zero3_bf16.json

{
  "zero_optimization": {
    "stage": 3,
    "offload_param": {
      "device": "cpu",
      "pin_memory": true
    },
    "offload_optimizer": {
      "device": "cpu"
    },
    "overlap_comm": true,
    "contiguous_gradients": true,
    "override_module_apply": false
  },
  "bf16": {
    "enabled": true,
    "gradient_checkpointing": true,
    "train_micro_batch_size_per_gpu": 1,
    "steps_per_print": 50
  }
}