what about having 2 agents, one on each GPU, on the same machine, serving the same queue ? So that when you enqueue, which ever agent (thus GPU) available will take the new task

Ah yes, genius, thanks @<1576381444509405184:profile|ManiacalLizard2>

--gpus 0,1 : I believe this basically say that your code launched by the agent has access to both GPUs and that is it. Now it is up to your code to choose which GPU to use and what not and how ...