Which component has the problem?

CUTLASS C++

Bug Report

Describe the bug
When configuring cutlass with CMake, I get the error cuDNN not found. However, I do have cuDNN installed, using the installer from NVIDIA. When installed that way, cudnn is NOT embedded inside CUDA directories.

Steps/Code to reproduce bug

1. Install CUDNN via the official installer. As a result, the include directory is C:\Program Files\NVIDIA\CUDNN\v9.13\include\12.9 (for CUDA 12.9) and C:\Program Files\NVIDIA\CUDNN\v9.13\include\13.0 (for CUDA 13.0). Please note that these paths do not end with /include
2. In this repo, launch cmake: cmake -B build -DCUTLASS_ENABLE_CUDNN=ON .
3. I get the error cuDNN not found.

Expected behavior
I expect CMake configure to go well, and CUDNN to be correctly found.

Environment details (please complete the following information):

• Regular Win11 PC

Additional context
It seems like cuDNN.cmake does a find_path but force cudnn.h to be under a path ending with /include. This is just not the case with the official installer...

I can propose a patch, if you want - mainly adding naked $ENV{CUDNN_PATH} in find_path/find_library