train/deepspeed_config/train_dp_config.json

{ 
    "train_micro_batch_size_per_gpu": "auto",
    
    "zero_optimization": {
      "stage": 2,
      "offload_optimizer": {
        "device": "cpu",
        "pin_memory": true
      },
      "overlap_comm": true,
      "reduce_bucket_size": 1e8
    },
    "fp16": {
      "enabled": true 
    },
    "gradient_accumulation_steps": "auto",  
    "wall_clock_breakdown": false  
  }