jaketae/ensemble-transformers

Provides task-specific ensemble classes (e.g., `EnsembleModelForSequenceClassification`) that automatically detect and instantiate matching preprocessors for each backbone model, eliminating manual tokenizer management. Supports multi-device distribution via `to_multiple()` to load different models across GPUs, and includes modality validation to prevent mixing incompatible model types. Returns per-model predictions alongside aggregated outputs via mean pooling, enabling flexible inference strategies across heterogeneous Hugging Face transformer architectures.

No commits in the last 6 months. Available on PyPI.