ds_config.json

{
  "train_batch_size" : 16,
  "train_micro_batch_size_per_gpu": 16,
  "steps_per_print": 1,

  "zero_optimization": {
    "stage": 1
  },

  "bf16": {
    "enabled": true
  },

  "data_types": {
        "grad_accum_dtype": "fp32" 
  },

  "wall_clock_breakdown" : false
}