The CUDAService already checks for the availability of a suitable GPU at job startup, so there is no need to explicitly disable it in the python configuration.
This fixes the case where the configuration is expanded or pickled on a machine without GPUs, and then run on a machine with a GPU: in this case the CUDAService gets incorrectly disabled, breaking the configuration itself.
The
CUDAService
already checks for the availability of a suitable GPU at job startup, so there is no need to explicitly disable it in the python configuration.This fixes the case where the configuration is expanded or pickled on a machine without GPUs, and then run on a machine with a GPU: in this case the
CUDAService
gets incorrectly disabled, breaking the configuration itself.