DeepSpeed未按预期向CPU卸载运算，显存不足报错求助

调大batch size训练CLIP模型时触发CUDA OOM错误，DeepSpeed ZeRO Stage2配置的CPU卸载未按预期缓解显存压力，调小batch size则无报错。

报错信息 #

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 7.04 GiB (GPU 1; 79.15 GiB total capacity; 68.07 GiB already allocated; 5.90 GiB free; 72.14 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

环境与配置 #

• 模型：HuggingFace CLIP
• 框架：PyTorch 1.13
• 训练方式：HuggingFace Trainer集成DeepSpeed
• 服务器：Azure虚拟机（AMD EPYC 7V13 64核）
• 优化器：AdamW
• DeepSpeed配置（Stage2）：

{
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },

    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "betas": "auto",
            "eps": "auto",
            "weight_decay": "auto"
        }
    },

    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto"
        }
    },

    "zero_optimization": {
        "stage": 2,
        "offload_optimizer": {
            "device": "cpu",
            "pin_memory": true
        },
        "allgather_partitions": true,
        "allgather_bucket_size": 2e8,
        "overlap_comm": true,
        "reduce_scatter": true,
        "reduce_bucket_size": 2e8,
        "contiguous_gradients": true
    },

    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto"
}

• Trainer初始化代码：

with open("./Multi_Modal_Model/zero_config/stage_2_config.json") as f:
    z_optimiser = json.load(f)
        
training_args = TrainingArguments(
    ...
    deepspeed=z_optimiser,
    ...
)

1. ZeRO Stage2的卸载范围限制 #

ZeRO Stage2仅会将优化器状态（如AdamW的动量、方差）卸载到CPU，模型的参数、激活值、梯度仍然保留在GPU上。调大batch size时，模型前向/反向传播产生的激活值显存占用会大幅增加，超过GPU剩余空间就会触发OOM——这是Stage2的设计特性，并非配置错误。

如果需要进一步卸载模型参数/梯度以释放GPU显存，需升级到ZeRO Stage3，修改配置如下：

"zero_optimization": {
    "stage": 3,
    "offload_optimizer": {
        "device": "cpu",
        "pin_memory": true
    },
    "offload_param": {
        "device": "cpu",
        "pin_memory": true
    },
    "allgather_partitions": true,
    "allgather_bucket_size": 2e8,
    "overlap_comm": true,
    "reduce_scatter": true,
    "reduce_bucket_size": 2e8,
    "contiguous_gradients": true
}

2. 修正"auto"参数的模糊配置 #

DeepSpeed配置中多个"auto"设置可能导致行为不可控：

• 明确设置train_micro_batch_size_per_gpu为固定值（与TrainingArguments中的per_device_train_batch_size保持一致），避免自动调整的micro batch导致显存峰值过高。
• 将fp16.enabled手动设为true，确保混合精度生效（混合精度可将显存占用降低约50%），替代依赖"auto"的自动判断。

3. 缓解显存碎片问题 #

报错提示reserved内存远大于allocated内存，说明存在显存碎片。设置环境变量缓解：

export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512

该设置会限制PyTorch分配的显存块大小，减少碎片产生。

4. 检查HuggingFace Trainer的DeepSpeed集成一致性 #

• 确保TrainingArguments中的per_device_train_batch_size与DeepSpeed配置的train_micro_batch_size_per_gpu数值一致，避免参数冲突。
• 查看训练日志，确认DeepSpeed已正确初始化（日志中会出现"DeepSpeed config info"等相关输出），若未初始化需检查Trainer的启动方式（需用deepspeed命令启动训练脚本，而非直接python）。

内容的提问来源于stack exchange，提问作者paragon00