimum": 95.0,
                "sample_count": 27
              },
              "host_mem_available_bytes": {
                "maximum": 96136097792,
                "mean": 96113169445.92592,
                "median": 96115363840,
                "minimum": 96087658496,
                "sample_count": 27
              },
              "process_pss_bytes": {
                "maximum": 913487872,
                "mean": 912757797.925926,
                "median": 912799744,
                "minimum": 911089664,
                "sample_count": 27
              },
              "process_rss_bytes": {
                "maximum": 916484096,
                "mean": 915806587.2592592,
                "median": 915795968,
                "minimum": 915148800,
                "sample_count": 27
              }
            },
            "process_sample_count": 27,
            "response_union_seconds": 27.660772588045802,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 20480
              },
              "observed_endpoint_seconds": 22.504645432985853,
              "observed_intervals": 5,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 27.68297484901268,
          "response_seconds": {
            "maximum": 5.5716617869911715,
            "mean": 5.53215451760916,
            "median": 5.527461774996482,
            "minimum": 5.496413041022606,
            "sample_count": 5
          },
          "response_window_union_seconds": 27.660772588045802,
          "server_decode_tokens_per_second": {
            "maximum": 53.86729127375231,
            "mean": 53.507414673402344,
            "median": 53.55129058610313,
            "minimum": 53.12845334946771,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 0.7718379290308803,
            "mean": 0.7662811154150404,
            "median": 0.7656330339959823,
            "minimum": 0.7624987220042385,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        },
        {
          "aggregate_output_tokens_per_second": 30.775781184108897,
          "completed_output_tokens": 1280,
          "completed_request_count": 5,
          "context_tokens": 8192,
          "cpu_snapshot_intervals": {
            "host_capacity_percent": 3.1204527047148796,
            "interval_count": 5,
            "interval_seconds": 41.610548944969196,
            "meaning": "Weighted CPU counter rate across round-bracketing snapshots, not exact token time.",
            "one_logical_cpu_percent": 99.85448655087615,
            "process_cpu_seconds": 41.55
          },
          "failed_request_count": 0,
          "measured_rounds": 5,
          "mode": "completion",
          "resources": {
            "gpu_energy": {
              "available": true,
              "interpolation": "linear_at_bracketed_boundaries",
              "joules": 12474.529296473467,
              "reason": null,
              "response_union_seconds": 41.59114572405815,
              "scope": "GPU board only; not whole-server AC energy",
              "time_weighted_mean_watts": 299.93233125236196,
              "trapezoid_pieces": 46,
              "watt_hours": 3.4651470267981854
            },
            "gpu_sample_count": 41,
            "host_sample_count": 41,
            "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": 17240788992,
                "mean": 17237064629.07317,
                "median": 17237659648,
                "minimum": 17232809984,
                "sample_count": 41
              },
              "gpu_board_watts": {
                "maximum": 300.28,
                "mean": 299.93048780487806,
                "median": 300.0,
                "minimum": 297.03,
                "sample_count": 41
              },
              "gpu_memory_used_bytes": {
                "maximum": 22502440960.0,
                "mean": 22502440960.0,
                "median": 22502440960.0,
                "minimum": 22502440960.0,
                "sample_count": 41
              },
              "gpu_sm_clock_mhz": {
                "maximum": 1657.0,
                "mean": 1608.780487804878,
                "median": 1605.0,
                "minimum": 1590.0,
                "sample_count": 41
              },
              "gpu_utilization_percent": {
                "maximum": 100.0,
                "mean": 99.04878048780488,
                "median": 99.0,
                "minimum": 84.0,
                "sample_count": 41
              },
              "host_mem_available_bytes": {
                "maximum": 96128155648,
                "mean": 96106022462.43903,
                "median": 96104517632,
                "minimum": 96084488192,
                "sample_count": 41
              },
              "process_pss_bytes": {
                "maximum": 925034496,
                "mean": 923237450.9268292,
                "median": 923604992,
                "minimum": 920840192,
                "sample_count": 41
              },
              "process_rss_bytes": {
                "maximum": 928030720,
                "mean": 926233674.9268292,
                "median": 926601216,
                "minimum": 923836416,
                "sample_count": 41
              }
            },
            "process_sample_count": 41,
            "response_union_seconds": 41.59114572405815,
            "sampled_counter_deltas": {
              "deltas": {
                "major_faults": 0,
                "read_syscall_bytes": 0,
                "storage_read_bytes": 0,
                "storage_write_bytes": 24576
              },
              "observed_endpoint_seconds": 36.850058982905466,
              "observed_intervals": 5,
              "requested_intervals": 5
            }
          },
          "response_outer_span_seconds": 41.62048171600327,
          "response_seconds": {
            "maximum": 8.362828276003711,
            "mean": 8.31822914481163,
            "median": 8.326255902997218,
            "minimum": 8.262266772042494,
            "sample_count": 5
          },
          "response_window_union_seconds": 41.59114572405815,
          "server_decode_tokens_per_second": {
            "maximum": 51.867399264866066,
            "mean": 51.518764345880015,
            "median": 51.44895386118346,
            "minimum": 51.2462896681451,
            "sample_count": 5
          },
          "synthetic_first_token_seconds": {
            "maximum": 3.3867225820431486,
            "mean": 3.3683619036222807,
            "median": 3.3697669280227274,
            "minimum": 3.345733996015042,
            "sample_count": 5
          },
          "warmup_completed_rounds": 1,
          "warmup_requests_per_round": 1
        }
      ],
      "id": "baseline-32b",
      "model": "qwen3-32b-q4",
      "module": "baseline",
      "placement": {
        "buffers": [
          {
            "device": "CPU_Mapped",
            "reported_mib": 417.3
          },
          {
            "device": "CUDA0",
            "reported_mib": 18423.65
          }
        ],
        "gpu_layers": 65,
        "meaning": "Runtime buffer allocation is not measured resident process RAM.",
        "requested_cpu_expert_layers": 0,
        "total_layers": 65
      },
      "record_sha256": "31d7b3fec56be9142f89e611f4085991a36d0edab0bd028a37a8232b8093b11b",
      "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": 16632758272,
              "mean": 16624463872,
              "median": 16624463872.0,
              "minimum": 16616169472,
              "sample_count": 2
            },
            "gpu_board_watts": {
              "maximum": 16.33,
              "mean": 16.21,
              "median": 16.21,
              "minimum": 16.09,
              "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": 96517685248,
              "mean": 96516294656,
              "median": 96516294656.0,
              "minimum": 96514904064,
              "sample_count": 2
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1007562470040284,
          "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": 16893157376,
              "mean": 16893157376,
              "median": 16893157376,
              "minimum": 16893157376,
              "sample_count": 1
            },
            "gpu_board_watts": {
              "maximum": 94.6,
              "mean": 94.6,
              "median": 94.6,
              "minimum": 94.6,
              "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": 2302.0,
              "mean": 2302.0,
              "median": 2302.0,
              "minimum": 2302.0,
              "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": 96570462208,
              "mean": 96570462208,
              "median": 96570462208,
              "minimum": 96570462208,
              "sample_count": 1
            },
            "process_pss_bytes": null,
            "process_rss_bytes": null
          },
          "process_sample_count": 0,
          "response_union_seconds": 1.1624215979827568,
          "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": 234.14076826269599,
            "reason": null,
            "response_union_seconds": 3.614590512996074,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 64.77656802917369,
            "trapezoid_pieces": 4,
            "watt_hours": 0.06503910229519333
          },
          "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": 17189969920,
              "mean": 17152371370.666666,
              "median": 17133707264,
              "minimum": 17133436928,
              "sample_count": 3
            },
            "gpu_board_watts": {
              "maximum": 78.92,
              "mean": 70.08,
              "median": 78.4,
              "minimum": 52.92,
              "sample_count": 3
            },
            "gpu_memory_used_bytes": {
              "maximum": 19906166784.0,
              "mean": 19906166784.0,
              "median": 19906166784.0,
              "minimum": 19906166784.0,
              "sample_count": 3
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2310.0,
              "mean": 2300.0,
              "median": 2310.0,
              "minimum": 2280.0,
              "sample_count": 3
            },
            "gpu_utilization_percent": {
              "maximum": 76.0,
              "mean": 71.0,
              "median": 76.0,
              "minimum": 61.0,
              "sample_count": 3
            },
            "host_mem_available_bytes": {
              "maximum": 96221175808,
              "mean": 96199781034.66667,
              "median": 96212836352,
              "minimum": 96165330944,
              "sample_count": 3
            },
            "process_pss_bytes": {
              "maximum": 20099049472,
              "mean": 13691381760,
              "median": 20098906112,
              "minimum": 876189696,
              "sample_count": 3
            },
            "process_rss_bytes": {
              "maximum": 20102045696,
              "mean": 13694377984,
              "median": 20101902336,
              "minimum": 879185920,
              "sample_count": 3
            }
          },
          "process_sample_count": 3,
          "response_union_seconds": 3.614590512996074,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 13,
              "storage_read_bytes": 0,
              "storage_write_bytes": 8192
            },
            "observed_endpoint_seconds": 2.344856615003664,
            "observed_intervals": 1,
            "requested_intervals": 1
          }
        },
        "ready": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 548.1502251793638,
            "reason": null,
            "response_union_seconds": 3.31754442199599,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 165.22769719224166,
            "trapezoid_pieces": 8,
            "watt_hours": 0.15226395143871216
          },
          "gpu_sample_count": 5,
          "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": 17240248320,
              "mean": 17218326528,
              "median": 17224609792,
              "minimum": 17190121472,
              "sample_count": 3
            },
            "gpu_board_watts": {
              "maximum": 300.01,
              "mean": 183.016,
              "median": 115.73,
              "minimum": 87.89,
              "sample_count": 5
            },
            "gpu_memory_used_bytes": {
              "maximum": 22502440960.0,
              "mean": 22497407795.2,
              "median": 22502440960.0,
              "minimum": 22477275136.0,
              "sample_count": 5
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2340.0,
              "mean": 2029.4,
              "median": 2310.0,
              "minimum": 1575.0,
              "sample_count": 5
            },
            "gpu_utilization_percent": {
              "maximum": 99.0,
              "mean": 39.6,
              "median": 0.0,
              "minimum": 0.0,
              "sample_count": 5
            },
            "host_mem_available_bytes": {
              "maximum": 96170061824,
              "mean": 96126780757.33333,
              "median": 96106954752,
              "minimum": 96103325696,
              "sample_count": 3
            },
            "process_pss_bytes": {
              "maximum": 925034496,
              "mean": 904964096,
              "median": 913561600,
              "minimum": 876296192,
              "sample_count": 3
            },
            "process_rss_bytes": {
              "maximum": 928030720,
              "mean": 907960320,
              "median": 916557824,
              "minimum": 879292416,
              "sample_count": 3
            }
          },
          "process_sample_count": 3,
          "response_union_seconds": 3.31754442199599,
          "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": 20792.77551286387,
            "reason": null,
            "response_union_seconds": 69.31620680802735,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 299.9698983882642,
            "trapezoid_pieces": 70,
            "watt_hours": 5.77577097579552
          },
          "gpu_sample_count": 68,
          "host_sample_count": 68,
          "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": 17240788992,
              "mean": 17231835497.411766,
              "median": 17233764352.0,
              "minimum": 17221709824,
              "sample_count": 68
            },
            "gpu_board_watts": {
              "maximum": 301.79,
              "mean": 299.9689705882353,
              "median": 300.0,
              "minimum": 297.03,
              "sample_count": 68
            },
            "gpu_memory_used_bytes": {
              "maximum": 22502440960.0,
              "mean": 22502440960.0,
              "median": 22502440960.0,
              "minimum": 22502440960.0,
              "sample_count": 68
            },
            "gpu_sm_clock_mhz": {
              "maximum": 1687.0,
              "mean": 1608.3088235294117,
              "median": 1605.0,
              "minimum": 1575.0,
              "sample_count": 68
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 99.02941176470588,
              "median": 99.0,
              "minimum": 84.0,
              "sample_count": 68
            },
            "host_mem_available_bytes": {
              "maximum": 96136097792,
              "mean": 96108860235.29411,
              "median": 96106113024.0,
              "minimum": 96084488192,
              "sample_count": 68
            },
            "process_pss_bytes": {
              "maximum": 925034496,
              "mean": 919076412.2352941,
              "median": 921339904.0,
              "minimum": 911089664,
              "sample_count": 68
            },
            "process_rss_bytes": {
              "maximum": 928030720,
              "mean": 922093507.7647059,
              "median": 924336128.0,
              "minimum": 915148800,
              "sample_count": 68
            }
          },
          "process_sample_count": 68,
          "response_union_seconds": 69.31620680802735,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 0,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 0,
              "storage_write_bytes": 65536
            },
            "observed_endpoint_seconds": 67.56908865296282,
            "observed_intervals": 2,
            "requested_intervals": 2
          }
        },
        "warmup": {
          "gpu_energy": {
            "available": true,
            "interpolation": "linear_at_bracketed_boundaries",
            "joules": 3891.3219742636566,
            "reason": null,
            "response_union_seconds": 13.74247567699058,
            "scope": "GPU board only; not whole-server AC energy",
            "time_weighted_mean_watts": 283.1601863977834,
            "trapezoid_pieces": 16,
            "watt_hours": 1.0809227706287934
          },
          "gpu_sample_count": 14,
          "host_sample_count": 13,
          "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": 17231908864,
              "mean": 17226236534.153847,
              "median": 17225515008,
              "minimum": 17220960256,
              "sample_count": 13
            },
            "gpu_board_watts": {
              "maximum": 306.92,
              "mean": 284.82928571428573,
              "median": 299.99,
              "minimum": 106.16,
              "sample_count": 14
            },
            "gpu_memory_used_bytes": {
              "maximum": 22502440960.0,
              "mean": 22502440960.0,
              "median": 22502440960.0,
              "minimum": 22502440960.0,
              "sample_count": 14
            },
            "gpu_sm_clock_mhz": {
              "maximum": 2197.0,
              "mean": 1657.7142857142858,
              "median": 1623.5,
              "minimum": 1590.0,
              "sample_count": 14
            },
            "gpu_utilization_percent": {
              "maximum": 100.0,
              "mean": 99.28571428571429,
              "median": 99.0,
              "minimum": 99.0,
              "sample_count": 14
            },
            "host_mem_available_bytes": {
              "maximum": 96137310208,
              "mean": 96114068558.76923,
              "median": 96116252672,
              "minimum": 96089333760,
              "sample_count": 13
            },
            "process_pss_bytes": {
              "maximum": 920127488,
              "mean": 915585339.0769231,
              "median": 914270208,
              "minimum": 911812608,
              "sample_count": 13
            },
            "process_rss_bytes": {
              "maximum": 923123712,
              "mean": 918581563.0769231,
              "median": 917266432,
              "minimum": 914808832,
              "sample_count": 13
            }
          },
          "process_sample_count": 13,
          "response_union_seconds": 13.74247567699058,
          "sampled_counter_deltas": {
            "deltas": {
              "major_faults": 8,
              "read_syscall_bytes": 0,
              "storage_read_bytes": 1372160,
              "storage_write_bytes": 12288
            },
            "observed_endpoint_seconds": 11.253299642936327,
            "observed_intervals": 2,
            "requested_intervals": 2
          }
        }
      },
      "reviewed_export_sha256": "59f09a9b924c4cb4ae1370f85109347542e72457c07f0a7a8ff351b31fedab64",
      "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": 93,
        "host_samples": 90,
        "maximum_gpu_gap_seconds": 1.0044987400178798,
        "maximum_host_gap_seconds": 1.322291104006581,
        "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": 93,
            "unavailable": 0
          },
          "software": {
            "active": 0,
            "not_active": 93,
            "unavailable": 0
          }
        }
      }
    },
    {
      "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
          }
        }
      }
    }
  ],
  "comparisons": [],
  "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": "57f1fd00f0013a2be96aa79b857391f27e23df5b5f847072b524c897e24d0361",
      "effective_chat_template_rule": "llama.cpp-7fe450e19-chat-source-v1",
      "effective_chat_template_sha256": "57f1fd00f0013a2be96aa79b857391f27e23df5b5f847072b524c897e24d0361",
      "files": [
        {
          "bytes": 19762149024,
          "sha256": "efd971561896866f0e910cce52761ca77b1b138090c7f15fe284676d57d1f689"
        }
      ],
      "id": "qwen3-32b-q4",
      "name": "Qwen3-32B",
      "quantization": "Q4_K_M",
      "repository": "Qwen/Qwen3-32B-GGUF",
      "revision": "938a7432affaec9157f883a87164e2646ae17555",
      "tokenizer_sha256": "2629971f294a1a43ef545d689a2090a820e0e1a15e5371e1c88d1c21159ef849"
    },
    {
      "architecture": "dense",
      "chat_template_sha256": "57f1fd00f0013a2be96aa79b857391f27e23df5b5f847072b524c897e24d0361",
      "effective_chat_template_rule": "llama.cpp-7fe450e19-chat-source-v1",
      "effective_chat_template_sha256": "57f1fd00f0013a2be96aa79b857391f27e23df5b5f847072b524c897e24d0361",
      "files": [
        {
          "bytes": 5027783488,
          "sha256": "d98cdcbd03e17ce47681435b5150e34c1417f50b5c0019dd560e4882c5745785"
        }
      ],
      "id": "qwen3-8b-q4",
      "name": "Qwen3-8B",
      "quantization": "Q4_K_M",
      "repository": "Qwen/Qwen3-8B-GGUF",
      "revision": "7c41481f57cb95916b40956ab2f0b139b296d974",
      "tokenizer_sha256": "2629971f294a1a43ef545d689a2090a820e0e1a15e5371e1c88d1c21159ef849"
    },
    {
      "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": "2026-09-28T00:41:20.654174+00:00",
      "plan_id": "core-sustained-20260927t203551-sustained79182232",
      "plan_sha256": "1f23d7c3b12fbcd78b46e277e64c415da613fa72c7fc9d831d17ae6859cf87be",
      "protocol_sha256": "3ce8ce9f2b17e66a655c07fd7e7aa3f1762790ad9be38f87d29e5e1279dec3ac",
      "record_sha256": "8941ad01c40078a4ab873ea0b80a24088e0a3fe1cd6bf76c20f728bd2f8122a0",
      "reviewed_export_sha256": "7e9bf3881a4bd446b234cda9c158b651eb9d123d6af6090c39eed1e26ad459ac",
      "runner_sha256": "1bac2436c5aa2eca4f4b42cc9b8f0942b5c66256eb1a41a2169b88c12b044f4e",
      "started_at": "2026-09-27T20:36:24.408847+00:00",
      "task_corpus_sha256": "0642db889565ce25b980e24afa5048eb83e4d15c36754457474a65b2c1f71c21"
    },
    {
      "exporter_sha256": "4f4046b91d99ba35175c3c6c04473a0021f54c6baa4892625cb66435f4c7e2ec",
      "finished_at": "2026-09-27T13:53:28.036613+00:00",
      "plan_id": "core-pilot8-20260927t135100-pilot8",
      "plan_sha256": "2952d4f1f06fa8282ea9ea2126cf5954bf3d774d16f9bc55a160ed6d808c3d49",
      "protocol_sha256": "3ce8ce9f2b17e66a655c07fd7e7aa3f1762790ad9be38f87d29e5e1279dec3ac",
      "record_sha256": "6c09a429d3b733b934749454b62d1a4243f5dce7c508680647c7d703a2b5daca",
      "reviewed_export_sha256": "2d1a82644c4ba9ba504ddf8fd2c7fe3c253e15cc9811363c52186bd4b7510a07",
      "runner_sha256": "1bac2436c5aa2eca4f4b42cc9b8f0942b5c66256eb1a41a2169b88c12b044f4e",
      "started_at": "2026-09-27T13:51:57.017889+00:00",
      "task_corpus_sha256": "0642db889565ce25b980e24afa5048eb83e4d15c36754457474a65b2c1f71c21"
    },
    {
      "exporter_sha256": "4f4046b91d99ba35175c3c6c04473a0021f54c6baa4892625cb66435f4c7e2ec",
      "finished_at": "2026-09-27T15:11:44.063808+00:00",
      "plan_id": "core-core-20260927t135400-core",
      "plan_sha256": "d6b25477667ae18cdbc23760f8a5540291e3d88246734ade84d167a32fc9011c",
      "protocol_sha256": "3ce8ce9f2b17e66a655c07fd7e7aa3f1762790ad9be38f87d29e5e1279dec3ac",
      "record_sha256": "fb26d7e444a2a826e9f652b7e048338d0bb8ddb92c1672ae6b919fc2334d492c",
      "reviewed_export_sha256": "59f09a9b924c4cb4ae1370f85109347542e72457c07f0a7a8ff351b31fedab64",
      "runner_sha256": "1bac2436c5aa2eca4f4b42cc9b8f0942b5c66256eb1a41a2169b88c12b044f4e",
      "started_at": "2026-09-27T13:56:01.723397+00:00",
      "task_corpus_sha256": "0642db889565ce25b980e24afa5048eb83e4d15c36754457474a65b2c1f71c21"
    },
    {
      "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": "2026-09-28"
}
