{
  "version": 1,
  "startup_batch_seconds": [
    2.3260779750000324,
    0.017838069999982054,
    0.0158013189999906,
    0.015721255000016754,
    0.09197544800002788,
    0.014327330000014626,
    0.013824727999974584,
    0.10075297500003444
  ],
  "source_benchmarks": [
    {
      "source": "climbmix",
      "rows": 256,
      "tokens": 903410,
      "bytes": 372598,
      "generation_seconds": 1.5044533369999726,
      "write_seconds": 0.005470204000005197,
      "read_seconds": 0.004841493000014907,
      "tokens_per_second": 598315.063954628,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 77638.02404567161,
      "required_pages_per_second_upper_estimate": 0.00537117864851449
    },
    {
      "source": "rewrite6",
      "rows": 256,
      "tokens": 149279,
      "bytes": 66870,
      "generation_seconds": 17.489590872999997,
      "write_seconds": 0.0027630569999814725,
      "read_seconds": 0.003109784999992371,
      "tokens_per_second": 8533.957213384612,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 9704.753005708952,
      "required_pages_per_second_upper_estimate": 0.004063177425202537
    },
    {
      "source": "openmath",
      "rows": 256,
      "tokens": 300086,
      "bytes": 97823,
      "generation_seconds": 0.7871533299999669,
      "write_seconds": 0.0022467860000006112,
      "read_seconds": 0.0028803160000165917,
      "tokens_per_second": 380144.35761751566,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 19409.506011417903,
      "required_pages_per_second_upper_estimate": 0.004042488239083526
    },
    {
      "source": "ultra_qa",
      "rows": 256,
      "tokens": 1042562,
      "bytes": 369139,
      "generation_seconds": 23.827993708999998,
      "write_seconds": 0.0062629279999555365,
      "read_seconds": 0.006445001999964006,
      "tokens_per_second": 43742.16556775435,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 29114.259017126853,
      "required_pages_per_second_upper_estimate": 0.0017453553731772578
    },
    {
      "source": "ultra_style",
      "rows": 256,
      "tokens": 803904,
      "bytes": 337820,
      "generation_seconds": 36.50458967899999,
      "write_seconds": 0.00519779599994763,
      "read_seconds": 0.005391688000031536,
      "tokens_per_second": 22018.86276523721,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 29114.259017126853,
      "required_pages_per_second_upper_estimate": 0.002263505578490004
    },
    {
      "source": "cortex",
      "rows": 128,
      "tokens": 77961,
      "bytes": 43256,
      "generation_seconds": 0.08065045199998622,
      "write_seconds": 0.004991313999994418,
      "read_seconds": 0.04061180899998362,
      "tokens_per_second": 910315.184299418,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 23,
      "required_token_positions_per_second": 29114.259017126853,
      "required_pages_per_second_upper_estimate": 0.37344645421591377
    }
  ],
  "gpu_measurement": "Selected forward/backward estimate; replaced by live successful update timings.",
  "layout": "Logical pages unchanged; multiple pages per immutable TAR shard.",
  "candidate_shard_mib": [
    16,
    32,
    64,
    128,
    256
  ],
  "local_shard_size_benchmark": [
    {
      "target_mib": 16,
      "actual_bytes": 15554560,
      "pages": 7,
      "packing_seconds": 0.03195600199990167,
      "read_seconds": 0.0013301719999390116,
      "io_bytes_per_second": 467297923.7588089,
      "target_reached": true
    },
    {
      "target_mib": 32,
      "actual_bytes": 18083840,
      "pages": 9,
      "packing_seconds": 0.0402841879999869,
      "read_seconds": 0.0016377959999545055,
      "io_bytes_per_second": 431368897.0451703,
      "target_reached": false
    },
    {
      "target_mib": 64,
      "actual_bytes": 18083840,
      "pages": 9,
      "packing_seconds": 0.0399535819999528,
      "read_seconds": 0.0013251859999172666,
      "io_bytes_per_second": 438090594.17802685,
      "target_reached": false
    },
    {
      "target_mib": 128,
      "actual_bytes": 18083840,
      "pages": 9,
      "packing_seconds": 0.04405787199993938,
      "read_seconds": 0.0014645830000290516,
      "io_bytes_per_second": 397250983.05907583,
      "target_reached": false
    },
    {
      "target_mib": 256,
      "actual_bytes": 18083840,
      "pages": 9,
      "packing_seconds": 0.038781767000045875,
      "read_seconds": 0.0013218610000649278,
      "io_bytes_per_second": 450927781.39548963,
      "target_reached": false
    }
  ],
  "shard_benchmark_scope": "Local packing/read measurements; upload capacity measured only on real data commits. Targets larger than available pages are sample-limited.",
  "estimated_fresh_generation_batch_seconds": 2.558590814491007,
  "selected": {
    "prefetch_batches": 64,
    "prefetch_capacity_limit": 512,
    "batch_memory_estimate_bytes": 720750,
    "producer_batch_seconds": 0.13055367253895156,
    "producer_p95_seconds": 0.26719631500236574,
    "gpu_update_seconds": 0.9496377697174327,
    "producer_batches_per_second": 7.659684944532245,
    "gpu_batches_per_second": 1.0530330952375164,
    "producer_headroom": 7.273926127463798,
    "sustainable": true,
    "buffer_seconds": 60.776817261915696,
    "buffer_drain_seconds": null,
    "shard_target_mib": 16,
    "upload_pages_per_commit": 86,
    "max_pending_pages": 4096,
    "generation_bytes_per_second": 276929.98004016274,
    "upload_bytes_per_second": 4630155.953714088,
    "upload_sustainable": true,
    "logical_page_rows": 4096,
    "logical_cortex_page_rows": 128,
    "estimated_full_shards_per_hour": 59.4227271166197,
    "selection_policy": "Smallest shard within 90% of measured local peak, enlarged to amortize 30 seconds of measured page production; bounded to 256 MiB."
  },
  "upload_measurements": [
    {
      "seconds": 1.2958958110029926,
      "bytes": 5955386,
      "pages": 27,
      "shards": 2
    },
    {
      "seconds": 1.584322863000125,
      "bytes": 6087710,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.2062028339969402,
      "bytes": 5756694,
      "pages": 35,
      "shards": 3
    },
    {
      "seconds": 1.087463227002445,
      "bytes": 6477149,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.7014945069968235,
      "bytes": 13448331,
      "pages": 37,
      "shards": 4
    },
    {
      "seconds": 1.4291834129980998,
      "bytes": 4870018,
      "pages": 31,
      "shards": 2
    },
    {
      "seconds": 1.6493652240023948,
      "bytes": 6440305,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.0658796659990912,
      "bytes": 2378111,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 1.2226023729999724,
      "bytes": 6024557,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 2.320106137998664,
      "bytes": 11400278,
      "pages": 37,
      "shards": 4
    },
    {
      "seconds": 1.142467518999183,
      "bytes": 6818785,
      "pages": 33,
      "shards": 2
    },
    {
      "seconds": 1.1884660100004112,
      "bytes": 6414581,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 0.931925422999484,
      "bytes": 2347633,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.7378346169971337,
      "bytes": 10782839,
      "pages": 36,
      "shards": 3
    },
    {
      "seconds": 1.3063333719983348,
      "bytes": 7915290,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 1.0522420070010412,
      "bytes": 1136593,
      "pages": 36,
      "shards": 1
    },
    {
      "seconds": 1.2345661709987326,
      "bytes": 11661576,
      "pages": 28,
      "shards": 3
    },
    {
      "seconds": 1.2733937859993603,
      "bytes": 6517566,
      "pages": 37,
      "shards": 3
    },
    {
      "seconds": 1.2599434170006134,
      "bytes": 6272622,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.9948137850005878,
      "bytes": 7735186,
      "pages": 36,
      "shards": 3
    },
    {
      "seconds": 1.6032572549993347,
      "bytes": 5165320,
      "pages": 25,
      "shards": 2
    },
    {
      "seconds": 1.8941422179996152,
      "bytes": 6019433,
      "pages": 39,
      "shards": 2
    },
    {
      "seconds": 1.3825668060017051,
      "bytes": 13491032,
      "pages": 40,
      "shards": 5
    },
    {
      "seconds": 0.9105340399983106,
      "bytes": 1249369,
      "pages": 35,
      "shards": 1
    },
    {
      "seconds": 1.295263830001204,
      "bytes": 5736433,
      "pages": 23,
      "shards": 2
    },
    {
      "seconds": 1.218638172998908,
      "bytes": 7112834,
      "pages": 43,
      "shards": 3
    },
    {
      "seconds": 1.2346669860016846,
      "bytes": 6391793,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.537398492997454,
      "bytes": 7291338,
      "pages": 30,
      "shards": 3
    },
    {
      "seconds": 1.5484525030005898,
      "bytes": 10816659,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 2.108060000999103,
      "bytes": 5732876,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.102688891998696,
      "bytes": 2347731,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.3670132119987102,
      "bytes": 6121191,
      "pages": 38,
      "shards": 2
    },
    {
      "seconds": 1.156715154000267,
      "bytes": 10313432,
      "pages": 31,
      "shards": 3
    },
    {
      "seconds": 1.0891685620008502,
      "bytes": 2577323,
      "pages": 38,
      "shards": 2
    },
    {
      "seconds": 1.423857304998819,
      "bytes": 12061949,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 1.17036508000092,
      "bytes": 7649153,
      "pages": 31,
      "shards": 3
    },
    {
      "seconds": 1.2898340729989286,
      "bytes": 5540548,
      "pages": 38,
      "shards": 2
    },
    {
      "seconds": 1.2250885649991687,
      "bytes": 5931268,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.6209707380003238,
      "bytes": 2622297,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.044743833001121,
      "bytes": 6332578,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.5906551469997794,
      "bytes": 16693261,
      "pages": 23,
      "shards": 5
    },
    {
      "seconds": 1.596418191002158,
      "bytes": 1389732,
      "pages": 43,
      "shards": 1
    },
    {
      "seconds": 1.2193843360000756,
      "bytes": 5952773,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.8950818810008059,
      "bytes": 7230636,
      "pages": 33,
      "shards": 3
    },
    {
      "seconds": 1.5622936799991294,
      "bytes": 5942154,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 2.6776525460008997,
      "bytes": 7165592,
      "pages": 30,
      "shards": 3
    },
    {
      "seconds": 1.7035883720018319,
      "bytes": 12069614,
      "pages": 41,
      "shards": 3
    },
    {
      "seconds": 1.154275818997121,
      "bytes": 1133519,
      "pages": 32,
      "shards": 1
    },
    {
      "seconds": 1.6667711109985248,
      "bytes": 11320824,
      "pages": 31,
      "shards": 4
    },
    {
      "seconds": 1.2526901849996648,
      "bytes": 1223359,
      "pages": 35,
      "shards": 1
    },
    {
      "seconds": 1.1693511949997628,
      "bytes": 7842479,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 0.9436758240008203,
      "bytes": 855631,
      "pages": 27,
      "shards": 1
    },
    {
      "seconds": 1.5082043110014638,
      "bytes": 16602529,
      "pages": 32,
      "shards": 4
    },
    {
      "seconds": 1.5457586960001208,
      "bytes": 7678717,
      "pages": 40,
      "shards": 4
    },
    {
      "seconds": 0.9852680410003813,
      "bytes": 917014,
      "pages": 26,
      "shards": 1
    },
    {
      "seconds": 3.7047144000025582,
      "bytes": 6068293,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.0213589359991602,
      "bytes": 2097340,
      "pages": 21,
      "shards": 2
    },
    {
      "seconds": 1.7385368430004746,
      "bytes": 10700049,
      "pages": 47,
      "shards": 3
    },
    {
      "seconds": 1.4434581130008155,
      "bytes": 11762868,
      "pages": 29,
      "shards": 3
    },
    {
      "seconds": 1.317350670000451,
      "bytes": 2491706,
      "pages": 34,
      "shards": 2
    },
    {
      "seconds": 1.2309653659976902,
      "bytes": 5725164,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 2.086732669002231,
      "bytes": 12324271,
      "pages": 40,
      "shards": 4
    },
    {
      "seconds": 1.4604756749977241,
      "bytes": 1030943,
      "pages": 27,
      "shards": 1
    },
    {
      "seconds": 1.2873325260006823,
      "bytes": 5911958,
      "pages": 38,
      "shards": 2
    }
  ],
  "cache_key": {
    "version": 1,
    "kind": "data",
    "gpu": {
      "name": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
      "total_memory": 101975851008,
      "capability": [
        12,
        0
      ]
    },
    "torch": "2.8.0+cu128",
    "cuda": "12.8",
    "python": "3.13.11",
    "packages": {
      "transformers": "4.57.1",
      "numpy": "2.2.6",
      "pyarrow": "21.0.0",
      "huggingface_hub": "0.36.0"
    },
    "settings": {
      "context": 2048,
      "global_batch": 90,
      "page_rows": 4096,
      "cortex_page_rows": 128,
      "mix": {
        "climbmix": 40,
        "ultra_qa": 15,
        "ultra_style": 15,
        "cortex": 15,
        "openmath": 10,
        "rewrite6": 5
      },
      "cpu_threads": 4
    },
    "source": "f8722b66ff2f92bca174f36564336149e85c65c48b964d05ff3a4d5d85da308f",
    "cpu": "unknown",
    "cpu_count": 20,
    "microbatch": 90
  },
  "cache_reused": true,
  "resume_probe_seconds": [
    2.975958035000019,
    0.01771132999999736
  ]
}