{
  "cache_key": {
    "version": 1,
    "kind": "gpu",
    "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,
      "microbatch_candidates": [
        8,
        16,
        32,
        48,
        64,
        72,
        76,
        80,
        86,
        90
      ],
      "batch_selection": "max_batch",
      "gradient_checkpointing": false,
      "compile_model": true,
      "autotune_microbatch": true,
      "attention_backend": "flash",
      "vram_fraction": 0.95
    },
    "source": "e6d135a4487a9fd1907e12903cda79e75fd61305fad149fda7ec5281279dc18b",
    "requested_microbatch_ceiling": 90
  },
  "cache_reused": true,
  "attention": "PyTorch FlashAttention (forced; forward/backward verified)",
  "selected": {
    "microbatch": 86,
    "mode": "compiled",
    "microbatch_seconds": 0.8974938420000171,
    "peak_gb": 85.39913082122803,
    "fits_budget": true,
    "estimated_update_seconds": 0.9533413040000482,
    "input_tokens_per_second": 196244.2434228943
  },
  "measurements": [
    {
      "microbatch": 8,
      "mode": "eager",
      "microbatch_seconds": 0.15782159200000478,
      "peak_gb": 14.079077243804932,
      "fits_budget": true,
      "estimated_update_seconds": 1.793421506500053,
      "input_tokens_per_second": 103813.42497165725
    },
    {
      "microbatch": 16,
      "mode": "eager",
      "microbatch_seconds": 0.3450626039999918,
      "peak_gb": 27.397329807281494,
      "fits_budget": true,
      "estimated_update_seconds": 1.9234990809999601,
      "input_tokens_per_second": 94962.47817106481
    },
    {
      "microbatch": 32,
      "mode": "eager",
      "microbatch_seconds": 0.7818640030000097,
      "peak_gb": 54.14767265319824,
      "fits_budget": true,
      "estimated_update_seconds": 2.185234111500023,
      "input_tokens_per_second": 83820.20370363462
    },
    {
      "microbatch": 48,
      "mode": "eager",
      "microbatch_seconds": 1.1685763079999845,
      "peak_gb": 80.94732475280762,
      "fits_budget": true,
      "estimated_update_seconds": 2.190914332499986,
      "input_tokens_per_second": 84122.87612457851
    },
    {
      "microbatch": 8,
      "mode": "compiled",
      "microbatch_seconds": 0.08803204299999834,
      "peak_gb": 8.428553104400635,
      "fits_budget": true,
      "estimated_update_seconds": 1.0060347614999898,
      "input_tokens_per_second": 186114.0494035826
    },
    {
      "microbatch": 16,
      "mode": "compiled",
      "microbatch_seconds": 0.1704058820000114,
      "peak_gb": 16.191233158111572,
      "fits_budget": true,
      "estimated_update_seconds": 0.9576335650000658,
      "input_tokens_per_second": 192293.8317352086
    },
    {
      "microbatch": 32,
      "mode": "compiled",
      "microbatch_seconds": 0.34082671699997036,
      "peak_gb": 31.810625553131104,
      "fits_budget": true,
      "estimated_update_seconds": 0.9493586249999737,
      "input_tokens_per_second": 192285.39527904938
    },
    {
      "microbatch": 48,
      "mode": "compiled",
      "microbatch_seconds": 0.5061537779999981,
      "peak_gb": 47.43971490859985,
      "fits_budget": true,
      "estimated_update_seconds": 0.9468006980000041,
      "input_tokens_per_second": 194217.6553308279
    },
    {
      "microbatch": 64,
      "mode": "compiled",
      "microbatch_seconds": 0.6658512800000267,
      "peak_gb": 63.073320388793945,
      "fits_budget": true,
      "estimated_update_seconds": 0.9295456565000109,
      "input_tokens_per_second": 196848.76178355434
    },
    {
      "microbatch": 72,
      "mode": "compiled",
      "microbatch_seconds": 0.7430151979999664,
      "peak_gb": 70.97307014465332,
      "fits_budget": true,
      "estimated_update_seconds": 0.9250642054999503,
      "input_tokens_per_second": 198456.23669195347
    },
    {
      "microbatch": 76,
      "mode": "compiled",
      "microbatch_seconds": 0.78068104700003,
      "peak_gb": 74.84185981750488,
      "fits_budget": true,
      "estimated_update_seconds": 0.9241739715000392,
      "input_tokens_per_second": 199374.63654089966
    },
    {
      "microbatch": 80,
      "mode": "compiled",
      "microbatch_seconds": 0.8210239410000213,
      "peak_gb": 78.72572612762451,
      "fits_budget": true,
      "estimated_update_seconds": 0.9229792625000073,
      "input_tokens_per_second": 199555.69115370748
    },
    {
      "microbatch": 86,
      "mode": "compiled",
      "microbatch_seconds": 0.8818629400000191,
      "peak_gb": 84.78368663787842,
      "fits_budget": true,
      "estimated_update_seconds": 0.9349941874999956,
      "input_tokens_per_second": 199722.64624250587
    },
    {
      "microbatch": 90,
      "mode": "compiled",
      "microbatch_seconds": 0.9259720020000373,
      "peak_gb": 88.6924238204956,
      "fits_budget": false,
      "estimated_update_seconds": 0.9259720020000373,
      "input_tokens_per_second": 199055.69455867045
    }
  ],
  "failures": [
    {
      "microbatch": 64,
      "mode": "eager",
      "reason": "oom"
    },
    {
      "microbatch": 72,
      "mode": "eager",
      "reason": "oom"
    },
    {
      "microbatch": 76,
      "mode": "eager",
      "reason": "oom"
    },
    {
      "microbatch": 80,
      "mode": "eager",
      "reason": "oom"
    },
    {
      "microbatch": 86,
      "mode": "eager",
      "reason": "oom"
    },
    {
      "microbatch": 90,
      "mode": "eager",
      "reason": "oom"
    }
  ],
  "selection_policy": "max_batch",
  "budget_gb": 89.6964394569397,
  "reserve_gb": 2.1791439056396484,
  "note": "Forward/backward throughput; excludes optimizer, data, logging and checkpoint time.",
  "legacy_seed_verified": false
}