Windows pip wheels for paddlepaddle-gpu 3.2.2/3.3.1 warn that Paddle was compiled with cuDNN 9.9 while installed dependency provides cuDNN 9.5.1.17

[benchan111999-cmd]

Environment

• OS: Windows
• GPU: NVIDIA GeForce RTX 4080
• NVIDIA driver: 596.49
• nvidia-smi CUDA version: 13.2
• Paddle runtime log:
  • GPU Compute Capability: 8.9
  • Driver API Version: 13.2
  • Runtime API Version: 12.6
• Python: 3.12.13
• Main environment:
  • paddlepaddle-gpu==3.2.2
  • paddleocr==3.7.0
  • paddlex==3.7.2
  • installed with paddleocr[doc-parser]
• Comparison environment:
  • paddlepaddle-gpu==3.3.1
  • paddleocr==3.7.0
  • paddlex==3.7.2
  • installed with paddleocr[doc-parser]
• In both environments, pip check passes.
• In both environments, the installed Paddle dependency set includes:
  • nvidia-cudnn-cu12==9.5.1.17

Steps to reproduce

1. Create a clean Windows Python 3.12 virtual environment.
2. Install PaddleOCR with the GPU Paddle wheel, for example:

   python -m pip install paddlepaddle-gpu==3.2.2 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
   python -m pip install "paddleocr[doc-parser]"

3. Verify Paddle:

   import paddle
   paddle.utils.run_check()

4. Run a minimal PaddleOCR GPU prediction, for example:

   from paddleocr import PaddleOCR
   
   ocr = PaddleOCR(device="gpu:0")
   result = ocr.predict(r"C:\tmp\paddleocr_smoke.png")
   print(result)

5. Repeat with paddlepaddle-gpu==3.3.1 from the same cu126 package index.

Observed behavior

• pip check passes.

• paddle.utils.run_check() passes and reports that PaddlePaddle works well on 1 GPU.

• PaddleOCR prediction succeeds. In my smoke test, the image text is recognized correctly:
  ECG COURSE SINUS RHYTHM 75 BPM

• However, the prediction path prints this warning:

  WARNING: device: 0. The installed Paddle is compiled with CUDNN 9.9, but CUDNN version in your machine is 9.5, which may cause serious incompatible bug. Please recompile or reinstall Paddle with compatible CUDNN version.
  

• The same warning appears with both paddlepaddle-gpu==3.2.2 and paddlepaddle-gpu==3.3.1.

• Both versions still install/use nvidia-cudnn-cu12==9.5.1.17, so the warning appears to be triggered by the cuDNN version provided by Paddle's own pip dependency set, not by a manually installed system cuDNN.

Expected behavior / question

• If the official Windows pip wheel is compiled with cuDNN 9.9, should the wheel dependency require a matching nvidia-cudnn-cu12 version instead of 9.5.1.17?
• Or is cuDNN 9.5.1.17 intentionally supported for these cu126 wheels, and the warning is overly strict / cosmetic?
• What is the officially supported cuDNN runtime version for Windows pip wheels of paddlepaddle-gpu==3.2.2 and 3.3.1?

Why this matters

• The warning says this mismatch "may cause serious incompatible bug", but the environment is otherwise clean:
  • pip check passes
  • paddle.utils.run_check() passes
  • PaddleOCR GPU inference succeeds
• This makes it unclear whether users should ignore the warning, manually override the cuDNN package, downgrade/upgrade Paddle, or wait for a corrected wheel/dependency.
• This is especially confusing for Windows users installing PaddleOCR through pip, because the reported cuDNN 9.5 runtime appears to come from Paddle's installed dependency set.

What has been tried

• Upgraded the main environment from paddlepaddle-gpu==3.2.0 to 3.2.2.
• Created a separate comparison environment with paddlepaddle-gpu==3.3.1.
• Installed paddleocr[doc-parser] in both environments.
• Confirmed pip check passes in both environments.
• Confirmed paddle.utils.run_check() passes in both environments.
• Confirmed PaddleOCR GPU prediction works and recognizes a smoke-test image correctly.
• Confirmed the warning remains in both Paddle versions.
• Confirmed both environments use nvidia-cudnn-cu12==9.5.1.17.

Related official issues

• PaddleOCR lstm_unit使用报错 #18164 mentions the same cuDNN 9.9 vs 9.5 warning on Windows / CUDA 12.6 / CC 8.9 with paddlepaddle-gpu==3.2.1, but that issue is mainly about PaddleOCR-VL throughput and a different warning.
• PaddleOCR 怎么通过numpy创建一个constant tensor #17420 also contains the same warning in logs, but the issue topic is recognition output difference, not Paddle wheel packaging or cuDNN compatibility.