CHANGELOG.md

v0.2.3 — 2026-04-23

Fixed

• Removed IO PSI row — false CRITICAL on GB10 with VLLM idle
• IO PSI references removed from anomaly logger and pressure detection

v0.2.2 — 2026-04-19

Added

• PWR row — GB10 power rail monitor via spark_hwmon (antheas/spark_hwmon)
  • GPU instantaneous draw, session peak, DC input, SysPL1 cap
  • PROCHOT status, active PL level, Tj thermal rise
  • Hidden on non-GB10 systems — shown only when spbm hwmon device present
• PROCHOT triggers anomaly logger — captured in anomaly.log.gz and summary.json
• PROCHOT turns ⚡UMA indicator red — visible at a glance
• peak_gpu_w added to summary.json

Why PWR matters

• DC input < SysPL1 cap = power brick not delivering enough = PD throttle
• PROCHOT ACTIVE = hardware brake fired = clock drop to 611MHz
• These two signals explain why CLOCK shows THROTTLED

Changelog

v0.2.0 — 2026-04-18

Added

• TEMP row — current and session peak for GPU and CPU, color-coded green (<60°C) / yellow (60–80°C) / red (>80°C)
• INFO row — live time (12hr), driver version, CUDA version, kernel, uptime
• Anomaly auto-logger — starts automatically when issues are detected, stops and compresses on clear
  • Logs saved to ~/sparkview_logs/<timestamp>/anomaly.log.gz
  • Machine-readable summary.json per event — trigger, duration, peak temps, driver, CUDA, kernel
• ⚡UMA indicator turns red when PSI HIGH, CLOCK THROTTLED, or temp > 80°C

Fixed

• UMA memory reporting — now uses vm.total - vm.available for used memory, accurate under heavy inference loads
• UMA detection — tightened to 0.9 ≤ mem.total/vm.total ≤ 1.1, prevents false positives on large discrete GPUs
• PROC list CPU% capped at 100% — prevents psutil multi-core artifact
• CLOCK / PSI alignment — consistent 8-char status field width
• Terminal height scaling for PROC list — dynamic, no hardcoded row limit

Changed

• INFO row sits tight between PSI and PROC — no blank line separator
• Removed ConnectX-7 from feature list pending GB10 validation

v0.1.0 — 2026-04-17

Added

• Initial release
• GPU utilization, temperature, power, memory via NVML
• GB10 / DGX Spark coherent UMA detection at runtime
• PSI memory pressure — LOW / MOD / HIGH / CRITICAL
• Load-gated clock states — IDLE / PASS / LOCKED / THROTTLED
• CPU utilization and active core count
• SWAP monitoring
• Process list sorted by GPU memory usage
• venv install instructions per community feedback