ChemPropFeaturizer and PairwiseFeaturizer default n_jobs=4 is an odd default for DataLoader num_workers

[smcolby]

Problem

ChemPropFeaturizer (chemprop.py:121) and PairwiseFeaturizer (pairwise.py:99) both default to n_jobs=4, which is passed directly as num_workers to a PyTorch DataLoader.

# chemprop.py
n_jobs: int = 4  # → DataLoader(num_workers=self.n_jobs, ...)

# pairwise.py
n_jobs: int = 4  # → DataLoader(num_workers=self.n_jobs, ...)

The value 4 is arbitrary: it is too low to be "use all available cores" and too high to be "single-threaded". This can cause surprising behaviour:

• On machines with fewer than 4 physical cores, spawning 4 workers adds overhead with no benefit.
• On CI / restricted environments, spawning extra processes can fail or be slow.
• For small molecule sets (e.g., unit/integration tests), multiprocessing overhead exceeds the compute savings.

Expected behaviour

The default should be either:

• 1 — single-threaded, zero overhead, consistent cross-platform behaviour (recommended for a library default), or
• -1 — use all available cores, consistent with the convention used elsewhere in the codebase (DescriptorFeaturizer, FingerprintFeaturizer)

The docstring for ChemPropFeaturizer.n_jobs already says "Number of parallel workers to use, by default 4" — the description is accurate but the default itself should be reconsidered.

Affected files

• openadmet/models/features/chemprop.py — ChemPropFeaturizer.n_jobs
• openadmet/models/features/pairwise.py — PairwiseFeaturizer.n_jobs