MiaAI-Lab / sparkLimiter

# sparkLimiter
Cap the GPU graphics clock on all your **NVIDIA DGX Spark** machines with one command, and keep the cap after reboots.
It finds your Sparks automatically, then runs
 [code] 
on each one and installs a small systemd service so the limit is re-applied at boot.
## Why
Limiting the maximum GPU clock lowers power draw, heat and fan noise, and can keep sustained workloads steadier. `nvidia-smi -lgc` does not survive a reboot, so this tool also sets up the boot-time service.
## Quick start
Run it from your laptop/workstation, or from one of the Sparks:
 [code] 
 [code] 
Press Enter to apply to all, pick a subset by number, add hosts by hand, or quit.
## Discovery
Candidates are gathered from several places, then each is probed over key-based `ssh` (read-only, no changes). Only hosts that report a GB10 GPU / DGX Spark product name count as Sparks, and duplicates (the same Spark via alias, IP and `.local`) are merged by machine-id.
1. This machine, if it is itself a Spark (listed as `localhost`)
2. Hosts in `~/.ssh/config`
3. Unhashed, port-22 entries in `~/.ssh/known_hosts` that look local (private/CGNAT addresses, `.local`/`.lan`/`.home` names, single-label names). Internet hosts are never probed.
4. mDNS `_ssh._tcp` services on Linux (`avahi-browse`, from `avahi-utils`, if installed). Only the advertised IP is used, never the advertised name. Not available on macOS.
5. The ARP/neighbour table (multicast/broadcast excluded)
6. With `--scan` (or when nothing else is found and you say yes): a sweep of the /24 around each of your local addresses
Hosts that answer on ssh but refuse your key are skipped and counted, not touched. Host names are validated and passed to `ssh` after `--`, so a name can't be turned into an ssh option.
## Install behaviour
- Uses plain `ssh`, so your keys and `~/.ssh/config` apply. Nothing needs to be installed beforehand.
- On each Spark it writes `/etc/systemd/system/spark-gpu-clock-limit.service` (a `oneshot` unit, `WantedBy=multi-user.target`), enables it and starts it immediately, which applies `nvidia-smi -lgc 0,2200`.
- `sudo` runs over `ssh -t`, so passwordless sudo just works and a password prompt appears if a host needs one.
- Prints a result per host and a summary; exits non-zero if any host failed.
- Safe to re-run: the unit is overwritten and restarted.
- **Host keys:** discovery probes are read-only, never send a password, don't forward agents or ports, and deliberately skip host-key checking (they use no `known_hosts`) so brand-new Sparks can be found. Applying uses normal ssh checking, so for a Spark you haven't connected to before you are asked to verify its fingerprint, *before* any sudo password is sent. Verify it, or ssh to the Spark once beforehand.
- Before touching a host it checks `nvidia-smi` and applies the clocks, so a host without a working driver is left unchanged.
- The unit retries on failure (`Restart=on-failure`). Note that `systemctl stop`/`restart` of the service briefly resets the