Releases · onnx/onnx · GitHub

ONNX v1.18.0 is now available with exciting new features! We would like to thank everyone who contributed to this release! Please visit onnx.ai to learn more about ONNX and associated projects. Key Updates ai.onnx Opset 23 Attention, Cast, CastLike, Constant, ConstantOfShape, DequantizeLinear, Flatten, Identity, If, Loop, Pad, QuantizeLinear, RMSNormalization, Reshape, RotaryEmbedding, Scan, Shape, Size, Squeeze, Transpose, Unsqueeze IR Version 11 Added FLOAT4E2M1 and multi-device configuration support Relaxed naming requirements (#6652) Python support Support Python 3.13 Experimental support for Python 3.13t (Windows, Mac) Removed support for Python 3.8 wheels for Windows Arm64 Build Minimum protobuf version is upgraded to v25.1 A new option ONNX_BUILD_CUSTOM_PROTOBUF is added for CMAKE (#6495) What's Changed Breaking Changes and Deprecations Remove/raise exception when external file exists during onnx.save by @tonypottera24 in #6497 Remove python 3.8 workflows (python 3.8 is eol) by @cyyever in #6434 Deprecate all type casting functions by @justinchuby in #6639 => scheduled to remove in 1.20 GroupNormalization-18 is now deprecated and replaced by GroupNormalization-23 due to an incorrect definition (#6358) helper.split_complex_to_pairs is now private. Users can consider duplicating the underlying logic for their own use. Spec and Operator Add FLOAT4E2M1 data type by @yuanyao-nv in #6318 Add FLOAT4E2M1 support to relevant operators by @yuanyao-nv in #6283 Fix typo in the GroupNorm description by @yuanyao-nv in #6358 clarify clip with min>max by @AlexandreEichenberger in #6395 Update IR spec to clarify optional input/output availability by @justinchuby in #6435 Fix GlobalLpPool input types by @justinchuby in #6503 Fix reference implementation for TopK by @neNasko1 in #6593 Clarify that FLOAT4E2M1 can be in int32_data by @justinchuby in #6640 Relax naming requirements in IR spec by @justinchuby in #6652 Add Rotary Embedding op to ONNX opset 23 by @shubhambhokare1 in