constella-gpu 0.1.1

Lightweight realtime heterogeneous accelerator monitoring with workload history.

Constella

constella-gpu is the PyPI distribution for Constella, a lightweight Web-based accelerator monitoring and workload history platform for standalone servers and small clusters.

Constella natively supports heterogeneous accelerator clusters across multiple hardware backends. It shows current accelerator activity in realtime, records workload history automatically, and makes completed training and inference jobs easy to review without requiring a Prometheus/Grafana stack.

The current release supports:

• NVIDIA GPUs: NVML with nvidia-smi fallback.
• Ascend NPUs: DCMI with npu-smi fallback.

The wheel includes the production Web interface and the constella command, so an installed package does not require a source checkout or a separate frontend build.

Install

Constella requires Python 3.10 or newer.

pip install constella-gpu

The distribution name is constella-gpu; the installed command and Python package are both named constella.

Quick start

Start the manager, SQLite history, and a local NVIDIA agent:

constella service start \
  --run-dir ~/.constella/run \
  --log-dir ~/.constella/logs

On an Ascend host, select the backend explicitly:

constella service start \
  --device ascend \
  --run-dir ~/.constella/run \
  --log-dir ~/.constella/logs

Open http://127.0.0.1:8765/overview in a browser.

Inspect or stop the managed processes with the same runtime directory:

constella service status --run-dir ~/.constella/run --log-dir ~/.constella/logs
constella service stop --run-dir ~/.constella/run --log-dir ~/.constella/logs

Constella binds to 127.0.0.1 by default. Keep that default for local access, or use SSH port forwarding when the service runs on a remote host.

Common commands

Print one accelerator snapshot:

constella probe --pretty
constella probe --device ascend --pretty

Run only the manager for a custom supervisor or a manager-only node:

constella serve --host 127.0.0.1 --port 8765
constella service start --no-local-agent

Manage agents defined in a cluster manifest:

constella cluster start --nodes nodes.yaml
constella cluster status --nodes nodes.yaml
constella cluster stop --nodes nodes.yaml

Optional history and job curves

constella service start enables SQLite history by default. Disable it for a realtime-only deployment:

constella service start --no-db

The high-resolution job-curve sidecar is optional:

constella service start --highres-sidecar

Project links

• Source repository
• Installed CLI reference
• Operations guide
• Cluster manifest example
• Issue tracker
• MIT license