deepspeed_config.json

{
  "bf16": {
    "enabled": true
  },
  "gradient_clipping": 0.0,
  "steps_per_print": 100,
  "train_batch_size": 4096,
  "train_micro_batch_size_per_gpu": 128,
  "wall_clock_breakdown": false,
  "zero_optimization": {
    "gather_16bit_weights_on_model_save": true,
    "stage": 0
  }
}