We need fixed shapes in order to support torch.compile in dtensor
We need fixed shapes in order to support torch.compile in dtensor