Switch between single and multigpu training


Is there a easy way to switch between using single gpu or multiple gpu in the same training config file ? Currently you need to set a extra distributed key but it needs to have atleast two gpus. Do I need to use environment variable to somehow add/subtract this from the config file ? Is there a recommended way ?

We recently made it possible to auto-detect CPU vs. GPU and act accordingly. It’d be interested to try to get this working for multi-GPU cases, but that seems like it’d be a lot more work. Currently, you’re stuck with changing the configuration file, or using the --overrides command line parameter.