nvitop-cluster

Multi-node GPU monitor for training clusters: read MPI/OpenMPI hostfile, query each host (local NVML / remote SSH), show nvitop-style UTIL/MEM bars, temperatures, power, and resolved process command lines.

Designed for multi-node jobs (e.g. DeepSpeed / torchrun / MPI) where plain nvitop only sees the current machine.

Screenshot

Sample from a 2-node × 8-GPU job (/etc/mpi/hostfile). Bars = GPU util (left) / memory (right). Second line per GPU is the process command (Ctrl-A / Ctrl-E switch head vs tail when truncated).

Features

• Hostfile discovery: /etc/mpi/hostfile, /etc/mpi/mpi-hostfile, or -f PATH
• Local + remote: local probe without SSH; peers via ssh (optional KML-style ssh_config)
• Wide dual bars: GPU-Util and Memory-Usage (width scales with terminal)
• Process names: optional host-PID → container-PID remap for Kubernetes/container jobs
• Watch mode by default (2s); keys:
  • Ctrl-A / a — CMD head (like nvitop)
  • Ctrl-E / e — CMD tail
  • v — verbose full cmdline
  • q — quit

Install

Zero third-party Python deps (stdlib only). Needs Python ≥ 3.9, nvidia-smi on each node, and SSH to peers for multi-node.

pip / uv (recommended)

# one-liner from GitHub (works now, no PyPI account needed on your side)
pip install "git+https://github.com/Micuks/nvitop-cluster.git"
# or
uv pip install "git+https://github.com/Micuks/nvitop-cluster.git"

# one-shot (no persistent install)
uvx --from "git+https://github.com/Micuks/nvitop-cluster.git" nvitop-cluster

# PyPI (after a release is published)
pip install nvitop-cluster

Then:

nvitop-cluster
# or
python -m nvitop_cluster

Multi-node tip: either install the package in the same image/env on every rank, or put the repo on a shared path and set NVITOP_CLUSTER_HOME so remote SSH probes can find the script.

From source (no install)

git clone https://github.com/Micuks/nvitop-cluster.git
cd nvitop-cluster
python3 -m nvitop_cluster
# or
./nvitop-cluster

Optional (local interactive nvitop + container PID fix):

pip install 'nvitop-cluster[nvitop]'
# then: nvitop_cluster/nvitop --cluster

Usage

# multi-node table (default: watch 2s)
nvitop-cluster

# once
nvitop-cluster -1

# custom hostfile / SSH config
nvitop-cluster -f /path/to/hostfile
export NVITOP_CLUSTER_SSH_CONFIG=~/.ssh/config

# via nvitop-compatible wrapper (optional extra)
python3 -m nvitop_cluster  # same as nvitop-cluster

Hostfile format

10.0.0.1 slots=8
10.0.0.2 slots=8

(Also accepts bare hostnames / host:slots.)

Environment

Container / Kubernetes note

On many GPU pods, NVML reports host PIDs while /proc only has namespace PIDs, so process names become N/A. nvitop_pid_map.py remaps host→local PIDs (heuristic: rank / device / RSS). Works on each node when share storage holds the same scripts (or install the package everywhere).

Layout (vs plain nvitop)

License

Apache-2.0 (see LICENSE).

Related

• nvitop — single-node interactive viewer
• all-smi — multi-host accelerator TUI
• gpustat-web — multi-node web dashboard