-
-
Notifications
You must be signed in to change notification settings - Fork 1
FAQ and Troubleshooting
On an installed system, start with ai-2 doctor, which checks the engine, model, tuning, services and repository key. For a bug report, ai-2 report writes ~/ai2-report.txt to attach to an issue. Questions and ideas go to Discussions.
The computer refuses to boot the stick and mentions Secure Boot. Turn Secure Boot off in the firmware setup (the key shown at power-on, often F2, Del, F10 or Esc). The image is not signed for Secure Boot.
The screen stays black after the boot menu. Reboot and pick "Start AI-2 (safe graphics, if the screen stays black)". Since ISO 20260825, AMD GPUs of the 2012-2016 era (SI/CIK) are steered to the mature radeon driver automatically, because the newer amdgpu driver probe-crashes on them under current kernels.
Is my machine too old? Above the floor (a 64-bit x86_64 CPU, 2 GB of RAM, about 6 GB of disk) AI-2 does not turn a computer away. CPUs without AVX or even SSE4.1 are served by dedicated engine builds, see Validated Hardware. 32-bit machines are not supported, and the boot menu says so.
Is the download corrupted? Verify with sha256sum -c ai-2-x86_64.iso.sha256, and use the boot menu's "Check this stick first" entry to verify the written stick.
"Erase disk" is missing in the installer. Fixed in ISO 20260825. Older ISOs activated any swap partition found at live boot, which made the installer refuse whole-disk installs on machines that previously ran Linux. With an older ISO, run sudo swapoff -a in a terminal before starting the installer, or download the current ISO.
An install made with ISO 20260816 or 20260817 boots to a text-mode GRUB without Windows in the menu. Those two ISOs lost Artix's GRUB defaults. On an affected install, set GRUB_DISABLE_OS_PROBER=false in /etc/default/grub, then run sudo grub-mkconfig -o /boot/grub/grub.cfg. ISOs from 20260819 on are fixed.
The chat page says "Not Secure". Expected. The page is served by your own computer at 127.0.0.1, nothing leaves the machine, there is nothing to secure.
The chat page shows an error the first time. Reload it. On a slow disk the server needs a moment to load the model.
The AI engine crashes instantly ("illegal instruction"). With the one ai2-llama-cpp package this should not happen, the engine picks the CPU build at start (x64 as the floor) and every file that must run on a pure-SSE2 machine is checked before it ships. If it does happen, run ai-2 report and attach ~/ai2-report.txt to an issue; it says which module was loaded. A machine still on one of the old per-CPU packages (ai2-llama-cpp-baseline, -noavx, -avx2) gets the new package with the next update.
Why does the benchmark download a model I did not ask for? The AI Score is always measured on one fixed model so scores stay comparable across machines. Benchmarking whatever model happens to be present would produce incomparable numbers, so AI-2 refuses to.
Why is the recommended model so small? I have the RAM for a bigger one. RAM says what fits, not what runs. The recommendation is gated by measured speed, and AI-2 would rather tell you a big model is unusably slow on this machine than pretend, see Models.
Getting help from someone remotely, in the live session. Double-click "Remote help (SSH)" on the live desktop, it shows the exact command and password the helper types on their computer. Only machines on your own network can connect, and it ends at shutdown.
SSH on an installed system. Off by default. To allow it:
sudo ln -s /etc/runit/sv/sshd /run/runit/service/