DeepStream 8.0 on Docker WSL – Segmentation Fault, CUDA Init Failure, and TensorRT Version Downgrade (10.9 → 10.3)

Accelerated Computing Intelligent Video Analytics DeepStream SDK
,
1

Hardware Platform: NVIDIA GPU (GeForce RTX 5080, running via WSL2 on PC)

DeepStream Version: 8.0.0

JetPack Version: N/A (running on PC, not Jetson)

TensorRT Version:

  • Initially detected: 10.9

  • After running DeepStream apps continuously or building models: changes to 10.3 (verified via deepstream-app --version-all)

NVIDIA GPU Driver Version: 577.00

CUDA Driver Version: 12.9

CUDA Runtime Version: 12.8

cuDNN Version: 9.13

libNVWarp360 Version: 2.0.1d3

Issue Type: Bug

How to Reproduce the Issue:

  1. Install DeepStream 8.0 Docker on WSL2.

  2. Run the following command:

deepstream-app -c samples/configs/deepstream-app/source30_1080p_dec_infer-resnet_tiled_display.txt

  1. Initially, deepstream-app --version-all shows TensorRT 10.9.

  2. After running sample apps continuously or after building models, deepstream-app --version-all unexpectedly shows TensorRT 10.3.

  3. Eventually, the following error occurs:

ERROR: [TRT]: createInferRuntime: Error Code 6: API Usage Error (CUDA initialization failure with error: 35 In checkCudaInstallation at runtime/dispatch/runtime.cpp:689)
ERROR: [TRT]: [checkMacros.cpp::catchCudaError::226] Error Code 1: Cuda Runtime (In catchCudaError at common/dispatch/checkMacros.cpp:226)
Segmentation fault (core dumped)

Additional Notes:

  • The TensorRT version downgrades automatically inside the container without any manual changes.

  • The error occurs intermittently — sometimes apps run fine, sometimes fail immediately.

  • Running on WSL2 with GPU passthrough (nvidia-smi confirms RTX 5080 with driver 577.00).

  • Possible issue with version mismatch between container libs and host driver/runtime.

Expected Behavior:

  • TensorRT version should remain stable (10.9, as initially reported).

  • DeepStream apps should run consistently without CUDA/TensorRT initialization failures.

2

Additional Observations / Continuation:

  • After creating a new DeepStream 8.0 Docker container and running apps for several hours or after building/loading a .pt model to ONNX, the custom DeepStream app fails with:
 Unable to set device in gst_nvstreammux_change_state
** ERROR: <main:802>: Failed to set pipeline to PAUSED
Quitting

Running deepstream-app --version-all at this stage shows:

deepstream-app version 8.0.0
DeepStreamSDK 8.0.0
CUDA Driver Version: 0.0
CUDA Runtime Version: 12.8
TensorRT Version: 10.9
cuDNN Version: 9.8
libNVWarp360 Version: 2.0.1d3

  • This indicates that the CUDA driver inside the container becomes unavailable after long runtime or model operations.

  • The loss of CUDA driver may be related to WSL2 GPU passthrough instability, DeepStream/TensorRT initialization, or container runtime handling of /dev/nvidia* devices.

  • This explains the intermittent TensorRT version mismatch (10.9 → 10.3) observed earlier.

Impact:

  • DeepStream apps fail to initialize pipelines.

  • TensorRT runtime becomes unstable.

  • Segmentation faults and CUDA initialization errors occur.

4

After starting the container, do you just run deepstream-app? Are there any other operations?

Is your installation environment exactly the same as that in the documentation?

5

The installation is correct and follows the documentation. Initially everything runs fine, but after keeping the container running for a few hours, or when I include a custom model (converted from .pt → ONNX → engine), the reported versions change unexpectedly (for example, TensorRT or CUDA Driver showing 0.0), and then the errors appear.

6

Thanks for your feedback, I’ll try to reproduce it

7

I set file-loop to true in the configuration file and ran it in docker for several hours without any similar problems.

8

Thanks for reply,Now i have no issues in docker

9

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.