config/deepspeed_zero2.json

{
  "train_batch_size": 8,
  "gradient_accumulation_steps": 1,
  "steps_per_print": 100,
  "optimizer": {
    "type": "Adam",
    "params": {
      "lr": 0.0005
    }
  },
  "zero_optimization": {
    "stage": 2,
    "offload_optimizer": {
      "device": "cpu",
      "pin_memory": true
    },
    "overlap_comm": true,
    "contiguous_gradients": true
  },
  "fp16": {
    "enabled": true
  }
}