Sys Health

sys-health

One-shot system health snapshot for this UniLab dev laptop. RL training here is CPU-only and runs inside WSL2. A long run stresses three things, in order of risk for a laptop:

1. Heat — sustained 100%-on-all-cores cooks a mobile CPU.
2. SSD write endurance — every model_*.pt checkpoint and tfevents line is

a write; small save_interval × many iters = real wear.

1. C: drive space — the WSL2 ext4 filesystem is a .vhdx that lives on C:;

if C: fills, the vhdx can't grow and training crashes mid-run.

This skill measures all three in one command (heat needs one-time setup, below).

How to invoke

wsl -d Ubuntu -u u20174 -- bash -c \
  "tr -d '\r' /dev/null

• tr -d '\r' makes the script immune to CRLF a Windows editor may introduce.
• 2>/dev/null mutes the WSL localhost-proxy Chinese-garbage warning.
• The script is read-only and idempotent — safe to run anytime, including

mid-training.

After it prints, give the user a one-paragraph health verdict + the single most important action (see Interpretation). Do not just dump the raw output.

Sections & how to read them

| Section | Healthy | Watch / Act | |---|---|---| | Load (1/5/15 min) | below 16 sustained | 15-min load pinned near/over 16 = box fully committed; nothing else runs well | | MEMORY & SWAP | Swap 0B used, available > 2GB | Swap in use = training spilling to SSD (slow + wear). Low available RAM → reduce algo.num_envs | | DISK USAGE | Use% 85% → clean up; this ext4 is the WSL vhdx | | TOP CPU PROCESSES | see what's running | a python at ~cores×100% for a long ELAPSED = a UniLab run is active (note the run path under TRAINING WRITE ACTIVITY) | | TRAINING WRITE ACTIVITY | few files in last 10 min | every listed file = an SSD write; checkpoint size × save_interval × iters = cumulative wear | | NVMe SMART | Data Units Written low, temp 60GB | C: < 40GB risky — vhdx + pagefile live here and grow with training data | | WINDOWS CPU TEMP | < 85°C (laptop) | needs LibreHardwareMonitor; without it, temp is the one signal this skill can't see |

Known gaps (and the one-time fixes)

• CPU temperature is invisible by default. WSL2's virtual kernel does not

expose host CPU temp, and stock Windows WMI hides it. This is the #1 laptop risk signal, so enable it once:

1. Install LibreHardwareMonitor (free), launch it, enable its WMI

provider (and ideally run at startup).

1. Re-run this skill — the WINDOWS CPU TEMP section auto-detects it via the

root/LibreHardwareMonitor namespace. No script edit needed.

• SSD wear numbers (bytes written, % life left) need nvme-cli + sudo.

Until then the script reports proxies: checkpoint count/size and recent write activity. For real numbers: sudo apt install nvme-cli, then re-run.

Verdict guidance tuned to THIS box

• It's a laptop (8745HS mobile APU, no discrete GPU). Heat is the real risk

under long CPU-only runs. Until LibreHardwareMonitor is on, recommend: elevate the laptop, keep vents clear, run in a cool room, and consider not pegging all 16 threads 100% (smaller num_envs).

• RAM ~14GB in WSL / ~28GB on host — memory is not the bottleneck; do not

let training swap.

• The Samsung 1TB NVMe is roomy but an OEM low-power part with modest write

endurance — checkpoint spam (very small save_interval) is the main wear source. Move old checkpoints off the SSD (see checkpoints/ convention in the project CLAUDE.md).

• If C: drops under ~40GB free, move/clean before starting a long run,

not after it crashes.

Files

• check_sys.sh — the probe script (read-only, idempotent, safe anytime).

Edit the probe list here if new hardware is added or a section misbehaves.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: phanghonghao
• Source: phanghonghao/THU-Awesome-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.