Day to day

Status	aisathome-node status, or the local UI
Pause (model stays loaded)	Pause in the local UI
Pause and free the GPU	sudo systemctl stop aisathome-node aisathome-llama
Resume	sudo systemctl start aisathome-llama aisathome-node
Logs	journalctl -u aisathome-node -u aisathome-llama -f
Start at boot	sudo systemctl enable aisathome-llama aisathome-node
Remove	sudo ./node/packaging/linux/uninstall.sh (keeps the model; --purge removes it too)

Local UI

Nodes installed from release v0.1.0 or later serve a status page on your computer at http://127.0.0.1:7381/; an older node gets it when you run the new installer (desktops also get an “AIs@home node” menu entry). It shows whether the node is connected, paused or serving, the model, slots in use, the GPU’s memory and temperature, requests and tokens served today and since install, and recent connections and errors. It never shows prompts or answers. Your nodes opens your account page.

Pause, resume and restarting the engine need the local password the installer created:

sudo cat /etc/aisathome-node/ui-password

Enter it once and the browser remembers it. Pause lets requests in flight finish and takes no new ones, while the model stays in GPU memory; it lasts until you resume, also across restarts. To free the GPU, stop the node instead (table above).

The page is reachable from this computer only. On a machine without a screen, open an SSH tunnel from your own computer and browse to http://localhost:7381/ there:

ssh -N -L 7381:127.0.0.1:7381 you@node-machine

In a terminal, aisathome-node status prints the same; it exits with 0 when the node is connected and 1 when it is not, so scripts and monitoring can use it.

Privacy hardening

These measures make it less likely that prompts end up on disk by accident:

• No core dumps (LimitCORE=0): a crash dump would contain prompts from memory.
• No swap (MemorySwapMax=0, needs cgroup v2): prompts are never paged out.
• A smaller prompt cache in RAM (--cache-ram 2048 MiB instead of 8192). Override with CACHE_RAM_MIB; 0 disables it.
• No prompt logging: llama-server runs without verbose logging.

They do not make the node confidential: anyone with root on the machine can still read prompts from RAM or VRAM while they are being served.

Troubleshooting

NVIDIA driver ... is too old for the prebuilt llama-server. The bundle is built with CUDA 12.8 and needs driver 525.60.13 or newer (570.26 or newer for an RTX 50). Update the driver (Ubuntu: sudo ubuntu-drivers install), reboot, and run the installer again. Or build llama.cpp yourself for your driver.

has compute capability ...; the prebuilt llama-server supports 8.6 8.9 12.0. The bundle is built for the RTX 30, 40 and 50 series only. Build llama.cpp yourself for your GPU.

not enough disk space. The model needs 16.5 GB plus about 2 GB in /var/lib. Free some space and run the installer again; a started download resumes.

model checksum mismatch. The downloaded model is not the pinned file (a corrupt download or a changed mirror). The installer deletes it; run it again to download it again. If it happens twice, tell us.

the prebuilt llama-server cannot run here. Your system’s C library is older than glibc 2.35 (for example Ubuntu 20.04). Build llama.cpp yourself.

no kernel image is available for execution on the device. llama-server was built for another GPU generation. Rebuild with "-DCMAKE_CUDA_ARCHITECTURES=86;89" (see install).

CMake: CMAKE_CUDA_COMPILER-NOTFOUND. nvcc is not on PATH: export PATH=/usr/local/cuda/bin:$PATH and configure again from an empty build/.

no /dev/nvidia-uvm and no nvidia-modprobe. The GPU device nodes are missing and the helper that creates them is not installed: sudo apt install nvidia-modprobe, then start the services again.

llama-server did not become healthy. Check journalctl -u aisathome-llama. Out of memory usually means another program uses the GPU: close it, or on a desktop whose display runs on the same card, a smaller context may fit (CTX_PER_SLOT=65536 sudo -E ./install.sh); the node then gets only requests that fit it.

The node does not connect. Check the gateway_url in /etc/aisathome-node/node.json and that the machine can reach api.aisathome.org on port 443. A rejected token shows up in journalctl -u aisathome-node.

After suspend. The node stops before suspend and starts again after resume. If the GPU was reset, the watchdog restarts llama-server within a few minutes.