ds_config.json

{
  "train_batch_size": 8,
  "train_micro_batch_size_per_gpu": 1,
  "gradient_accumulation_steps": 4,
  "gradient_clipping": 1.0,
  "bf16": {
    "enabled": true
  },
  "zero_optimization": {
    "stage": 2,
    "offload_optimizer": {
      "device": "none"
    },
    "contiguous_gradients": true
  },
  "steps_per_print": 100,
  "wall_clock_breakdown": false
}