Skip to content

0.12.0

Latest
Latest

Choose a tag to compare

View all tags
@enpasos enpasos released this 09 Feb 19:20
· 2 commits to main since this release
v0.12.0
48c8840

jax2onnx 0.12.0 - Layout controls, opset 23 defaults, and regression hardening

  • NCHW boundary layout support:
    Added inputs_as_nchw / outputs_as_nchw for to_onnx(...) and allclose(...), with layout-optimization docs/tests and transpose-cleanup improvements for Conv-heavy graphs.

  • Depth-to-space and residual-stack coverage:
    Added dm_pix.depth_to_space lowering to ONNX DepthToSpace and expanded NNX regression examples/tests for depth-to-space and nested residual groups.

  • Primitive and IR improvements:
    Added jax.numpy.mean lowering to ReduceMean; fixed symbolic dim_as_value handling; and stabilized dynamic reshape folding used by CLIP/MaxText exports.

  • ONNX opset 23 path for attention models:
    Added opset >= 23 RotaryEmbedding/Attention support and made opset 23 the default in to_onnx(...).

  • Gather/scatter regression fixes:
    Fixed scatter-add broadcast window handling and issue #52 lowering edge cases; fixed gather indexing and vmap(dynamic_slice_in_dim) gather lowering regressions.

  • Compatibility refresh:
    Expanded tested Python versions to 3.11-3.14 and updated runtime dependency floors (onnx, onnxruntime, dm-pix) for the new paths.

Assets 2
Loading