Is this page helpful?

Prerequisites#

Before installing TensorRT, ensure your system meets the following requirements. This page is organized by category to help you quickly find the information you need.

Quick Checklist:

  • NVIDIA GPU (Turing architecture or later)

  • CUDA Toolkit installed (refer to Which CUDA should I install? below)

  • Appropriate GPU drivers: r580+ on Linux and Windows for the CUDA 13.x packages, which is every Debian, RPM, tar, and zip package in this release. The lower r535+ on Linux and r537+ on Windows minimum applies only to CUDA 12.x pip wheels

  • Python 3.10-3.14 supported. Python 3.8 and 3.9 bindings are deprecated, scheduled for removal, and not supported by the samples

On Linux, collect the installed GPU, driver, toolkit, and Python versions with:

nvidia-smi --query-gpu=name,driver_version --format=csv
nvcc --version
python3 --version

nvidia-smi reports the installed driver and GPU. nvcc --version reports the installed CUDA Toolkit; the CUDA version displayed in the standard nvidia-smi header is the maximum version supported by the driver, not necessarily the installed toolkit. On Windows, run nvidia-smi, nvcc --version, and py --version in PowerShell.

Important

Which CUDA should I install for TensorRT 11.4.0?

  • Debian / RPM / tar / zip packages: Install CUDA Toolkit 13.4 update 1 and download the TensorRT package whose filename contains cuda-13.4. This matches the path used by Build Your First Engine.

  • Pip wheels: Use a CUDA 12.x or 13.x driver/toolkit that matches the wheel suffix (tensorrt / tensorrt-cu13 for CUDA 13, or tensorrt-cu12 for CUDA 12). Pip does not include trtexec.

  • Containers: The NGC image pins the CUDA toolkit for you.

Before You Begin#

Review Release Information

Before installation, familiarize yourself with the NVIDIA TensorRT Release Notes to understand:

  • New features in this release

  • Known issues and limitations

  • Platform-specific considerations

  • Compatibility changes

Choose Your API

TensorRT provides both C++ and Python APIs:

  • C++ API - Full functionality, no Python dependency

  • Python API - Convenient for rapid prototyping and integration

  • Both - Most users install both (default)

The installation instructions assume you want both APIs. For C++ only installation, skip Python-specific packages.

Required Software#

NVIDIA CUDA Toolkit

TensorRT requires the NVIDIA CUDA Toolkit. If not already installed, refer to the NVIDIA CUDA Installation Guide.

Note

TensorRT 11.4.0 Debian, RPM, tar, and zip packages are published for CUDA 13.4 update 1. When downloading from the TensorRT download page, select the cuda-13.4 variant to match the package filename. Prefer CUDA Toolkit 13.4 update 1 for those packages; use CUDA 12.x only when you intentionally install matching pip wheels (tensorrt-cu12) or an older toolkit path documented in the support matrix.

Supported CUDA Versions:

Driver Requirements:

The driver minimum follows the CUDA major version you install, not the operating system.

  • CUDA 13.x (every Debian, RPM, tar, and zip package in TensorRT 11.4.0, and the default tensorrt and tensorrt-cu13 pip wheels): NVIDIA driver r580 or later on both Linux and Windows.

  • CUDA 12.x (the tensorrt-cu12 pip wheels only): NVIDIA driver r535 or later on Linux, r537 or later on Windows.

For more information, refer to the TensorRT Support Matrix.

Optional Dependencies#

The following libraries are optional and only needed for specific use cases:

cuBLAS (Optional)

cuBLAS is optional and only used for a few specific layers.

  • When needed: If your model requires cuBLAS-accelerated layers

  • Installation: Refer to the NVIDIA cuBLAS website

CUDA-Python (Optional)

CUDA-Python enables direct CUDA kernel calls from Python.

NCCL (Optional)

Required only for the multi-device inference feature.

For more information, refer to the TensorRT Support Matrix.

Framework and Model Support#

PyTorch Integration

If you plan to use TensorRT with PyTorch:

  • Tested with: PyTorch >= 2.0

  • Compatibility: May work with older versions

  • Use case: Examples and integration samples

ONNX Model Support

The ONNX-TensorRT parser supports:

  • ONNX version: 1.20.0

  • Opset support: Up to opset 25

  • Backward compatibility: Official support is provided for opset 9 and above

For more information, refer to the TensorRT Support Matrix and the ONNX Opset Guide.

TensorRT Installation Modes#

TensorRT offers three runtime package families:

  • Full: tensorrt for pip or the tensorrt and tensorrt-libs package family for Debian/RPM. Use Full to build and run engines.

  • Lean: tensorrt-lean for pip or libnvinfer-lean for Debian/RPM. Use Lean to run compatible version-compatible engines.

  • Dispatch: tensorrt-dispatch for pip or libnvinfer-dispatch for Debian/RPM. Use Dispatch to select a compatible Lean runtime for a version-compatible engine.

Lean and Dispatch have smaller footprints than Full. Exact installed sizes vary by platform, CUDA variant, and package revision; use your package manager’s metadata for byte counts. The Installation Method Comparison summarizes install paths, and Understanding TensorRT Runtime Options describes when to choose each runtime.

For a development-to-production workflow, use Full Installation during development, serialize optimized engines to plan files, then deploy with Lean or Dispatch in production.

Next Steps#

After verifying prerequisites:

  1. Proceed to Installing TensorRT to choose your installation method

  2. Follow the step-by-step installation instructions for your platform

  3. After verification, follow Build Your First Engine to confirm your install with a 10-minute end-to-end build