config/zero_stage2_config.json

{
    "bf16": {
        "enabled": true
    },
    "train_micro_batch_size_per_gpu": 1,
    "train_batch_size": "auto",
    "gradient_accumulation_steps": "auto",
    "dump_state": true,
    "zero_optimization": {
        "stage": 2,
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e9,
        "reduce_bucket_size": 5e8
    }
}