{
  "train_batch_size" : 60,
  "train_micro_batch_size_per_gpu": 1,
  "steps_per_print": 1,
  "zero_optimization": {
    "stage": 1
  },
  "bf16": {
    "enabled": true
  },
    "data_types": { 
    "grad_accum_dtype": "fp32" 
  },
  "overlap_comm":true,
  "wall_clock_breakdown": true,
  "zero_allow_untested_optimizer": true
}