pypi onnx 1.19.0
v1.19.0

6 days ago

ONNX v1.19.0 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

IR Version 12

  • Added FLOAT8E8M0 type

ai.onnx Opset 24

  • Added Swish op
  • Added TensorScatter op and updated Attention op for in-place KV cache updates
  • Enabled FLOAT8E8M0 for QuantizeLinear, DequantizeLinear, Cast, CastLike, Constant, ConstantOfShape, Identity, Reshape, Shape, Size, If, Loop, Scan, Flatten, Pad, Squeeze, Unsqueeze, and Transpose.
  • Enabled BF16 for TopK and SplitToSequence.
  • Other
    • Added dependency on ml-dtypes
    • BUILD_ONNX_PYTHON symbol is deprecated (we be removed for 1.20). Please, use ONNX_BUILD_PYTHON instead.

What's Changed

Breaking Changes and Deprecations

Spec and Operator

  • Clarify the axes input of [un]Squeeze to be 1D tensors by @justinchuby in #6888
  • Clarify that variable shadowing is not allowed by @justinchuby in #6955
  • Clarify Mod operator by @cbourjau in #6973
  • Fix typo regarding Attention scale in the spec by @yuanyao-nv in #6984
  • Clarify default value for ratio input of Dropout operator by @robertknight in #7032
  • Correct dtype attribute docs for EyeLike operator by @robertknight in #7031
  • Update float8 table for the Cast op spec by @justinchuby in #7085
  • Document Multi-Device Configuration proto specifications in IR.md by @Copilot in #7056
  • Add FLOAT8E8M0 data type by @yuanyao-nv in #7030
  • Enable float8e8m0 for Q/DQ, and other ops by @yuanyao-nv in #7120
  • Update the saturating behavior for E4M3FNUZ/E5M2FNUZ in Cast and CastLike by @justinchuby in #7130
  • Fix ELU and Softplus operators to support tensors of any shape by @Copilot in #7136
  • Fix Shape operator specification: correct range bounds and document start > end behavior by @Copilot in #7132
  • Fix Attention 3D, reference implementation and c++ expansion by @xadupre in #7142
  • Fix RMS norm function definition by @justinchuby in #7135
  • Fix spec for ReduceSumSquare and other reduce ops when noop_with_empty_axes is set by @Copilot in #7137
  • Add bf16 support to TopK and SplitToSequence by @gramalingam in #7158
  • Add Swish operator by @isdanni in #7172
  • Add TensorScatter op for in-place kv cache update by @yuanyao-nv in #7114
  • Fix Resize operator document by @kcvlex in #6686

Reference Implementation

Utilities and Tools

  • Support set schema inference function in python by @OYCN in #5940
  • Improve model Extractor by @justinchuby in #6920
  • Fix: prefixing of graphs when rename_inputs=False/rename_outputs=False by @KarelZe in #6994
  • Fix Einsum shape inference segfault for scalar inputs by @Copilot in #7055
  • Add support for constructing functions with graph attributes by @Copilot in #7112
  • Make some op-level shape inference functions public by @titaiwangms in #7091
  • Implement saturate_cast in numpy helper by @justinchuby in #7143
  • Add .txtpb as a support text proto format in serialization by @justinchuby in #7161

Build, CI and Tests

Documentation

Other Changes

New Contributors

Full Changelog: v1.18.0...v1.19.0

Don't miss a new onnx release

NewReleases is sending notifications on new releases.