configs/zero2.json

{
    "bf16": {
        "enabled": "auto"
    },
    "train_micro_batch_size_per_gpu": "auto",
    "train_batch_size": "auto",
    "gradient_accumulation_steps": "auto",
    "zero_optimization": {
        "stage": 2,
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e9
    }
}