WAN2.1-14B LoRA Training problem

I attempted to fine-tune WAN2.1 14B using LoRA on 8 L40s GPUs (48GB each).

When using ZeRO-2, I encountered OOM (Out of Memory) errors regardless of parameter settings.
When switching to ZeRO-3, the following error occurred:

[rank5]: torch.utils.checkpoint.CheckpointError: torch.utils.checkpoint: Recomputed values for the following tensors have different metadata than during the forward pass.  
[rank5]: tensor at position 14:                                                                                                                                             
[rank5]: saved metadata: {'shape': torch.Size([5120]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}                                                     
[rank5]: recomputed metadata: {'shape': torch.Size([0]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}                                                   
[rank5]: tensor at position 24:                                                                                                                                             
[rank5]: saved metadata: {'shape': torch.Size([5120]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}                                                     
[rank5]: recomputed metadata: {'shape': torch.Size([0]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}                                                   
[rank5]: tensor at position 60:                                                                                                                                             
[rank5]: saved metadata: {'shape': torch.Size([5120]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}                                                     
[rank5]: recomputed metadata: {'shape': torch.Size([0]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}
[rank5]: tensor at position 70:
[rank5]: saved metadata: {'shape': torch.Size([5120]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}
[rank5]: recomputed metadata: {'shape': torch.Size([0]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}
[rank5]: tensor at position 85:
[rank5]: saved metadata: {'shape': torch.Size([5120]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}
[rank5]: recomputed metadata: {'shape': torch.Size([0]), 'dtype': torch.bfloat16, 'device': device(type='cuda', index=5)}

This occurred while using the default LoRA configuration.

Start by reproducing WAN2.1 14B LoRA training with the default LoRA configuration on 8 L40 GPUs, comparing ZeRO-2 and ZeRO-3. Investigate the reported checkpoint metadata mismatch and OOM behavior; done means the training run completes without either failure.