That seems like a good ideas, since, bizarrely, I can't see how to set the memory required as an input environment variable. SBATCH_MEM_PER_GPU exists, but SLURM_MEM_PER_CPU and SLURM_MEM_PER_NODE are only available as input variables. Strange.
Åke Sandgren <[email protected]> writes: > If you're building a single easyconfig you could also just make a small > submit file that runs eb the same way you would do manually, and send it > to the right node. > > On 12/4/19 10:23 AM, Loris Bennett wrote: >> Hi, >> >> With Kenneth's help I have realised/remembered that I need to compile >> Keras / TensorFlow on a machine with the appropriate CUDA drivers. This >> seemed like a good scenario in which to use the >> >> --job >> >> option. However, if I write >> >> ... --job --job-backend=Slurm --job-cores=4 >> >> I would still need to specify the partition with the GPU nodes, so that >> the job is indeed scheduled to a machine on which the drivers are >> installed. >> >> How would I do that? >> >> Cheers, >> >> Loris >> -- Dr. Loris Bennett (Mr.) ZEDAT, Freie Universität Berlin Email [email protected]
