v0.7.2

What's Changed

Optimizer and rewriter

• Fix constant folding dropping subgraph initializers when inlining If branches by Justin Chu (@justinchuby) with @Copilot in #2825
• Improve Reshape and Slice folding by Justin Chu (@justinchuby) in #2807

Torch Lib

• [torchlib] Add prims::remainder by Justin Chu (@justinchuby) in #2953
• Generate new opsets up to 27 by Justin Chu (@justinchuby) in #2958
• Fix incorrect rank handling in aten_repeat_interleave_self_int ONNX export #2932 by pratik wayase (@PratikWayase) in #2936
• Export aten::relu6 as a single Clip op by Justin Chu (@justinchuby) with @Copilot in #2987
• Fix FFT export truncation by preserving dft_length in rFFT lowering by Justin Chu (@justinchuby) with @Copilot in #2991
• Fix invalid ONNX graph from None start/end in aten_slice_scatter by fruet (@gabrielfruet) in #3006
• Honor the dtype argument of aten::mean when exporting without dim by om singhal (@Om-singhaI) in #3009
• Fix dynamo export for scalar mul on converted real tensors by Justin Chu (@justinchuby) with @Copilot in #3005
• Materialize the DFT axis default when converting opset 19 to 20 by Mohammed Alkindi (@MohammedAlkindi) in #3023
• Fix aten_isclose handling of infinities and equal_nan by om singhal (@Om-singhaI) in #3026
• Fix float16 trunc division rounding by Cpthimself (@Alphaxiaoteng) in #3025
• Add converter for aten::_grouped_mm.default by Sid (@Sid-V5) in #2805
• Fix aten_roll for negative dims and shifts past the dimension length by om singhal (@Om-singhaI) in #3024
• [torchlib] Reimplement as_strided without an ONNX loop by Justin Chu (@justinchuby) with @Copilot in #2928
• Skip GQA key/value repeat when query and key head counts are equal by Justin Chu (@justinchuby) with @Copilot in #2957
• Fix SplitToSequence with scalar uneven split producing incorrect equal-split output by Justin Chu (@justinchuby) with @Copilot in #2858
• Fix ReshapeReshape initializer name collisions when resolving shared -1 shapes by Justin Chu (@justinchuby) with @Copilot in #2947
• [torchlib] add aten::bincount ONNX lowering and bool-mask aten::index_put fix with e2e coverage by Justin Chu (@justinchuby) with @Copilot in #2921
• Set kernel_shape attribute on Conv/ConvTranspose in torchlib by Ti-Tai Wang (@titaiwangms) with @Copilot in #2972
• Disable Publish Docs workflow on forks by Takeshi Watanabe (@take-cheeze) in #2980
• Fix aten_linear rank mismatch for 1D weight (missing squeeze) by fruet (@gabrielfruet) in #2984
• Improve integer aten_floor_divide export to drop Sign ops by Justin Chu (@justinchuby) with @Copilot in #2960
• Fix full_like Dynamo export when memory_format is passed by Justin Chu (@justinchuby) with @Copilot in #2994
• Use the query dtype's lowest finite value for the causal attention mask by Subhajit Mitra (@SubhajitMitra-GH) in #2992

Other changes

• Accept memory_format kwarg on rand/randn/randint _like ops by fruet (@gabrielfruet) in #3011
• Emit scalar Range bounds in aten_unfold by om singhal (@Om-singhaI) in #3013
• docs(_framework_apis): say check_model is a deliberate no-op by Mohammed Alkindi (@MohammedAlkindi) in #3012
• Fix arange and logit behavior for torch-nightly CI by Justin Chu (@justinchuby) with @Copilot in #2974
• Fold Gemm beta into the fused BatchNorm bias by Mohammed Alkindi (@MohammedAlkindi) in #3027
• Fix aten_pow_scalar type promotion when the exponent is not floating point by om singhal (@Om-singhaI) in #3036
• Update Codecov test-results upload to supported action path by Justin Chu (@justinchuby) with @Copilot in #2970
• Fix protobuf serialization failure in ort_fusions ort_run for large models by Justin Chu (@justinchuby) with @Copilot in #2955
• Pin GitHub Actions to full-length commit SHAs by Dan Fiedler (@danfiedler-msft) in #2995

New Contributors

• pratik wayase (@PratikWayase) made their first contribution in #2936
• fruet (@gabrielfruet) made their first contribution in #2984
• Subhajit Mitra (@SubhajitMitra-GH) made their first contribution in #2992
• Dan Fiedler (@danfiedler-msft) made their first contribution in #2995
• om singhal (@Om-singhaI) made their first contribution in #3009
• Mohammed Alkindi (@MohammedAlkindi) made their first contribution in #3012
• Cpthimself (@Alphaxiaoteng) made their first contribution in #3025
• Sid (@Sid-V5) made their first contribution in #2805

Full Changelog: v0.7.1...v0.7.2