When using TFTrainer there must be provided a fixed numer of model replicas.
Is there any chance of using ray autoscaler sometime in the future to automatically use as many resources as are currently available or is it not viable to implement?
When using TFTrainer there must be provided a fixed numer of model replicas.
Is there any chance of using ray autoscaler sometime in the future to automatically use as many resources as are currently available or is it not viable to implement?