I am not entirely sure if this is the full solution to the issue, but I assumed that it would be only applicable to a Ray cluster. Therefore, the new behavior is to query Ray for the number of available GPUs and assign a fraction of them to each trial in the exact same way CPUs are assigned.
Solves https://github.com/ray-project/tune-sklearn/issues/143.
I am not entirely sure if this is the full solution to the issue, but I assumed that it would be only applicable to a Ray cluster. Therefore, the new behavior is to query Ray for the number of available GPUs and assign a fraction of them to each trial in the exact same way CPUs are assigned.