LLNL
flux: `gpus` key semantics should be total number of gpus
In the 0.26.0 adaptor, parts of it assume that gpus means total number of gpus and other parts assume it means gpus_per_task. The adaptor should be modified to assume it means total number of gpus for consistency with the slurm adaptor
Just a note for this issue as it was related.
Talking to @Jmast, he mentioned that he prefers to have the keys be in terms of cores per task and gpus per task, which is contra what you and I had discussed related to running math in the background where a user can't see it. I think favoring a gpus per task key would put this more in line with cpus per task and make things consistent -- meaning we would need to perform that math in the SLURM case.
