Installation#

TensorRT Edge-LLM has two separate components that need to be installed on different systems:

  1. Experimental Quantization and llm_loader (runs on x86 host with GPU)

  2. C++ Runtime (Jetson Thor, NVIDIA DRIVE / DriveOS, or optional x86 developer build)

Part 1: Experimental Quantization and llm_loader (x86 Host with GPU)#

The experimental quantization package and llm_loader convert and quantize models. This must run on an x86 Linux system with an NVIDIA GPU.

System Requirements#

  • Platform: x86-64 Linux system

  • Recommended OS: Ubuntu 22.04, 24.04

  • GPU: NVIDIA GPU with Compute Capability 8.0+ (Ampere or newer)

  • CUDA: 12.x or 13.x

  • Python: 3.10+

Memory Requirements#

GPU Memory (VRAM):

  • General rule: ~2-3x model size for most operations, ~5-6x model size for FP8 ONNX export

  • Small models (0.6B-3B): 8-16GB

  • Large models (7B-8B): 20-48GB

  • Very large models (13B+): 48GB+

CPU Memory (RAM):

  • General rule: ~2-3x model size for most operations, ~18-20x model size for FP8 ONNX export

  • Small models (0.6B-3B): 8-16GB (48GB+ for FP8 ONNX export)

  • Large models (7B-8B): 20-48GB (128GB+ for FP8 ONNX export)

  • Very large models (13B+): 48GB+

Note: FP8 ONNX export currently requires significantly higher CPU (up to 20x model size) and GPU (up to 6x model size) memory due to internal processing. This is a known issue and is being actively optimized.

Verify Your Prerequisites:

# Check CUDA installation
nvcc --version
# Should show CUDA 12.x or 13.x

# Check GPU and available memory
nvidia-smi
# Look for GPU memory (e.g., "24576MiB" for 24GB)

# Check Python version
python3 --version
# Should show Python 3.10 or higher

If CUDA is not installed:

Download and install CUDA Toolkit from NVIDIA CUDA Downloads. Choose version 12.x or 13.x for your system.

After installation, verify with nvcc --version and nvidia-smi.

Installing#

For a containerized environment for clean installation, it is recommended to use the NVIDIA PyTorch Docker image:

# Pull the recommended Docker image
docker pull nvcr.io/nvidia/pytorch:25.12-py3

# Run the container with GPU support
docker run --gpus all -it --rm \
    -v $(pwd):/workspace \
    -w /workspace \
    nvcr.io/nvidia/pytorch:25.12-py3 \
    bash

1. Clone Repository

git clone https://github.com/NVIDIA/TensorRT-Edge-LLM.git
cd TensorRT-Edge-LLM
git submodule update --init --recursive

2. Install Python Dependencies

If you are not using container, it is recommended to use a virtual environment:

# Create virtual environment (recommended)
python3 -m venv venv
source venv/bin/activate

Install the base and checkpoint-based loader dependencies. Add the standalone quantization dependencies only when you run experimental.quantization. Do not install the repository as a pip package for this workflow; the examples run directly from the checkout through PYTHONPATH.

# Required for llm_loader export, high-level API, and server workflows
pip3 install -r requirements.txt
pip3 install -r experimental/llm_loader/requirements.txt

# Required only when running experimental.quantization
pip3 install -r experimental/quantization/requirements.txt

This installs all required Python dependencies including:

  • PyTorch

  • Transformers

  • NVIDIA Model Optimizer

  • ONNX

  • ONNX Script and ONNX GraphSurgeon

  • And all other required dependencies

Note: For specific version requirements, please refer to requirements.txt, experimental/llm_loader/requirements.txt, experimental/quantization/requirements.txt, and pyproject.toml.

3. Configure and Verify the Experimental Export Workflow

Use the virtual environment created in Step 2 for this checkout. Do not mix packages from older release branches into the same environment.

The recommended export path is experimental.quantization -> llm_loader. Use the quantization package only when you need to create a unified quantized checkpoint from an FP16/BF16 source checkpoint before export. Pre-quantized HuggingFace checkpoints can be exported directly with llm_loader.

export EDGE_LLM_PATH=/path/to/TensorRT-Edge-LLM
export PYTHONPATH=$EDGE_LLM_PATH:$EDGE_LLM_PATH/experimental:$PYTHONPATH

# Verify the recommended quantization, export, LoRA, and vocabulary tools
python -m experimental.quantization --help
python -m llm_loader.export_all_cli --help
python -m llm_loader.lora.merge_lora_cli --help
python -m llm_loader.vocab_reduction --help

The deprecated tensorrt_edgellm/ export package remains available in 0.7.1 for compatibility and for components that are not yet fully covered by llm_loader. It will be removed in 0.8.0 after the experimental/quantization -> experimental/llm_loader workflow reaches full feature parity for all models and features.

4. Configure HuggingFace Access (Optional)

Some models on HuggingFace require you to accept terms before downloading. This is not required for the quick start example (Qwen3-0.6B).

Models that require HuggingFace login:

  • Llama family (Llama 3.x)

  • Phi-4-Multimodal

  • Alpamayo-R1-10B

  • Other models marked as “gated” on HuggingFace

To configure access:

# Install HuggingFace CLI and login
hf auth login
# Enter your HuggingFace access token when prompted

How to get a token: Visit HuggingFace Settings - Tokens, create a new token (read access is sufficient), and copy it.

For the quick start guide: You can skip this step and proceed to verification.

You’re done with export pipeline setup! You can now quantize and export models with the experimental workflow. The ONNX files will be transferred to the Edge device for runtime deployment.

Part 2: C++ Runtime (Edge Device)#

The C++ runtime builds and executes models on the target. Jetson Thor: follow the steps below on the device and use EMBEDDED_TARGET=jetson-thor. NVIDIA DRIVE / DriveOS: run the same flow inside the DriveOS SDK Docker image with EMBEDDED_TARGET=auto-thor, then copy build/ to the DRIVE system. x86: optional local build using the Alternative cmake block (no toolchain).

System Requirements#

Target Platform:

  • NVIDIA Jetson Thor

  • JetPack 7.1

  • CUDA 13.x (included in JetPack)

  • TensorRT 10.x+ (included in JetPack)

  • Disk Space: ~20-50GB for ONNX files and TensorRT engines

Build Instructions#

1. Install System Dependencies (on Edge device)

sudo apt update
sudo apt install -y \
    cmake \
    build-essential \
    git

2. Verify CUDA and TensorRT Installation

After JetPack is installed, TensorRT should be installed in /usr

# Check CUDA version
nvcc --version  # Should show CUDA 13.x

# Check TensorRT version
dpkg -l | grep tensorrt  # Should show TensorRT 10.x+

3. Clone Repository (on Edge device)

# Clone to home directory (used in all examples)
cd ~
git clone https://github.com/NVIDIA/TensorRT-Edge-LLM.git
cd TensorRT-Edge-LLM
git submodule update --init --recursive

4. Configure Build

On your Jetson Thor device, configure the build with the following command:

mkdir build
cd build

cmake .. \
    -DCMAKE_BUILD_TYPE=Release \
    -DTRT_PACKAGE_DIR=/usr \
    -DCMAKE_TOOLCHAIN_FILE=cmake/aarch64_linux_toolchain.cmake \
    -DEMBEDDED_TARGET=jetson-thor

NVIDIA DRIVE / DriveOS: The cmake line is the same except EMBEDDED_TARGET=auto-thor.

Alternative: Building on x86 GPU Systems (Optional for Developers)

If you want to build and test on an x86 workstation with NVIDIA GPU (for development purposes before deploying to Edge devices), you can use this configuration instead:

mkdir build
cd build

cmake .. \
    -DCMAKE_BUILD_TYPE=Release \
    -DTRT_PACKAGE_DIR=/usr/local/TensorRT-10.x.x \
    -DCUDA_CTK_VERSION=<YOUR_CUDA_VERSION>

Note: Replace /usr/local/TensorRT-10.x.x with your actual TensorRT installation path. Use dpkg -l | grep tensorrt to find it, or download from NVIDIA TensorRT downloads. Replace <YOUR_CUDA_VERSION> with your actual CUDA version (e.g., 13.0). Use nvcc --version to check your CUDA version.

CMake Options:

Option

Description

Default

TRT_PACKAGE_DIR

Path to TensorRT installation. Auto-detected; manual hint to disambiguate multiple versions.

N/A

CMAKE_TOOLCHAIN_FILE

Required for Edge devices: Use cmake/aarch64_linux_toolchain.cmake for Edge device builds. Not needed for GPU builds

N/A

EMBEDDED_TARGET

Required for Edge devices: jetson-thor (Jetson) or auto-thor (DRIVE / DriveOS). Not needed for GPU builds

N/A

CUDA_CTK_VERSION

CUDA Toolkit version (such as 13.0). Important for matching target platform.

13.0

BUILD_UNIT_TESTS

Build unit tests

OFF

ENABLE_COVERAGE

Enable gcov code coverage instrumentation (see Code Coverage)

OFF

ENABLE_CUTE_DSL

Enable prebuilt CuTe DSL kernels: OFF (default), ALL, or a group list such as gdn, fmha, gemm, or ssd

OFF

CUTE_DSL_ARTIFACT_TAG

Optional artifact tag under cpp/kernels/cuteDSLArtifact/<arch>/, for example sm_110 or sm_121. Required when multiple local artifact tags exist for the same CPU architecture.

auto

Building with CuTe DSL Kernels (Optional)

CuTe DSL binaries are prebuilt and shipped with the repository. Add -DENABLE_CUTE_DSL=ALL (or a group selection such as gdn, fmha, gemm, or ssd) to the CMake configure command when a model or kernel path needs them. Qwen3.5 GDN requires -DENABLE_CUTE_DSL=gdn or -DENABLE_CUTE_DSL=ALL.

If you have multiple local artifact tags for the same CPU architecture, also pass -DCUTE_DSL_ARTIFACT_TAG=<tag>.

For supported model families, precisions, and hardware notes, see Supported Models.

5. Build Project

make -j$(nproc)

Build time: ~1-2 minutes depending on hardware.

6. Verify Build

# Test C++ examples
./examples/llm/llm_build --help
./examples/llm/llm_inference --help

You’re done with C++ runtime setup! You can now build engines and run inference on the Edge device.

Next Steps#

After installation, proceed to the Quick Start Guide for a complete end-to-end workflow, or see the Examples for detailed pipeline stages and advanced use cases.

Troubleshooting#

Common Installation Issues#

Issue: Python module import errors

Solution: Ensure the virtual environment is activated and PYTHONPATH points to both the repository root and experimental/:

source venv/bin/activate
export EDGE_LLM_PATH=/path/to/TensorRT-Edge-LLM
export PYTHONPATH=$EDGE_LLM_PATH:$EDGE_LLM_PATH/experimental:$PYTHONPATH
python -m llm_loader.export_all_cli --help

Issue: nvcc: command not found

Solution: Ensure JetPack 7.1 is properly installed with CUDA support:

# Verify CUDA installation
nvcc --version
# Should show CUDA 13.x

Issue: TensorRT not found during CMake

Solution: Specify TensorRT package directory. This directory should contain lib and include directories, and we are looking for the nvinfer library and header:

cmake .. \
    -DTRT_PACKAGE_DIR=/usr/local/TensorRT-10.x.x \
    -DCMAKE_TOOLCHAIN_FILE=cmake/aarch64_linux_toolchain.cmake \
    -DEMBEDDED_TARGET=jetson-thor

Issue: Thread issue during C++ build

Solution: Reduce parallel jobs or even use sequential build:

make -j  # Instead of make -j$(nproc)

Getting Help#

  • Documentation: Check the docs/source/developer_guide directory

  • Issues: Report bugs on GitHub Issues

  • Discussions: Ask questions on GitHub Discussions

  • Community: Join the NVIDIA Developer Forums

Uninstalling#

Experimental Quantization and llm_loader (x86 Host):

  • Deactivate and remove virtual environment: deactivate && rm -rf venv

  • Remove repository (optional): rm -rf TensorRT-Edge-LLM

C++ Runtime (Edge Device):

  • Remove build directory: rm -rf build

  • Remove repository (optional): rm -rf TensorRT-Edge-LLM