{
  "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": 76188.9364639213,
      "required_pages_per_second_upper_estimate": 0.005270927407262573
    },
    {
      "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": 9523.617057990163,
      "required_pages_per_second_upper_estimate": 0.003987339586441396
    },
    {
      "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": 19047.234115980325,
      "required_pages_per_second_upper_estimate": 0.003967036557016223
    },
    {
      "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": 28570.851173970488,
      "required_pages_per_second_upper_estimate": 0.001712778902715767
    },
    {
      "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": 28570.851173970488,
      "required_pages_per_second_upper_estimate": 0.002221258008883095
    },
    {
      "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": 28570.851173970488,
      "required_pages_per_second_upper_estimate": 0.3664762018697873
    }
  ],
  "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": 63,
    "prefetch_capacity_limit": 512,
    "batch_memory_estimate_bytes": 720750,
    "producer_batch_seconds": 0.14993500997690035,
    "producer_p95_seconds": 0.3327517670040834,
    "gpu_update_seconds": 0.967699556154237,
    "producer_batches_per_second": 6.669556364147803,
    "gpu_batches_per_second": 1.033378587021502,
    "producer_headroom": 6.454126733331496,
    "sustainable": true,
    "buffer_seconds": 60.96507203771693,
    "buffer_drain_seconds": null,
    "shard_target_mib": 16,
    "upload_pages_per_commit": 79,
    "max_pending_pages": 4096,
    "generation_bytes_per_second": 226320.4537732936,
    "upload_bytes_per_second": 3763872.806813212,
    "upload_sustainable": true,
    "logical_page_rows": 4096,
    "logical_cortex_page_rows": 128,
    "estimated_full_shards_per_hour": 48.563100909224566,
    "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": 1.2861647640020237,
      "bytes": 6463355,
      "pages": 21,
      "shards": 2
    },
    {
      "seconds": 1.3757250710004882,
      "bytes": 11835703,
      "pages": 44,
      "shards": 4
    },
    {
      "seconds": 1.5062627519982925,
      "bytes": 6228251,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.0129257990010956,
      "bytes": 2288801,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.3661311769974418,
      "bytes": 11305021,
      "pages": 40,
      "shards": 3
    },
    {
      "seconds": 1.2991957800004457,
      "bytes": 5765469,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.2353659159998642,
      "bytes": 8486468,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 1.681274881997524,
      "bytes": 6344982,
      "pages": 32,
      "shards": 2
    },
    {
      "seconds": 1.540335444999073,
      "bytes": 2545551,
      "pages": 34,
      "shards": 2
    },
    {
      "seconds": 1.1612642559994129,
      "bytes": 9600274,
      "pages": 33,
      "shards": 3
    },
    {
      "seconds": 1.406910452999,
      "bytes": 5890041,
      "pages": 38,
      "shards": 2
    },
    {
      "seconds": 1.0961665050017473,
      "bytes": 2428850,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 6.228057981999882,
      "bytes": 11900976,
      "pages": 33,
      "shards": 3
    },
    {
      "seconds": 1.5676271939992148,
      "bytes": 12360812,
      "pages": 44,
      "shards": 4
    },
    {
      "seconds": 1.7888847590002115,
      "bytes": 1223080,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 4.01536119299999,
      "bytes": 5846921,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 2.7826384430009057,
      "bytes": 7726579,
      "pages": 35,
      "shards": 3
    },
    {
      "seconds": 1.4537843559992325,
      "bytes": 4298944,
      "pages": 20,
      "shards": 2
    },
    {
      "seconds": 1.740884680999443,
      "bytes": 961930,
      "pages": 26,
      "shards": 1
    },
    {
      "seconds": 2.1066386819984473,
      "bytes": 13346856,
      "pages": 36,
      "shards": 4
    },
    {
      "seconds": 1.6676214010003605,
      "bytes": 6129871,
      "pages": 39,
      "shards": 2
    },
    {
      "seconds": 1.2406462889994145,
      "bytes": 5768030,
      "pages": 34,
      "shards": 2
    },
    {
      "seconds": 2.4877858859981643,
      "bytes": 7529407,
      "pages": 32,
      "shards": 3
    },
    {
      "seconds": 1.8479857339989394,
      "bytes": 6111417,
      "pages": 38,
      "shards": 2
    },
    {
      "seconds": 1.0848596139985602,
      "bytes": 2384316,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 2.864481454998895,
      "bytes": 16026880,
      "pages": 38,
      "shards": 4
    },
    {
      "seconds": 1.1737483289980446,
      "bytes": 1071557,
      "pages": 32,
      "shards": 1
    },
    {
      "seconds": 1.4824117819953244,
      "bytes": 7847757,
      "pages": 41,
      "shards": 3
    },
    {
      "seconds": 1.698374239000259,
      "bytes": 5912160,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.2937959129994852,
      "bytes": 7777018,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 1.521280608001689,
      "bytes": 6127630,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.3758144730018103,
      "bytes": 6889142,
      "pages": 32,
      "shards": 2
    },
    {
      "seconds": 1.881114518000686,
      "bytes": 7983801,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 1.3452542180020828,
      "bytes": 6178324,
      "pages": 24,
      "shards": 2
    },
    {
      "seconds": 1.2824458319955738,
      "bytes": 4898067,
      "pages": 41,
      "shards": 2
    },
    {
      "seconds": 1.4413858800035086,
      "bytes": 7671053,
      "pages": 37,
      "shards": 3
    },
    {
      "seconds": 1.4289778950042091,
      "bytes": 1006990,
      "pages": 33,
      "shards": 1
    },
    {
      "seconds": 1.5516541759934626,
      "bytes": 17080870,
      "pages": 28,
      "shards": 5
    },
    {
      "seconds": 1.6339488890007487,
      "bytes": 6146671,
      "pages": 38,
      "shards": 2
    },
    {
      "seconds": 1.0191766480056685,
      "bytes": 906488,
      "pages": 27,
      "shards": 1
    },
    {
      "seconds": 1.433525214997644,
      "bytes": 7753087,
      "pages": 40,
      "shards": 3
    },
    {
      "seconds": 2.9534675750037422,
      "bytes": 5286995,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.3232240569996065,
      "bytes": 8048155,
      "pages": 41,
      "shards": 3
    },
    {
      "seconds": 3.3732278300012695,
      "bytes": 11426245,
      "pages": 26,
      "shards": 3
    },
    {
      "seconds": 2.9584781199955614,
      "bytes": 1202687,
      "pages": 36,
      "shards": 1
    },
    {
      "seconds": 2.338399190994096,
      "bytes": 11297335,
      "pages": 36,
      "shards": 5
    },
    {
      "seconds": 2.0653895860014018,
      "bytes": 6188816,
      "pages": 38,
      "shards": 2
    },
    {
      "seconds": 1.478956710001512,
      "bytes": 1024661,
      "pages": 32,
      "shards": 1
    },
    {
      "seconds": 1.288585857997532,
      "bytes": 7415692,
      "pages": 34,
      "shards": 3
    },
    {
      "seconds": 1.3701110699985293,
      "bytes": 11305656,
      "pages": 28,
      "shards": 3
    },
    {
      "seconds": 1.8134901779994834,
      "bytes": 7862983,
      "pages": 41,
      "shards": 3
    },
    {
      "seconds": 1.568131148997054,
      "bytes": 886900,
      "pages": 26,
      "shards": 1
    },
    {
      "seconds": 2.2302413529978367,
      "bytes": 6385571,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 2.218881031003548,
      "bytes": 11370155,
      "pages": 33,
      "shards": 4
    },
    {
      "seconds": 1.2318224780028686,
      "bytes": 1293663,
      "pages": 36,
      "shards": 1
    },
    {
      "seconds": 1.8874604160009767,
      "bytes": 5227490,
      "pages": 13,
      "shards": 2
    },
    {
      "seconds": 1.7371134120039642,
      "bytes": 13601006,
      "pages": 50,
      "shards": 4
    },
    {
      "seconds": 1.710468639001192,
      "bytes": 4673486,
      "pages": 32,
      "shards": 2
    },
    {
      "seconds": 1.66814742799761,
      "bytes": 318500,
      "pages": 9,
      "shards": 1
    },
    {
      "seconds": 1.7101979840008426,
      "bytes": 7394765,
      "pages": 40,
      "shards": 3
    },
    {
      "seconds": 1.701770421997935,
      "bytes": 5790962,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.3935572619957384,
      "bytes": 5831979,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 2.100949927000329,
      "bytes": 13094733,
      "pages": 28,
      "shards": 4
    },
    {
      "seconds": 1.6310563769948203,
      "bytes": 6709713,
      "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": [
    2.576500048000014,
    0.012398547999964649
  ]
}