Installation

This guide will help you install trtutils and its dependencies. The recommended method is to install trtutils into a virtual environment to ensure dependency isolation.

System Requirements

  • Python 3.8 or later

  • CUDA toolkit

  • TensorRT

  • NVIDIA GPU or Jetson device

Basic Installation

The simplest way to install trtutils is using pip:

$ pip install trtutils

For development or to get the latest features, install from source:

$ git clone https://github.com/justincdavis/trtutils.git
$ cd trtutils
$ pip install -e .

Optional Dependencies

trtutils provides several optional dependency groups that can be installed using pip’s extras feature:

JIT Compiler Support

Install support for the JIT compiler:

$ pip install "trtutils[jit]"

This installs: - Numba - LLVM-Lite

This enables the use of trtutils.enable_jit() to accelerate some CPU operations.

Development Tools

For development or contributing to trtutils:

$ pip install "trtutils[dev]"

This installs: - Testing frameworks - Linting tools - Documentation generators - Development utilities

Troubleshooting

Common Installation Issues

  1. CUDA/TensorRT Not Found - Ensure CUDA and TensorRT are properly installed - Check environment variables (LD_LIBRARY_PATH, etc.) - Verify CUDA version compatibility

  2. Dependency Conflicts - Use a virtual environment - Check package versions - Update pip: pip install --upgrade pip

  3. Jetson-Specific Issues - Install Jetson-specific TensorRT version - Use compatible CUDA version - Check Jetpack installation

  4. libnvrtc.so.* Not Found - Ensure the version of cuda-python installed matches the version of CUDA installed - If using a custom CUDA path, ensure it is correctly set in the environment variables

Getting Started

After installation, verify your setup:

from trtutils import TRTEngine

# Create a test engine
engine = TRTEngine("test.engine")
print("Installation successful!")

For more detailed examples, see the Examples section.