Troubleshooting

Common install / runtime snags and their fixes.

Install

torch / CUDA mismatch. The core package depends on torch>=2.0 but does not pin a CUDA build. If import torch fails or silently runs on CPU when you expect GPU, install the matching wheel from the PyTorch site before installing dotime.

ModuleNotFoundError: pyarrow. Reading/writing frozen suites (parquet) needs the evaluation extra:

pip install 'dotime[evaluation]'

The 'gdp' encoder backend requires the optional [gdp] extra. The default encoder backend is transformer and runs on CPU with no extra dependency. The gdp (GatedDeltaProduct / TempoPFN) backend is GPU-only and needs:

pip install 'dotime[gdp]'   # requires a CUDA GPU + flash-linear-attention

Unless you explicitly pass backend="gdp", you never need this.

Runtime

DoOverTimePFN baseline needs a trained checkpoint. Pass a checkpoint path:

from dotime import baselines
model = baselines.get("DoOverTimePFN", checkpoint="/path/to/best.pt")

The [models] extra (pfns) must be installed for the model to import.

RuntimeWarning: SCM diverged ... returning zeros. The diverse prior occasionally samples an unstable SCM; those trajectories are zeroed and flagged rather than dropped. The empirical divergence rate is < 1% on the released suites; the warnings are safe to ignore (or filter with warnings.simplefilter("ignore")).

Benchmark cache

load_benchmark caches downloaded suites under ~/.cache/dotime (override with $DOTIME_CACHE or the cache_dir= argument). If a cached suite is corrupt you will see a checksum mismatch error — delete the suite directory and reload with force_download=True.