About using partial GPU


I want to use “partial GPUs” (not fractional GPU)
For example, I have 8 GPU in one machine.
I want to assign one job (including many actors and sub-jobs) to GPU 0-3 and assign another job (with hyperparameter changed) to GPU 4-8, similar to other deep learning frameworks using “CUDA_VISIBLE DEVICES.”
How can I set ray to achieve this?


Hmm, interesting!

Will this work if you just do ray.remote(num_gpus=4) on each function?