hybrid_engine_config.json

{
  "train_batch_size" : 32,
  "train_micro_batch_size_per_gpu": 2,
  "steps_per_print": 10,
  "zero_optimization": {
    "stage": 0,
    "offload_param": {
        "device": "cpu"
    },
    "stage3_param_persistence_threshold": 0
  },
  "fp16":{
    "enabled": true,
    "loss_scale_window": 100
  },
  "gradient_clipping": 1.0,
  "prescale_gradients": false,
  "wall_clock_breakdown" : false
}