vramwatch does not write to the GPU or control a loader. Live observation shells
out to nvidia-smi/amd-smi, reads OS counters, and calls configured loader APIs.
It writes only prediction-ledger records and report files requested by the user.
Remote fit requests contact the selected Hugging Face or Ollama registry (or an
explicit HTTPS GGUF URL) for metadata. GGUF range reads and JSON responses share a
16 MiB limit, range bodies are closed as soon as architecture parsing completes,
unknown sizes fail closed, and parser counts/depths are bounded.
doctor --online also makes explicit registry probes. There is no telemetry.
HF_TOKEN is used as an authorization header for Hugging Face metadata and is not
printed or placed in SVG output. Local ledger records and raw JSON can retain the
original model reference, so treat the state directory as private diagnostic data.
The shareable SVG path is separately scrubbed of local paths, URL queries, host
identity, PIDs, bus addresses, and serial-number fields.
If you find a security issue (for example command execution through crafted provider output, a GGUF parser resource bypass, ledger path traversal, token disclosure, or private data in an SVG), please report it privately:
- Use GitHub's "Report a vulnerability" (Security → Advisories) on this repo, or
- email the maintainer listed on the GitHub profile.
Please do not open a public issue for security reports. You'll get an acknowledgement within a few days.
Only the latest released version is supported.