If you’re running Debian family system and frequently putting your system to sleep, you might have encountered an issue where CUDA becomes unavailable after the computer wakes up. This happens because the NVIDIA uvm (Unified Virtual Memory) module doesn’t reload automatically on resume. While rebooting the computer resolves the issue, it’s not an ideal solution for day-to-day use.
In this article, I’ll walk you through an automated solution to ensure that your NVIDIA uvm module is reloaded every time your system wakes up—without any manual intervention.
The Problem
On many Debian systems, especially those using systemd, the NVIDIA drivers include a sleep script located in the /lib/systemd/system-sleep/ directory called nvidia. This script manages certain NVIDIA-related tasks when your computer suspends and resumes. However, it does not handle the reloading of the nvidia_uvm module, which is critical for CUDA functionality.
The workaround many users apply is to manually run:
sudo modprobe -r nvidia_uvm && sudo modprobe nvidia_uvm
after waking up the system. But why do it manually every time when we can automate the process?
The Automated Solution
Systemd provides a convenient mechanism to execute scripts during sleep and wake cycles. Scripts placed in /lib/systemd/system-sleep/ are executed in lexical (alphabetical) order. By naming our script appropriately, we can ensure that it runs after the existing NVIDIA sleep script.
Step 1: Create Your Script
Create a new file in the /lib/systemd/system-sleep/ directory with a name that sorts after the NVIDIA script. For example, naming it nvidia-uvm.sh ensures it runs later in the sequence:
[ . . . ]