"CUDA Error of work_lda: named symbol not found" error after gpu4pyscf 1.3.1 update

[JeongheonSeok]

After updating gpu4pyscf to version 1.3.1 (and updating other dependent packages using pip), I get an error like the title. I'm getting the same error even when I create a new environment and install only gpu4pyscf, cutensor and dependency packages.

Even a very simple example like below is giving me the error right away. Can you give me some advice on how to fix it?

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='b3lyp').density_fit()
energy = mf.kernel()

Error

CUDA Error of work_lda: named symbol not found

+) above error occurs at (gpu4pyscf-cuda12x==1.3.1, gpu4pyscf-libxc-cuda12x==0.6), (gpu4pyscf-cuda12x==1.3.0, gpu4pyscf-libxc-cuda12x==0.5). (gpu4pyscf-cuda12x==1.3.1, gpu4pyscf-libxc-cuda12x==0.5), (gpu4pyscf-cuda12x==1.3.0, gpu4pyscf-libxc-cuda12x==0.5) are okay. I think there might be a problem with gpu4pyscf-libxc-cuda12x==0.6 version.

[wxj6000]

Yes, it is a bug introduced by gpu4pyscf-libxc-cuda12x==0.6. You can revert its version to 0.5 at this point. I am fixing the issue right now.

[JeongheonSeok]

Yes, thank you for checking.

[wxj6000]

@JeongheonSeok The issue has been fixed. You can upgrade libxc via pip3 install gpu4pyscf-libxc-cuda12x --upgrade for the fix. It will take a few mins for the first run, as it needs to compile the binary for the corresponding GPU. You can also keep gpu4pyscf-libxc-cuda12x==0.5 for the same using experience as before.

[JeongheonSeok]

Yes, I checked. Thank you for your kind reply and correction!

[sunqm]

The precompiled pip wheels only include ptxas code for sm-70, 80, and 90. You might need to compile additional sm archs for your system from source code.