gpu-tools

gpu-tools

A pure-Go, no-cgo GPU infrastructure diagnostics + tuning-advice + benchmark + monitoring CLI — single self-contained binary, portable across glibc distributions. More than a monitor: it detects inventory, renders reports, prints read-only tuning advice, runs benchmarks, watches live, and serves Prometheus metrics.

简体中文 · English

Table of Contents

• Features
• Install
• Quickstart
• Backends
• Exporter
• Requirements
• Using with an LLM or agent
• Configuration
• Development
• License

Features

• Detect — collect a point-in-time NVIDIA GPU inventory with gpu-tools detect.
• Report — render Markdown, table, or JSON snapshots with gpu-tools report.
• Tune — print deterministic, read-only tuning recommendations; it never mutates hardware settings.
• Bench — run supported external benchmark tools and normalize their parsed throughput.
• Watch — refresh live with gpu-tools detect --watch 2s: the screen re-renders each tick until Ctrl-C, or -o json --watch streams NDJSON (one compact object per line).
• Per-process GPU usage — detect and report add a GPU Processes section (GPU, PID, Type, Process, User, Mem) when compute processes are present.
• Richer device metrics — encoder/decoder utilization (Enc/Dec %) and PCIe link (genXxwN) alongside utilization, memory, temperature, power, and clocks.
• Prometheus exporter — gpu-tools export --listen :9835 serves a headless /metrics endpoint for scraping and Grafana dashboards.
• AMD backend (best-effort) — --backend amd reads a subset of metrics via rocm-smi; auto still prefers NVIDIA.

See Architecture for the collector model and package layout.

Install

POSIX one-liner

curl -fsSL https://raw.githubusercontent.com/sunerpy/gpu-tools/main/scripts/install.sh | sh

The installer verifies checksums.txt before extracting the release archive. Override the destination with GPU_TOOLS_INSTALL_DIR or --dir.

PowerShell

irm https://raw.githubusercontent.com/sunerpy/gpu-tools/main/scripts/install.ps1 | iex

The PowerShell installer defaults to $env:LOCALAPPDATA\Programs\gpu-tools; override with $env:GPU_TOOLS_INSTALL_DIR or -Dir.

Go toolchain

go install github.com/sunerpy/gpu-tools@latest

Prebuilt releases

Download Linux, macOS, or Windows archives from GitHub Releases. Release artifacts are built with CGO_ENABLED=0 for amd64 and arm64.

Quickstart

# Detect local NVIDIA GPU inventory.
gpu-tools detect

# Print a point-in-time report to stdout.
gpu-tools report --out -

# Show read-only tuning recommendations.
gpu-tools tune

# Run an external benchmark tool and parse the result.
gpu-tools bench --tool gpu-burn

Live watch, per-process usage, exporter, and AMD:

# Refresh the inventory table every 2 seconds until Ctrl-C.
gpu-tools detect --watch 2s

# Stream one compact NDJSON object per tick (machine-readable watch).
gpu-tools --output json detect --watch 2s

# detect/report show a "GPU Processes" section when compute processes exist:
#
#   GPU Processes
#   GPU  PID   Type     Process  User   Mem
#   0    1234  compute  python   alice  512 MiB

# Serve GPU metrics for Prometheus on http://<host>:9835/metrics.
gpu-tools export --listen :9835

# Read AMD GPUs (best-effort subset) via rocm-smi.
gpu-tools --backend amd detect

Common global flags:

gpu-tools --output json detect
gpu-tools --output markdown tune
gpu-tools --backend nvidia-smi report --out -
gpu-tools --config ./config.yaml config show

Backends

gpu-tools selects collectors through --backend auto|nvml|nvidia-smi|amd:

1. purego NVML (nvml) — primary NVIDIA backend; loads NVML dynamically without cgo.
2. nvidia-smi (nvidia-smi) — NVIDIA fallback backend; shells out to nvidia-smi and parses CSV.
3. rocm-smi (amd) — best-effort AMD backend; parses rocm-smi --json for a subset of metrics (index, name, utilization, memory, temperature, power).
4. DCGM — deferred; not implemented in v1.

auto prefers NVML, then falls back to nvidia-smi (it does not auto-select AMD — pass --backend amd explicitly). If no requested backend is available, commands fail gracefully with no NVIDIA GPU detected and exit code 1. See FAQ for no-GPU, watch, exporter, and benchmark exit behavior.

Exporter

gpu-tools export --listen :9835 runs a headless Prometheus exporter. Scrape /metrics; a bare / returns gpu-tools exporter. The endpoint always answers HTTP 200 — on a host with no GPU it emits gpu_tools_up 0 and no device series, so a Prometheus target never flaps just because a node lacks a GPU.

Exposed metrics:

[!NOTE] All per-GPU/per-process series use the bare gpu_ prefix; only gpu_tools_up carries the gpu_tools namespace.

Grafana: point a Prometheus scrape job at <host>:9835, then chart gpu_utilization_percent and gpu_memory_used_bytes by the index/name labels; use gpu_tools_up for target health and gpu_process_used_memory_bytes for per-process breakdowns.

Requirements

• A recent Go toolchain is required only when building from source.
• Runtime GPU data requires an installed NVIDIA driver and either NVML or nvidia-smi on the host.
• The binary itself is pure Go and built with CGO_ENABLED=0; no C toolchain is needed to build it, and it can start on hosts without NVIDIA GPUs.
• The purego NVML backend loads NVML via the system dynamic loader at runtime, so the binary is not fully static and requires the system loader plus an NVIDIA driver for real GPU data.
• Benchmarks use external tools (gpu-burn, nvbandwidth, or bandwidthTest); some tools may require elevated privileges depending on the environment.

Using with an LLM or agent

Install via the Install section, then drive the CLI with this compact command contract.

Configuration

Configuration lives at ~/.gpu-tools/config.yaml by default and supports:

• default_output: table, json, or markdown
• backend: auto, nvml, nvidia-smi, or amd
• report_dir: default directory for report files
• nvidia_smi_path: optional nvidia-smi binary override

See Configuration for field details, flag overrides, backend selection, and report output rules.

Development

Common contributor commands:

make build
make test
make coverage-gate
make fmt
make lint

Install local hooks with:

pre-commit install --hook-type pre-commit --hook-type pre-push

See Development for the coverage gate, formatting tools, Conventional Commits, and release-please flow.

License

MIT