Hey everyone, I noticed that the number of resources to use per trial can be set with the tune.with_resources() function. What I am looking is to find a way to dynamically change this according to the resources actually needed. So, suppose I start with train_func = tune.with_resources(train_func, {“cpu”: 1}). Is it possible for ray to dynamically increase the resources per trial if the process needs.
Thanks in advance.
I do not think that this is possible, as tune makes use of the placement group concept. The a priori fixed resource requirement per trial is put to a placement group, which is requested then at the head node and worker nodes.