Compatibility of PaddleDetection with RTX 5070 (CUDA 12.9) and PaddlePaddle 3.3.0 – Export Model Issues

问题确认 Search before asking

• 我已经搜索过问题，但是没有找到解答。I have searched the question and found no related answer.

请提出你的问题 Please ask your question

Hello,

I would like to ask about the compatibility between PaddleDetection and newer GPU architectures.

I am currently using the following setup:

• GPU: NVIDIA RTX 5070 (Compute Capability 12.0)
• CUDA: 12.9
• PaddlePaddle: 3.3.0
• PaddleDetection: 2.9.0

Training appears to run, and I am able to load model weights successfully. However, I consistently encounter errors when trying to export the model using tools/export_model.py.

The error seems to be related to internal API incompatibilities (e.g., Variable vs Value types and issues with functions like sigmoid during dy2static conversion). From what I understand, this may be due to differences between PaddlePaddle 2.x and 3.x APIs.

My questions are:

• Is PaddleDetection 2.9.0 officially compatible with PaddlePaddle 3.3.0?
• Is it currently possible to reliably export models using this setup (RTX 5070 + CUDA 12.9 + PaddlePaddle 3.3.0)?
• If not, is there a recommended version of PaddleDetection that supports PaddlePaddle 3.x and newer GPUs?
• Are there any workarounds or flags to make export work under PaddlePaddle 3.x? If needed, I can provide the full error log and configuration files. Thank you in advance for your help.