What's the problem?
Your system has a Strix Halo GPU (gfx1151), but your Linux kernel is missing a critical fix that exports CWSR (Context Wave Save/Restore) properties. Without this fix, ROCm may have incorrect VGPR counts, causing crashes for various workloads.
This requirement applies to any ROCm backend on gfx1151 — llamacpp:rocm,
sd-cpp:rocm-stable, and
vllm:rocm all share the same kernel prerequisite.
Technical details:
- Strix Halo optimization guide
- The fix is included in kernel 6.18.4+ but may be backported to older kernels
How to fix this
Install the OEM Kernel
Ubuntu provides an OEM kernel with the fix included. Open a terminal and run:
sudo apt update && sudo apt install linux-oem-24.04
After installation, reboot your system for the new kernel to take effect.
Upgrade to a Kernel with the CWSR Fix
Upgrade your kernel to version 6.18.4 or later, or a kernel with the CWSR fix backported. The fix exports cwsr_size and ctl_stack_size properties to userspace.
Check Current Kernel Version
Open a terminal and run:
uname -r
Update Your Kernel
Use your distribution's kernel update process. For example:
- Fedora/RHEL:
sudo dnf update kernel - Arch Linux:
sudo pacman -Syu linux - openSUSE:
sudo zypper update kernel-default
Reboot
Reboot your system to load the new kernel:
sudo reboot
Verify
After rebooting, verify the CWSR properties are exported:
grep -E "cwsr_size|ctl_stack_size" /sys/class/kfd/kfd/topology/nodes/*/properties
You should see both cwsr_size and ctl_stack_size listed.
Still seeing ROCm failures after updating your kernel?
If your kernel is up to date but Lemonade still refuses to install or run the ROCm backend,
you may have the AMDGPU DKMS
module installed. On Ryzen AI / Strix Halo systems, the DKMS module conflicts with the
in-tree amdgpu driver and breaks GPU access for ROCm workloads.
Reinstall ROCm without DKMS
AMD publishes a Ryzen-specific ROCm install guide that explicitly avoids the DKMS module found here:
In short, when running amdgpu-install, pass a use case that does
not include dkms, for example:
sudo amdgpu-install -y --usecase=rocm --no-dkms
If DKMS is already installed, remove it first:
sudo apt remove --purge amdgpu-dkms && sudo reboot
After reinstalling ROCm without DKMS and rebooting, retry installing the Lemonade ROCm
backend (llamacpp:rocm, sd-cpp:rocm-stable, or
vllm:rocm).
Need Help?
If you continue to experience issues after updating your kernel:
- Check the GitHub Issues page
- Contact us on Discord