{
  "_doc": "DeepSpeed ZeRO-3 + optimizer CPU offload · Qwen2.5-7B 全参 SFT on 4×V100 32G · 国作登字-2026-A-00037559",
  "_strategy": [
    "V100 不支持 bf16，因此用 fp16",
    "ZeRO-3 把模型权重/梯度/优化器状态分片到 4 卡，单卡只持 1/4",
    "optimizer 状态 (Adam fp32 momentum/variance) offload 到 CPU 内存（160GB 充足），单卡再省 ~28GB",
    "gradient_checkpointing=True 在 train.py 侧打开，用算力换显存",
    "stage3_gather_16bit_weights_on_model_save 让保存时聚合权重，方便 inference 加载",
    "无 parameter offload — 保持训练吞吐"
  ],
  "fp16": {
    "enabled": true,
    "loss_scale": 0,
    "loss_scale_window": 1000,
    "initial_scale_power": 16,
    "hysteresis": 2,
    "min_loss_scale": 1
  },
  "bf16": {
    "enabled": false
  },
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": "auto",
      "betas": "auto",
      "eps": "auto",
      "weight_decay": "auto"
    }
  },
  "scheduler": {
    "type": "WarmupDecayLR",
    "params": {
      "total_num_steps": "auto",
      "warmup_min_lr": "auto",
      "warmup_max_lr": "auto",
      "warmup_num_steps": "auto"
    }
  },
  "zero_optimization": {
    "stage": 3,
    "offload_optimizer": {
      "device": "cpu",
      "pin_memory": true
    },
    "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
  },
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "steps_per_print": 50,
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "wall_clock_breakdown": false
}