{
  "avg_time_ms": 26.93254135781899,
  "throughput_tokens_per_sec": 152083.67994618745,
  "memory_allocated_gb": 2.2425241470336914,
  "memory_cached_gb": 4.14453125,
  "memory_increase_gb": 1.2922062873840332,
  "device": "cuda",
  "dtype": "torch.bfloat16",
  "tokens": 4096,
  "warmup_iters": 10,
  "timing_iters": 50
}