adam_zero1_amp.json

{
  "train_batch_size": 16,
  "gradient_accumulation_steps": 1,
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": 0.00004
    }
  },
  "amp": {
    "enabled": true,
    "opt_level": "O1"
  },
  "log": {
    "steps_per_print": 50
  },
  "wall_clock_breakdown": false
}