GPU use for embedding nets

[sgb67]

When running SNPE with an embedding net, the SBI documentation suggests that GPUs might offer a speed advantage. However, we are having trouble setting the device to "gpu" or "cuda". We are able to set the embedding net and the prior device to gpu, but an error message appears, which seems to flag that the training device is still "cpu". We would be very grateful if anybody could help us solve this issue so that we can make SBI use GPUs. Thanks very much in advance.

[michaeldeistler]

Are you using the following?

inference = SNPE(prior=prior, device="gpu")

If yes, then please share your code and I can have a look.

Michael