How to make the code run on multiple GPUs

[NagabhushanSN95]

By default, the train code runs on only 1 GPU. How can make it run on multiple GPUs?

I tried wrapping the frame_predictor, posterior and prior modules with DataParallel. Then I got the error, No such attribute init_hidden for DataParallel. So, I changed frame_predictor.init_hidden() to frame_predictor.module.init_hidden(). Then I got the error something like, trying to do backward a second time. Graph not saved.

Is it possible to make the code run on multiple GPUs?

[Pixie8888]

Hi, I am also trying to push the model to multi-gpus. I change all LSTM.hidden = lstm.init_hidden into lstm.module.hidden = lstm.module.init_hidden. I didn't encounter your error, but I encountered another error regarding hidden states: hidden states ont on the same device as input. How do you push hidden_states onto the same device as input?

[NagabhushanSN95]

Hi @Pixie412 good that you didn't encounter the error I faced. I'm not sure how to solve yours. I couldn't solve mine either. But the error you're getting seems generic. You can ask that question in Stack Overflow. I think you've a better chance of getting an answer there. Looks like the authors are not checking the raised issues :/

If you ask a question or get an answer there, please share the link here too, so that it may benefit future users.