About the running environment for sionna-rk with Jetson AGX Orin for 5G OAI with L1 on GPU?

[hitpony2]

Hello, Sir

I have been following the guidance from this link (sionna-rk) with 5G OAI with L1 GPU:
https://nvlabs.github.io/sionna/rk/index.html
Platform used is the "Jetson AGX Orin 64G Development Kit", and using SDK Manager to install the JetPack into the Orin and installed kernal with SCTP, built docker images for OAI and also tried to install sionna-rk plugins, but failed ...

My quesions are:
Q1: What is the tested/exact JetPack version for OAI (25w34 with UHD 4.8.0) with run sionna-rk (Latest 1.1.0) to support GPU for L1 PDCP and Nerual Demmapper, should I use JetPack 6.2.1, 6.1, 6.0 (Ubuntu 22.04), or 5.x (Ubuntu 20.04)?
Q2: What the TensorRT version should I use? can I use libnvinfer.so.10.3.0 or must use libnvinfer.so.8.6.2?
Q3: For OAI, after I check the docker files generated after applied the sionna-rk's patch (./patches/openairinterface5g.patch), the docker files seem are for X86+RTX GPU, NOT for Jetson AGX Orin (arm), my quesion is: do we have the docker file for Jetson AGX Orin (with Jetpack / L4T)? or can I build OAI on top of JerPack host directly and do not use OAI docker image? what is the tested version of OAI support sionna-rk 1.1.0 for L1 GPU?

Thanks in advance!

Detailed steps and errors met are given below:
Firstly, I installed the Jetpack 6.2.1 (L4T Release 36.4.7) with SDK Manager:

hitpony@ubuntu:~$ cat /etc/lsb-release
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=22.04
DISTRIB_CODENAME=jammy
DISTRIB_DESCRIPTION="Ubuntu 22.04.5 LTS"
hitpony@ubuntu:~$ cat /etc/nv_tegra_release
# R36 (release), REVISION: 4.7, GCID: 42132812, BOARD: generic, EABI: aarch64, DATE: Thu Sep 18 22:54:44 UTC 2025
# KERNEL_VARIANT: oot
TARGET_USERSPACE_LIB_DIR=nvidia
TARGET_USERSPACE_LIB_DIR_PATH=usr/lib/aarch64-linux-gnu/nvidia
hitpony@ubuntu:~/sionna-rk$ dpkg-query --show nvidia-l4t-core
nvidia-l4t-core	36.4.7-20250918154033

and I can rebuilt the kernal successfully with the latest available version of L4T Release 36.4.4 with following command:

./scripts/build-custom-kernel.sh --verbose --clean --l4t-version 36.4.4
./scripts/install-custom-kernel.sh

Note: "--l4t-version 36.4.4" must be added, or the script will failed as the is no 36.4.7 kernal available to be downloaded

And SCTP module can be see:

hitpony@ubuntu:~$ lsmod | grep -E "sctp"
sctp                  364544  10
ip6_udp_tunnel         20480  1 sctp
udp_tunnel             28672  1 sctp
libcrc32c              16384  4 nf_conntrack,nf_nat,nf_tables,sctp
ipv6                  503808  176 bridge,sctp

After that I can build the 5 docker images successfully follow the links:
https://nvlabs.github.io/sionna/rk/setup/OAI.html#manual-build-steps

hitpony@ubuntu:~/sionna-rk$ docker images
IMAGE                   ID             DISK USAGE   CONTENT SIZE   EXTRA
oai-flexric:latest      7400be001277       1.46GB          359MB        
oai-gnb-cuda:latest     d77b0de317c7       26.4GB         9.96GB        
oai-nr-ue-cuda:latest   0a55cc790ed7       12.5GB         4.36GB        
ran-base-cuda:latest    7a0caa18929d       27.5GB         10.2GB        
ran-build-cuda:latest   f1a8d59b9b58       32.5GB         11.5GB  

But failed at the following steps of "Build Plugins":

hitpony@ubuntu:~/sionna-rk$ ./plugins/common/build_all_plugins.sh --container
Building plugins using build.container section
Searching for plugins in: /home/hitpony/sionna-rk/plugins
Processing plugin: data_acquisition
  Build skipped (not enabled or no command configured)
----------------------------------------
Processing plugin: neural_demapper
  Build enabled. Executing: docker run --runtime=nvidia -v ../../plugins/:/mnt/plugins -v ../../ext/neural_demapper:/mnt/neural_demapper --rm -t ran-build-cuda:latest /mnt/plugins/neural_demapper/scripts/build-trt-plans.sh --plan /mnt/plugins/neural_demapper/models/neural_demapper_qam16_2.plan
  Working directory: /home/hitpony/sionna-rk/plugins/neural_demapper
docker: Error response from daemon: unknown or invalid runtime name: nvidia

I also tried Sionna Installation with following links:
https://nvlabs.github.io/sionna/rk/setup/sionna.html#jetson-agx-orin-orin-nano
and can NOT find any GPU with following commands:

hitpony@ubuntu:~$ python3 - << 'EOF'
> import tensorrt as trt
> print(trt.__version__)
> EOF
10.3.0

hitpony@ubuntu:~$ python3 - << 'EOF'
> import tensorflow as tf
> print(tf.config.list_physical_devices('GPU'))
> EOF
2025-12-27 17:27:23.540475: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT: INTERNAL: Cannot dlopen all TensorRT libraries: FAILED_PRECONDITION: Could not load dynamic library 'libnvinfer.so.8.6.2'; dlerror: libnvinfer.so.8.6.2: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-12.6/lib64:
/home/hitpony/.local/lib/python3.10/site-packages/matplotlib/projections/__init__.py:63: UserWarning: Unable to import Axes3D. This may be due to multiple versions of Matplotlib being installed (e.g. as a system package and as a pip package). As a result, the 3D projection is not available.
  warnings.warn("Unable to import Axes3D. This may be due to multiple versions of "
2025-12-27 17:27:25.283360: I external/local_xla/xla/stream_executor/cuda/cuda_executor.cc:984] could not open file to read NUMA node: /sys/bus/pci/devices/0000:00:00.0/numa_node
Your kernel may have been built without NUMA support.
2025-12-27 17:27:25.325430: W tensorflow/core/common_runtime/gpu/gpu_device.cc:2251] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform.
Skipping registering GPU devices...
[]

The TensorRT version is as below:

hitpony@ubuntu:~$ dpkg -l | grep nvinfer
ii  libnvinfer-bin                               10.3.0.30-1+cuda12.5                        arm64        TensorRT binaries
ii  libnvinfer-dev                               10.3.0.30-1+cuda12.5                        arm64        TensorRT development libraries
ii  libnvinfer-dispatch-dev                      10.3.0.30-1+cuda12.5                        arm64        TensorRT development dispatch runtime libraries
ii  libnvinfer-dispatch10                        10.3.0.30-1+cuda12.5                        arm64        TensorRT dispatch runtime library
ii  libnvinfer-headers-dev                       10.3.0.30-1+cuda12.5                        arm64        TensorRT development headers
ii  libnvinfer-headers-plugin-dev                10.3.0.30-1+cuda12.5                        arm64        TensorRT plugin headers
ii  libnvinfer-lean-dev                          10.3.0.30-1+cuda12.5                        arm64        TensorRT lean runtime libraries
ii  libnvinfer-lean10                            10.3.0.30-1+cuda12.5                        arm64        TensorRT lean runtime library
ii  libnvinfer-plugin-dev                        10.3.0.30-1+cuda12.5                        arm64        TensorRT plugin libraries
ii  libnvinfer-plugin10                          10.3.0.30-1+cuda12.5                        arm64        TensorRT plugin libraries
ii  libnvinfer-samples                           10.3.0.30-1+cuda12.5                        all          TensorRT samples
ii  libnvinfer-vc-plugin-dev                     10.3.0.30-1+cuda12.5                        arm64        TensorRT vc-plugin library
ii  libnvinfer-vc-plugin10                       10.3.0.30-1+cuda12.5                        arm64        TensorRT vc-plugin library
ii  libnvinfer10                                 10.3.0.30-1+cuda12.5                        arm64        TensorRT runtime libraries
ii  python3-libnvinfer                           10.3.0.30-1+cuda12.5                        arm64        Python 3 bindings for TensorRT standard runtime
ii  python3-libnvinfer-dev                       10.3.0.30-1+cuda12.5                        arm64        Python 3 development package for TensorRT standard runtime
ii  python3-libnvinfer-dispatch                  10.3.0.30-1+cuda12.5                        arm64        Python 3 bindings for TensorRT dispatch runtime
ii  python3-libnvinfer-lean                      10.3.0.30-1+cuda12.5                        arm64        Python 3 bindings for TensorRT lean runtime

hitpony@ubuntu:~$ ls -l /usr/lib/aarch64-linux-gnu/libnvinfer.so*
lrwxrwxrwx 1 root root        20 Sep 15  2024 /usr/lib/aarch64-linux-gnu/libnvinfer.so -> libnvinfer.so.10.3.0
lrwxrwxrwx 1 root root        20 Sep 15  2024 /usr/lib/aarch64-linux-gnu/libnvinfer.so.10 -> libnvinfer.so.10.3.0
-rw-r--r-- 1 root root 232442880 Sep 15  2024 /usr/lib/aarch64-linux-gnu/libnvinfer.so.10.3.0

hitpony@ubuntu:~$ cat /etc/ld.so.conf.d/tensorrt.conf 
/usr/lib/aarch64-linux-gnu

Best Regards
Yuchu

[gmarcusm]

Hi Yuchu,

Q1: What is the tested/exact JetPack version for OAI (25w34 with UHD 4.8.0) with run sionna-rk (Latest 1.1.0) to support GPU for L1 PDCP and Nerual Demmapper, should I use JetPack 6.2.1, 6.1, 6.0 (Ubuntu 22.04), or 5.x (Ubuntu 20.04)?

We have validated SRK 1.1.0 against Jetpack 6.0 in AGX Orin, which is L4T 36.3.0 (Ubuntu 22.04). SRK 1.1.0 uses OAI 2025.w34.

Q2: What the TensorRT version should I use? can I use libnvinfer.so.10.3.0 or must use libnvinfer.so.8.6.2?

The version is mostly relevant for tests only. The L4T docker image used in Orin as a base to build the OAI images is version 36.3.0, and that uses 8.6.2.3-1+cuda12.2

Q3: For OAI, after I check the docker files generated after applied the sionna-rk's patch (./patches/openairinterface5g.patch), the docker files seem are for X86+RTX GPU, NOT for Jetson AGX Orin (arm), my quesion is: do we have the docker file for Jetson AGX Orin (with Jetpack / L4T)? or can I build OAI on top of JerPack host directly and do not use OAI docker image? what is the tested version of OAI support sionna-rk 1.1.0 for L1 GPU?

ah, I see the issue. You should not use the manual setup in this case. When building the images, running scripts/quickstart-oai.sh performs some extra configuration of the Dockerfile specific for the Orin platform, including a different base image. Running make sionna-rk should do all the neccessary steps to build the required Docker images, generate the configuration files, and build the TRT plans. If you try to build the docker images manually without the specific base images, the generated image will not work on Orin.

Regarding the TRT plans: the --host and --container addresses the mismatch between the TRT versions in the host and the container versions, creating two different plans. This is because the unit tests are executed in the host directly, while the integration and common use case is inside the container.

Thanks for pointing out the issues with L4T 36.4.x, and with the documentation of the manual steps, I will look into it in more detail.

Guillermo

[hitpony2]

Hello Guillermo

Thanks so much for your feedback, these are very useful and I have switched back to JetPack 6.0 on my Jetson AGX Orin, and followed your guidance and OAI gNB docker with sionna-rk plugins can be successfully run, and I can access to the gNB with COTS Mobile/UE now!

Yuchu