[Carlos-A-Ramos-G]

Two bugs in PySCFTheory.run_BS_SCF (BS=True) make the broken-symmetry procedure fail with platform="GPU". Neither appears on CPU, which is probably why they went unnoticed.

1. The spin change never reaches the SCF. run_BS_SCF sets self.mol.spin before the high-spin and the BS steps. On GPU, prepare_run does self.mf = self.mf.to_gpu(), which deep-copies the Mole. The SCF then uses self.mf.mol, which keeps the spin it was built with. So the "high-spin" step silently runs at the target (BS) spin, from scratch. In our [4Fe-4S] case (HSmult 19) it ran at Ms = 0 with ⟨S²⟩ = 0 and diverged. Fix: a small helper, set_BS_spin, that sets the spin on self.mol and, when it is a different object, on self.mf.mol.

2. The flipped density is passed as a Python list. After the α/β swap, dm = [dma, dmb] goes to mf.run(dm). On GPU, dma/dmb are cupy arrays, and gpu4pyscf calls cupy.asarray(dm0), which cannot build an array from a list of cupy arrays and raises AssertionError (cupy 14.2). Fix: pass one (2, nao, nao) array, cupy.stack([dma, dmb]) on GPU and np.array([dma, dmb]) on CPU.

Changes

• ash/interfaces/interface_pyscf.py: set_BS_spin helper, used for both spin changes in run_BS_SCF; stacked initial density (+16/−3).
• ash/tests/test_pyscf_broken_symmetry.py (new, CPU only, about 1 s): BS on Li2 stretched to 6 Å (UHF/def2-SVP, HSmult 3, flip atom 1). It checks the spin the SCF actually uses at each step (2, then 0), that the BS step gets a single array, and that ⟨S²⟩ ≈ 1. A second case copies the Mole into mf, the way to_gpu() does, so the GPU failure is covered without a GPU.

Testing

• New tests: both fail on master (spins [0, 0] instead of [2, 0]; list passed to the SCF) and pass with this PR. test_pyscf_scf_convergence.py still passes.
• Real GPU (gpu4pyscf 1.8.1, cupy 14.2, A100), same Li2 system through PySCFTheory(platform="GPU", BS=True, ...):

	HS step: mf.mol.spin, ⟨S²⟩	BS step	result
master	0, 0.000	AssertionError in cupy.asarray	crash
this PR	2, 2.000	0, ⟨S²⟩ 0.989, ndarray passed	E = −14.850294 Eh

• With both fixes applied (plus PySCFTheory: use total density for the point-charge gradient of unrestricted SCF #527), UKS B3LYP/def2-SVP BS-DFT/MM MD of a 36k-atom [4Fe-4S] enzyme (APS reductase) runs on GPU. The SCF gives ⟨S²⟩ = 90.09 at HS (pure S = 9: 90.00) and 8.18 at BS.

Note for the test molecule: H2 can't be used on GPU. Its triplet has no β electrons, and gpu4pyscf's get_occ logging then fails (TypeError: must be real number, not NoneType). That's why the test uses Li2.

🤖 Generated with Claude Code

[RagnarB83]

Good catch, indeed this was never tested on the GPU