jetson-memory-audit

Jetson Memory Audit

Read-only memory-focused snapshot for a Jetson, plus the drop_caches verify-loop helper that confirms freed memory actually shows up as free instead of cached.

Purpose

Measure current Jetson memory consumers, capture before/after baselines, and verify whether user-approved changes actually reclaimed DRAM. Use live device data rather than estimates from container size, model size, or generic process memory.

CRITICAL: Memory appears stuck after stopping vLLM / sglang (JetPack below 7.2 / L4T below r39.0)

This is the most common memory confusion on Jetson releases before JetPack 7.2 or before L4T r39.0.

After you stop a vLLM, sglang, or Ollama server (or any CUDA workload), the memory shown as free by free -h or tegrastats may not recover — even though the process is gone. nvidia-smi may also show misleadingly low free GPU memory.

Root cause: The Thor RM (resource manager) holds freed sysmem pages in its own pool after a CUDA context exits. On Unified Memory Architecture (UMA) devices like Jetson, cudaMemGetInfo reads RM pool state and reports far less free memory than is actually available to a new process.

Workaround (for JetPack below 7.2 or L4T below r39.0):

sudo sync && sudo sysctl -w vm.drop_caches=3

Run this on the host, not inside a container. The important operation is sudo sysctl -w vm.drop_caches=3; keep sudo sync immediately before it so dirty data is flushed before reclaimable page/dentry/inode caches are dropped. After running it, free -h and tegrastats will reflect the true available memory.

For affected releases, recommend this command when a user says:

• "Memory didn't free after I stopped vLLM/sglang"
• "Why does tegrastats still show high usage after my container exited?"
• "OOM even though nothing is running"
• "Memory was fine yesterday but now it's full"

On JetPack below 7.2 or L4T below r39.0, drop_caches is the reliable workaround when memory appears stuck after a CUDA workload exits; on newer releases, use it only if the same symptom is observed and the user approves.

When to use

• "How much memory is in use on this Jetson? What's holding it?"
• "I disabled the GUI / stopped vLLM / quit my container — did memory actually free?"
• "Why does free -h still show low free memory after I stopped my workload?"
• As the baseline before applying jetson-headless-mode or other memory-related changes, and again after to compute the actual delta.

Prerequisites

• Run on the Jetson host, or in a sandbox/container with host-visible /proc, /etc/nv_tegra_release, tegrastats, and process data.
• NvMap debugfs reads may require root. If unavailable, report that GPU memory attribution is limited rather than guessing.
• drop_caches.sh requires root or passwordless sudo -n; run it only after the user explicitly authorizes cache dropping.

Available Scripts

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

Instructions

For "how much memory is in use right now?" questions, run scripts/audit.sh and report only values from the JSON snapshot.

Reporting guidance

Do not only print or mention the path to a helper. Invoke the helper and then summarize the returned data.

• For "how much memory is in use" prompts, run scripts/audit.sh and quote mem_total_gb, memory_kb.available, and the leading procrank_top process or nvmap.top_clients consumer.
• For GUI/desktop memory prompts, run scripts/audit.sh and report default_systemd_target plus any display manager in candidate_services (gdm3, gdm, lightdm, sddm, or display-manager). Do not disable anything; hand off to jetson-headless-mode for a plan.
• For prompts that explicitly authorize cache dropping after a stopped workload, run scripts/drop_caches.sh (equivalent to sudo sync && sudo sysctl -w vm.drop_caches=3 by default) and report its before/after free, available, and cached deltas. If root is unavailable, explain that it must be run on the host with sudo.

If your agent runtime does not execute helper scripts relative to this skill directory, resolve script paths with the AgentSkills {baseDir} placeholder:

{baseDir}/scripts/audit.sh
{baseDir}/scripts/drop_caches.sh

Do not call jetson-memory-audit as a tool name unless the runtime explicitly registers skills as callable tools; Agent Skills are normally instructions plus files, not direct tool functions.

Sandbox note for agents: seeing this skill file does not guarantee access to Jetson host memory data. If /proc/device-tree/model, /etc/nv_tegra_release, tegrastats, /sys/kernel/debug/nvmap, or host process data are missing inside a NemoClaw/OpenClaw sandbox, say the sandbox lacks Jetson host visibility and ask the user to run on the Jetson host or relaunch with a host-visible sandbox profile. Do not fabricate memory totals, available memory, PSS, NvMap, or reclamation deltas.

For "how much memory did this change free?" questions, use a before/after delta. Do not estimate freed memory from container size, image size, RSS, or a single post-change snapshot.

1. Before the change, run scripts/audit.sh and save the JSON baseline.
2. Make the user-approved change (stop the container, switch mode, apply a tuning recommendation, etc.).
3. On JetPack below 7.2 / L4T below r39.0, or when the same stuck-memory symptom is observed on a newer release, flush reclaimable page cache on the host (not inside a container) so freed pages show up as free instead of cached:

   sudo sync && sudo sysctl -w vm.drop_caches=3
   

4. Re-run scripts/audit.sh and compare memory_kb.available before vs after — that delta is the real reclamation.

If the user already made the change and no baseline exists, say that the exact freed amount cannot be recovered from the current snapshot alone. Capture a new baseline now so the next change can be measured.

Use live audit data as the source of truth. Memory totals, available memory, NvMap totals, PSS values, display-manager state, and savings deltas must come from scripts/audit.sh, free -h, or tegrastats on the actual device. If a number is not present in those outputs, do not guess it.

Output contract for audit.sh

{
  "sku": "orin-nano",
  "variant": "orin-nano-8gb",
  "mem_total_gb": 8,
  "l4t_version": "36.4.0",
  "product_model": "nvidia jetson orin nano developer kit",
  "memory_kb": { "total": 8123456, "available": 4123456, "free": 1023456, "cached": 1234567, "swap_total": 0, "swap_free": 0 },
  "default_systemd_target": "graphical.target",
  "candidate_services": { "gdm3": { "active": "active", "enabled": "enabled" } },
  "tegrastats_sample": "RAM 4011/8138MB (lfb 8x4MB) ...",
  "nvmap": { "readable": false, "total_kb": 0, "top_clients": [] },
  "procrank_top": [ { "pid": 4321, "pss_kb": 4000000, "cmd": "vllm" } ]
}

Limitations

• Exact freed-memory deltas require a before snapshot, the user-approved change, cache flush when appropriate, and an after snapshot.
• NvMap attribution depends on host-visible debugfs access; if it is unavailable, report limited GPU memory attribution instead of guessing.
• Sandbox/container runs may not see host /proc, tegrastats, systemd, or NvMap data unless the runtime exposes them.

Error handling

• If scripts/audit.sh cannot access host Jetson data, report the missing visibility and ask to rerun on the Jetson host or in a host-visible sandbox.
• If scripts/drop_caches.sh lacks root or passwordless sudo -n, report that cache dropping must be run on the host with sudo approval.
• If no before snapshot exists, say the exact reclaimed amount cannot be recovered from the current state alone and capture a new baseline for the next change.

Safety

Read-only. drop_caches is non-destructive (kernel only releases pages it could reclaim under pressure anyway; sync runs first to preserve dirty data).

Hand off to

• jetson-headless-mode — biggest single user-space win on systems still booting graphical.target.
• jetson-inference-mem-tune — when a model server is the top NvMap / PSS consumer.
• If runtime changes cannot hit the target, report that further reclamation is outside this skill's scope rather than suggesting unsafe boot-time edits.