Succesfully compiled on RTX 5070 Ti (Blackwell) with CUDA 13.0

Description
I successfully compiled and ran gpu4pyscf on the new NVIDIA RTX 5070 Ti (Blackwell architecture) with CUDA 13.0.

Since the official pre-built wheels do not support the new architecture yet, I had to compile from source. I encountered nvcc fatal : Unsupported gpu architecture 'compute_70' errors because CUDA 13 seems to have dropped support for older compute capabilities (or defaults strictly).

Environment

• GPU: NVIDIA RTX 5070 Ti
• CUDA Version: 13.0
• OS: Linux (Ubuntu/WSL)
• Python: 3.10

The Fix
I made two changes to get it working:

1. Modified gpu4pyscf/lib/CMakeLists.txt:
   Changed the default architecture handling to use native (or 90 for Hopper/Blackwell compatibility) instead of the hardcoded 70-real;80-real;90-real.

   # Original
   # set(CMAKE_CUDA_ARCHITECTURES "70-real;80-real;90-real")
   
   # Modified
   set(CMAKE_CUDA_ARCHITECTURES "native")
   

2. Modified setup.py:
   Relaxed the libxc version requirement because 0.5 is not available for CUDA 13 on PyPI, and I had to compile libxc from source using -DBUILD_LIBXC=ON.

   # Modified dependency check
   f'gpu4pyscf-libxc-cuda{CUDA_VERSION}>=0.5',
   

Build Command
I built it using:

export CMAKE_CONFIGURE_ARGS="-DCUDA_ARCHITECTURES=native -DBUILD_LIBXC=ON"
pip install -v -e .

Review gpu4pyscf/lib/CMakeLists.txt and setup.py, starting with the CUDA architecture setting and libxc dependency check described in the issue. Reproduce the editable install with CMAKE_CONFIGURE_ARGS="-DCUDA_ARCHITECTURES=native -DBUILD_LIBXC=ON" on CUDA 13.0, then verify that the RTX 5070 Ti build completes and runs without the reported architecture or dependency errors.