You can try using the LegacyMultiProc plugin, which is better able to handle accumulating memory within Python processes.
Nipype moved to a new MultiProc backend a few versions back, which doesn’t run into the failure where a process killed by the OS simply disappears and Python receives no signal and thus waits forever. However, it is less tweakable, and a lot of the optimizations that went into LegacyMultiProc became impossible to replicate. Unfortunately, it appears that whichever we choose as the default will cause problems for a different set of users.