Issue with YOLOX-ROS TensorRT Conversion: CUDA Compatibility Error

[YuminosukeSato]

trafficstars

Hello,

I am working on converting the YOLOX Tiny model to TensorRT format using the YOLOX-ROS package. However, I encountered a CUDA compatibility error during the conversion process. The command I used was:

./src/YOLOX-ROS/weights/tensorrt/convert.bash yolox_tiny 16

The script seemed to initiate correctly, but then it terminated with the following error message:

Cuda failure: forward compatibility was attempted on non supported HW
./src/YOLOX-ROS/weights/tensorrt/convert.bash: line 37:  8669 Aborted                 (core dumped) /usr/src/tensorrt/bin/trtexec --onnx=$SCRIPT_DIR/../onnx/$MODEL.onnx --saveEngine=$SCRIPT_DIR/$MODEL.trt --fp16 --verbose --workspace=$((1<<$TRT_WORKSPACE))

This error suggests a compatibility issue with my hardware and CUDA. I am wondering if you could provide some insights or guidance on how to resolve this issue. Here are some details about my setup:

Any help or suggestions to fix this compatibility issue would be greatly appreciated. Thank you for your time and assistance.

[Ar-Ray-code]

First of all you should tell me what environment you are using.

• Ubuntu version
• CUDA(-toolkit) version
• TensorRT version
• GPU (RTX4090? Jetson Orin?)

Errors during conversion are usually due to GPU, CUDA and TensorRT version mismatch.

[YuminosukeSato]

Thanks for letting me know.

• Ubuntu version:22.04
• CUDA(-toolkit) version:11.4
• TensorRT version:8.4
• GPU :RTX3090

[Ar-Ray-code]

The following shows the environment in which I have worked at hand. Please refer

• Ubuntu22.04
• GPU-driver : 525.89.02
• CUDA : CUDA-11.8
• TensorRT (trtexec) : v8.5.1
• GPU : RTX3060