Earlier technical tool. The public source release is under review. Do not use the current public repository to approve a rerun or declare hardware healthy. See current Ashiba work →
scanprobe v0 · alpha
A tiny read-only scan for the first few minutes of GPU troubleshooting.
Status: v0 alpha · NVIDIA local evidence only · seeking incident reports from real GPU operators · one-pager (PDF)
Use scanprobe when a GPU node has become a suspect and you need obvious local NVIDIA evidence before rerunning, draining, or filing a support ticket.
It answers one narrow question: is this node or GPU obviously weird from local NVIDIA evidence, or should I keep looking elsewhere?
What it checks
scanprobe gathers the first-pass checks an operator often runs by hand:
nvidia-smiGPU visibility and basic GPU state- ECC counters exposed by
nvidia-smi - temperature and hardware throttle flags
- readable current-boot NVIDIA Xid / kernel logs
It prints a primary issue, the visible evidence, a next action, and one conservative triage label:
CLEAR— no local drain or watch evidence was visibleWATCH— inspect the visible evidence before rerunning expensive workDRAIN— do not launch new work until the listed evidence is resolvedUNKNOWN— this shell cannot see enough local GPU or kernel-log state
Release status
A safer local branch uses bounded evidence labels and explicit limits. Ashiba has not published that branch to the public repository. The source link and run instructions will return after a reviewed release.
Constraints
No daemon. No telemetry. No mutation. No stress workload. No API key. No hidden benchmark. No claim that a GPU is healthy.
scanprobe is not a DCGM replacement. DCGM is the serious datacenter GPU management and diagnostics stack. scanprobe is the smaller first-pass scan before rerun, drain, escalation, or deeper diagnostics.
The goal is modest: save a few minutes, catch obvious local evidence, and make the first response easier to paste into Slack, Jira, or a provider ticket. In alpha, we are testing whether that means roughly 2–10 minutes saved per first-pass incident.
Use it for the first look after something feels off. Do not use it as a routine heartbeat or a health certificate.
Roadmap
One command gathers obvious local NVIDIA evidence and prints a pasteable summary. Current alpha.
Better fixtures from real operator reports, clearer issue tags, and stronger handling for common visibility, Xid, ECC, throttle, and nvidia-smi failure cases.
A read-only runbook pilot that helps operators decide which deeper tool to run next, including DCGM or provider diagnostics when appropriate.
Scope
scanprobe currently ships NVIDIA local evidence only. Ashiba will add AMD support once real AMD SMI, ROCm, and kernel-log fixtures show which read-only signals change an operator's next action.
Resources
- Public source release — under review
- Kernel Contracts — earlier work on bounded ML-kernel checks