{
  "avg_time_ms": 62.308485079556704,
  "throughput_tokens_per_sec": 65737.43519474348,
  "memory_allocated_gb": 1.329831600189209,
  "memory_cached_gb": 1.8359375,
  "memory_increase_gb": 0.3795137405395508,
  "device": "cuda",
  "dtype": "torch.bfloat16",
  "tokens": 4096,
  "warmup_iters": 10,
  "timing_iters": 50
}