when and how the trainer or module move the data to gpu?

[FutureWithoutEnding]

Automatically move data to GPU is fine.
But, not knowing what moves and what doesn't, When the batch in training_step(self, batch, batch_idx) is a dict, no movement will occur.

Also don't know when to move and how to move to GPU?
I didn't find where to implement in the package!!

[akihironitta]

Hi @FutureWithoutEnding

When the batch is transferred to device
I think the pseudocode in the docs explains when a batch is transferred to each device very well:
https://pytorch-lightning.readthedocs.io/en/1.6.5/common/lightning_module.html#hooks

What is transferred to device
See the list of supported data structures in the documentation. If you have custom data structure, you need to override this hook in your LightningModule:
https://pytorch-lightning.readthedocs.io/en/1.6.5/common/lightning_module.html#transfer-batch-to-device

Also, if you need to manually transfer tensors to device, you can utilise self.device so that your code stays hardware-agnostic. your_tensor.to(self.device):
https://pytorch-lightning.readthedocs.io/en/1.6.5/accelerators/gpu.html#init-tensors-using-type-as-and-register-buffer

[FutureWithoutEnding]

thank you very much!
I need to read the documentation properly.