[UNEXPECTED BEHAVIOR] Large memory cost in generating 3c2e integral

[RHSong]

Here is my script

import numpy as np
from pyscf.pbc import gto, scf

atom = '''
Cu      0.0 0.0 0.0
Cu      0.0    , 1.8075, 1.8075
Cu      1.8075 , 0.0, 1.8075
Cu      1.8075 , 1.8075, 0.0
'''
a = np.eye(3) * 3.615

cell = gto.Cell(atom=atom, a=a, basis='cc-pvdz')
cell.verbose = 6
cell.build()
cell.max_memory = 80 * 1024

nk = [2,1,1]
tw = np.array([0.81,0.57,0.27])
kpts = cell.make_kpts(nk, scaled_center=tw)
kmf = scf.KRHF(cell, kpts).density_fit()
kmf.kernel()

I ran this code with 24 cores/96 GB mem, but it still had an OOM issue and got stuck at generating PBCint3c2e_sph. I checked the function 'gen_int3c_kernel' in pbc/df/incore.py and found that it costs a lot of memory when calculating int3c.

Also, when I make the change 'tw = np.array([0.0,0.0,0.0])' or 'tw = np.array([0.1,0.0,0.2])', the code works with no problem.

Edit: using rs_density_fit() can avoid the problem.