Install the probe on a Linux host

Download one archive and run the probe on the host. Use this on a machine with GPUs that you manage yourself — bare metal or a VM.

Before you start

Beyond the requirements, this mode needs:

Root access on the host. The probe loads eBPF programs and reads GPU state.
/sys/fs/bpf and /sys/kernel/tracing mounted.

Install the probe

Download the release

The probe and the CLI ship in one archive.

VERSION=v0.0.4
ARCH=$(uname -m | sed 's/x86_64/amd64/; s/aarch64/arm64/')
BASE=https://dl.warpscale.ai/stable/${VERSION}

curl --proto '=https' --tlsv1.2 -fLO ${BASE}/warpscale_${VERSION}_linux_${ARCH}.tar.gz
curl --proto '=https' --tlsv1.2 -fLO ${BASE}/warpscale_${VERSION}_linux_${ARCH}.tar.gz.sha256

Check the download

sha256sum -c warpscale_${VERSION}_linux_${ARCH}.tar.gz.sha256

Expect OK. This confirms the archive arrived intact — it catches a corrupt or truncated transfer, not a tampered one.

Extract and install

tar -xzf warpscale_${VERSION}_linux_${ARCH}.tar.gz
sudo install -m 0755 warpscaled /usr/local/bin/

Configure and start it

The probe requires two settings: your API address and your API key. See the Quickstart.

Run it directly

sudo warpscaled \
  --api-address <your-organization-host>:443 \
  --api-token <your-api-key>

The probe logs one JSON object per line. Expect a startup line naming the version and pid file, then one line per GPU it finds:

{"level":"info","ts":1786353441.8652174,"caller":"warpscaled/main.go:43","msg":"starting warpscaled","version":"v0.0.4","revision":"4345178af9662f256c3df2bae6c461f2728233f2","pidfile":"/var/run/warpscale/warpscaled.pid"}
{"level":"info","ts":1786353442.1411493,"logger":"nvml","caller":"nvml/nvml.go:171","msg":"discovered GPU","agent_instance_id":"9bf0171c-f381-4ed3-9f6a-7419313fdb0b","index":0,"name":"NVIDIA L4","uuid":"GPU-12e260f8-8a17-2297-1d41-ea2f8fe53e7a","compute_capability":"8.9","gpm":false}

Press Ctrl+C to stop.

Verify it’s working

The socket should exist — that is what the CLI and the vLLM plugin connect to.

ls /var/run/warpscale/warpscaled.sock

Upgrade

Stop the probe, repeat the install steps with the new version, and start it again. If you have also installed the CLI, replace both from the same release — do not upgrade one on its own. Let any in-flight runs finish first — stopping the probe mid-run loses that run’s telemetry.

Uninstall

Stop the probe, then remove the binary and its state:

sudo rm -f /usr/local/bin/warpscaled
sudo rm -rf /var/run/warpscale
sudo rm -f /dev/shm/ws-cg-*

Troubleshooting

warpscaled: version 'GLIBC_2.32' not found

The host’s glibc is older than the version the probe is built against. Check with ldd --version. Ubuntu 20.04, Debian 11, and RHEL 9 or newer all satisfy the 2.31 floor.

warpscaled exits immediately with code 2

--api-address and --api-token are both required. The error names the one that is missing.

warpscaled starts but reports no GPUs

The probe cannot reach the NVIDIA driver. Confirm nvidia-smi works as root on the same host, and that the driver is installed rather than only the CUDA toolkit.

Failed to load eBPF programs

Confirm /sys/fs/bpf and /sys/kernel/tracing are mounted and the kernel has BTF enabled (ls /sys/kernel/btf/vmlinux).

Getting started

Install

Training

Inference

Observability

Install the probe on a Linux host

Before you start

Install the probe

Configure and start it

Verify it’s working

Upgrade

Uninstall

Troubleshooting

Next steps

Install the CLI

Install the vLLM plugin

​Before you start

​Install the probe

​Configure and start it

​Verify it’s working

​Upgrade

​Uninstall

​Troubleshooting

​Next steps

Install the CLI

Install the vLLM plugin

Before you start

Install the probe

Configure and start it

Verify it’s working

Upgrade

Uninstall

Troubleshooting

Next steps