---
id: perf-analyzer
version: "2.60.0"
license: unclear
license_treatment: unclear
maintenance: active
---

# perf-analyzer — Triton Performance Analyzer

License: unclear · Maintenance: active · Downloads: 434.4K/mo

## What it is and what it does

Perf Analyzer is a command-line benchmarking tool for Triton Inference Server that measures model inference performance under realistic load. It supports multiple load modes (concurrency, request rate, custom intervals) and measurement strategies (time windows, count windows) to help developers identify performance bottlenecks and validate optimization changes. The tool works with standard models, sequence models, ensemble models, and decoupled models, and can auto-generate or accept custom input data for testing.

You run it against a live Triton server to collect latency, throughput, and other performance metrics as you adjust model configurations or server settings. It's designed for iterative optimization workflows where you need to measure the impact of each change before moving to the next experiment.

Use it for:

- Benchmark latency and throughput of a model before and after applying optimization techniques.
- Profile ensemble or sequence models to identify which component is the performance bottleneck.
- Validate that model changes (quantization, batching, etc.) actually improve end-to-end inference speed.
- Load-test a Triton deployment to find the maximum sustainable request rate or concurrency.
- Compare inference performance across different hardware or Triton configurations.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Perf Analyzer is a CLI tool that measures and optimizes inference performance on Triton Inference Server by testing different optimization strategies under controlled load conditions.

Yes, if you are actively optimizing models on Triton Inference Server and need a structured way to measure performance changes. The tool is actively maintained, has no runtime dependencies, and directly addresses the workflow of iterative model tuning. However, verify the unclear license terms first, and note that it requires an external Triton server to be useful—it is not a standalone profiler.

## Install

pip install perf-analyzer
uv add perf-analyzer
poetry add perf-analyzer

## Installing perf-analyzer

Before you install: Medium install friction due to platform-specific wheels (manylinux_2_38 for aarch64 and x86_64). Active maintenance with recent release (108 days ago), though no runtime dependencies simplifies deployment once installed.

License in practice: License treatment is unclear—no SPDX identifier or raw license text provided in metadata. Verify licensing terms before use in proprietary or commercial contexts.

Quickstart:

pip install perf-analyzer
perf_analyzer -m simple

Requires a running Triton Inference Server instance and a model deployed in its repository; the tool is a client that connects to an external server, not a standalone profiler.

Verify before relying:

- Python version requirements are unspecified; verify compatibility with your environment.
- Whether genai-perf deprecation notice affects current perf-analyzer maintenance or feature roadmap.
- Exact license terms and any restrictions on commercial use or redistribution.

## Package facts

- License: not declared (unclear)
- Python support: unspecified
- Install friction: medium
- Maintenance: active
- Downloads: 434.4K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

triton inference server benchmarking, model performance profiling tool, inference load testing cli, optimize model inference latency, triton server performance measurement, inference-benchmarking, triton-ecosystem, performance-profiling

[View on SkillFed](https://skillfed.io/packages/perf-analyzer) · [View on PyPI](https://pypi.org/project/perf-analyzer/)