{
  "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": 0.98
    },
    "source": "e6d135a4487a9fd1907e12903cda79e75fd61305fad149fda7ec5281279dc18b",
    "requested_microbatch_ceiling": 90
  },
  "cache_reused": true,
  "attention": "PyTorch FlashAttention (forced; forward/backward verified)",
  "selected": {
    "microbatch": 90,
    "mode": "compiled",
    "microbatch_seconds": 0.9387972919998901,
    "peak_gb": 89.37766551971436,
    "fits_budget": true,
    "estimated_update_seconds": 0.9387972919998901,
    "input_tokens_per_second": 196336.31410178967
  },
  "measurements": [
    {
      "microbatch": 8,
      "mode": "eager",
      "microbatch_seconds": 0.15844926500000156,
      "peak_gb": 14.764304637908936,
      "fits_budget": true,
      "estimated_update_seconds": 1.8007989265000077,
      "input_tokens_per_second": 103402.18365796673
    },
    {
      "microbatch": 16,
      "mode": "eager",
      "microbatch_seconds": 0.34631374599999276,
      "peak_gb": 28.084960460662842,
      "fits_budget": true,
      "estimated_update_seconds": 1.929804849999968,
      "input_tokens_per_second": 94619.40329680326
    },
    {
      "microbatch": 32,
      "mode": "eager",
      "microbatch_seconds": 0.7831460159999892,
      "peak_gb": 54.827155113220215,
      "fits_budget": true,
      "estimated_update_seconds": 2.188403235499976,
      "input_tokens_per_second": 83682.98971210101
    },
    {
      "microbatch": 48,
      "mode": "eager",
      "microbatch_seconds": 1.1706683780000162,
      "peak_gb": 81.63183498382568,
      "fits_budget": true,
      "estimated_update_seconds": 2.1949380675000043,
      "input_tokens_per_second": 83972.54239321277
    },
    {
      "microbatch": 8,
      "mode": "compiled",
      "microbatch_seconds": 0.08827481200000875,
      "peak_gb": 9.113307476043701,
      "fits_budget": true,
      "estimated_update_seconds": 1.008851873500106,
      "input_tokens_per_second": 185602.20779624404
    },
    {
      "microbatch": 16,
      "mode": "compiled",
      "microbatch_seconds": 0.17047579999996287,
      "peak_gb": 16.87287473678589,
      "fits_budget": true,
      "estimated_update_seconds": 0.9571819019997747,
      "input_tokens_per_second": 192214.96540862185
    },
    {
      "microbatch": 32,
      "mode": "compiled",
      "microbatch_seconds": 0.34263174000000163,
      "peak_gb": 32.49641752243042,
      "fits_budget": true,
      "estimated_update_seconds": 0.9525396829999977,
      "input_tokens_per_second": 191272.41393339593
    },
    {
      "microbatch": 48,
      "mode": "compiled",
      "microbatch_seconds": 0.5075005140000144,
      "peak_gb": 48.123157024383545,
      "fits_budget": true,
      "estimated_update_seconds": 0.9488392155000156,
      "input_tokens_per_second": 193702.2668709991
    },
    {
      "microbatch": 64,
      "mode": "compiled",
      "microbatch_seconds": 0.6643874849999634,
      "peak_gb": 63.757601737976074,
      "fits_budget": true,
      "estimated_update_seconds": 0.926719353999971,
      "input_tokens_per_second": 197282.46386219515
    },
    {
      "microbatch": 72,
      "mode": "compiled",
      "microbatch_seconds": 0.7446585950000326,
      "peak_gb": 71.65071392059326,
      "fits_budget": true,
      "estimated_update_seconds": 0.9283865335000314,
      "input_tokens_per_second": 198018.26097232322
    },
    {
      "microbatch": 76,
      "mode": "compiled",
      "microbatch_seconds": 0.7858128550000174,
      "peak_gb": 75.5242338180542,
      "fits_budget": true,
      "estimated_update_seconds": 0.9300540555000225,
      "input_tokens_per_second": 198072.60597689834
    },
    {
      "microbatch": 80,
      "mode": "compiled",
      "microbatch_seconds": 0.8269688110000288,
      "peak_gb": 79.40062236785889,
      "fits_budget": true,
      "estimated_update_seconds": 0.9293605420000404,
      "input_tokens_per_second": 198121.13567121493
    },
    {
      "microbatch": 86,
      "mode": "compiled",
      "microbatch_seconds": 0.8882591199999865,
      "peak_gb": 85.4687147140503,
      "fits_budget": true,
      "estimated_update_seconds": 0.9413692904999778,
      "input_tokens_per_second": 198284.4825730612
    },
    {
      "microbatch": 90,
      "mode": "compiled",
      "microbatch_seconds": 0.9302335780000135,
      "peak_gb": 89.3772840499878,
      "fits_budget": true,
      "estimated_update_seconds": 0.9302335780000135,
      "input_tokens_per_second": 198143.7827650609
    }
  ],
  "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": 92.52170804977416,
  "reserve_gb": 2.1791439056396484,
  "note": "Forward/backward throughput; excludes optimizer, data, logging and checkpoint time.",
  "legacy_seed_verified": false
}