[NCCU-Schultz-Lab]

Cluster-workflow follow-ups from M-BATCH2 (B2.1, B2.3, B2.5, B2.6), a fix for batch checkpoint resume, and GPU offload for frequency jobs.

Changes

SLURM batch scripts (B2.1)

• Scripts ask for --ntasks=1 --cpus-per-task=N (one multi-threaded worker) instead of N tasks. An inherited SLURM_CPUS_PER_TASK no longer sets the thread count.
• New optional settings QUANTUI_SLURM_GRES, QUANTUI_SLURM_QOS and QUANTUI_SLURM_ACCOUNT add #SBATCH lines. They are validated before the job folder is written.
• apptainer exec --nv is used only when the GRES asks for a GPU. QUANTUI_SLURM_NV=1/0 overrides that choice.
• Inside the image, the worker runs plain python from the image's PATH, not the submitting process's interpreter path.
• The quantui-batch launcher port and its parity tests follow the same rules.

Batch geometry optimization results (B2.3)

• result.json now records final_molecule, the dipole moment and Mulliken charges.
• --from chaining uses the recorded geometry.

Pre-optimization before Frequency (B2.5)

• It now converges to 0.01 eV/Å instead of 0.05.
• If the frequencies still show an imaginary mode below 20i cm⁻¹ (typically a methyl rotor at its eclipsed saddle point), QuantUI displaces along the mode, re-optimizes and repeats the frequencies once. This applies in the app and in batch jobs.
• The retry only runs after a pre-optimization that succeeded. If it fails, the first result is kept.
• In batch, the retry resumes after a resubmission.
• The result card, the History card and spectra.ir.imaginary_mode_retry record what was done.

Batch checkpoint resume (fix)

• Since job folders got one attempt folder per run, the worker's checkpoint and cached pre-opt geometry were kept inside the attempt folder.
• A resubmitted job never resumed: killed optimizations, PES scans and frequency runs started over.
• Both now live in the job folder.

Lazy import quantui (B2.6)

• Every public name is resolved on first use.
• import quantui, quantui.cli and quantui.backends.* no longer load numpy, RDKit, ASE, Plotly, PySCF, IPython or requests.
• import quantui went from 0.44 s to 0.01 s here.

Frequency jobs on the GPU

• The reference SCF and the analytical Hessian are now offloaded like single points and optimizations.
• If the GPU Hessian fails, it is recomputed on CPU.
• After the Hessian, the IR/Raman steps and thermochemistry use the CPU object as before.

Testing

• pytest -m "not network": 3810 passed, 19 skipped, 0 failed.
• ruff and black: clean.
• mypy, run as in CI: clean.
• New tests cover:
  • script directives and the --nv decision, including parity with the launcher;
  • the geometry-opt payload;
  • the frequency pre-opt threshold;
  • the saddle retry, including its failure and resume paths;
  • checkpoint resume across attempt folders (this test fails on the old code);
  • the lazy-import guarantee;
  • frequency GPU offload with a stand-in GPU object.
• Real worker run, toluene RHF/STO-3G with pre-opt: the first pass gave 39.4i cm⁻¹; after the retry the lowest mode is 10.1 cm⁻¹ and no imaginary modes remain.

Not checked yet (needs cluster hardware):

• GPU and CPU jobs on NCShare with the new directives.
• The Lab 4 MoO2(acac)2 frequency job on an H200, including whether the Mo ECP Hessian runs on the GPU.
• Resume after a resubmission.
• quantui submit --help timing on the login node.

Authorship

• Claude (Opus 5.5): code edits, review, and conceptual discussion
• Jonathan Schultz: overall vision, planning, review, and orchestration