When you specify cuda device in SupervisedRunner (and SupervisedWandbRunner), it doesn't influence what GPUs will be used. Manually setting CUDA_VISIBLE_DEVICES works though.

Example

import torch
from catalyst.dl.runner import SupervisedRunner, SupervisedWandbRunner
...
runner = SupervisedRunner(device=torch.device('cuda:1'))
runner.train(...)

Expected behavior: Only GPU with ID 1 should be used.
Actual behavior: Catalyst uses all available devices.