System telemetry
Strong recommendation: collect system telemetry with the Metrum AI all-smi fork. It exposes supported GPU, NPU and APU power, utilization, memory and deep GPM metrics for Bench to scrape alongside the serving engine. Bench timestamps telemetry and requests on the same run clock, so their intervals can be joined. Metric availability depends on the device and exporter; a scrape is a sample, not a measurement of every request.
Install and verify
Download the appropriate binary and checksum from GitHub Releases.
The latest recorded release is v0.28.1-metrum.3. Verify the SHA-256 before
installing and record the tag with the benchmark. For example, on Linux x86_64:
TAG=v0.28.1-metrum.3
ASSET=all-smi-linux-x86_64.tar.gz
BASE="https://github.com/metrum-ai/all-smi/releases/download/${TAG}"
curl -fsSLO "${BASE}/${ASSET}"
curl -fsSLO "${BASE}/${ASSET}.sha256"
sha256sum -c "${ASSET}.sha256"
tar -xzf "${ASSET}"
# Keep the glibc Linux plugins beside the binary for AMD and DCGM support.
sudo install -m 0755 all-smi liball_smi_amd.so liball_smi_dcgm.so /usr/local/bin/
Use a different asset for other platforms; the musl build omits the AMD and
DCGM plugins. DCGM enrichment requires an external DCGM hostengine; without it,
DCGM fields are omitted while supported NVML/GPM collection continues. Verify the actual listening address,
protect the port as appropriate, and confirm the /metrics endpoint and
available series before a run.
Collect with Bench
Start the API:
all-smi api --port 9090
Add these options to your LLM, VLM, ASR, image-generation or strategic run:
--ndjson run.ndjson --telemetry examples/telemetry/all-smi-vllm.yaml
The preset is included in Bench release archives under examples/telemetry/;
it also includes vLLM metrics, so configure both endpoints for your setup.
Adjust its include patterns to the live exporter series. The startup probe
reports matched series; add --require-telemetry to fail on sustained scrape
errors. Do not infer absent metrics as zero. For sampling, exporter freshness,
privacy and analysis caveats, see the repository's telemetry guide.