{
  "_comment_purpose": "ZeRO-3 + CPU offload for 32B QLoRA on Kaggle T4×2 (16GB×2). Each T4 holds ~8GB sharded weights + LoRA grads. Optimizer states are paged to CPU RAM (Kaggle has 30GB).",

  "bf16": {
    "enabled": true
  },
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": "auto",
      "betas": "auto",
      "eps": "auto",
      "weight_decay": "auto"
    }
  },
  "scheduler": {
    "type": "WarmupDecayLR",
    "params": {
      "warmup_min_lr": "auto",
      "warmup_max_lr": "auto",
      "warmup_num_steps": "auto",
      "total_num_steps": "auto"
    }
  },
  "zero_optimization": {
    "stage": 3,
    "overlap_comm": true,
    "contiguous_gradients": true,
    "sub_group_size": 1000000000,
    "reduce_bucket_size": "auto",
    "stage3_prefetch_bucket_size": "auto",
    "stage3_param_persistence_threshold": "auto",
    "stage3_max_live_parameters": 1000000000,
    "stage3_max_reuse_distance": 1000000000,
    "stage3_gather_16bit_weights_on_model_save": true,
    "offload_param": {
      "device": "cpu",
      "pin_memory": true
    },
    "offload_optimizer": {
      "device": "cpu",
      "pin_memory": true
    }
  },
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "wall_clock_breakdown": false
}