[BUG]AttributeError in ZeRO-3 with gradient_checkpointing: 'NoneType' object has no attribute 'next_functions' (DeepSpeed 0.18.5)

[Hylbcxs]

Describe the bug
When using DeepSpeed ZeRO-3 with gradient_checkpointing=True, training fails with:
AttributeError: 'NoneType' object has no attribute 'next_functions'
This occurs even in DeepSpeed 0.18.5.

Environment

• Deepspeed version: 0.18.5
• PyTorch version: 2.3.0+cu121
• Transformers version: 4.57.5
• CUDA version: 12.8 (nvcc)
• Python version: 3.10
• OS: Ubuntu 22.04

DeepSpeed Config

{
    "fp16": {
        "enabled": "auto"
    },
    "bf16": {
        "enabled": "auto"
    },
    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "weight_decay": "auto"
        }
    },
    "zero_optimization": {
        "stage": 3,
        "offload_optimizer": {
            "device": "cpu",
            "pin_memory": true
        },
        "offload_param": {
            "device": "none",
            "pin_memory": true
        },
        "overlap_comm": true,
        "contiguous_gradients": true,
        "reduce_bucket_size": "auto",
        "stage3_prefetch_bucket_size": "auto",
        "stage3_param_persistence_threshold": "auto",
        "stage3_gather_16bit_weights_on_model_save": true
    },
    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "steps_per_print": 2000,
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
}