Installing TensorRT

This guide provides step-by-step instructions for installing TensorRT using various methods. Choose the installation method that best fits your development environment and deployment needs.

Before You Begin: Ensure you have reviewed the Prerequisites to confirm your system meets all requirements.

What Is TensorRT, and What Does Installing It Give You?

NVIDIA TensorRT is an SDK for high-performance deep learning inference (running a trained model on new inputs to produce predictions) on NVIDIA GPUs. It compiles a trained model into a hardware-specific binary called an engine (also referred to as a plan file), then runs that engine inside your application.

Installing TensorRT puts the following on your system:

• The TensorRT libraries (libnvinfer, libnvonnxparser, and friends), which your C++ or Python application links against to load and execute engines.

• Python bindings (the tensorrt Python package), so you can build, save, and run engines from Python without writing C++.

• The trtexec command-line tool (Debian/RPM, tar/zip, and container installs only), which builds an engine from an ONNX file, benchmarks it, and is the fastest way to confirm a fresh non-pip install works. Pip wheels do not include trtexec.

• C++ headers (NvInfer.h and friends), included with every method except pip, for applications that build against the C++ API.

After installation:

• Non-pip methods: run trtexec --help and import tensorrt from Python, then follow Build Your First Engine.

• Pip only: verify with import tensorrt (refer to Method 1: pip). For the ~10-minute CLI first-engine tutorial, install Debian/RPM, tar/zip, or container so trtexec is on your PATH.

If you only need to run engines built by someone else (for example, a deployment server), the Lean or Dispatch runtimes described below are smaller alternatives to the Full Runtime.

Installation Method Comparison

Quick Comparison Table:

Method	Best For	Requires Root	C++ Headers	Multi-Version	Installation Time
pip (Python)	Python development	No	No	Yes (venv)	⚡ Fastest (~2 min)
Debian/RPM	System-wide install	Yes	Yes	No	🔵 Fast (~5 min)
Tar/Zip	Multiple versions	No	Yes	Yes	🟡 Moderate (~10 min)
Container (NGC)	Isolated environments	Docker access; initial setup may require administrator privileges	Yes	Yes	⚡ Fastest (~5 min)

The comparison groups Debian with RPM and tar with zip because each pair has the same operational tradeoffs. The Installation Methods section provides a separate page for every platform-specific method.

Choosing Your Installation Method

Use the Installation Method Comparison table above to pick pip, Debian/RPM, Tar/Zip, or Container. Follow the matching page under Installation Methods for step-by-step instructions.

Understanding TensorRT Runtime Options

TensorRT ships Full, Lean, and Dispatch runtime packages with different capabilities and footprints. For package names and a development-to-production workflow, refer to TensorRT Installation Modes in the prerequisites.

Downloading TensorRT

Before installing with Debian (local repo), RPM (local repo), Tar, or Zip methods, you must download TensorRT packages.

Tip

For pip installation: Skip this section. The pip method downloads packages automatically from PyPI.

1. Go to https://developer.nvidia.com/tensorrt/download/11x.

2. Accept the license agreement.

3. Select TensorRT version 11.4.0 (or your target version).

4. For TensorRT 11.4.0, select the CUDA 13.4 update 1 package variant. Its filename contains cuda-13.4.

5. Download the package for your platform:

   • Linux x86-64: Debian local repo (.deb), RPM local repo (.rpm), or Tar (.tar.zst)

   • Linux ARM SBSA: Debian local repo (.deb) or Tar (.tar.zst)

   • Windows x64: Zip (.zip)

   • Windows Arm64: Zip (.zip)

Installation Methods

• Method 1: Python Package Index (pip)
• Method 2: Debian Package Installation
• Method 3: RPM Package Installation
• Method 4: Tar File Installation
• Method 5: Zip File Installation (Windows)
• Method 6: Container Installation
• Alternative Installation Methods

Verifying Your Installation

Use this checklist first. Each install-method page adds package-manager or path-specific steps.

Table 8 Installation verification checklist

Install method	Quick check	Expected result
Debian / RPM	Locate trtexec and add its directory to PATH by using the verification command on the matching package-manager page, then run trtexec --help	Prints trtexec usage (CLI present)
Tar / zip / container	trtexec --help	Prints trtexec usage (CLI present)
Debian / RPM / tar / zip / container	python3 -c "import tensorrt as trt; print(trt.__version__)"	Prints the installed TensorRT version
Pip	python3 -c "import tensorrt as trt; print(trt.__version__); assert trt.Builder(trt.Logger())"	Prints the version; Builder constructs without error
Pip	trtexec --help	Not expected to work (no CLI in pip wheels)

Debian / RPM / tar / zip / container (includes trtexec):

trtexec --help

If a Debian or RPM shell reports trtexec: command not found, use the trtexec_path command in Debian Installation or RPM Installation to locate the installed executable and add its directory to PATH. Tar and zip installs require the explicit PATH step on their method pages.

import tensorrt as trt
print(trt.__version__)

If both succeed, follow Build Your First Engine to build and run a test engine in about 10 minutes.

Pip (Python bindings and libraries only; no trtexec):

import tensorrt as trt
print(trt.__version__)
assert trt.Builder(trt.Logger())

Do not expect trtexec --help to work after a pip-only install. For the CLI first-engine tutorial, install a non-pip method above, or stay on the Python API path documented in the Quick Start Guide and samples.

Each method page adds method-specific verification (package queries, LD_LIBRARY_PATH / PATH, or Lean/Dispatch imports). If verification fails, refer to Troubleshooting.