throughput_base.json

{
  "label": "base",
  "checkpoint": null,
  "device": "cuda",
  "dtype": "torch.bfloat16",
  "seq_len": 512,
  "repeats": 6,
  "mean_forward_s": 0.09076352066673128,
  "min_forward_s": 0.0884767360000751,
  "tokens_per_s_mean": 5641.032831681131,
  "tokens_per_s_best": 5786.83191929193,
  "apply_elapsed_s": 0.0,
  "apply_replaced": 0,
  "apply_missing_count": 0,
  "apply_shape_mismatches_count": 0,
  "checkpoint_method": null,
  "checkpoint_avg_bpw": null,
  "checkpoint_weighted_rmse": null
}