{
  "model_name": "Qwen/Qwen3-4B",
  "gpu_name": "NVIDIA A100-SXM4-80GB",
  "attention_implementation": "flash_attention_2",
  "batch_size": 3,
  "sequence_length": 2048,
  "num_steps": 10,
  "forward": {
    "stage_name": "forward",
    "duration_ms": 3359.0412912890315,
    "tokens_processed": 61440,
    "tokens_per_second": 18290.933237210196,
    "energy_joules": 1292.2280000448227,
    "energy_per_token": 0.021032356771562868,
    "avg_power_watts": 387.19580415542595,
    "peak_memory_gb": 79.66021728515625,
    "avg_gpu_util_percent": 97.8
  },
  "backward": {
    "stage_name": "backward",
    "duration_ms": 6954.944152384996,
    "tokens_processed": 61440,
    "tokens_per_second": 8834.003358449821,
    "energy_joules": 2729.588000059128,
    "energy_per_token": 0.0444268880217957,
    "avg_power_watts": 394.24766095856324,
    "peak_memory_gb": 79.66021728515625,
    "avg_gpu_util_percent": 97.8
  },
  "optimizer": {
    "stage_name": "optimizer",
    "duration_ms": 1153.845101594925,
    "tokens_processed": 61440,
    "tokens_per_second": 53248.048559614595,
    "energy_joules": 362.6529998779297,
    "energy_per_token": 0.005902555336554845,
    "avg_power_watts": 299.1223537953503,
    "peak_memory_gb": 79.66021728515625,
    "avg_gpu_util_percent": 97.8
  },
  "total_duration_ms": 11467.830545268953,
  "total_tokens": 61440,
  "total_tokens_per_second": 5357.595733340081,
  "total_energy_joules": 4384.46899998188,
  "total_energy_per_token": 0.07136180012991342,
  "timestamp": 1768519431.5985208
}