Troubleshooting and FAQ

Troubleshooting and FAQ

The launcher cannot import MaterialsFramework

Make sure the correct Python environment is active:

conda activate materialsframework-main
python -c "import materialsframework"
export PAIPAI_PYTHON="$CONDA_PREFIX/bin/python"

CUDA is not found

Check:

nvidia-smi
echo $CUDA_VISIBLE_DEVICES

On clusters, load the CUDA/cuDNN modules required by your MLIP environment.

finiteT rejects many interstitial moves

finiteT rejects trials if relaxation moves an interstitial atom to a different reference site. Possible actions:

• increase --interstitial-site-cutoff cautiously;
• improve relaxation stability;
• use local hop moves;
• start from a better relaxed state.

Search mode seems stuck

Check:

ls waiting_pool
ls waiting_work
ls reports
cat counters/fast_count
cat counters/slow_count

Also inspect fast_*.log and slow_*.log.

Prefast logs are very large

prefast_weights.log writes the full weight vector after every learning update. For very long runs, this can become large. It is written only with --prefast-diagnostics full. For ordinary runs, keep the default --prefast-diagnostics summary; use updates when you need per-active-feature weight changes.

What does prefast change physically?

Nothing in the final acceptance criterion. During warmup, prefast only learns from ordinary single-trial proposals. After warmup, prefast only changes which trial is sent to the fast worker first. The slow-worker relaxed energy remains the source of truth.

Which file should I resume from?

Use an accepted state directory under mcprocess/, or point to the mcprocess directory itself:

paipai --mode finiteT --resume-state search/mcprocess

Use SAVE, not REFERENCE_SAVE, as the accepted reference state. The launcher handles this automatically.