{
  "date": "2026-09-08",
  "gpu": "NVIDIA RTX 5000 Ada Generation",
  "torch": "2.14.0+cu130",
  "synthetic": true,
  "effective_batch_tokens": 32768,
  "context_length": 1024,
  "selected": [
    {
      "preset": "20m",
      "micro_batch_size": 16,
      "compile": true,
      "tokens_per_second": 352067.7821058689,
      "peak_memory_bytes": 3110031872,
      "attention_kernels": [
        "aten::_flash_attention_backward",
        "aten::_flash_attention_forward",
        "aten::_scaled_dot_product_flash_attention",
        "aten::_scaled_dot_product_flash_attention_backward",
        "aten::scaled_dot_product_attention"
      ]
    },
    {
      "preset": "50m",
      "micro_batch_size": 16,
      "compile": true,
      "tokens_per_second": 204059.3034399494,
      "peak_memory_bytes": 5258774016,
      "attention_kernels": [
        "aten::_flash_attention_backward",
        "aten::_flash_attention_forward",
        "aten::_scaled_dot_product_flash_attention",
        "aten::_scaled_dot_product_flash_attention_backward",
        "aten::scaled_dot_product_attention"
      ]
    },
    {
      "preset": "90m",
      "micro_batch_size": 8,
      "compile": true,
      "tokens_per_second": 114909.36361201691,
      "peak_memory_bytes": 5157307904,
      "attention_kernels": [
        "aten::_flash_attention_backward",
        "aten::_flash_attention_forward",
        "aten::_scaled_dot_product_flash_attention",
        "aten::_scaled_dot_product_flash_attention_backward",
        "aten::scaled_dot_product_attention"
      ]
    }
  ]
}