v0.7.1

What's Changed

Breaking Changes

The op.op() interface in rewrite rules will now require inputs and attributes to be specified as args and kwargs, instead of a list and a dictionary.

• Unify BuilderBase and GraphBuilder: shared call_op with feature flags by @gramalingam in #2910

Optimizer and rewriter

• Rewriter: fix fuse_relu_clip with None max by @AyoubMDL in #2843

Torch Lib

• Fix incorrect float output dtype in aten_arange_start_step default step by @titaiwangms with @Copilot in #2927
• Set exclude_outside=1 on Resize for antialiased upsample by @ssam18 in #2916
• Add torchvision.ops.deform_conv2d and opset19-specific tests by @f-fuchs in #2923
• Fix aten_gru: add linear_before_reset=1 to match PyTorch GRU semantics by @justinchuby with @Copilot in #2853
• Fix upsample half pixel mode for size one by @PrithviBadiga in #2918
• Fix aten_stft ONNX spec violations by @justinchuby in #2943
• [torchlib] Implement quantize_per_channel and dequantize_per_channel by @justinchuby with @Copilot in #2390

Other Changes

• Bumped version to 0.7.1 by @tadani3 in #2897
• Add more fusion test-cases (part 1) by @gramalingam in #2896
• Add remaining high-priority rewriter extended tests by @gramalingam in #2899
• Remove dead Opset function_defs registration by @gramalingam in #2901
• chore(deps): bump onnxruntime from 1.24.2 to 1.25.1 in /requirements/ci by @dependabot[bot] in #2902
• chore(deps): bump onnx-weekly from 1.22.0.dev20260330 to 1.22.0.dev20260421 in /requirements/ci by @dependabot[bot] in #2898
• Fix nn.PixelUnshuffle wrong channel ordering (replace SpaceToDepth with Reshape→Transpose→Reshape) by @justinchuby with @Copilot in #2892
• Fix two breaking CI failures: reshape-reshape rule test and logsumexp float16 tolerance by @justinchuby with @Copilot in #2907
• Refactor: Introduce OpBuilderBase/TapeBuilder as unified op-building interface by @gramalingam in #2905
• Bump lintrunner mypy dependency to 2.0.0 by @justinchuby with @Copilot in #2912
• Expose build_function and build_graph in public API by @gramalingam in #2913
• Fix fuse_pad_into_conv incorrectly fusing when Pad contains negative values by @justinchuby with @Copilot in #2841
• Fix incorrect dtype default in aten_full causing ONNX shape-inference errors by @titaiwangms with @Copilot in #2926
• Fix aten_masked_scatter for broadcasting masks by @gonultasbu in #2929
• fix: invoke NameFixPass after graph-modifying operations in rewriter, version-converter, and optimizer by @gramalingam in #2922
• chore(deps): bump codecov/codecov-action from 6 to 7 by @dependabot[bot] in #2934

New Contributors

• @ssam18 made their first contribution in #2916
• @f-fuchs made their first contribution in #2923
• @gonultasbu made their first contribution in #2929
• @PrithviBadiga made their first contribution in #2918

Full Changelog: v0.7.0...v0.7.1

v0.7.0

What's Changed

Optimizer and Rewriter

• [Rewriter] Extend list of supported commutative operations by @iksnagreb in #2741
• fix(fuse_batchnorm): support convtranpose + bn fusion with group != 1 by @AyoubMDL in #2879
• optimizer: Prevent constant folding of DynamicQuantizeLinear by @Copilot in #2865
• Unify failure-handling in rewrite-rule by @gramalingam in #2866
• Fix non-deterministic rewriter behavior in multi-output pattern matching by @Copilot in #2880
• Fix BatchNorm fusion producing invalid ONNX when Conv nodes share weight initializers by @Copilot in #2883
• Add fusion rule to remove Expand before broadcast-capable binary operators by @Copilot in #2862
• Use logging instead of print in RewritePass by @tgolsson in #2835
• fix: normalize cache key dtype to prevent initializer name collisions by @gramalingam in #2888

ONNX IR

• Refactor schemas to use the onnx_ir version by @justinchuby in #2797

Torch Lib

• Add converter torch aten::histc by @xadupre in #2796
• Trace single op.SplitToSequence by @titaiwangms in #2817
• Fix bicubic antialias export: use cubic_coeff_a=-0.5 instead of -0.75 by @Copilot in #2849
• [torchlib] Add missing dtype parameter to aten_mean_dim by @linusjuni in #2885
• [torchlib] Fix torchvision::roi_align lowering to accept 7-arg schema by @FraGirla in #2830

Core ONNX Script

• First version of GraphBuilder by @gramalingam in #2814
• Allow GraphBuilder to call script functions by @gramalingam in #2820
• Add GraphBuilder.subgraph() and TensorType.to_ir() for control-flow ops by @gramalingam in #2824
• Support outer scope value reference in script by @gramalingam in #2831
• Expose OpBuilder as onnxscript.OpBuilder by @gramalingam in #2840
• Extend graph construction utility by @gramalingam in #2842
• Implement schema-based input/attribute partitioning in GraphBuilder by @gramalingam in #2837
• Replace input-constraint asserts with TranslationError in converter by @gramalingam in #2845
• Support None as op input in GraphBuilder by @justinchuby in #2868
• Add parent/root tracking to GraphBuilder for subgraph Parameter realization by @gramalingam in #2877
• Add input() and add_output() methods to GraphBuilder by @justinchuby in #2828
• Handling initializers in GraphBuilder by @gramalingam in #2889
• Make GraphBuilder.init use keyword-only args after graph by @gramalingam in #2893
• Trivial cleanup of onnxscript converter by @gramalingam in #2839

New Features

• Add onnxscript.nn module with Module and Parameter classes by @justinchuby in #2819
• Add nn.Sequential: callable ModuleList that chains forward calls by @justinchuby in #2823
• Update nn.Sequential signature to match PyTorch (*modules varargs) by @Copilot in #2827

Other Changes

• Raise version converter error when function attribute is RefAttr by @titaiwangms in #2806
• Update ruff version to 0.15.1 by @justinchuby in #2809
• Add back inliner to version converter by @titaiwangms in #2813
• Remove unused function before version converter by @titaiwangms in #2816

New Contributors

• @FraGirla made their first contribution in #2830
• @linusjuni made their first contribution in #2885
• @tadani3 made their first contribution in #2894

Full Changelog: v0.6.2...v0.7.0

v0.6.2

Note

This patch release includes important bug fixes to version converter.

What's Changed

Optimizer and rewriter

• Fix SlicesSplit rewrite rule by @justinchuby in #2802

Torch Lib

• Fix {lstm,gru}: slice hidden/cell state with sequence by @tgolsson in #2803

Other Changes

• Bump version from 0.6.1 to 0.6.2 by @justinchuby in #2798
• Fix version converter regression by @titaiwangms in #2799

New Contributors

• @tgolsson made their first contribution in #2803

Full Changelog: v0.6.1...v0.6.2

v0.6.1

What's Changed

Note

This update includes a slight change in how PyTorch SDPA's Boolean attention mask is exported in opset<23

• Fix attention mask to use float_lowest instead of -inf and add NaN-safe softmax handling by @Aniketsy in #2654

Other Changes

• Fix compatibility with the latest onnx-ir version (0.1.16) by supporting onnx functions in version converter by @titaiwangms in #2791

New Contributors

• @Aniketsy made their first contribution in #2654

Full Changelog: v0.6.0...v0.6.1

v0.6.0

What's Changed

This release introduces breaking changes to ONNXFunction. The .param_schemas and schema properties are removed, and are replaced by .op_signature which is more flexible and captures the complete signature of the ONNXFunction.

Additionally, onnxscript's AST converter for the scripting mode has been migrated to fully leverage onnx-ir.

Breaking Changes

• Remove ParamSchema by @justinchuby in #2768
• Replace op_schema with op_signature by @justinchuby in #2771

Core ONNX Script

• Fix call to ast.Expression by @justinchuby in #2758
• Remove creation of duplicate value by @gramalingam in #2775
• Remove duplicated return_var by @justinchuby in #2776
• Fix loop body value creation by @justinchuby in #2777

Optimizer and rewriter

• Add test for unused initializer check with graph outputs by @Copilot in #2733
• Add OutputFixPass and NameFixPass in optimize by @justinchuby in #2779

Torch Lib

• Add support for more complex operators by @simonbyrne in #2734
• [torchlib] Consolidate all overloads and prevent new ones from being created by @justinchuby in #2621
• [torchlib] Fix aten__native_batch_norm_legit_functional by @justinchuby in #2753
• [torchlib] Implement signbit by @justinchuby in #2754
• feat: modify aten_bilinear from einsum to matmul by @fw7th in #2746
• [torchlib] Migrate torchvision implementations by @justinchuby in #2569
• [torchlib] Fix linspace implementation for int64 by @Aravind-11 in #2693
• [torchlib] Fix prod and normal.float_float by @justinchuby in #2762
• Fix conversion when enable_gqa is False and dimension are different by @xadupre in #2763
• [torchlib] prims::sum by @justinchuby in #2778
• [torchlib] Fix irfft by @simonbyrne in #2770

Other Changes

• Bump version from 0.5.7 to 0.6.0 by @justinchuby in #2735
• Migrate onnxscript converter to use onnx ir by @gramalingam in #2706
• Added padding_idx=None option and new test cases for aten_embedding_bag by @crypto-a in #2549
• Move converter implementation files to _internal folder by @Copilot in #2738
• Minor cleanup of onnxscript converter by @gramalingam in #2748
• Update onnx_ir and onnx dependency versions by @justinchuby in #2749
• Add read permissions for contents in optional lint by @justinchuby in #2750
• Cleanup SymbolValue by @gramalingam in #2752
• Remove select_ir_versions from public API by @justinchuby in #2739
• Fix unicode issue by @gramalingam in #2747
• Clean up onnxscript/_internal/converter.py by @justinchuby in #2759
• Remove 3rd-party code by @justinchuby in #2760
• Create torch apis 2.11 by @justinchuby in #2767
• [ez] Expose FOLDED_FROM_KEY to onnxscript.optimizer public API by @Copilot in #2773
• Temporarily disable mypy by @justinchuby in #2786
• Support metadata_prop merge and version 25 in version converter by @titaiwangms in #2782

New Contributors

• @simonbyrne made their first contribution in #2734
• @crypto-a made their first contribution in #2549
• @fw7th made their first contribution in #2746
• @Aravind-11 made their first contribution in #2693

Full Changelog: v0.5.7...v0.6.0

v0.5.7

What's Changed

Optimizer and rewriter

• Improve constant folding error messages and allow Identity to skip shape merging by @justinchuby in #2670
• Fix scalar constant check by @gramalingam in #2672
• Capture rewrite rule name as metadata by @gramalingam in #2675
• Keep creating constants when constants are folded inside ir.Function by @titaiwangms in #2679
• Avoid initializer name collision in _fuse_batchnorm.py by @titaiwangms in #2680
• Merge metadata props in rewriter by @gramalingam in #2682
• Implement SDPA via MHA by @gramalingam in #2683
• Don't constant fold Quantize/DequantizeLinear nodes by default by @ruro in #2713
• Fix unused initializer check by @gramalingam in #2732

ONNX IR

• Add option to clear metadata in ort fusion by @gramalingam in #2685

Torch Lib

• feat: implement LSTM and GRU operators for torchlib by @ombrdr47 in #2674
• [torchlib] Fix unbind.int if num_outputs=1 by @sebimarkgraf in #2684
• [torchlib] Fix mod on SymInt by @justinchuby in #2686
• Implement aten.stft by @moatom in #2645
• Add converter for unique_consecutive by @xadupre in #2694
• Add missing output_size kwarg to repeat_interleave by @yuanyao-nv in #2691
• add converter for aten::sym_storage_offset by @xadupre in #2697
• Implement ONNX export for fake_quantize_per_*_affine by @ruro in #2696
• Fix aten_unbind for torch >= 2.7 dynamo export by @afshin-paydar in #2719
• Update aten_index_put implementation by @gramalingam in #2712
• [torchlib] Fix and implement overloads for aten::remainder by @justinchuby in #2727

Documentation

• Utility and example for custom op expansion by @gramalingam in #2701

Other Changes

• Add GQA fusion test cases by @gramalingam in #2669
• chore(deps): bump ruff from 0.14.2 to 0.14.3 in /requirements/lintrunner by @dependabot[bot] in #2676
• chore(deps): bump editorconfig-checker from 3.4.0 to 3.4.1 in /requirements/lintrunner by @dependabot[bot] in #2677
• chore(deps): bump onnx-weekly from 1.20.0.dev20251027 to 1.21.0.dev20251103 in /requirements/ci by @dependabot[bot] in #2678
• Bump version by @gramalingam in #2702
• Provide inplace replacement util by @gramalingam in #2708
• chore(deps): bump ruff from 0.14.3 to 0.14.6 in /requirements/lintrunner by @dependabot[bot] in #2716
• chore(deps): bump actions/checkout from 5 to 6 by @dependabot[bot] in #2715
• chore(deps): bump onnxruntime from 1.23.1 to 1.23.2 in /requirements/ci by @dependabot[bot] in #2652
• chore(deps): bump github/codeql-action from 3 to 4 by @dependabot[bot] in #2626
• chore(deps): bump ruff from 0.14.6 to 0.14.7 in /requirements/lintrunner by @dependabot[bot] in #2721
• support opset23 by @titaiwangms in #2725
• chore(deps): bump actions/upload-artifact from 5 to 6 by @dependabot[bot] in #2730
• chore(deps): bump ruff from 0.14.7 to 0.14.9 in /requirements/lintrunner by @dependabot[bot] in #2731

New Contributors

• @ombrdr47 made their first contribution in #2674
• @moatom made their first contribution in #2645
• @ruro made their first contribution in #2696
• @afshin-paydar made their first contribution in #2719

Full Changelog: v0.5.6...v0.5.7

v0.5.6

What's Changed

Optimizer and rewriter

• Clear initializers in constant folding pass by @justinchuby in #2668

Full Changelog: v0.5.5...v0.5.6

v0.5.5

What's Changed

Breaking Changes

• Create initializers not constant nodes in constant folding pass by @titaiwangms in #2650

Core ONNX Script

• Add support for traced if statements in onnxscript script by @gramalingam in #2644

Optimizer and rewriter

• Add RMS Normalization rule variant by @gramalingam in #2638
• Extend GQA fusion for Qwen by @gramalingam in #2662

Torch Lib

• Unsqueeze unbatched input of avg_pool by @wodesuck in #2646
• Support math trunc by @titaiwangms in #2653
• [torchlib] Fix concat when input tensor has shape (0,) by @justinchuby in #2661

Other Changes

• Extend GQA fusion for Gemma3 by @gramalingam in #2639
• Bump version to 0.5.5 by @titaiwangms in #2640
• Add Gemma3 GQA fusion test case by @gramalingam in #2642
• [Rewriter]: introduce remove_optional_bias by @AyoubMDL in #2635
• Add a verbose mode to torch api for external data save by @justinchuby in #2643
• [version converter] Fix DFT opset 20 by @titaiwangms in #2659
• Declare support for Python 3.14 in pyproject.toml by @justinchuby in #2663

Full Changelog: v0.5.4...v0.5.5

v0.5.4

What's Changed

Optimizer and rewriter

• Fix constant in constant folding by @titaiwangms in #2622
• Create helper for comparing semantic equivalence of shapes by @justinchuby in #2620
• Fix GQA fusion to produce present key/value by @justinchuby in #2634

Torch Lib

• Separated implementation of aten::scatter overloads by @linshokaku in #2605
• Enhanced type annotations and simplified implementation of scatter.value by @linshokaku in #2612
• support for scalar args to aten::scatter by @linshokaku in #2613
• [torchlib] Implement aten_bilinear function using Einsum by @Copilot in #2574
• Simplify aten_unbind when shape is static by @justinchuby in #2597
• Consolidate overloads in torchlib by @justinchuby in #2604
• [torchlib] Fix implementations for bitwise_* overloads by @justinchuby in #2618
• [torchlib] Deprecate Rank and IsScalar by @justinchuby in #2624
• [torchlib] Fix operator add by @justinchuby in #2630
• Remove redundant registration of operator::add and fix sub.Scalar by @justinchuby in #2631

Other Changes

• Update torch api error message to include value names by @justinchuby in #2599
• Remove beartype by @justinchuby in #2603
• Allow opset_version to be set explicitly when exporting by @NoRaincheck in #2615
• Merge shapes only in identity op and nodel-level shape inference by @titaiwangms in #2623

Full Changelog: v0.5.3...v0.5.4

v0.5.3

What's Changed

Optimizer and rewriter

• Fix Onnx 23 Rotary Fusion by @gramalingam in #2576
• Record names of contributing values in the constant folding pass by @justinchuby in #2575
• Merge output shape with input shape instead of override by @wodesuck in #2578
• Extend utilities for checking a scalar value by @gramalingam in #2587
• Merge input and output shape when removing identity by @wodesuck in #2588
• Add NaN handling in softmax pattern in SDPA fusion by @gramalingam in #2593
• Fix collapse slices rewrite rules to handle unknown dims by @justinchuby in #2583
• Expose the should_fold option to optimize() by @justinchuby in #2594

Torch Lib

• [torchlib] Add trace_only flag to aten_copy, aten_tril, aten_triu by @justinchuby in #2572
• [torchlib] Support integers in logical_and/or ops and update other logical ops by @justinchuby in #2582
• [torchlib] Add back operator and/or by @justinchuby in #2590
• Improve aten_floor_divide for int inputs by @justinchuby in #2592

Other Changes

• Remove usages of ir.Input in test by @justinchuby in #2591

New Contributors

• @wodesuck made their first contribution in #2578

Full Changelog: v0.5.2...v0.5.3