Force single core per process

I’m trying to parallelise my experiments on a CPU cluster. Each experiment uses comparatively tiny symbolic expressions (max 16 operators). Yet all my CPU cores are used when I run a single experiment using SGD or Adam—resulting in no gains when parallelising. The BFGS and LM implementations from Scipy for example, only use a single CPU core. How can I force Pytorch to use only a single CPU core per process?