Molecular DF (.density_fit()) fails in cart2sph for lanthanide basis sets with g functions (auxbasis reaches l=8) — molecular analogue of #761

[rhjvanworkum]

Describe the bug

Molecular density-fitted SCF (.density_fit()) crashes with RuntimeError: failed in cart2sph kernel for lanthanide elements whose orbital basis contains g functions (l=4) — e.g. Eu with def2-TZVP. The auto-generated DF auxiliary basis for such a system reaches l=8, and the molecular DF path (gpu4pyscf/df/int3c2e.py) calls cart2sph(..., ang=lk) with that high angular momentum, which the cart2sph CUDA kernel cannot handle.

This appears to be the molecular-path analogue of #761, which fixed the same "lanthanide g-shells → auxbasis reaches l=8" problem for the periodic path by raising L_AUX_MAX. The molecular DF path seems not to have received an equivalent fix, and instead of a graceful assert it fails inside the CUDA kernel.

• CPU PySCF runs the identical system fine.
• GPU without density fitting (direct SCF) also gets past this point (no cart2sph error) — so the gap is specific to the DF integral transform, consistent with the core DFT path supporting high l (up to i, Support I orbital in DFT #340).

Minimal reproducer

import re
import basis_set_exchange as bse
from pyscf import gto
from gpu4pyscf.dft import rks

# PySCF's built-in library has no Eu def2-TZVP, so fetch it (basis + ECP) from BSE.
nw = bse.get_basis("def2-tzvp", elements=["Eu"], fmt="nwchem", header=False)
m = re.search(r"(?im)^\s*ECP\s*$", nw)
basis = {"Eu": gto.basis.parse(nw[: m.start()])}
ecp = {"Eu": gto.basis.parse_ecp(nw[m.start():])}

mol = gto.M(atom="Eu 0 0 0", basis=basis, ecp=ecp, spin=7, verbose=4)
# orbital lmax = 4 (g); auto DF aux lmax = 8
mf = rks.RKS(mol, xc="pbe").density_fit()
mf.kernel()   # -> RuntimeError: failed in cart2sph kernel

(The same failure occurs for a real Eu coordination complex at wB97M-V/def2-TZVPD — that was the original context; the single atom is the minimal trigger.)

Traceback

File ".../gpu4pyscf/df/df.py", line 282, in cholesky_eri_gpu
File ".../gpu4pyscf/df/df.py", line 339, in _cderi_task
File ".../gpu4pyscf/df/int3c2e.py", line 1527, in get_int3c2e_slice
    int3c_blk = cart2sph(int3c_blk, axis=0, ang=lk, out=out)
File ".../gpu4pyscf/lib/cupy_helper.py", line 615, in cart2sph
    raise RuntimeError('failed in cart2sph kernel')
RuntimeError: failed in cart2sph kernel

Diagnostics

For the reproducer:

• orbital basis lmax = 4 (g functions on Eu; all lighter atoms are ≤ f)
• auto DF auxiliary basis lmax = 8 (via pyscf.df.addons.make_auxmol) — the ang=lk passed into cart2sph reaches this value
• CPU PySCF: builds and runs (this system has 762 basis functions, matching the reference)
• GPU direct SCF (no .density_fit()): does not hit the cart2sph error

Expected behavior

Density-fitted SCF should support auxiliary-basis angular momenta produced by lanthanide (g-function) orbital bases — i.e. the molecular DF cart2sph should handle ang up to at least 8, mirroring the L_AUX_MAX fix in #761 for the periodic path (or fall back gracefully).

Environment

• gpu4pyscf-cuda12x 1.7.4
• pyscf 2.13.1
• cupy-cuda12x 14.1.1
• CUDA 12.8, NVIDIA L4 (24 GB)
• Python 3.13, Linux

Related

• Extend L_AUX_MAX to support lanthanide GTH basis sets #761 — same "lanthanide g-shells → auxbasis reaches l=8" issue, fixed for the periodic path by extending L_AUX_MAX.
• Support I orbital in DFT #340 — I-orbital (l=6) support in the core DFT path (which is why the non-DF path works).

[sunqm]

As a temporary solution, you can modify the auxiliary basis set by removing basis functions with l=7 and l=8. For DFT, these high angular momentum auxiliary basis functions typically have negligible contributions to the results.