RL training asks for one GPU
RL training runs on one GPU. A run that asked for more than one GPU on one machine used to start, and its graded
results could overwrite each other. Nodus now refuses it before any spend, with the message “RL training runs on one
GPU for now; set resources.gpu.count to 1”. Runs that use the presets, rl.train(...) included, already ask for one
GPU and are unaffected; rl.train(..., gpu="RTX 4090:2") or any other count above one is now refused the same way.
Evaluating a model against an Environment on more than one GPU had the same problem and is refused too, with
“Environment evaluation runs on one GPU for now”. Benchmark evaluations and every other kind of training keep their
multi-GPU support, with or without an Environment bound.