{
  "train_micro_batch_size_per_gpu": 32,
  "gradient_accumulation_steps": 1,
  "gradient_clipping": 1.0,
  "zero_optimization": {
    "stage": 1,
    "allgather_partitions": true,
    "allgather_bucket_size": 5e8,
    "overlap_comm": true,
    "reduce_scatter": true,
    "reduce_bucket_size": 5e8,
    "contiguous_gradients": true
  },
  "wall_clock_breakdown": false
}