What's Changed

• Add quantize/compile support for ~1.9x GPU speedup by @maxwbuckley in #342
• Feature/prompt compression by @Ingvarstep in #345
• Feature/serving by @Ingvarstep in #346
• Load from_pretrained at target dtype; restrict quantize= to int8 by @maxwbuckley in #348
• Implement a proper gliner client, fix bugs by @Ingvarstep in #352
• Fix 35 pre-existing test failures; add CI workflow by @maxwbuckley in #351
• Add variant= to from_pretrained for selective half-precision downloads by @maxwbuckley in #354
• Add low_cpu_mem_usage=True to from_pretrained for ~2x faster cold starts by @maxwbuckley in #355
• Fix/train eval by @Ingvarstep in #356
• Add github workflow for automatic PyPI releases by @Ingvarstep in #357

Full Changelog: v0.2.26...v0.2.27

What's Changed

• Vectorize CPU-path preprocessing and decoding hot loops by @maxwbuckley in #333
• Vectorize relation decoding: replace BRC .item() calls with torch.where by @maxwbuckley in #335
• Batch-level span decoding: reduce CUDA kernel launches from B*8 to ~8 by @maxwbuckley in #334
• Restructure the repo / update documentation by @Ingvarstep in #336
• Fix normalization factor for mean loss reduction by @Ingvarstep in #337
• Improve GLiNER-relex architecture by @Ingvarstep in #338
• fix relex model decoding by @Ingvarstep in #340
• Fix onnx formatting by @BryanBradfo in #341

New Contributors

• @maxwbuckley made their first contribution in #333
• @BryanBradfo made their first contribution in #341

Full Changelog: v0.2.25...v0.2.26

What's Changed

• prepare documentation deployment by @Ingvarstep in #310
• Fix: move apex.amp import to avoid crash when importing GLiNER by @tannonk in #313
• Fix ONNX model loading when torch file is absent by @Ingvarstep in #317
• loss fc fixed by @BioMikeUkr in #322
• Introduce new token-level architectures by @Ingvarstep in #321
• Add handling of empty strings by @Ingvarstep in #325
• fix(onnx): use explicit max_width in SpanMLP.view() for ONNX export by @ajroetker in #318
• Fix training issues and compatability with transformers v4.x and v5 by @Ingvarstep in #326
• Feature/candidate labels by @Ingvarstep in #328
• Update README.md by @m-newhauser in #330

New Contributors

• @tannonk made their first contribution in #313
• @ajroetker made their first contribution in #318
• @m-newhauser made their first contribution in #330

Full Changelog: v0.2.24...v0.2.25

What's Changed

• Fix load by @urchade in #309

Full Changelog: v0.2.23...v0.2.24

What's Changed

• fix masking interface for proper training span models by @Ingvarstep in #296
• Add inference-time sequence packing support by @vivekkalyanarangan30 in #292
• resolved packing params not propagating; added support for other backbones by @vivekkalyanarangan30 in #301
• Refactoring GLiNER, making it more modular and easier to use by @Ingvarstep in #305
• update transformer version by @urchade in #306

New Contributors

• @vivekkalyanarangan30 made their first contribution in #292

Full Changelog: v0.2.22...v0.2.23

What's Changed

• jieba3 added by @oleksandrlukashov in #269
• Update requirements.txt to not use onnxruntime-gpu for mac by @anny-lanai in #271
• Add StanzaWordsSplitter for language-aware tokenization by @anthonyyazdani in #276
• Implement Assymetric loss by @Ingvarstep in #272
• Fixed ALL_LAYERNORM_LAYERS bug by @dcrey7 in #278
• Add generation with constraints by @Ingvarstep in #283
• loss reduction argument fixed by @BioMikeUkr in #286
• Labels decoder by @Ingvarstep in #282
• Fix Cython modules usage by @Ingvarstep in #287
• Converts texts given as a single str to list by @jackboyla in #288
• fix SpanBiEncoderProcessor label assignment by @cceyda in #289

New Contributors

• @anny-lanai made their first contribution in #271
• @anthonyyazdani made their first contribution in #276
• @dcrey7 made their first contribution in #278
• @BioMikeUkr made their first contribution in #286
• @jackboyla made their first contribution in #288
• @cceyda made their first contribution in #289

Full Changelog: 0.2.21...v0.2.22

What's Changed

• Add return index for relation extraction by @rafmacalaba in #260
• Multilang splitter by @oleksandrlukashov in #265
• Fix multi-gpu training for both span and token-level models by @Ingvarstep in #266

New Contributors

• @oleksandrlukashov made their first contribution in #265

Full Changelog: v0.2.20...0.2.21

What's Changed

• Update relation_extraction.py by @rafmacalaba in #255
• add cache_dir parameter to AutoConfig, AutoTokenizer by @bit2244 in #250
• add load from config by @urchade in #257
• Add support of Flash DeBERTa by @Ingvarstep in #259

New Contributors

• @rafmacalaba made their first contribution in #255
• @bit2244 made their first contribution in #250

Full Changelog: v0.2.19...v0.2.20

What's Changed

• docs: update README_Extended.md by @eltociear in #239
• add neg sampling by @urchade in #251
• modify trainer by @urchade in #252
• new v by @urchade in #253
• add data_type in config by @urchade in #254

Full Changelog: v0.2.17...v0.2.19

What's Changed

• Fix bug in text classification evaluation by @seanbenhur in #228

New Contributors

• @seanbenhur made their first contribution in #228

Full Changelog: v0.2.16...v0.2.17