v4.1.0
📦 sentence-transformersView on GitHub →
✨ 5 features🐛 1 fixes🔧 2 symbols
Summary
Version 4.1.0 adds ONNX and OpenVINO backends for CrossEncoder, a new `backend` argument, and utilities for model optimization, while remaining backward compatible.
Migration Steps
- Install the new version using the appropriate extra, e.g., `pip install sentence-transformers[onnx]==4.1.0` for ONNX support.
- If you want to use the new backends, pass `backend="onnx"` or `backend="openvino"` when constructing a `CrossEncoder`.
- After first export, push the exported model to the hub or save it with `model.save_pretrained` to avoid re‑export on subsequent loads.
- Update any custom code that relied on implicit model loading to handle the new `model_kwargs` passed to the underlying Optimum model if needed.
✨ New Features
- Added ONNX and OpenVINO backends for CrossEncoder with a new `backend` argument (values: "torch" (default), "onnx", "openvino").
- Enabled automatic export of ONNX/OpenVINO models when not present in the repository.
- Added helper function `export_optimized_onnx_model` for ONNX optimization and quantization.
- Improved hard negatives mining strategies.
- Provided pip extras for ONNX GPU, ONNX CPU, and OpenVINO installations.
🐛 Bug Fixes
- Minor improvements and bug fixes.
🔧 Affected Symbols
CrossEncoderexport_optimized_onnx_model