{
  "version": 1,
  "startup_batch_seconds": [
    17.862166522000052,
    0.01599195200003578,
    0.012736249999989013,
    0.10922379700002693,
    0.015912060000005113,
    0.014246264000007614,
    0.01397435499995936,
    0.10559937200002878
  ],
  "source_benchmarks": [
    {
      "source": "climbmix",
      "rows": 256,
      "tokens": 721064,
      "bytes": 302776,
      "generation_seconds": 0.6739276440000594,
      "write_seconds": 0.00442227799999273,
      "read_seconds": 0.04565144100001817,
      "tokens_per_second": 1062967.6168812835,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 75964.15602325292,
      "required_pages_per_second_upper_estimate": 0.006584380514702311
    },
    {
      "source": "rewrite6",
      "rows": 256,
      "tokens": 156675,
      "bytes": 68694,
      "generation_seconds": 0.12810849499999222,
      "write_seconds": 0.001732062999963091,
      "read_seconds": 0.002349199000036606,
      "tokens_per_second": 1206672.263377472,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 9495.519502906614,
      "required_pages_per_second_upper_estimate": 0.0037879047003776187
    },
    {
      "source": "openmath",
      "rows": 256,
      "tokens": 270174,
      "bytes": 90063,
      "generation_seconds": 0.6564187429999038,
      "write_seconds": 0.002473517999987962,
      "read_seconds": 0.0032515120000198294,
      "tokens_per_second": 410042.7581134458,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 18991.03900581323,
      "required_pages_per_second_upper_estimate": 0.00439324264312379
    },
    {
      "source": "ultra_qa",
      "rows": 256,
      "tokens": 1026361,
      "bytes": 365167,
      "generation_seconds": 0.46967248600003586,
      "write_seconds": 0.004728889000034542,
      "read_seconds": 0.0050225659999796335,
      "tokens_per_second": 2163486.5624068812,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 28486.55850871984,
      "required_pages_per_second_upper_estimate": 0.0017346819557592212
    },
    {
      "source": "ultra_style",
      "rows": 256,
      "tokens": 811870,
      "bytes": 338806,
      "generation_seconds": 0.5208254859999215,
      "write_seconds": 0.004081093000081637,
      "read_seconds": 0.004787273000033565,
      "tokens_per_second": 1546694.2737633225,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 1,
      "required_token_positions_per_second": 28486.55850871984,
      "required_pages_per_second_upper_estimate": 0.0021929741298422037
    },
    {
      "source": "cortex",
      "rows": 128,
      "tokens": 80537,
      "bytes": 22745,
      "generation_seconds": 0.07537080299994159,
      "write_seconds": 0.003962095000019872,
      "read_seconds": 0.042780299000014566,
      "tokens_per_second": 1015177.8395898147,
      "note": "Bounded fresh sample; full-page latency is an extrapolation.",
      "estimated_pages_for_buffer": 22,
      "required_token_positions_per_second": 28486.55850871984,
      "required_pages_per_second_upper_estimate": 0.3537077182999099
    }
  ],
  "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": 16107520,
      "pages": 6,
      "packing_seconds": 0.03508346299997811,
      "read_seconds": 0.0015898530000413302,
      "io_bytes_per_second": 439216350.1111124,
      "target_reached": true
    },
    {
      "target_mib": 32,
      "actual_bytes": 18657280,
      "pages": 8,
      "packing_seconds": 0.04401178199998412,
      "read_seconds": 0.0016850460000341627,
      "io_bytes_per_second": 408283918.5247723,
      "target_reached": false
    },
    {
      "target_mib": 64,
      "actual_bytes": 18657280,
      "pages": 8,
      "packing_seconds": 0.04403130000002875,
      "read_seconds": 0.0015284270000393008,
      "io_bytes_per_second": 409512550.4148024,
      "target_reached": false
    },
    {
      "target_mib": 128,
      "actual_bytes": 18657280,
      "pages": 8,
      "packing_seconds": 0.04190044699998907,
      "read_seconds": 0.001583749000019452,
      "io_bytes_per_second": 429058870.0316856,
      "target_reached": false
    },
    {
      "target_mib": 256,
      "actual_bytes": 18657280,
      "pages": 8,
      "packing_seconds": 0.042332877000035296,
      "read_seconds": 0.0015621619999137693,
      "io_bytes_per_second": 425043021.37700915,
      "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": 0.1434574083060457,
  "selected": {
    "prefetch_batches": 62,
    "prefetch_capacity_limit": 512,
    "batch_memory_estimate_bytes": 688946,
    "producer_batch_seconds": 0.13243488976564777,
    "producer_p95_seconds": 0.3283783129991207,
    "gpu_update_seconds": 0.9705630110263002,
    "producer_batches_per_second": 7.55088029876089,
    "gpu_batches_per_second": 1.0303298071730267,
    "producer_headroom": 7.328605118664539,
    "sustainable": true,
    "buffer_seconds": 60.17490668363061,
    "buffer_drain_seconds": null,
    "shard_target_mib": 16,
    "upload_pages_per_commit": 82,
    "max_pending_pages": 4096,
    "generation_bytes_per_second": 286918.1228622732,
    "upload_bytes_per_second": 3925264.591360211,
    "upload_sustainable": true,
    "logical_page_rows": 4096,
    "logical_cortex_page_rows": 128,
    "estimated_full_shards_per_hour": 61.56595005417964,
    "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": 1.1223962459989707,
      "bytes": 6954584,
      "pages": 39,
      "shards": 2
    },
    {
      "seconds": 1.6785005409983569,
      "bytes": 5701009,
      "pages": 27,
      "shards": 2
    },
    {
      "seconds": 0.9818456940010947,
      "bytes": 2688719,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.0817985579997185,
      "bytes": 6075598,
      "pages": 20,
      "shards": 2
    },
    {
      "seconds": 1.3699396060001163,
      "bytes": 12184416,
      "pages": 47,
      "shards": 4
    },
    {
      "seconds": 0.945849695999641,
      "bytes": 1129028,
      "pages": 33,
      "shards": 1
    },
    {
      "seconds": 1.1423426910005219,
      "bytes": 5919036,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.1784255179991305,
      "bytes": 13145531,
      "pages": 35,
      "shards": 5
    },
    {
      "seconds": 1.339017384001636,
      "bytes": 6203121,
      "pages": 35,
      "shards": 2
    },
    {
      "seconds": 1.3240333100002317,
      "bytes": 6490741,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 0.9587836459977552,
      "bytes": 2593270,
      "pages": 35,
      "shards": 2
    },
    {
      "seconds": 0.8898652169991692,
      "bytes": 5769808,
      "pages": 27,
      "shards": 2
    },
    {
      "seconds": 2.3888570920025813,
      "bytes": 6457509,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 1.143103587000951,
      "bytes": 11836705,
      "pages": 33,
      "shards": 3
    },
    {
      "seconds": 1.1624738930004241,
      "bytes": 6035324,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.368994181000744,
      "bytes": 7247693,
      "pages": 32,
      "shards": 3
    },
    {
      "seconds": 0.8721044880003319,
      "bytes": 6235096,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.1694963700028893,
      "bytes": 895144,
      "pages": 30,
      "shards": 1
    },
    {
      "seconds": 1.171581717000663,
      "bytes": 7061925,
      "pages": 24,
      "shards": 3
    },
    {
      "seconds": 1.0121167409997724,
      "bytes": 6826316,
      "pages": 33,
      "shards": 2
    },
    {
      "seconds": 1.0728631890015095,
      "bytes": 9404752,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 32.4843381790015,
      "bytes": 7401467,
      "pages": 32,
      "shards": 3
    },
    {
      "seconds": 1.1189806990005309,
      "bytes": 7079866,
      "pages": 44,
      "shards": 2
    },
    {
      "seconds": 1.104166561999591,
      "bytes": 7255454,
      "pages": 33,
      "shards": 3
    },
    {
      "seconds": 1.0669172650013934,
      "bytes": 6210420,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.0871770170015225,
      "bytes": 6926231,
      "pages": 33,
      "shards": 2
    },
    {
      "seconds": 1.2097226180012512,
      "bytes": 7368748,
      "pages": 29,
      "shards": 3
    },
    {
      "seconds": 1.1227261329986504,
      "bytes": 5278941,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.216369551999378,
      "bytes": 7462357,
      "pages": 34,
      "shards": 3
    },
    {
      "seconds": 1.5228881869988982,
      "bytes": 5895199,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.0030821439977444,
      "bytes": 1174580,
      "pages": 34,
      "shards": 1
    },
    {
      "seconds": 1.8239342810011294,
      "bytes": 17807257,
      "pages": 31,
      "shards": 5
    },
    {
      "seconds": 1.1584871390004992,
      "bytes": 5888359,
      "pages": 33,
      "shards": 2
    },
    {
      "seconds": 0.8503742749999219,
      "bytes": 1173601,
      "pages": 34,
      "shards": 1
    },
    {
      "seconds": 1.4903823460008425,
      "bytes": 7681898,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 0.8942368999996688,
      "bytes": 850966,
      "pages": 25,
      "shards": 1
    },
    {
      "seconds": 1.4387539480012492,
      "bytes": 11588272,
      "pages": 38,
      "shards": 4
    },
    {
      "seconds": 1.6636693889995513,
      "bytes": 11864781,
      "pages": 37,
      "shards": 3
    },
    {
      "seconds": 1.1722596879990306,
      "bytes": 1208158,
      "pages": 35,
      "shards": 2
    },
    {
      "seconds": 1.2069239850025042,
      "bytes": 7251816,
      "pages": 27,
      "shards": 3
    },
    {
      "seconds": 1.169119137000962,
      "bytes": 6109087,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.4141185940025025,
      "bytes": 7704490,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 1.2197817560008843,
      "bytes": 5784149,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.3641967180010397,
      "bytes": 7009240,
      "pages": 38,
      "shards": 2
    },
    {
      "seconds": 1.4704428500008362,
      "bytes": 11028707,
      "pages": 30,
      "shards": 4
    },
    {
      "seconds": 0.9246157949964982,
      "bytes": 1204930,
      "pages": 34,
      "shards": 1
    },
    {
      "seconds": 1.0427682160006952,
      "bytes": 5985007,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.1093607039983908,
      "bytes": 7338185,
      "pages": 30,
      "shards": 3
    },
    {
      "seconds": 0.9098767819996283,
      "bytes": 5901703,
      "pages": 35,
      "shards": 2
    },
    {
      "seconds": 1.1044195929971465,
      "bytes": 13283238,
      "pages": 34,
      "shards": 4
    },
    {
      "seconds": 0.8937192140001571,
      "bytes": 6207827,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 0.9678468759993848,
      "bytes": 1151038,
      "pages": 35,
      "shards": 1
    },
    {
      "seconds": 1.3402203920013562,
      "bytes": 12178587,
      "pages": 33,
      "shards": 4
    },
    {
      "seconds": 0.8894950190006057,
      "bytes": 1180133,
      "pages": 34,
      "shards": 1
    },
    {
      "seconds": 1.1408973270008573,
      "bytes": 5952533,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.4523514239990618,
      "bytes": 13500548,
      "pages": 42,
      "shards": 4
    },
    {
      "seconds": 0.8281688100032625,
      "bytes": 661491,
      "pages": 19,
      "shards": 1
    },
    {
      "seconds": 1.42470605500057,
      "bytes": 11031230,
      "pages": 41,
      "shards": 4
    },
    {
      "seconds": 0.8167821940005524,
      "bytes": 1161839,
      "pages": 34,
      "shards": 1
    },
    {
      "seconds": 0.9240367780002998,
      "bytes": 5199805,
      "pages": 10,
      "shards": 2
    },
    {
      "seconds": 1.4070672229972843,
      "bytes": 1010316,
      "pages": 29,
      "shards": 1
    },
    {
      "seconds": 1.181387053002254,
      "bytes": 11521167,
      "pages": 38,
      "shards": 4
    },
    {
      "seconds": 1.134901017001539,
      "bytes": 6770130,
      "pages": 32,
      "shards": 2
    },
    {
      "seconds": 1.1226096480022534,
      "bytes": 7935280,
      "pages": 38,
      "shards": 3
    }
  ],
  "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": 86
  },
  "cache_reused": true,
  "resume_probe_seconds": [
    2.5993740069999944,
    0.014666526000041813
  ]
}