training/deepspeed_zero3.json

{
  "bf16": { "enabled": true },
  "zero_optimization": {
    "stage": 3,
    "offload_optimizer": { "device": "cpu" },
    "offload_param": { "device": "cpu" }
  },
  "gradient_accumulation_steps": "auto",
  "train_micro_batch_size_per_gpu": "auto"
}