Omarchy will hand freed GPU memory straight back to the system on the DGX Spark
A stopped 100 GiB model shouldn’t keep 100 GiB of RAM hostage. A GB10 fix merged by DHH turns off NVIDIA’s driver memory pools, matching what DGX OS does.
Image: Omarchy (omarchy.org)
Running big local models on an NVIDIA DGX Spark under Omarchy is about to get less wasteful. DHH merged a pull request on November 18, 2026 that makes GB10 machines give freed GPU memory back to the system as soon as a process lets go of it. It’s coming in the next version.
The culprit is a feature of NVIDIA’s 615 driver. According to the PR, the driver keeps the GPU memory a process frees in its own page pools (NVreg_EnableSystemMemoryPools, with 4K, 64K and 2M pools by default) so it can hand it out faster next time, and only gives it back when the kernel’s shrinker runs under memory pressure. On GB10, the GPU’s memory is the machine’s RAM.
The PR puts a number on it: after stopping a 100 GiB local model on a DGX Spark, 106.7 GiB stayed in the driver’s pools, showing up as KReclaimable in /proc/meminfo instead of as free memory. For comparison, a Spark running NVIDIA’s own DGX OS, whose 580 driver has no pools, got all of it back the moment the model stopped.
The fix matches DGX OS by setting options nvidia NVreg_EnableSystemMemoryPools=0 on GB10 machines only:
- New installs get it from a new
install/hardware/gb10/nvidia-memory-pools.shstep, run alongside the other GB10 hardware steps. - Existing machines get a migration that writes the same modprobe file, rebuilds the boot image (the
nvidiamodule loads from it, along with its options) and asks for a reboot.
The PR description is upfront about the caveats. The pool figures describe the problem, but the change itself had “not yet run on a Spark” when it was merged. It only takes effect after a reboot, and the thing to check, it says, is “that model load time doesn’t suffer without the pools.”
It’s the latest piece of Spark tuning since DHH posted on X on November 9, 2026 that NVIDIA had sent him a pair of DGX Sparks, that Omarchy was “fully up and running” on them, and that he was “working to tune the out-of-box local AI offerings.”
