v1.13.0: ONNX weight deduplication, ONNX export and ORT extension

Deduplicate Embedding / LM head weight in the ONNX export

Workaround for a bug in the PyTorch ONNX export that does not deduplicate the Embedding and LM head shared weight: pytorch/pytorch#108342. For small enough models, this results in up to 50% ONNX serialized model size decrease.

• Fix PyTorch tied weights being duplicated in the exported ONNX models by @fxmarty in #1326
• Fix initializer detection for weight deduplication by @fxmarty in #1333

Extended ONNX Runtime support

ONNX Runtime integration now supports Pix2Struct and MPT architectures. Donut now supports IO Binding. Encoder-Decoder models are now supported as well.

• Pix2Struct onnxruntime support by @krathul in #1296
• Add MPT onnx and ORT support by @jiqing-feng in #1161
• Donut iobinding by @IlyasMoutawwakil in #1209
• Add encoder decoder model by @mht-sharma in #851

Extended ONNX export: MPT, TIMM models, Encoder-Decoder

Additionally, the model SAM is now be default exported as a vision_encoder.onnx, and prompt_encoder_mask_decoder.onnx.

• Add MPT onnx and ORT support by @jiqing-feng in #1161
• Adds ONNX Export Support for Timm Models by @mht-sharma in #965
• Add encoder decoder model by @mht-sharma in #851
• Fix SAM ONNX export requirements with transformers 4.32, export vision encoder separately by @fxmarty in #1301

BetterTransformer supports Falcon

• [BetterTransformer] Add falcon to BetterTransformer by @younesbelkada in #1343

Major bugfix: ability to set GPTQ Exllama kernel maximum length in the transformers integration

The function exllama_set_max_input_length from auto-gptq can now be used with Transformers GPTQ models.

• Version bump + add max_input_length to gptq by @SunMarc in #1329

Other changes and bugfixes

• Update version to 1.12.1.dev0 following release by @fxmarty in #1312

• Add GPTQ prefill benchmark by @fxmarty in #1313

• Precise ORTModel documentation by @fxmarty in #1268

• Improve BetterTransformer backward compatibility by @fxmarty in #1314

• Improve ORTModel documentation by @fxmarty in #1245

• Add bitsandbytes benchmark by @fxmarty in #1320

• fix typo in log message by @AAnirudh07 in #1322

• Support customize dtype for dummy generators by @JingyaHuang in #1307

• Fix opset custom onnx export by @mht-sharma in #1331

• Replace mpt to ernie custom export by @mht-sharma in #1332

• Fix BT benchmark script by @fxmarty in #1344

• Add name_or_path for donut generation by @fxmarty in #1345

• send both negative prompt embeds to ORT SDXL by @ssube in #1339

• add vae image processor by @echarlaix in #1219

• add negative prompt test by @echarlaix in #1347

• Add GPT BigCode to the BT documentation by @fxmarty in #1356

• Add BT dummy objects by @fxmarty in #1355

• Add text2text-generation-with-past test for encoder-decoder model by @mht-sharma in #1338

• Fix sentence transformer export by @mht-sharma in #1366

New Contributors

• @krathul made their first contribution in #1296
• @AAnirudh07 made their first contribution in #1322
• @jiqing-feng made their first contribution in #1161
• @ssube made their first contribution in #1339

Full Changelog: v1.12.0...v1.13.0