Install the probe - Warpscale

The probe, warpscaled, is a background daemon that collects GPU, kernel, and CUDA telemetry from each node and sends it to Warpscale. Install it first — the CLI and the vLLM plugin report through it.

Where it runs


Linux host	A machine with GPUs that you manage yourself, bare metal or a VM. Runs as a binary on the host.
Kubernetes	A GPU node pool. Runs as a DaemonSet. Coming soon
Docker	A single Docker host. Runs as a container. Coming soon

Configuration

The settings are the same in every mode — only how you supply them differs. Each one can be given three ways, in order of precedence:

A flag — --api-address
A WS_-prefixed environment variable — WS_API_ADDRESS
A key in a YAML file passed with --config — api-address

Flag	Default
`--api-address`	—	Your organization’s host and port. Required.
`--api-token`	—	An organization API key. Required.
`--no-tls`	`false`	Disable TLS to the Warpscale API. Local development only.
`--config`	—	Path to a YAML config file.
`--pidfile`	`/var/run/warpscale/warpscaled.pid`	Where the pid is written.
`--user-event-socket`	`/var/run/warpscale/warpscaled.sock`	Socket the CLI, the Python SDK, and the vLLM plugin report to.
`--node-name`	hostname	Node name attached to telemetry. On Kubernetes, the downward-API `spec.nodeName`.
`--logfile`	stderr	Log to a file instead of stderr.
`--debug`	`false`	Enable debug logging.
`--bpf-loglevel`	`1`	eBPF log level: 1 debug, 2 info, 3 warn, 4 error.
`--pprof-addr`	—	Bind a debug server with pprof and `/metrics`, for example `127.0.0.1:6060`.
`--vllm-metrics-url`	—	vLLM’s Prometheus endpoint. Enables the metrics collector. See Set up vLLM.
`--vllm-poll-interval`	`10s`	Scrape cadence for the vLLM collector.
`--enable-otlp-receiver`	`false`	Receive per-request traces from the inference engine. gRPC only.
`--otlp-receiver-endpoint`	`127.0.0.1:4317`	Where that receiver listens. Widen it when the engine cannot reach loopback.
`--version`		Print the version and exit.

Run warpscaled --help for the authoritative list.