{
  "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": 74551.87473408281,
      "required_pages_per_second_upper_estimate": 0.006461967551951249
    },
    {
      "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": 9318.984341760352,
      "required_pages_per_second_upper_estimate": 0.003717482185160504
    },
    {
      "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": 18637.968683520703,
      "required_pages_per_second_upper_estimate": 0.004311566037886858
    },
    {
      "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": 27956.95302528106,
      "required_pages_per_second_upper_estimate": 0.0017024317604430275
    },
    {
      "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": 27956.95302528106,
      "required_pages_per_second_upper_estimate": 0.0021522036336852775
    },
    {
      "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": 21,
      "required_token_positions_per_second": 27956.95302528106,
      "required_pages_per_second_upper_estimate": 0.34713179067113326
    }
  ],
  "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": 61,
    "prefetch_capacity_limit": 512,
    "batch_memory_estimate_bytes": 690670,
    "producer_batch_seconds": 0.7983091551562893,
    "producer_p95_seconds": 0.3582648830015387,
    "gpu_update_seconds": 0.9889489736237824,
    "producer_batches_per_second": 1.2526475407941733,
    "gpu_batches_per_second": 1.0111745162500383,
    "producer_headroom": 1.2388044997807528,
    "sustainable": true,
    "buffer_seconds": 60.32588739105073,
    "buffer_drain_seconds": null,
    "shard_target_mib": 16,
    "upload_pages_per_commit": 80,
    "max_pending_pages": 4096,
    "generation_bytes_per_second": 263295.85887776804,
    "upload_bytes_per_second": 4407349.330201067,
    "upload_sustainable": true,
    "logical_page_rows": 4096,
    "logical_cortex_page_rows": 128,
    "estimated_full_shards_per_hour": 56.49716210126668,
    "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": 1.154916157000116,
      "bytes": 2367312,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.8538454429981357,
      "bytes": 11918988,
      "pages": 37,
      "shards": 3
    },
    {
      "seconds": 1.6541002780031704,
      "bytes": 6139021,
      "pages": 33,
      "shards": 2
    },
    {
      "seconds": 1.9353761299971666,
      "bytes": 7270095,
      "pages": 29,
      "shards": 3
    },
    {
      "seconds": 2.698638427998958,
      "bytes": 6167283,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 2.6425060760011547,
      "bytes": 2623268,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 2.5355882220028434,
      "bytes": 9765836,
      "pages": 33,
      "shards": 3
    },
    {
      "seconds": 1.361022408997087,
      "bytes": 6242966,
      "pages": 18,
      "shards": 3
    },
    {
      "seconds": 1.2147977610002272,
      "bytes": 6193785,
      "pages": 35,
      "shards": 2
    },
    {
      "seconds": 0.9691767799995432,
      "bytes": 2430413,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 1.0292037019971758,
      "bytes": 5575102,
      "pages": 18,
      "shards": 2
    },
    {
      "seconds": 1.288112906000606,
      "bytes": 5599143,
      "pages": 33,
      "shards": 2
    },
    {
      "seconds": 1.0466477340014535,
      "bytes": 7416772,
      "pages": 30,
      "shards": 3
    },
    {
      "seconds": 1.263268196998979,
      "bytes": 1135440,
      "pages": 34,
      "shards": 1
    },
    {
      "seconds": 1.37565966299735,
      "bytes": 13270818,
      "pages": 36,
      "shards": 4
    },
    {
      "seconds": 0.7815917669977352,
      "bytes": 913950,
      "pages": 28,
      "shards": 1
    },
    {
      "seconds": 1.2434416440009954,
      "bytes": 10283141,
      "pages": 41,
      "shards": 3
    },
    {
      "seconds": 1.034428533999744,
      "bytes": 7372684,
      "pages": 29,
      "shards": 3
    },
    {
      "seconds": 1.070391503002611,
      "bytes": 1123754,
      "pages": 35,
      "shards": 1
    },
    {
      "seconds": 1.18286236700078,
      "bytes": 5852988,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.7732943640003214,
      "bytes": 17450620,
      "pages": 33,
      "shards": 5
    },
    {
      "seconds": 0.8563921040004061,
      "bytes": 1024816,
      "pages": 34,
      "shards": 1
    },
    {
      "seconds": 1.4308035579997522,
      "bytes": 7307496,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 1.6164238569981535,
      "bytes": 892499,
      "pages": 26,
      "shards": 1
    },
    {
      "seconds": 1.2368712060015241,
      "bytes": 10018327,
      "pages": 37,
      "shards": 3
    },
    {
      "seconds": 1.856085086998064,
      "bytes": 7579977,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 1.2339968710002722,
      "bytes": 6738522,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 2.4403519309998956,
      "bytes": 5986840,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 2.588106572999095,
      "bytes": 12282184,
      "pages": 39,
      "shards": 4
    },
    {
      "seconds": 2.1326607650007645,
      "bytes": 1150711,
      "pages": 35,
      "shards": 1
    },
    {
      "seconds": 2.922968579998269,
      "bytes": 7646925,
      "pages": 36,
      "shards": 3
    },
    {
      "seconds": 2.558325848000095,
      "bytes": 890978,
      "pages": 28,
      "shards": 1
    },
    {
      "seconds": 2.274885417002224,
      "bytes": 15797221,
      "pages": 34,
      "shards": 4
    },
    {
      "seconds": 3.0481821479988866,
      "bytes": 2407311,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.2172690299994429,
      "bytes": 6177200,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.8479619009995076,
      "bytes": 5907251,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 1.345177949002391,
      "bytes": 6941210,
      "pages": 37,
      "shards": 3
    },
    {
      "seconds": 1.2569710710013169,
      "bytes": 6280484,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.521943290001218,
      "bytes": 13325077,
      "pages": 36,
      "shards": 4
    },
    {
      "seconds": 1.0711962380009936,
      "bytes": 889212,
      "pages": 26,
      "shards": 2
    },
    {
      "seconds": 1.3221969680016628,
      "bytes": 10781457,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 1.231661190999148,
      "bytes": 2563678,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.1248420010015252,
      "bytes": 5951433,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.6543720410008973,
      "bytes": 7466727,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 1.3103912280021177,
      "bytes": 10646431,
      "pages": 28,
      "shards": 3
    },
    {
      "seconds": 1.3748210780031513,
      "bytes": 6494530,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 0.9434360529994592,
      "bytes": 2587569,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.083952794000652,
      "bytes": 5627431,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.2045630370012077,
      "bytes": 11050050,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 1.3764796949981246,
      "bytes": 2478260,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.1128817630014964,
      "bytes": 6162674,
      "pages": 35,
      "shards": 2
    },
    {
      "seconds": 1.2386194030004845,
      "bytes": 6597417,
      "pages": 26,
      "shards": 2
    },
    {
      "seconds": 1.1337901490005606,
      "bytes": 11234450,
      "pages": 36,
      "shards": 4
    },
    {
      "seconds": 0.9795328820000577,
      "bytes": 930296,
      "pages": 29,
      "shards": 1
    },
    {
      "seconds": 1.2613436630017532,
      "bytes": 7588531,
      "pages": 35,
      "shards": 3
    },
    {
      "seconds": 0.8843855299965071,
      "bytes": 5949979,
      "pages": 34,
      "shards": 2
    },
    {
      "seconds": 1.1865619210002478,
      "bytes": 5319050,
      "pages": 35,
      "shards": 2
    },
    {
      "seconds": 1.3752340050014027,
      "bytes": 13115816,
      "pages": 30,
      "shards": 4
    },
    {
      "seconds": 1.269822567999654,
      "bytes": 1251620,
      "pages": 34,
      "shards": 1
    },
    {
      "seconds": 0.9414359830007015,
      "bytes": 5894779,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 1.1529198649986938,
      "bytes": 11606253,
      "pages": 40,
      "shards": 4
    },
    {
      "seconds": 0.9286228500022844,
      "bytes": 849903,
      "pages": 26,
      "shards": 1
    },
    {
      "seconds": 1.0139027430013812,
      "bytes": 7512493,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 0.9002696629977436,
      "bytes": 6383764,
      "pages": 19,
      "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": 86
  },
  "cache_reused": true,
  "resume_probe_seconds": [
    2.844574209999962,
    0.2503580399999805
  ]
}