adam_zero2_fp16.json

{
  "train_batch_size": 16,
  "gradient_accumulation_steps": 1,
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": 0.00004
    }
  },
  "fp16": {
    "enabled": true
  },
  "zero_optimization": {
    "stage": 2,
    "allgather_partitions": true,
    "allgather_bucket_size": 1e9,
    "overlap_comm": true,
    "reduce_scatter": true,
    "reduce_bucket_size": 1e9,
    "contiguous_gradients": true
  },
  "log": {
    "steps_per_print": 50
  }
}