metadata/performance.json

{
  "hardware": "8 × NVIDIA A100-SXM4-80GB",
  "batch_size_per_gpu": 1,
  "steady_samples": 4726,
  "mean_seconds": 39.404512504240735,
  "median_seconds": 39.90122409281321,
  "neutral_mean_seconds": 31.685628565184853,
  "reference_edit_mean_seconds": 40.2569268490143,
  "timings_include": "50 denoising steps with real CFG4, reference VAE encode when applicable, VAE decode; CUDA synchronized",
  "timings_exclude": "Positive prompt preencoding, model loading, compile warmup, PNG/TAR writes and upload",
  "pilot_excluded": "Pilot GPU7 was shared with background production and used CPU offload; it is not a dedicated-GPU benchmark"
}