lucifer1004/veloq

VeloQ is a command-line tool written in Rust that lets you query GPU performance profiles from a terminal rather than opening a graphical interface. You run one command, get back JSON, and move on. The tool is built with coding agents and shell scripts in mind, so every response follows a stable, versioned format that programs can parse reliably. It reads profiling data from three sources. The first is NVIDIA Nsight Systems, which captures a timeline of GPU activity during a program run. The second is NVIDIA Nsight Compute, which collects detailed metrics about individual GPU kernels (the small programs that execute on the GPU). The third, still experimental, is PyTorch's built-in profiler output, which uses a Chrome trace format. Across all three, VeloQ exposes around 37 commands covering summaries, kernel searches, performance counters, memory copies, and more. The problem it solves is that existing options are either a GUI you click through, raw text output that varies by tool version, or SQLite queries you write yourself. VeloQ gives you a single binary with a predictable JSON envelope: every list response uses the same row format with stable keys, errors come back through the same structure with a non-zero exit code, and results are easy to diff across separate captures. Installing on Linux or macOS takes one curl command, which drops a binary in your local bin directory and also installs a set of skills for Claude Code (Anthropic's AI coding assistant), allowing AI agents to call VeloQ as part of an automated profiling workflow. Windows users can install via cargo binstall or download a release binary directly. The tool is not meant for interactive, visual exploration of a trace. The Nsight GUI handles that use case better. VeloQ is for scripted, repeatable queries where you need structured output on demand.