sparkDash v2.9 is out, and it's a big one 🔥 This one is all about testing your models, with quality scores, and keeping your DGX Sparks clean with a new storage manager! ⚡️ Quality Benchmarks - Rebuilt test suite with bigger categories and grades next to scores - New fidelity test: per-token loss, made for comparing a quant with the original - New divergence test: compares full replies one by one - Every score shows how uncertain it is and how much of the test it covers - Compare runs across units and pin a reference run per model family - Pause, resume, or continue a run that got cut off ⚡️ NEW: Storage Manager - See what's using your Spark's disk and browse every folder and file - Bulk delete with a safety check first (programs using the files, private keys, git work) - One-click Docker cleanup for unused images - Quick picks for whole disk, home folder or cache ⚡️ GPU controls - Per-card power limits right on the unit page - Fleet-wide clock caps that can stay after a reboot - Custom temperature warning and critical limits - One bar per GPU on multi-GPU hosts ⚡️ Also new - Pico engine, with decode tok/s support - Mac support over SSH, with Start and Stop - Kiosk page: full-screen wall display for every unit - API Keys page for LiteLLM spend and budgets - MQTT + Home Assistant integration - Harness sessions (OpenClaw, OpenCode, oh-my-pi) on each LLM card - English and Simplified Chinese - Lots of fixes, including TensorFold cache and prefill numbers Many of the improvements and changes are due to awesome community PRs and issues. Thank you! Get it here: