deepspeed_config/zero2_offload_cpu.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,
        "offload_optimizer": {
            "device": "cpu",
            "pin_memory": true
        }
    },

    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": 5e-5,
            "betas": [0.9, 0.95],
            "weight_decay": 0.01
        }
    },
    "scheduler": {
        "type": "WarmupDecayLR",
        "params": {
            "warmup_min_lr": 1e-6,
            "warmup_max_lr": 5e-5,
            "total_num_steps": 10000
        }
    }
}