Run nvcc --version in your terminal to check the installed CUDA toolkit version. Then, choose the proper package based on your CUDA toolkit version.

One can compile the package with

git clone https://github.com/pyscf/gpu4pyscf.git
cd gpu4pyscf
cmake -S gpu4pyscf/lib -B build/temp.gpu4pyscf
cmake --build build/temp.gpu4pyscf -j 4
CURRENT_PATH=`pwd`
export PYTHONPATH="${PYTHONPATH}:${CURRENT_PATH}"

Then install cutensor and cupy for acceleration (please switch the versions according to your nvcc version!)

pip3 install cutensor-cu12 cupy-cuda12x

There shouldn't be cupy or cutensor compilation during pip install process. If you see the following warning at the beginning of a gpu4pyscf job, it implies problems with cupy and cutensor installation (likely a version mismatch, or multiple versions of same package installed).

<repo_path>/gpu4pyscf/lib/cutensor.py:<line_number>: UserWarning: using cupy as the tensor contraction engine.

The package also provides multiple dockerfiles in dockerfiles. One can use them as references to create the compilation envrionment.

• Density fitting scheme and direct SCF scheme;
• SCF, analytical Gradient, and analytical Hessian calculations for Hartree-Fock and DFT;
• LDA, GGA, mGGA, hybrid, and range-separated functionals via libXC;
• Spin-conserved and spin-flip TDA and TDDFT for excitated states
• Geometry optimization and transition state search via geomeTRIC;
• Dispersion corrections via DFTD3 and DFTD4;
• Nonlocal functional correction (vv10) for SCF and gradient;
• ECP is supported and calculated on CPU;
• PCM models, SMD model, their analytical gradients, and semi-analytical Hessian matrix;
• Unrestricted Hartree-Fock and Unrestricted DFT, gradient, and Hessian;
• MP2/DF-MP2 and CCSD (experimental);
• Polarizability, IR, and NMR shielding (experimental);
• QM/MM with PBC;
• CHELPG, ESP, and RESP atomic charge

• Rys roots up to 9 for density fitting scheme and direct scf scheme;
• Atomic basis up to g orbitals;
• Auxiliary basis up to i orbitals;
• Density fitting scheme up to ~168 atoms with def2-tzvpd basis, bounded by CPU memory;
• meta-GGA without density laplacian;
• Double hybrid functionals are not supported;

import pyscf
from gpu4pyscf.dft import rks

atom ='''
O       0.0000000000    -0.0000000000     0.1174000000
H      -0.7570000000    -0.0000000000    -0.4696000000
H       0.7570000000     0.0000000000    -0.4696000000
'''

mol = pyscf.M(atom=atom, basis='def2-tzvpp')
mf = rks.RKS(mol, xc='LDA').density_fit()

e_dft = mf.kernel()  # compute total energy
print(f"total energy = {e_dft}")

g = mf.nuc_grad_method()
g_dft = g.kernel()   # compute analytical gradient

h = mf.Hessian()
h_dft = h.kernel()   # compute analytical Hessian

to_gpu is supported since PySCF 2.5.0

import pyscf
from pyscf.dft import rks

atom ='''
O       0.0000000000    -0.0000000000     0.1174000000
H      -0.7570000000    -0.0000000000    -0.4696000000
H       0.7570000000     0.0000000000    -0.4696000000
'''

mol = pyscf.M(atom=atom, basis='def2-tzvpp')
mf = rks.RKS(mol, xc='LDA').density_fit().to_gpu()  # move PySCF object to GPU4PySCF object
e_dft = mf.kernel()  # compute total energy

Find more examples in gpu4pyscf/examples

Speedup with GPU4PySCF v0.6.0 on A100-80G over Q-Chem 6.1 on 32-cores CPU (Desity fitting, SCF, def2-tzvpp, def2-universal-jkfit, B3LYP, (99,590))

Find more benchmarks in gpu4pyscf/benchmarks

@misc{li2024introducting,
      title={Introducing GPU-acceleration into the Python-based Simulations of Chemistry Framework}, 
      author={Rui Li and Qiming Sun and Xing Zhang and Garnet Kin-Lic Chan},
      year={2024},
      eprint={2407.09700},
      archivePrefix={arXiv},
      primaryClass={physics.comp-ph},
      url={https://arxiv.org/abs/2407.09700}, 
}

@misc{wu2024enhancing,
      title={Enhancing GPU-acceleration in the Python-based Simulations of Chemistry Framework}, 
      author={Xiaojie Wu and Qiming Sun and Zhichen Pu and Tianze Zheng and Wenzhi Ma and Wen Yan and Xia Yu and Zhengxiao Wu and Mian Huo and Xiang Li and Weiluo Ren and Sheng Gong and Yumin Zhang and Weihao Gao},
      year={2024},
      eprint={2404.09452},
      archivePrefix={arXiv},
      primaryClass={physics.comp-ph},
      url={https://arxiv.org/abs/2404.09452}, 
}