Hey @Jian_1_2,
This issue is coming up in a few other places on the forum (e.g., here: Quantum Transfer Learning). You’re not alone!
I spoke with someone internally and it looks like this issue might be due to the fact that putting the torch tensor on a gpu is problematic when lightning-gpu is being used as well. Lightning-gpu and torch’s gpu pipeline are entirely differerent, and lightning-gpu expects the data to be on the host right now, so that should fix it!