Is there multi-gpu versions for inference ?

[fgbxx]

HI, all
For the image volume is very large , for example, it is 800512512 pixels, maybe patches number will be 700 or more. so the prediction is very slow for single GPU.
I just want to know if have multi-GPUs, is there someone that had multi-gpus versions or share some implementation steps or examples?
very thankful for your feedback.

Best,
Gang

[FabianIsensee]

Hi Gang,
you can use multiple GPUs for inference by parallelizing the inference on a per-data level.
Example:

CUDA_VISIBLE_DEVICES=0 nnUNet_predict [...] --part_id 0 --num_parts 2
CUDA_VISIBLE_DEVICES=1 nnUNet_predict [...] --part_id 1 --num_parts 2

(run these two at the same time)
This will run half of the images on GPU0 and the other half on GPU1.

[Peaceandmaths]

Can i do the same for training ? Do I understand correctly, I can define an arbitrary number of parts depending on the volume of my data ?