The documentation here is intended to help customers build the Open Source DeepStream Dockerfiles. This information is useful for both x86 systems with dGPU setup and on NVIDIA Jetson devices.

Since DS 6.3, deepStream docker containers do not package libraries necessary for certain multimedia operations like audio data parsing, CPU decode, and CPU encode. This change could affect processing certain video streams/files like mp4 that include audio tracks.

Please run the below script inside the docker images to install additional packages that might be necessary to use all of the DeepStreamSDK features :

/opt/nvidia/deepstream/deepstream/user_additional_install.sh

Please refer to the Prerequisites section at DeepStream NGC page NVIDIA NGC to build and run deepstream containers.

1. Please download the DeepStreamSDK release x86 tarball and place it locally in the $ROOT/x86_64 folder of this repository.

cp deepstream_sdk_v6.4.0_x86_64.tbz2 x86_64/

2. image_url is the desired docker name:TAG

3. ds_pkg and ds_pkg_dir shall be the tarball file-name with and without the tarball extension respectively. Refer to Section 2.1.3 x86 Build Command for sample command.

4. base_image is the desired container name. Please feel free to use the sample name provided in the command above. This name is used in the triton build steps alone. Refer to Section 2.1.3 x86 Build Command for sample command.

Note: These packages (uff-converter-tf and graphsurgeon-tf) are now included by default.

Download file link: nv-tensorrt-local-repo-ubuntu2204-8.6.1-cuda-12.0_1.0-1_amd64.deb from TensorRT download page. Note: You may have to login to developer.nvidia.com to download the file.
Quick Steps:
$ROOT is the root directory of this git repo.
cd $ROOT/x86_64

cp nv-tensorrt-local-repo-ubuntu2204-8.6.1-cuda-12.0_1.0-1_amd64.deb x86_64/

Also the CuDNN file.

Download file link: cudnn-local-repo-ubuntu2204-8.9.4.25_1.0-1_amd64.deb from TensorRT download page. Note: You may have to login to developer.nvidia.com to download the file. Quick Steps: $ROOT is the root directory of this git repo. cd $ROOT/x86_64

cp cudnn-local-repo-ubuntu2204-8.9.4.25_1.0-1_amd64.deb x86_64/

⚠️ NOTE: Docker ADD image size workaround by using wget instead ⚠️

Docker ADD method is used by default for ease to building the x86 Dockers. Docker ADD also increases the image size by approximately 2GB.

To workaround this problem you can host the nv-tensorrt-local-repo-ubuntu2204-8.6.1-cuda-12.0_1.0-1_amd64.deb a server and pull it in during the docker build using wget. This section of code is commented out by default.

To use this wget method you can uncomment this code section of the code in the dockerfile. Then also add the complete URL to the file.

# install TensorRT repo from a hosted file on a server

The commented out this code section in the dockerfile that starts with this.

# Add TensorRT repo

sudo image_url=deepstream:6.4.0-triton-local \
     ds_pkg=deepstream_sdk_v6.4.0_x86_64.tbz2 \
     ds_pkg_dir=deepstream_sdk_v6.4.0.0_x86_64/ \
     base_image=dgpu-any-custom-base-image make -f Makefile_x86_triton triton-devel -C x86_64/

sudo image_url=deepstream:6.4.0-triton-local \
     ds_pkg=deepstream_sdk_v6.4.0_x86_64.tbz2 \
     ds_pkg_dir=deepstream_sdk_v6.4.0.0_x86_64/ \
     base_image=dgpu-any-custom-base-image make -f Makefile_x86_triton triton-devel -C x86_64/

There is an example build script is called $ROOT/buildx86.sh with the same contents.

sudo image_url=deepstream:6.4.0-samples-local \
     ds_pkg=deepstream_sdk_v6.4.0_x86_64.tbz2 \
     ds_pkg_dir=deepstream_sdk_v6.4.0.0_x86_64/ \
     base_image=dgpu-any-custom-base-image make -f Makefile runtime -C x86_64/

Must be built on a Jetson device (e.g. Jetson AGX Orin 64GB ).

Please refer to the Prerequisites section at DeepStream NGC page NVIDIA NGC to build and run deepstream containers.

Download DeepStreamSDK tarball from DeepStreamSDK release Jetson tarball and place it locally in the $ROOT/jetson folder of this repository.

cp deepstream_sdk_v6.4.0_jetson.tbz2 jetson/

The following VPI 3.0.10 *.deb (vpi-dev-3.0.10-aarch64-l4t.deb, vpi-lib-3.0.10-aarch64-l4t.deb) files are required from JetPack 6.0 DP. These need to be copied into the Jetson folder so they can be used by the build.

cp vpi-dev-3.0.10-aarch64-l4t.deb jetson/
cp vpi-lib-3.0.10-aarch64-l4t.deb jetson/

sudo image_url=deepstream-l4t:6.4.0-triton-local \
     ds_pkg=deepstream_sdk_v6.4.0_jetson.tbz2 \
     ds_pkg_dir=deepstream_sdk_v6.4.0_jetson \
     base_image=jetson-any-custom-base-image make -f Makefile triton -C jetson/

     There is an example build script is called $ROOT/buildjet.sh with the same contents. 

sudo image_url=deepstream-l4t:6.4.0-triton-local \
     ds_pkg=deepstream_sdk_v6.4.0_jetson.tbz2 \
     ds_pkg_dir=deepstream_sdk_v6.4.0_jetson \
     base_image=jetson-any-custom-base-image make -f Makefile triton -C jetson/

sudo image_url=deepstream-l4t:6.4.0-samples-local \
     ds_pkg=deepstream_sdk_v6.4.0_jetson.tbz2 \
     ds_pkg_dir=deepstream_sdk_v6.4.0_jetson \
     base_image=jetson-any-custom-base-image make -f Makefile runtime -C jetson/     

Steps:

1. Open Triton Docker file:

x86_64/trtserver_base_devel/Dockerfile

2. Edit the FROM command in Dockerfile.

Change the FROM command to use the desired Triton version.

Current: Triton 23.08

FROM nvcr.io/nvidia/tritonserver:23.08-py3

Example Migration to: Triton 23.11

FROM nvcr.io/nvidia/tritonserver:23.11-py3

3. Edit the Triton client libraries URL in Dockerfile.

Client libraries are available for download from Triton Inference Server Releases page.

Current: Triton 23.08

wget https://github.com/triton-inference-server/server/releases/download/v2.37.0/v2.37.0_ubuntu2204.clients.tar.gz

Example Migration to: Triton 23.11

wget https://github.com/triton-inference-server/server/releases/download/v2.40.0/v2.40.0_ubuntu2204.clients.tar.gz

4. Build the DS x86 triton docker following instructions here.

Steps:

1. Open Triton Docker file:

jetson/ubuntu_base_devel/Dockerfile

2. Edit the FROM command in Dockerfile.

Change the FROM command to use the desired Triton version.

Current: Triton 23.11

FROM nvcr.io/nvidia/tritonserver:23.11-py3-igpu

Note: Newer version of Triton is unavailable to verify migration at the time of DS 6.4 release.

3. Edit the Triton client libraries URL in Dockerfile.

Client libraries are available for download from Triton Inference Server Releases page.

Current: Triton 23.08

wget https://github.com/triton-inference-server/server/releases/download/v2.40.0/tritonserver2.40.0-igpu.tar.gz

Note: Newer version of Triton is unavailable to verify migration at the time of DS 6.4 release.

4. Build the DS x86 triton docker following instructions here.

DeepStream 6.4 Triton Server API is based on Triton 23.08 (x86) and Triton 23.11 (Jetson) release.

Regarding API compatibility, if a customer wants to upgrade triton, they need to make sure:

a) new version's tritonserver.h is compatible with the:

23.08 version of tritonserver.h for x86, 23.11 version of tritonserver.h for jetson, and

b) new version’s model_config.proto is compatible with:

23.08 version for x86, 23.11 version for jetson,

To build specific Tritonserver version libs, users can follow instructions at https://github.com/triton-inference-server/server/blob/master/docs/build.md.

Gst-nvinferserver plugin’s config file kept backward compatibility. Triton model/backend’s config.pbtxt file must follow rules of 23.08’s model_config.proto for x86 and 23.11's model_config.proto for jetson.

DeepStream 6.4 release package inherently support Ubuntu 22.04.

Thus, the only thing to consider is API/ABI compatibility between the new Triton version and the Triton version supported by current DS release.