Hi! I am reading Parallel processing with Dask scheduler and it doesn’t seems to support automatic distribution of GPU resources. Is it true that I must specify the GPU ID to use when I stage a job instead of specifying how many GPUs?
Currently, that’s correct, you need to explicitly schedule the runs on the GPUs using either the
--gpus option or
CUDA_VISIBLE_DEVICES with the
This is an issue we’ll fix in the next release. Other users have asked about this as well.