Skip to content
cccskills
FreePublished on GitHub

jetson-headless-mode

Plan and apply safe Jetson headless-mode changes to reclaim GUI and daemon memory.

The English summary is being prepared. Showing the original description.

See how to install

Files included (7)

  • SKILL.md7.5 KB
  • BENCHMARK.md3.2 KB
  • evals/evals.json6.4 KB
  • scripts/apply.sh5.7 KB
  • scripts/plan.sh5.1 KB
  • skill-card.md3.6 KB
  • skill.oms.sig4.9 KB

SKILL.md (original)

Review the instructions the agent will receive before you install.

Jetson Headless Mode

Plan-then-apply for safe, reversible user-space memory reclamation: switch the default systemd target away from graphical.target and disable a curated set of non-essential daemons. This is the highest-yield, lowest-risk memory win on Jetson.

Purpose

Build a user-approved headless-mode plan from live audit data, then apply only safe, reversible user-space changes that reduce desktop and daemon memory use on Jetson.

When to use

  • "Free as much memory as possible — I don't need the GUI."
  • "I'm shipping this Jetson as an inference appliance / edge node."
  • After jetson-memory-audit shows default_systemd_target=graphical.target or shows gdm3 / lightdm / sddm active on a system the user describes as headless.

When NOT to use

  • The user needs the local desktop, display output, kiosk UI, or any X/Wayland session. In that case, do not recommend disabling the graphical target or display manager; use jetson-memory-audit for a read-only view and suggest non-GUI memory options instead.
  • You do not have current audit data. Run jetson-memory-audit first, or ask the user for its output, before proposing changes or estimating savings.

Use live device data as the source of truth. Jetson family, SKU/variant, memory totals, active display services, and savings estimates must come from jetson-diagnostic/scripts/detect_jetson.sh, audit.json, or a fresh jetson-memory-audit run. If a value is not available, say it is unknown instead of guessing. The savings numbers below are upper bounds; the real delta is whatever a before/after audit reports.

Prerequisites

  • Start from a current jetson-memory-audit JSON snapshot.
  • Confirm the user does not need the local desktop, display output, kiosk UI, or X/Wayland session.
  • Mutating changes require sudo and explicit user approval; dry-run first unless approval was already given in the same prompt.
  • Run on the Jetson host or in a host-visible sandbox with access to systemd state.

Available Scripts

ScriptPurposeArguments
scripts/plan.shReads a memory audit JSON and emits a plan containing safe, reversible recommendations.--audit PATH or --audit -, plus --human.
scripts/apply.shPrints or applies the safe commands from a plan JSON. Dry-run by default.--plan PATH or --plan -, --apply, --reboot, --drop-caches.

If your agent runtime supports run_script, use it to run scripts/plan.sh and scripts/apply.sh and summarize the returned output. Otherwise run the scripts with bash from the repository root.

Instructions

  1. Run scripts/plan.sh to read audit.json (from jetson-memory-audit) and emit a plan with only safety: safe knobs (target switch, display managers, audio, print, modem, etc.).
  2. Show the plan to the user and confirm.
  3. Run scripts/apply.sh --plan plan.json for a dry run. Re-run with --apply to execute. Add --drop-caches to flush the page cache afterward, or --reboot to take effect immediately.
  4. Re-run jetson-memory-audit/scripts/audit.sh to verify the actual delta.

Expected workflow

Use the scripts for estimates and application so recommendations are based on the current device state rather than the static upper-bound table alone.

  • For "what would headless save", "estimate", "plan", or production planning prompts, run scripts/plan.sh --audit <audit.json> and report estimated_total_savings_mb, the top recommendations[*].knob, and whether any display manager or graphical.target is active. Do not run apply.sh.
  • For prompts where the user explicitly says to apply headless mode now, run scripts/apply.sh --plan <plan.json> once as a dry run first. If the user has already approved mutation in the same prompt, re-run the same command with --apply and mention the reversible command(s).
  • If direct execution fails in an agent runtime, invoke scripts with bash {baseDir}/scripts/<script-name> .... Do not try to chmod installed skill files.

Plan / apply contract

  • plan.sh emits the same JSON shape as jetson-inference-mem-tune/scripts/recommend.py: an array of recommendations with {layer, knob, estimated_savings_mb, safety, command, reversible_command, rationale}.
  • apply.sh filters entries to safety == "safe" with a non-empty command, then re-checks the filtered safety marker in the shell loop before execution. Anything else, such as kernel command-line changes, device-tree changes, or accuracy tradeoffs, is out of scope for this skill.
  • Default mode is dry-run. --apply is required to mutate the system.

Knobs covered

KnobActionEstimated savingsReversible?
disable-graphical-targetsystemctl set-default multi-user.targetup to 865 MByes
stop-gdm3 / gdm / lightdm / sddm / display-managersystemctl disable --now <svc>~200 MB / svcyes
stop-pulseaudiodisable audio daemon~8 MByes
stop-bluetoothdisable Bluetooth stack~6 MByes
stop-ModemManagerdisable WWAN manager~4 MByes
stop-cups / stop-cups-browseddisable print stack~5 / ~3 MByes
stop-snapddisable Snap daemon~30 MByes
stop-whoopsie / kerneloopsdisable crash reporters~4 / ~2 MByes
stop-avahi-daemondisable mDNS~3 MByes
stop-unattended-upgrades / packagekitdisable background package work~6 / ~8 MByes

Do NOT disable these services

  • nvargus-daemon — required for any libargus camera pipeline.
  • nvgetty.service — serial console; disabling can lock you out of recovery.
  • nvpmodel — power-mode service; required for clock/power tuning.
  • containerd / docker — leave on if you run containers (most inference workloads do).
  • nvfb / nvdisplay-related kernel services — tied to boot-time display configuration, so this skill does not change them.

Safety

  • Does not edit /boot/extlinux/extlinux.conf, the device tree, or boot-time memory reservations.
  • Does not disable services it does not have an explicit entry for (no blanket "disable everything not whitelisted").
  • Every applied change has a documented reversible_command. Re-running the plan with the reverts is sufficient to restore.
  • Dry-run by default. --apply is the only way to mutate.
  • Report only device facts and savings figures that came from live detection or audit output.

Cross-platform behavior

The same set of knobs applies to every Jetson family in the matrix above. The script reads JETSON_GENERATION / JETSON_PRODUCT_LINE / JETSON_VARIANT from jetson-diagnostic/scripts/detect_jetson.sh (and still exports legacy JETSON_SKU) so the agent can attribute the savings correctly in its summary, but it does not branch on product line.

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV/Parquet I/O, nullable semantics, and multi-GPU DataFrame workloads.

The English summary is being prepared. Showing the original description.

NVIDIA/skills3.6KUpdated Oct 18, 2026

Customize NVIDIA Nemotron Voice Agent's Generic Pipecat example for healthcare appointment, five-field patient intake, or custom tool-calling workflows without a separate backend.

The English summary is being prepared. Showing the original description.

NVIDIA/skills3.6KUpdated Oct 18, 2026

Calibrate a new dataset from live RTSP camera streams via the AutoMagicCalib REST API. Use when the user provides RTSP URLs or asks to calibrate live cameras; VIOS records clips, AMC ingests them, then runs calibration.

The English summary is being prepared. Showing the original description.

NVIDIA/skills3.6KUpdated Oct 18, 2026

Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.

The English summary is being prepared. Showing the original description.

NVIDIA/skills3.6KUpdated Oct 18, 2026

Calibrates pre-recorded `cam_*.mp4` datasets through the AutoMagicCalib REST API. Use for user-supplied local MP4s; route live RTSP streams to `amc-run-rtsp-calibration`.

The English summary is being prepared. Showing the original description.

NVIDIA/skills3.6KUpdated Oct 18, 2026

Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.

The English summary is being prepared. Showing the original description.

NVIDIA/skills3.6KUpdated Oct 18, 2026

See all skills by NVIDIA

Report a problem with this skill