NPDM using extremely large amount of memory

[cfengno1]

Dear Huanchen,

I am trying to measure the charge density using the following code. But it using an unexpectedly large amount of memory (more than 700G), although the MPS file is only 14G (see attached) and the DMRG sweep only uses 180G.

Can you help?

Thank you so much!

file.txt

cpu    mem    io │ maxrss   maxvm │ tasks  procs │ command
   %      %     % │    GiB     GiB │              │
══════════════════╪════════════════╪══════════════╪════════
91.6  100.0  99.5 │  748.5  1529.8 │   192      1 │ python
     
Summary:  
     
    Elapsed time:   1%   (0.1 out of 12.0 h timelimit)
             CPU:   1%   (1.0 out of 192 physical CPU cores)
Max. main memory: 100% (750.0 out of 750.0 GiB min. available per node)

from pyblock2.driver.core import DMRGDriver, SymmetryTypes, MPOAlgorithmTypes
import numpy as np

Lx = 24 
Ly = 6 
L = Lx * Ly
n_elec = 8/24

N_ELEC = int(np.round(L*n_elec))

TWO_S = 0
Nthreads = 192
Nmem = round(Nthreads/8) * 32

MPS_dir = "./MPS_M6000" 

driver = DMRGDriver(scratch=MPS_dir, symm_type=SymmetryTypes.SU2 | SymmetryTypes.CPX , n_threads=Nthreads, stack_mem=Nmem << 30)
driver.initialize_system(n_sites=L, n_elec=N_ELEC, spin=TWO_S)
mps = driver.load_mps(tag="KET", nroots=1)

M = mps.info.bond_dim
row = driver.get_npdm(mps, npdm_expr='n', mask=[0])[0]

[hczhai]

npdm_expr='n' is not valid in the SU2 mode. For electron occupation at each site, you need use driver.get_1pdm(mps, mask=[0, 0]).