How can I deactivate the GPU locking by cntk from the python API (2.0 beta 12.0) such that multiple users may use the same GPU?
If I execute the cntk tutorial 201B from https://github.com/Microsoft/CNTK/tree/master/Tutorials the training works fine, but all other users / processes get the error message: "Failed to lock GPU 0 for exclusive use" when executing cntk based optimizations.
You can set the device explicitly using GPU id:
cntk.device.set_default_device(cntk.device.gpu(0))
That way you can run multiple jobs on the same GPU.