Upstream notes "GPU support on Linux is not currently supported." In practice that's three separate, fixable issues. This PR fixes all three so the ALPR module reports inferenceDevice=GPU on a Linux / CUDA-12 host.

Tested on an RTX A5000 (CUDA 12.4 / 12.7 driver, cuDNN 8.9, Python 3.8).

Changes

• requirements.linux.cuda12.txt — paddlepaddle-gpu was unpinned, but PyPI only ships CUDA ≤ 11.7 builds, so the CUDA-12 build was never installed and ALPR silently ran on CPU. Pin paddlepaddle-gpu==2.6.2.post120 and add Paddle's cu120 package index via --extra-index-url. (Runs on CUDA 12.0–12.7 drivers via minor-version forward compatibility.)
• modulesettings.linux.json — "InstallGPU": false forced the adapter's enable_GPU off even when a capable GPU was present (use_gpu = enable_GPU and can_use_GPU). Set to true.
• install.sh — PaddlePaddle dlopens the unversioned libcudnn.so, but Ubuntu's cuDNN package ships only libcudnn.so.8, causing PreconditionNotMetError: Cannot load cudnn shared library. Symlink libcudnn.so → libcudnn.so.8 on Linux x86_64 during install (guarded: only if the versioned lib exists and the link is missing).

Notes

• Do not bump to PaddlePaddle 3.x to get a newer CUDA build — it drops Python 3.8 and breaks PaddleOCR 2.7.0.3. This stays on 2.6.2.post120 + paddleocr==2.7.0.3.
• The .post120 wheel is China-hosted (paddlepaddle.org.cn / bcebos.com); on networks that can't reach it, download the wheel elsewhere and pip install the local file. The InstallGPU flip and libcudnn.so symlink need no network and apply regardless.

Testing

python -c "import paddle; paddle.utils.run_check()"   # -> "PaddlePaddle works well on 1 GPU"

and the ALPR module status reports inferenceDevice: GPU.