Constella TUI

constella-gpu-tui installs the standalone keyboard-first Textual client. It connects to a centrally deployed Constella service and reads the realtime /ws/cluster stream and historical analytics APIs, so live monitoring and completed workloads stay available in one terminal workspace.

Constella natively supports heterogeneous accelerator clusters. Version 0.1.3rc1 supports NVIDIA GPUs and Ascend NPUs.

Views

• Overview — compact node/GPU navigation, selected-GPU processes, telemetry, and a live Braille dot curve.
• Cluster — complete node inventory and the selected node's hardware details.
• Rankings — user GPU-hours, jobs, and anomaly rankings from persisted analytics.
• History — all GPUs on the selected node overlaid in shared utilization and memory dot charts, plus a GPU-by-time utilization heatmap.
• Performance — compact high-resolution dot charts for SM activity, occupancy, Tensor Cores, DRAM bandwidth, and non-Tensor FP16/FP32/FP64 pipelines.

Rankings and History require the manager database. They show an explicit unavailable state when the manager was started with --no-db.

Run

Install the standalone TUI distribution:

uv tool install "constella-gpu-tui==0.1.3rc1"
constella-tui

Connect to a remote manager through an SSH tunnel or a directly reachable endpoint:

constella-tui --url https://gpu.example.com

CONSTELLA_URL can provide the default manager URL.

Keyboard

The selected node and GPU survive realtime refreshes and remain shared across views. The client reconnects automatically after a dropped connection. No agent token is needed because these read-only endpoints follow the same access boundary as the Web UI.

History assigns the first eight GPUs stable, high-contrast colors spaced across the full hue range. Press g to bring the next GPU to the foreground without hiding the other GPU curves.

Performance requires the manager's nvidia.gpm.v1 profile and high-resolution performance cache. Unsupported accelerators continue to work in the other TUI views and show an explicit unavailable state in Performance. The first metric page contains four Compute and Memory charts; the second contains three non-Tensor pipeline charts.

The TUI package deliberately does not install the backend, FastAPI, Uvicorn, or the server-side constella command. Use constella-gpu when both server and client components are wanted on one machine.