Run headless, and get the compositor’s VRAM back

A desktop session holds VRAM that the model could be using. On a 24 GB card that is not a rounding error: it is most of a 35B model’s cache, and for one entry in the catalogue it is three times the context.

Do it

sudo systemctl isolate multi-user.target

That drops to a text console immediately. No reboot and nothing uninstalled. Go back with:

sudo systemctl isolate graphical.target

Log in on the console (Ctrl-Alt-F3 if you need another one) and use lllm3090 exactly as before, or reach the panel over an SSH tunnel — see Reach the panel from another machine.

First, check the panel will survive it

Warning

The panel is a user unit. It runs inside user@UID.service, and the user manager stops when your last session ends — taking the panel and the engine in its cgroup with it. Isolating to multi-user.target ends the graphical session, so without lingering enabled, going headless stops the very thing you went headless for.

lllm3090 doctor          # the 'linger' check reports this

or ask directly:

loginctl show-user "$USER" -p Linger

If it says Linger=no:

sudo loginctl enable-linger "$USER"

With lingering on, the user manager is started at boot and persists with no session, so the panel keeps serving on a text console and across logouts. If you would rather not enable it, stay logged in on the console — a console session is a session, and the user manager lives as long as one exists.

What it is worth

Measured on the reference RTX 3090, per conversation:

model	desktop	headless
Qwen3.6-35B-A3B-Q4KS	61k × 2	181k × 2	+197%
Gemma-4-26B-A4B	138k × 2	256k × 2	+86%
Qwen3.8-27B	101k × 2	139k × 2	+38%
Muse-Glimmer-30B	128k × 3	128k × 4	one more slot
Qwen3.6-35B-A3B	212k × 2	256k × 2	+21%
gpt-oss-20b, Qwen3-8B, Gemma-4-12B-QAT	—	—	no change

The models that gain nothing are already stopped by their own RoPE ceiling rather than by the card, so extra cache buys them nothing — see What actually decides whether a model fits.

The models that gain most are the ones the desktop was squeezing hardest. Qwen3.6-35B-A3B-Q4KS is 20.9 GB of weights on a 24 GB card, so the compositor’s share was coming almost entirely out of its cache.

You do not have to tell it

lllm3090 asks systemd whether graphical.target is running and plans accordingly, so the figures in lllm3090 models and in the panel are already the ones that apply to the session you are in. Switch targets and the numbers change with it.

If it cannot tell — no systemctl, or the call fails — it assumes a desktop is running. The two mistakes are not symmetric: over-reserving costs you context, while under-reserving produces an engine that loads, reports itself healthy and then fails every request with ErrorOutOfDeviceMemory.

Warning

Context is chosen when the engine starts. Plan a model on a text console, then switch back to graphical.target without restarting it, and the engine is now sized for VRAM that the compositor has taken back. lllm3090 start compares the plan against actually-free VRAM and warns, but it cannot see a desktop that arrives afterwards. Restart the engine after switching.

Is it worth it?

For interactive use with an agent, usually yes when the model is a tight fit, and not otherwise. If your model is already RoPE-capped on a desktop, headless buys nothing at all and costs you a working desktop.

The place it pays is a long unattended run — an overnight batch, a large repo reviewed in one pass — where nothing needs the screen and the extra cache is the difference between one conversation and several, or between compaction and none.