first lasscf_async combine_o1 3-frag convergence

The problem in the previous commit was actually due to the
evaluation of the convergence tolerance. In the subproblem, the
gradient was lower than the tolerance, but the overall gradient
norm was larger than it, so the iteration just stopped making
progress. Enforce a minimum of 1 cycle through the subproblem
iteration brute-force solves this issue for now.

my_pyscf/mcscf/lasci.py

@@ -910,6 +910,7 @@ def __init__(self, mf, ncas, nelecas, ncore=None, spin_sub=None, frozen=None, fr
         self.ah_level_shift = 1e-8
         self.max_cycle_macro = 50
         self.max_cycle_micro = 5
+        self.min_cycle_macro = 0
         keys = set(('e_states', 'fciboxes', 'nroots', 'weights', 'ncas_sub', 'nelecas_sub',
                     'conv_tol_grad', 'conv_tol_self', 'max_cycle_macro', 'max_cycle_micro',
                     'ah_level_shift', 'states_converged', 'chkfile', 'e_lexc'))

my_pyscf/mcscf/lasci_sync.py

@@ -128,9 +128,11 @@ def kernel (las, mo_coeff=None, ci0=None, casdm0_fr=None, conv_tol_grad=1e-4,
         #    ('LASCI micro init : E = %.15g ; |g_orb| = %.15g ; |g_ci| = %.15g ; |x0_orb| = %.15g '
         #    '; |x0_ci| = %.15g'), H_op.e_tot, norm_gorb, norm_gci, norm_xorb, norm_xci)
         las.dump_chk (mo_coeff=mo_coeff, ci=ci1)
-        if (norm_gorb<conv_tol_grad and norm_gci<conv_tol_grad)or((norm_gorb+norm_gci)<norm_gx/10):
-            converged = True
-            break
+        if (((norm_gorb<conv_tol_grad and norm_gci<conv_tol_grad)
+             or ((norm_gorb+norm_gci)<norm_gx/10))
+            and (it>=las.min_cycle_macro)):
+                converged = True
+                break
         H_op._init_eri_() 
         # ^ This is down here to save time in case I am already converged at initialization
         t1 = log.timer ('LASCI Hessian constructor', *t1)

my_pyscf/mcscf/lasscf_async/combine.py

@@ -133,6 +133,8 @@ def relax (las, kf, freeze_inactive=False, frozen_frags=None):
         flas.__dict__.update (las.__dict__)
         flas.frozen = []
         flas.frozen_ci = frozen_frags
+        # TODO: ensure robust tolerance selection so things always make progress
+        flas.min_cycle_macro = 1
         if freeze_inactive:
             flas.frozen.extend (list (range (las.ncore)))
         for ifrag in frozen_frags:
@@ -196,7 +198,6 @@ def select_aa_block (las, frags1, frags2, fock1):
     gmax = np.argmax (gblk)
     i = frags1[gmax // len (frags2)]
     j = frags2[gmax % len (frags2)]
-    print (i, j, gblk[gmax])
     return i, j
 
 def combine_pair (las, kf1, kf2):