config/deepspeed/zero2_config.json

{
    "bf16": {
      "enabled": "auto"
    },
    "zero_optimization": {
      "stage": 2,
      "offload_optimizer": {
          "device": "none"
      },
      "contiguous_gradients": true,
      "overlap_comm": true
    },
    "train_micro_batch_size_per_gpu": 1,
    "gradient_accumulation_steps": "auto"
}