Arch Linux

So what starts nvidia_svm but doesn't shut it down? I don't even find a file called "nvidia_svm", so I have no idea what it is.

sorry I misspelled it, it was nvidia_uvm. Can I edit the task it self?

Can't find anything by that name in the repos, either. I'm just trying to figure out where to assign this.

optirun with my own opengl application works well. Maybe caused by python-pycuda? I don't know either. nvidia_uvm is part of nvidia driver, for unified memory feature of CUDA.

New comment, run the plain cuda executable with optirun also cause this problem, so this is not the bug of pycuda, as OpenGL applications runs smoothly, it should be a bug of cuda itself, so maybe we need to ignore it cause cuda is propriety.

This is not a bug with CUDA.
When you say "optirun blaa" what optirun does is to switch on the gpu and insert the nvidia module. When the program finishes, it removes the nvidia module and turns off the GPU.

On newer cuda, when running a cuda application, the nvidia module requests that the nvidia_uvm module is loaded as well, as it is needed by the cuda runtime. Bumblebee doesn't know about this module and tries to remove nvidia, but nvidia_uvm depends on nvidia so it can't remove the module and can't turn off the GPU.

Solution:
manually:
modprobe -r nvidia_uvm
modprobe -r nvidia
tee /proc/acpi/bbswitch <<< OFF

Should this be a bug?
If anything you could talk to the Bumblebee developers to check for nvidia_uvm before trying to remove the nvidia module and shutdown the GPU.

thomasking0529, do you still have this issue?

Sorry I don't use pycuda anymore so can't answer this question.
p.s. thomasking0529 is my old account