Closed cpandar closed 6 years ago
Thanks Chethan, this should now be fixed via 870d574. I've also added the flag oneTaskPerGPU
, which defaults to true. This defaults to run one model on each GPU, rather than squeezing as many onto each GPU as the available memory will support.
in RunCollection.m:
gpuListStr is calculated in lines 305-310, and is properly output to run_lfadsqueue.py so that the variable is defined in the script
However, this variable is never actually passed to the command lq.run_lfads_queue on line 345
The following replacement for line 345 fixes the issue:
fprintf(fid, 'tasks = lq.run_lfads_queue(queue_name, tensorboard_script, task_specs, gpu_list=gpu_list%s%s)\n\n', maxStr, donefileStr);