Unable to use onnxruntime-gpu package when multiple versions of CUDA are installed

[ivanst0]

Describe the bug
PATH and CUDA_PATH environment variables point to the most recently installed version of CUDA (usually 11.x), while onnxruntime-gpu package from PyPI requires CUDA 10.2. The only workaround is to manually set PATH/CUDA_PATH environment variable before using this package (and restore it afterwards).

Having to configure an environment variable before/after each use of onnxruntime-gpu is quite inconvenient. Dev machines very often have multiple versions of CUDA installed.

System information

• OS Platform and Distribution: Windows 10
• ONNX Runtime installed from: PyPI
• ONNX Runtime version: 1.6.0
• Python version: 3.6, 3.7, 3.8, 3.9
• CUDA/cuDNN version: multiple

To Reproduce

1. Install CUDA Toolkit 10.2.
2. Install CUDA Toolkit 11.1.
3. Install onnxruntime-gpu package.
4. Try to import onnxruntime.

An ImportError is reported.

Additional context
onnxruntime-gpu package for Python 3.6 and 3.7 relies on PATH environment variable (the first directory in PATH containing the target DLL) to locate CUDA/cuDNN DLLs, while with Python 3.8 and 3.9 it uses CUDA_PATH variable for the same purpose.