0.1.4

fix windows build

0.1.3

New Features

• Add SerializeOptions.max_serialized_size_bytes to hard-cap serialization output size (#3207)
• Show node indexes in pretty_onnx graph output (#3031)
• Add --token support to fillshape for binding symbolic dimension tokens to ranges (#3028)
• Add include_release option to to_svg and to_mermaid (#3020)
• Add fillshape --release-info to emit release metadata independently (#3019)
• Add verbose progress output for fillshape execution (#3012)
• Expose random input generation as onnx_light.onnx.tools.make_random_input (#3009)
• Add Logger to onnx_helpers (#2999)
• Add dim_ranges_from_expressions to infer dimension ranges from equality constraints (#2988)
• Add C++/nanobind shape, axes, and weight metadata tagging and tag-aware Mermaid/SVG rendering (#2959)
• Add runtime progress output for verbose ReferenceEvaluator execution (#2961)
• Add run subcommand to generate random inputs and execute a model from the CLI (#2956)
• Add --verbose [LEVEL] support to fillshape CLI (#2952)
• Add ChaCha20-Poly1305 support for encrypted model I/O (ONNXCRY2) (#2948)
• Add light-only ai.rt DelayedInitializer schema, runtime kernel, and shape inference (#2940)
• Add pre-serialization weight rewrite callback for model save paths (#2944)
• Add python -m onnx_light fillshape command (#2938)
• Implement ParseFromIstream for all proto classes (#2942)
• Add support for a callback function when parsing a model (#2928)
• Add include_inplace option to to_mermaid/to_svg renderers (#2930)
• Add function to record in-place reuse opportunities into node metadata (#2918)
• Guard graph inputs from in-place overwrite with an opt-in flag (#2916)
• Add ChaCha20 raw_data callback example for weight serialization/parsing (#2957)

Improvements

• Update onnxruntime integration and related compatibility changes (#3205)
• Mirror upstream onnx#8157: fix nanobind Python target resolution in cross-compile CMake path (#3217)
• Add C++ backend test for kRelease event in shape inference (#3105)
• Align floor-division simplification with Python // semantics (#3026)
• Propagate Reshape shape-tag metadata backward through producer chains (#3024)
• Reuse the core shape-inference API for fillshape shape-tag metadata writing (#3022)
• Seed weight tags for rank-2 FLOAT graph inputs in shape-tag inference (#3021)
• Fix empty-axis broadcasting and add half-precision RMSNormalization support (#3018)
• Extend ComputeContext with shape-tag release info (#3017)
• Allow Transpose in-place reuse detection when storage size is equal (#3015)
• Move tiny external-tensor inlining from fillshape into ParseOptions (#3011)
• Move inplace_reuse and value-tag helpers into onnx_optim.annotations (#3007)
• Show release, in-place, and shape-tag annotations in fillshape --show output (#3005)
• Fix symbolic shape inference for Reshape shapes built from Unsqueeze(Gather(Shape(...))) (#3003)
• Simplify nested floor divisions such as x//5//2 into x//10 (#3001)
• Slice shape inference now emits symbolic expressions instead of fresh names (#2954)
• Simplify divisible additive offsets for floor and exact division in expressions (#2950)
• Improve to_svg layout: shorten edges and reduce crossings (#2932)
• Accept arbitrary iterables in RepeatedField.extend/__init__ (#2934)
• Implement deleter in ByteSpan and TensorProto (#2926)
• Refactor ComputeInPlaceReuse into a ComputeContext class (#2924)

0.1.2

New Features

• Show node indexes in pretty_onnx graph output (#3031)
• Add --token support to fillshape for binding symbolic dimension tokens to ranges (#3028)
• Add include_release option to to_svg and to_mermaid (#3020)
• Add fillshape --release-info to emit release metadata independently (#3019)
• Add verbose progress output for fillshape execution (#3012)
• Expose random input generation as onnx_light.onnx.tools.make_random_input (#3009)
• Add Logger to onnx_helpers (#2999)
• Add dim_ranges_from_expressions to infer dimension ranges from equality constraints (#2988)
• Add C++/nanobind shape, axes, and weight metadata tagging and tag-aware Mermaid/SVG rendering (#2959)
• Add runtime progress output for verbose ReferenceEvaluator execution (#2961)
• Add run subcommand to generate random inputs and execute a model from the CLI (#2956)
• Add --verbose [LEVEL] support to fillshape CLI (#2952)
• Add ChaCha20-Poly1305 support for encrypted model I/O (ONNXCRY2) (#2948)
• Add light-only ai.rt DelayedInitializer schema, runtime kernel, and shape inference (#2940)
• Add pre-serialization weight rewrite callback for model save paths (#2944)
• Add python -m onnx_light fillshape command (#2938)
• Implement ParseFromIstream for all proto classes (#2942)
• Add support for a callback function when parsing a model (#2928)
• Add include_inplace option to to_mermaid/to_svg renderers (#2930)
• Add function to record in-place reuse opportunities into node metadata (#2918)
• Guard graph inputs from in-place overwrite with an opt-in flag (#2916)
• Add ChaCha20 raw_data callback example for weight serialization/parsing (#2957)

Improvements

• Add C++ backend test for kRelease event in shape inference (#3105)
• Align floor-division simplification with Python // semantics (#3026)
• Propagate Reshape shape-tag metadata backward through producer chains (#3024)
• Reuse the core shape-inference API for fillshape shape-tag metadata writing (#3022)
• Seed weight tags for rank-2 FLOAT graph inputs in shape-tag inference (#3021)
• Fix empty-axis broadcasting and add half-precision RMSNormalization support (#3018)
• Extend ComputeContext with shape-tag release info (#3017)
• Allow Transpose in-place reuse detection when storage size is equal (#3015)
• Move tiny external-tensor inlining from fillshape into ParseOptions (#3011)
• Move inplace_reuse and value-tag helpers into onnx_optim.annotations (#3007)
• Show release, in-place, and shape-tag annotations in fillshape --show output (#3005)
• Fix symbolic shape inference for Reshape shapes built from Unsqueeze(Gather(Shape(...))) (#3003)
• Simplify nested floor divisions such as x//5//2 into x//10 (#3001)
• Slice shape inference now emits symbolic expressions instead of fresh names (#2954)
• Simplify divisible additive offsets for floor and exact division in expressions (#2950)
• Improve to_svg layout: shorten edges and reduce crossings (#2932)
• Accept arbitrary iterables in RepeatedField.extend/__init__ (#2934)
• Implement deleter in ByteSpan and TensorProto (#2926)
• Refactor ComputeInPlaceReuse into a ComputeContext class (#2924)

0.1.1

Almost full parity with onnx (onnx.hub is not supported as deprecated). It is fully implemented in C++ included the backend test (+2000) and the reference evaluator (used to produce the backend test). The first release includes the wheels and C++ binaries. The reduced wheels are the package (only linux) without the backend tests and the kernels (~12Mb on linux).

• documentation
• wheel and binary sizes: onnx-light can be released with or with the backend test and kernels.
• number of backend test per operator
• onnx-light backend test coverage: evaluation of onnxruntime, onnx reference evaluation and onnx-light reference evaluation on onnx-light backend test
• onnx backend test coverage: same evaluation on the original backend tests
• backend tests on shape inference: evaluate shape inference algorithme on two dozens of examples, it compares the inferred shapes for the intermediates results

The results are generated for the current main branch not necessarily this release. ir-py unit tests are passing with branch https://github.com/onnx/ir-py/tree/xadupre/ol