PySCF 2.0.0 wheels for Python 3.9 from PyPI segfault

Dear developers, while trying out the new PySCF release in a fresh virtual environment (Python 3.9.7, Linux-5.10.70-1-MANJARO-x86_64-with-glibc2.33) I noticed that it always segfaults, even for simple cases like examples/scf/00-simple_hf.py.

2021-11-03 07:52:21 ⌚  dachshoehle in ~/pyscfvenv                                                                                                                                                                
○ → pip install pyscf                                                                                                                                                                                              
Collecting pyscf                                                                                                                                                                                                   
  Using cached pyscf-2.0.0-cp39-cp39-manylinux1_x86_64.whl (40.9 MB)                                                                                                                                               
Collecting scipy!=1.5.0,!=1.5.1                                                                                                                                                                                    
  Using cached scipy-1.7.1-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.whl (28.5 MB)                                                                                                                          
Collecting numpy!=1.16,!=1.17,>=1.13                                                                                                                                                                               
  Downloading numpy-1.21.3-cp39-cp39-manylinux_2_12_x86_64.manylinux2010_x86_64.whl (15.7 MB)                                                                                                                      
     |████████████████████████████████| 15.7 MB 6.8 MB/s                                                                                                                                                           
Collecting h5py>=2.7                                                                                                                                                                                               
  Using cached h5py-3.5.0-cp39-cp39-manylinux_2_12_x86_64.manylinux2010_x86_64.whl (4.5 MB)                                                                                                                        
Installing collected packages: numpy, scipy, h5py, pyscf                                                                                                                                                           
Successfully installed h5py-3.5.0 numpy-1.21.3 pyscf-2.0.0 scipy-1.7.1

python 00-simple_hf.py  | tee 00-simple_hf.log

yields

#INFO: **** input file is /home/johannes/Code/pyscf/examples/scf/00-simple_hf.py ****
#!/usr/bin/env python
#
# Author: Qiming Sun <[e-mail]>
#

'''
A simple example to run HF calculation.

.kernel() function is the simple way to call HF driver.
.analyze() function calls the Mulliken population analysis etc.
'''

import pyscf

mol = pyscf.M(
    atom = 'H 0 0 0; F 0 0 1.1',  # in Angstrom
    basis = 'ccpvdz',
    symmetry = True,
    verbose=4,
)

myhf = mol.HF()
myhf.kernel()

# Orbital energies, Mulliken population etc.
myhf.analyze()


#
# myhf object can also be created using the APIs of gto, scf module
#
from pyscf import gto, scf
mol = gto.M(
    atom = 'H 0 0 0; F 0 0 1.1',  # in Angstrom
    basis = 'ccpvdz',
    symmetry = True,
)
myhf = scf.HF(mol)
myhf.kernel()

#INFO: ******************** input file end ********************


System: uname_result(system='Linux', node='dachshoehle', release='5.10.70-1-MANJARO', version='#1 SMP PREEMPT Thu Sep 30 15:29:01 UTC 2021', machine='x86_64')  Threads 12
Python 3.9.7 (default, Aug 31 2021, 13:28:12) 
[GCC 11.1.0]
numpy 1.21.3  scipy 1.7.1
Date: Wed Nov  3 07:58:17 2021
PySCF version 2.0.0
PySCF path  /home/johannes/pyscfvenv/lib/python3.9/site-packages/pyscf

[CONFIG] conf_file None
[INPUT] verbose = 4
[INPUT] num. atoms = 2
[INPUT] num. electrons = 10
[INPUT] charge = 0
[INPUT] spin (= nelec alpha-beta = 2S) = 0
[INPUT] symmetry True subgroup None
[INPUT] Mole.unit = angstrom
[INPUT]  1 H      0.000000000000   0.000000000000   0.000000000000 AA    0.000000000000   0.000000000000   0.000000000000 Bohr
[INPUT]  2 F      0.000000000000   0.000000000000   1.100000000000 AA    0.000000000000   0.000000000000   2.078698737022 Bohr

nuclear repulsion = 4.32963172570909
point group symmetry = Coov
num. orbitals of irrep A1 = 9
num. orbitals of irrep E1x = 4
num. orbitals of irrep E1y = 4
num. orbitals of irrep E2x = 1
num. orbitals of irrep E2y = 1
number of shells = 8
number of NR pGTOs = 33
number of NR cGTOs = 19
basis = ccpvdz
ecp = {}
CPU time:         0.26


******** <class 'pyscf.scf.hf_symm.SymAdaptedRHF'> ********
method = SymAdaptedRHF-RHF
initial guess = minao
damping factor = 0
level_shift factor = 0
DIIS = <class 'pyscf.scf.diis.CDIIS'>
diis_start_cycle = 1
diis_space = 8
SCF conv_tol = 1e-09
SCF conv_tol_grad = None
SCF max_cycles = 50
direct_scf = True
direct_scf_tol = 1e-13
chkfile to save SCF result = /home/johannes/Code/pyscf/examples/scf/tmp3vx6zikp
max_memory 4000 MB (current use 82 MB)
Freeze 0 electrons in irreps []
    10 free electrons in irreps A1 E1x E1y E2x E2y
Set gradient conv threshold to 3.16228e-05
# The segfault in German below is printed to STDERR
Speicherzugriffsfehler (Speicherabzug geschrieben)

I just added verbose = 4, so some output is created. I also noticed issue #1093. There, the last comment says the new wheels are already online?! But it seems that the wheels on PyPI are still from the 1st of November, before you merged PR #1092 that was supposed to fix some build issues?

So my question is: are the wheels currently available on PyPI supposed to be correct and work?

All the best, Johannes

[eljost]

Thanks for the quick fix. It works flawlessly now.

[sunqm]

It’s a bug of the openblas library in the previous pip wheel environment. In PR #1099 I updated the openblas library to a relatively new version and added a new release tag 2.0.1. The new release should be able to fix the problem