Hello,

My goal is to perform point cloud classification on LAS format files using Open3D’s CUDA-enabled PyTorch backend. To this end, I am trying to build and use Open3D in the following environment:

Environment and versions
Operating System: Windows 10 (WSL2 Ubuntu 22.04) / Native Linux or Windows (mixed environment)

Python: 3.10.13 and 3.11.4 (tried both versions, similar results)

Conda environment name: o3d_cuda (freshly created)

PyTorch: 2.5.1 with CUDA 12.1 (cu121) support

CUDA: 12.1.105

Visual Studio 2019 (or 2022)

Open3D: Latest stable branch from source

OpenSSL: Both 1.1 and 3.x versions installed and mixed in the system

Steps performed
Created a clean conda environment and installed Python 3.10/3.11.

Installed PyTorch 2.5.1 (CUDA 12.1 supported) via pip.

Downloaded Open3D source, configured CMake with CUDA and PyTorch support enabled.

CMake output shows correct detection of PyTorch and CUDA libraries.

Started build with Visual Studio.

Issues encountered
Linker errors related to OpenSSL:

crypto.lib(a_strnid.obj) : error LNK2001: unresolved external symbol __imp_bsearch
ssl.lib(ssl_cipher.obj) : error LNK2001: unresolved external symbol __imp_bsearch
...
undefined reference to SSL_CTX_add_extra_chain_cert

Although build appears to complete, importing PyTorch ML module in Python fails with:

Exception: Open3D was not built with PyTorch support!
The libOpen3D.so or libOpen3D.dll does not contain or is missing the PyTorch module.

There is a conflict in OpenSSL versions:
If only OpenSSL 3.x remains on the system, I noticed that some functions required by Open3D that depend on OpenSSL 1.1 are not available. This may explain the linker errors and missing modules.

Questions / Help needed
What is the recommended solution for these OpenSSL linker errors?
Especially missing symbols like __imp_bsearch and SSL_CTX_add_extra_chain_cert.
What linker settings or OpenSSL versions should I pay attention to?

Why might the PyTorch-enabled Python module fail to build?
I have enabled -DBUILD_CUDA_MODULE=ON and -DBUILD_PYTORCH_OPS=ON in CMake, but the module is still missing.

Are there any special configurations or additional steps required in Open3D ML to handle classification of LAS point cloud files?

What are your recommendations regarding Python version compatibility with CUDA/PyTorch?
Is there any known difference or issue between 3.10 and 3.11?

Are there any additional dependencies or environment setup steps (e.g. addict, pyyaml, scikit-learn) required before or after build?

Finally, is there an alternative installation method or pre-built package you would recommend to avoid these build issues and still enable LAS point cloud classification with Open3D CUDA + PyTorch?

Additional info
nvcc and CUDA toolkit are correctly installed and on path.

PyTorch CUDA version verified via pip list.

CMake correctly detects PyTorch and CUDA paths.

Visual Studio build logs show OpenSSL-related warnings/errors.

Tried building Open3D ML PyTorch module independently; issues persist.

Thanks in advance for any guidance. I can provide logs and config files if needed.

Additional question:
Given the system specs (Intel i7-10700K, RTX 4060 Laptop GPU, 32 GB RAM, CUDA 12.1), are there any alternative installation methods or pre-built binaries you recommend that could avoid these OpenSSL and build issues while enabling LAS point cloud classification with Open3D CUDA + PyTorch?