johanjoensson / impurityModel / 32150669468

94%

Build Type

push

github

Committed by johanjoensson

Commit Message

Add fast regression coverage for the truncate_initial/ARPACK deadlock fix

Two tests, both in test_rank_independence.py:

test_truncate_initial_basis_is_rank_independent builds a basis directly
(no ~18s NiO workload) above truncation_threshold with dense_cutoff low
enough to engage TRLM, calls truncate_initial on COMM_WORLD and COMM_SELF,
and diffs the surviving determinant fingerprint. Uses a seeded random
one-body Hamiltonian rather than a symmetric physical one: an earlier
attempt with the eg/t2g-style Hop from test_manybody_basis.py gave the
ground eigenvector an exact fourfold weight tie, which truncate()'s
cutoff-at-target selection resolves by comparing floating-point scores --
under an exact tie, roundoff from a different MPI reduction order can push
different tied determinants across the cutoff on different rank counts.
That is a pre-existing indeterminacy of ranking within an exactly-
degenerate manifold, not something this fix's scope covers, so the test
avoids it instead of chasing it.

test_scipy_eigensystem_root_driven_matches_serial_with_an_empty_rank drives
the hardened root-driven ARPACK path directly against a column-distributed
matrix where the last rank deliberately owns zero columns whenever
comm.size > 1 -- the empty-rank edge case that has bitten this codebase
before -- and checks the returned low-lying spectrum against a dense
eigvalsh reference.

Verified at -n1 through -n4: both pass, and the file's full 5 tests
(including the two that previously hung at -n3) pass together in 52s.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Coverage Stats

52 of 52 new or added lines in 1 file covered. (100.0%)

303 existing lines in 19 files now uncovered.

27354 of 30049 relevant lines covered (91.03%)

0.91 hits per line