Skip to content

EXAMPLE_VALIDATION_OPERATIONS.md

Latest commit

 

History

History
168 lines (131 loc) · 7.25 KB

EXAMPLE_VALIDATION_OPERATIONS.md

File metadata and controls

168 lines (131 loc) · 7.25 KB
Raw

Example Validation Operations

版本: 2026-07-23 日期: 2026-07-23 状态: 维护 责任人: ATST-Tools maintainers

Scope

This guide is the sole active developer entry for maintaining repository examples, curated outputs, and execution evidence. It is not a user quick start: users should begin with the runnable learning paths in examples/README.md.

Keep the example README limited to ordinary prerequisites, configuration validation, workflow commands, and scientific limitations. Site setup, submission patterns, output provenance, and release-facing evidence belong here. The SAI partition names, QOS names, module names, and job IDs below are historical evidence for the recorded runs, not public requirements. Reconfirm current site policy and available resources before submitting new work.

Local maintainer checks

Start by checking that a changed configuration parses and that its selected workflow can be planned without launching a calculator:

atst config validate examples/06_relax_H2-Au/config.yaml --print-normalized
atst run --dry-run examples/06_relax_H2-Au/config.yaml

The short DP configurations used for P0/P1 interface coverage are:

cd examples/02_neb_H2-Au
atst run config_two_stage_dp.yaml

cd ../10_irc_H2
atst run config_descent_dp.yaml

cd ../12_ccqn_H2-Au
atst run config_auto_modes_dp.yaml

Their ABACUS counterparts are config_two_stage.yaml, config_descent.yaml, and config_auto_modes.yaml. Curated outputs for these cases live under the respective outputs/ directories. For ordinary NEB, the main 01/02 examples and example 13 use two_stage: true with stage1_fmax: 0.20 and stage1_steps: 20; the short configurations retain their recorded step limits.

Use the normal unit and example checks appropriate to the changed surface. The authoritative P0/P1 execution evidence is P0_P1_EXAMPLE_RUNTIME_VALIDATION_2026-05-28.md.

SAI and site-specific execution

Image-level NEB parallelism requires an MPI-aware Python environment. The recorded SAI pattern loaded abacus/LTSv3.10.1-sm70-auto, activated atst-neb-mpi, and used one outer Python rank per active image:

module load abacus/LTSv3.10.1-sm70-auto
conda activate atst-neb-mpi
cd examples/01_neb_Li-Si
mpirun -np 3 atst run config_parallel_smoke.yaml

cd ../03_autoneb_Cy-Pt
mpirun -np 4 atst run config_parallel_smoke.yaml

The outer mpirun schedules ASE images. calculator.abacus.mpi controls the ABACUS subprocess count per image; keep it at 1 for an initial parallel check unless the allocation is sized for nested MPI. Place the outer launch in the site's submission script, or use the launcher that provides an mpi4py world. Design constraints and validation evidence are in MPI4PY_ASE_NEB_PARALLEL_ATST_SUMMARY_2026-05-27.md.

The dedicated Cy-Pt image-parallel historical submission patterns were:

cd examples/13_neb_parallel_Cy-Pt
sbatch submit_huge_gpu.sbatch

cd ../14_autoneb_parallel_Cy-Pt
sbatch submit_rush_gpu.sbatch

Example 13 recorded five interior images using the 8V100V0 partition and huge-gpu QOS, with outer mpirun -np 5, calculator.abacus.mpi: 4, and an isolated inner launcher. Example 14 recorded n_simul: 4 using rush-gpu, outer mpirun -np 4, and calculator.abacus.mpi: 1. Both use calculator.abacus.omp: 8. Their input thresholds are respectively stage1_fmax: 0.20, stage1_steps: 20, fmax: 0.12; and fmax: [0.20, 0.20].

For experimental DMF staging, run the recorded P3 submission entry only after the local configuration checks and current site-resource review:

cd examples/18_dmf_production_validation
export ATST_DMF_ATST=/path/to/atst
sbatch submit_dmf_dp_rush_1gpu.sbatch

That script runs two staged DP-backed cases and writes dmf_validation_report.json with scripts/validate_dmf_candidates.py. The same directory contains the recorded D2S-DMF Sella, vibration, descent-IRC, ABACUS comparison, and refined ABACUS comparison submissions:

  • submit_d2s_dmf_sella_dp_rush_1gpu.sbatch
  • submit_d2s_dmf_sella_vib_dp_rush_1gpu.sbatch
  • submit_d2s_dmf_irc_dp_rush_1gpu.sbatch
  • submit_d2s_dmf_abacus_compare_rush_gpu.sbatch
  • submit_d2s_dmf_abacus_refined_compare_rush_gpu.sbatch

DMF-D2S remains experimental: rough_method: dmf is opt-in and has not replaced the default NEB rough stage. When it is selected, ATST uses the final DMF t_eval grid to choose the tmax neighbourhood for displacement vectors, CCQN references, and automatic vibration indices.

Curated-output provenance

Curated reference values and structures under examples/ are reviewable artifacts, not generic user baselines. The ABACUS records are from SAI 4V100 using ABACUS LTS 3.10.1 with GPU ks_solver: cusolver. The DP records are from SAI 4V100PX using temp_repos/dp_model/DPA-3.1-3M.pt with SHA-256 86dd3a804d78ca5d203ebf98747e8f16dff9713ba8950097ceb760b161e19907 and head Omat24. The model URL and identity are pinned in examples/dp_model_manifest.json; do not add the checkpoint as a normal Git blob.

The completed image-parallel outputs have the following historical provenance:

  • 13_neb_parallel_Cy-Pt/outputs/: job 461967 on 8V100V0; inspect with atst neb summary outputs/neb_parallel_nested_mpi.traj --n-max 5 --strict.
  • 14_autoneb_parallel_Cy-Pt/outputs/: job 462244 on 4V100; inspect with atst neb summary --autoneb-prefix outputs/run_autoneb_parallel_single_gpu.

The retained work directories for the latter runs are under validation_runs/cypt_parallel_e2e_20260528/. Do not overwrite curated inputs or outputs during ad hoc runs; write generated outputs to a disposable or ignored directory, then deliberately curate any accepted replacement.

Relevant active evidence reports are:

Evidence reports and lifecycle

Treat the records in this guide as site-specific historical evidence. When an example, curated output, or execution boundary changes, update the affected reference data, write or revise the appropriate evidence report, and update the active ledger in DOCUMENTATION_STATUS_REPORT.md. Classify the report and follow archival rules in DOCUMENTATION_STANDARDS.md; do not replace a current conclusion by silently editing an old historical report.

For the complete example and release checklist, return to HANDOVER.md. Keep site commands in this guide or evidence reports, never in the example learning map.