Hello NVIDIA DeepStream team,
I am validating DeepStream sample inference on a Jetson Orin Nano Developer Kit / Orin Nano Super setup and need confirmation about the correct DeepStream Docker image/tag for this platform.
Device / Host Setup
Device:
- NVIDIA Jetson Orin Nano Developer Kit
- JetPack: 7.2
- Jetson Linux / L4T: R39.2.0
- Ubuntu: 24.04.4 LTS
- Python: 3.12.3
- Docker: 29.1.3
- Docker Compose: v2.40.3
Docker was tested successfully with hello-world.
The goal is to validate a standard DeepStream sample inference pipeline on this Jetson before integrating DeepStream into an edge video analytics application.
DeepStream 8.0 Test
Docker image used:
nvcr.io/nvidia/deepstream:8.0-samples-multiarch
Version output from inside the container:
deepstream-app version 8.0.0
DeepStreamSDK 8.0.0
CUDA Driver Version: 13.2
CUDA Runtime Version: 13.0
TensorRT Version: 10.13
cuDNN Version: 9.12
libNVWarp360 Version: 2.0.1d3
The DS8 container starts correctly. NVIDIA runtime is visible. nvinfer is also registered successfully:
Long-name: NvInfer plugin
Description: Nvidia DeepStreamSDK TensorRT plugin
Version: 8.0.0
I tested the sample config:
/opt/nvidia/deepstream/deepstream-8.0/samples/configs/deepstream-app/source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display.txt
To avoid display/EGL issues, I changed only sink0 to fakesink/headless:
cd /opt/nvidia/deepstream/deepstream-8.0/samples/configs/deepstream-app
rm -f ds8_headless_test.txt
cp source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display.txt ds8_headless_test.txt
sed -i '/^\[sink0\]/,/^\[sink1\]/{s/^type=.*/type=1/; s/^sync=.*/sync=0/}' ds8_headless_test.txt
timeout 30s deepstream-app -c ds8_headless_test.txt 2>&1 | tee /tmp/ds8_headless_test.log
The pipeline reaches nvinfer and TensorRT engine creation, but fails here:
NvDsInferContextImpl::deserializeEngineAndBackend():
deserialize engine from file:
../../models/Secondary_VehicleMake/resnet18_vehiclemakenet_pruned.onnx_b16_gpu0_fp16.engine failed
NvDsInferContextImpl::generateBackendContext():
deserialize backend context from engine failed, try rebuild
NvDsInferContextImpl::buildModel():
Trying to create engine from model files
ERROR: [TRT]: IBuilder::buildSerializedNetwork: Error Code 9: API Usage Error
(Target GPU SM 87 is not supported by this TensorRT release.
In checkCurrentSMEnabled at optimizer/common/builderHelpers.cpp:707)
timeout: the monitored command dumped core
I also confirmed that the DS8 sample image does not contain prebuilt .engine or .plan files under the samples directory:
find /opt/nvidia/deepstream/deepstream-8.0/samples -type f \( -name "*.engine" -o -name "*.plan" \) | sort
This returned no engine/plan files.
The configs reference ONNX files and expected engine files, for example:
onnx-file=../../models/Secondary_VehicleMake/resnet18_vehiclemakenet_pruned.onnx
model-engine-file=../../models/Secondary_VehicleMake/resnet18_vehiclemakenet_pruned.onnx_b16_gpu0_fp16.engine
The ONNX files exist, but the .engine files do not. Therefore DeepStream tries to build the TensorRT engine at runtime and fails with SM87 unsupported.
I also installed common GStreamer packages inside the DS8 container and repeated the headless test. The result was unchanged. nvinfer still loads correctly, and the failure remains at TensorRT engine build with SM87 unsupported.
DeepStream 7.1 Fallback Test
I also tested:
nvcr.io/nvidia/deepstream:7.1-samples-multiarch
Image pulled successfully.
Version output:
deepstream-app version 7.1.0
DeepStreamSDK 7.1.0
CUDA Driver Version: 13.2
CUDA Runtime Version: 12.6
TensorRT Version: 10.3
cuDNN Version: 9.0
Dewarper: not found
I tried the DS7.1 INT8 source4 sample:
cd /opt/nvidia/deepstream/deepstream-7.1/samples/configs/deepstream-app
rm -f ds71_headless_test.txt
cp source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt ds71_headless_test.txt
sed -i '/^\[sink0\]/,/^\[sink1\]/{s/^type=.*/type=1/; s/^sync=.*/sync=0/}' ds71_headless_test.txt
timeout 30s deepstream-app -c ds71_headless_test.txt 2>&1 | tee /tmp/ds71_headless_test.log
This failed before inference. Many NVIDIA/DeepStream GStreamer plugins failed with undefined symbol errors, for example:
Failed to load plugin 'libgstnvvideo4linux2.so':
undefined symbol: g_once_init_leave_pointer
Failed to load plugin 'libnvdsgst_deepstream_bins.so':
libnvdsbufferpool.so.1.0.0: undefined symbol: g_once_init_enter_pointer
Failed to load plugin 'libnvdsgst_multistream.so':
libnvdsbufferpool.so.1.0.0: undefined symbol: g_once_init_enter_pointer
Then DeepStream failed to create the source muxer:
** ERROR: <create_multi_source_bin:1520>: Failed to create element 'src_bin_muxer'
** ERROR: <create_multi_source_bin:1608>: create_multi_source_bin failed
** ERROR: <create_pipeline:1927>: create_pipeline failed
** ERROR: <main:695>: Failed to create pipeline
Quitting
App run failed
I checked nvstreammux:
gst-inspect-1.0 nvstreammux
Result:
No such element or plugin 'nvstreammux'
But the plugin file exists:
find /usr/lib /opt/nvidia -name "*multistream*" -o -name "*deepstream_bins*" 2>/dev/null
Output included:
/opt/nvidia/deepstream/deepstream-7.1/lib/gst-plugins/libnvdsgst_deepstream_bins.so
/opt/nvidia/deepstream/deepstream-7.1/lib/gst-plugins/libnvdsgst_multistream.so
/opt/nvidia/deepstream/deepstream-7.1/lib/gst-plugins/libnvdsgst_multistreamtiler.so
Direct plugin inspection shows the actual load issue:
GST_DEBUG=3 GST_PLUGIN_PATH=/opt/nvidia/deepstream/deepstream-7.1/lib/gst-plugins gst-inspect-1.0 /opt/nvidia/deepstream/deepstream-7.1/lib/gst-plugins/libnvdsgst_multistream.so
Error:
Opening module failed:
/usr/lib/aarch64-linux-gnu/nvidia/libnvdsbufferpool.so.1.0.0:
undefined symbol: g_once_init_enter_pointer
So DS7.1 appears to have plugin ABI/library mismatch on this JetPack 7.2 / R39.2 host.
Laptop Environment Note
On my laptop, I also have a DeepStream 8.0 Docker image that can report versions:
DeepStream: 8.0.0
CUDA container banner: 12.8.1
CUDA runtime: 12.9
TensorRT: 10.9
cuDNN: 9.8
However, in that laptop Docker run, GPU is not available:
NVIDIA Driver was not detected
So I am not treating the laptop result as a valid GPU inference test. Also, I understand that TensorRT .engine files are not safely portable from laptop to Jetson because they are tied to GPU architecture, TensorRT version, CUDA stack, batch/profile, and platform.
Summary of Current Results
| Stack | Result |
|---|---|
| Jetson Orin Nano + JP7.2/R39.2 + DS8.0 samples multiarch | Container starts, nvinfer loads, pipeline reaches TensorRT, but engine build fails with Target GPU SM 87 is not supported by this TensorRT release |
| Jetson Orin Nano + JP7.2/R39.2 + DS7.1 samples multiarch | Container starts, but important NVIDIA/DeepStream GStreamer plugins fail with undefined symbol errors; nvstreammux is not registered; pipeline fails before inference |
| Laptop DS8 image | Version command works, but GPU is not available in Docker, so not a valid inference comparison |
Questions
-
What is the correct NVIDIA-supported DeepStream Docker image/tag for Jetson Orin Nano Developer Kit / Orin Nano Super running JetPack 7.2 / Jetson Linux R39.2?
-
Is
nvcr.io/nvidia/deepstream:8.0-samples-multiarchexpected to support TensorRT engine generation for Orin Nano SM87 on JetPack 7.2 / R39.2? -
If DeepStream 8.0 is supported for this device, should the TensorRT version inside the container be different from TensorRT 10.13?
-
Is the
Target GPU SM 87 is not supported by this TensorRT releaseerror expected with this DS8.0 image on Orin Nano, or does it indicate that I am using the wrong image/tag? -
For JetPack 7.2 / R39.2 on Orin Nano, should we use a different DeepStream 8.x image, a Jetson-specific L4T image, or another container tag?
-
Is DeepStream 7.1 expected to work on JetPack 7.2 / R39.2, or are the
undefined symbol: g_once_init_enter_pointererrors expected because DS7.1 is built for an older JetPack/L4T generation? -
Should we reflash the Jetson to an earlier JetPack version if we need DeepStream 7.1 support, or is there a supported DeepStream 8 image for Orin Nano on JP7.2/R39.2?
-
Are there any required additional host packages, runtime flags, or NVIDIA container runtime settings needed for this exact Orin Nano JP7.2/R39.2 setup?
The main goal is to identify the officially supported DeepStream + TensorRT + CUDA container stack for Jetson Orin Nano / Orin Nano Super on JetPack 7.2 / Jetson Linux R39.2.
