{
  "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": 77423.96757010218,
      "required_pages_per_second_upper_estimate": 0.005356369724855144
    },
    {
      "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": 9677.995946262772,
      "required_pages_per_second_upper_estimate": 0.00405197480316336
    },
    {
      "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": 19355.991892525544,
      "required_pages_per_second_upper_estimate": 0.004031342659380466
    },
    {
      "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": 29033.98783878832,
      "required_pages_per_second_upper_estimate": 0.0017405432386028551
    },
    {
      "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": 29033.98783878832,
      "required_pages_per_second_upper_estimate": 0.0022572648474497824
    },
    {
      "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": 29033.98783878832,
      "required_pages_per_second_upper_estimate": 0.3724168217286633
    }
  ],
  "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": 0.350425369203208,
    "producer_p95_seconds": 0.3773666150009376,
    "gpu_update_seconds": 0.9522632630941351,
    "producer_batches_per_second": 2.8536746705119698,
    "gpu_batches_per_second": 1.0501297684746933,
    "producer_headroom": 2.7174495535508094,
    "sustainable": true,
    "buffer_seconds": 60.94484883802465,
    "buffer_drain_seconds": null,
    "shard_target_mib": 16,
    "upload_pages_per_commit": 75,
    "max_pending_pages": 4096,
    "generation_bytes_per_second": 494837.46299891575,
    "upload_bytes_per_second": 5127793.568167325,
    "upload_sustainable": true,
    "logical_page_rows": 4096,
    "logical_cortex_page_rows": 128,
    "estimated_full_shards_per_hour": 106.18060033298116,
    "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": 1.425732910996885,
      "bytes": 11302182,
      "pages": 41,
      "shards": 4
    },
    {
      "seconds": 1.1526390029976028,
      "bytes": 5901438,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.9387289110018173,
      "bytes": 8423867,
      "pages": 37,
      "shards": 3
    },
    {
      "seconds": 1.6871719589980785,
      "bytes": 10804637,
      "pages": 35,
      "shards": 3
    },
    {
      "seconds": 1.1380872770023416,
      "bytes": 998627,
      "pages": 27,
      "shards": 1
    },
    {
      "seconds": 1.2768399569977191,
      "bytes": 7586837,
      "pages": 40,
      "shards": 3
    },
    {
      "seconds": 1.1069597620007698,
      "bytes": 5924616,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.2267345689979265,
      "bytes": 1209498,
      "pages": 33,
      "shards": 1
    },
    {
      "seconds": 1.4170462420006515,
      "bytes": 11019556,
      "pages": 29,
      "shards": 4
    },
    {
      "seconds": 1.1776043950012536,
      "bytes": 12312597,
      "pages": 48,
      "shards": 3
    },
    {
      "seconds": 1.1231176639994374,
      "bytes": 2410876,
      "pages": 29,
      "shards": 3
    },
    {
      "seconds": 1.1969935200031614,
      "bytes": 10533402,
      "pages": 40,
      "shards": 3
    },
    {
      "seconds": 0.9813283670009696,
      "bytes": 889426,
      "pages": 27,
      "shards": 1
    },
    {
      "seconds": 1.3786637550001615,
      "bytes": 7505060,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 0.9612054499957594,
      "bytes": 587735,
      "pages": 19,
      "shards": 1
    },
    {
      "seconds": 1.4943926799969631,
      "bytes": 17356969,
      "pages": 43,
      "shards": 5
    },
    {
      "seconds": 1.3928040120008518,
      "bytes": 6228637,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.2593115519994171,
      "bytes": 6168643,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.411603455999284,
      "bytes": 2470060,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.2517306329973508,
      "bytes": 10612343,
      "pages": 40,
      "shards": 3
    },
    {
      "seconds": 1.098978062000242,
      "bytes": 637051,
      "pages": 18,
      "shards": 1
    },
    {
      "seconds": 1.4846390739985509,
      "bytes": 13365694,
      "pages": 41,
      "shards": 4
    },
    {
      "seconds": 1.3245377540006302,
      "bytes": 6089742,
      "pages": 38,
      "shards": 2
    },
    {
      "seconds": 1.1078838069952326,
      "bytes": 6470653,
      "pages": 34,
      "shards": 3
    },
    {
      "seconds": 1.8197720390016912,
      "bytes": 6211915,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.0536710650048917,
      "bytes": 923281,
      "pages": 30,
      "shards": 1
    },
    {
      "seconds": 1.7383342659959453,
      "bytes": 12982783,
      "pages": 32,
      "shards": 4
    },
    {
      "seconds": 1.1493779860029463,
      "bytes": 10545016,
      "pages": 43,
      "shards": 3
    },
    {
      "seconds": 1.2370521369957714,
      "bytes": 2360341,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 1.5772486200003186,
      "bytes": 6280980,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 0.9224069299962139,
      "bytes": 730166,
      "pages": 20,
      "shards": 1
    },
    {
      "seconds": 1.1284890140013886,
      "bytes": 4366438,
      "pages": 9,
      "shards": 2
    },
    {
      "seconds": 1.7931644429991138,
      "bytes": 8105432,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 1.5261870840040501,
      "bytes": 11600268,
      "pages": 32,
      "shards": 3
    },
    {
      "seconds": 1.0267256459992495,
      "bytes": 1290324,
      "pages": 36,
      "shards": 1
    },
    {
      "seconds": 1.2432711480068974,
      "bytes": 11617672,
      "pages": 33,
      "shards": 4
    },
    {
      "seconds": 1.6245845399971586,
      "bytes": 5983952,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 0.9982556870018016,
      "bytes": 2742943,
      "pages": 39,
      "shards": 2
    },
    {
      "seconds": 1.3246486520001781,
      "bytes": 6072941,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.6959029329955229,
      "bytes": 5081198,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 1.6100998589972733,
      "bytes": 13518430,
      "pages": 43,
      "shards": 4
    },
    {
      "seconds": 1.2164938489950146,
      "bytes": 6185487,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.2006639809987973,
      "bytes": 2710534,
      "pages": 39,
      "shards": 3
    },
    {
      "seconds": 1.3522469360032119,
      "bytes": 11156444,
      "pages": 32,
      "shards": 3
    },
    {
      "seconds": 1.1713541569988593,
      "bytes": 1313572,
      "pages": 37,
      "shards": 1
    },
    {
      "seconds": 2.0110623340005986,
      "bytes": 13186052,
      "pages": 37,
      "shards": 4
    },
    {
      "seconds": 1.3507091479987139,
      "bytes": 5962622,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.3216858209998463,
      "bytes": 5170754,
      "pages": 39,
      "shards": 2
    },
    {
      "seconds": 1.4559352279975428,
      "bytes": 7649400,
      "pages": 31,
      "shards": 3
    },
    {
      "seconds": 1.5415065289998893,
      "bytes": 6305077,
      "pages": 39,
      "shards": 2
    },
    {
      "seconds": 1.1717002670047805,
      "bytes": 2453796,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 1.5481451480009127,
      "bytes": 15819770,
      "pages": 39,
      "shards": 4
    },
    {
      "seconds": 1.0710374850023072,
      "bytes": 956267,
      "pages": 29,
      "shards": 1
    },
    {
      "seconds": 1.2827229300019098,
      "bytes": 7827828,
      "pages": 38,
      "shards": 3
    },
    {
      "seconds": 1.3130354629975045,
      "bytes": 10468764,
      "pages": 33,
      "shards": 3
    },
    {
      "seconds": 1.019322487001773,
      "bytes": 2703084,
      "pages": 37,
      "shards": 2
    },
    {
      "seconds": 1.0808033600042108,
      "bytes": 6031307,
      "pages": 29,
      "shards": 2
    },
    {
      "seconds": 1.2031739800004289,
      "bytes": 6971821,
      "pages": 36,
      "shards": 2
    },
    {
      "seconds": 1.2199941110011423,
      "bytes": 11488272,
      "pages": 36,
      "shards": 4
    },
    {
      "seconds": 1.293346582002414,
      "bytes": 5756789,
      "pages": 30,
      "shards": 2
    },
    {
      "seconds": 1.215349038997374,
      "bytes": 2821977,
      "pages": 39,
      "shards": 2
    },
    {
      "seconds": 1.116951727999549,
      "bytes": 5803988,
      "pages": 28,
      "shards": 2
    },
    {
      "seconds": 1.751644532996579,
      "bytes": 10468749,
      "pages": 36,
      "shards": 3
    },
    {
      "seconds": 1.4208347750027315,
      "bytes": 8470292,
      "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": 90
  },
  "cache_reused": true,
  "resume_probe_seconds": [
    3.415684295999995,
    0.013631696000004467
  ]
}