System Requirements

TensorRT-Model-Connect requires a compatible NVIDIA GPU software stack. Choose one installation path:

Path	Current boundary
Release wheel	Linux aarch64, Python 3.10 or 3.12, glibc 2.39 or newer, and official TensorRT 11.1.0.106.
Source build	Linux x86_64 or aarch64 with Docker, NVIDIA Container Toolkit, and enough disk space for the image and bundle.

x86_64 release wheels are not published yet. x86_64 users should use the source-build path.

Check the host

All users:

uname -m
nvidia-smi

Source users also need Docker:

docker --version

Wheel users following the first-time Python 3.12 path instead check:

getconf GNU_LIBC_VERSION
python3.12 --version

Confirm that:

• the host architecture matches the selected path;
• the NVIDIA driver sees the target GPU;
• Docker and NVIDIA Container Toolkit are available for a source build; and
• the wheel's Python and glibc requirements are met for a wheel install.

Do not mix binaries, bundles, or TensorRT libraries from different TensorRT cohorts.

Source users should continue to Build from Source, where the target GPU and compute capability are selected once. Wheel users should continue to Installation.

Common boundaries

Symptom	Check first
nvidia-smi fails	Host driver or GPU access.
Docker cannot see the GPU	NVIDIA Container Toolkit configuration.
Wheel is incompatible	Architecture, Python, glibc, and TensorRT cohort.
Hugging Face returns 401/403/not found	Model ID, network, authentication, and gated access.
CMake cannot find CUDA or TensorRT	Use the repository source container.
TensorRT reports an ABI mismatch	Build and run in a compatible TensorRT cohort.