Release notes

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Version 1.0.0 (August 27, 2026)

Breaking

• Removed the ::AbstractVector interface to ExaModels. The input x to add_predictor for ExaModels must now be an ExaModels.Variable or ExaModels.Expression. No other input types are allowed. (#324), (#325)

  This is a breaking change only if you used the ExaModels extension, and only if your model did not follow ExaModels best practice and used something like

  import ExaModels, MathOptAIcore = ExaModels.ExaCore(; concrete = Val(true))core, x = ExaModels.add_var(core, 2)x_vec = [x[i] for i in 1:2](core, y), _ = MathOptAI.add_predictor(core, MathOptAI.ReLU(), x_vec)

  instead of

  import ExaModels, MathOptAIcore = ExaModels.ExaCore(; concrete = Val(true))core, x = ExaModels.add_var(core, 2)(core, y), _ = MathOptAI.add_predictor(core, MathOptAI.ReLU(), x)

Other

• Fixed some typos in the documentation (#323)

Version 0.2.9 (August 22, 2026)

Other

• Documentation updates (#315), (#316), (#318), (#319), (#320), (#321)
• Update minimum Julia version to v1.10 (#322)

Version 0.2.8 (August 10, 2026)

Added

• Added SoftPlusEpigraph (#310), (#312)

Other

• Documentation updates (#303), (#305), (#307)
• Bumped actions/setup-python from 6 to 7 (#304)
• Fixed formatting for latest JuliaFormatter (#306)
• Added more tests for ExaModels (#311)

Version 0.2.7 (July 23, 2026)

Added

• Added gray-box support for ExaModels (#300), (#301)
• Add ReLUEpigraph for ExaModels (#294), (#296)

Other

• Added test for LeakyReLU{ReLUEpigraph} (#293)
• Explicitly include predictors in ExaModelsExt (#298)
• Updated printing because of JuMP@1.31.0 release (#299)

Version 0.2.6 (July 3, 2026)

Added

• Added SoftPlusConicEpigraph (#288)

Other

• Updated GitHub actions (#282)
• Updated documentation (#286), (#287)

Version 0.2.5 (June 25, 2026)

Added

• Added ReLUEpigraph (#274)
• Updated to ExaModels@0.11 (#277)
• Enabled build_predictor on activation functions (#278)
• Added an input convex tutorial using ReLUEpigraph (#279), (#280)

Other

• Updated GitHub actions (#269), (#271), (#272)

Version 0.2.4 (March 16, 2026)

Added

• Added support for Flux.gelu (#263)

Fixed

• Don't construct eval_hessian_lagrangian if predictor.hessian=false (#266)

Other

• Added ExaModels examples to the tutorials (#264)

Version 0.2.3 (March 5, 2026)

Added

• Added ; filter keyword argument to replace_weights_with_variables (#257)
• Added ExaModels.jl extension (#259)

Fixed

• Added an error when ::Array passed as input for ::AbstractPredictor (#252)
• Removed build_predictor fallback and simplified add_predictor fallbacks (#253)
• Removed import JuMP from extensions where possible (#260)

Other

• Updated GitHub action versions (#250), (#255), (#261)
• Added ParallelTestRunner.jl (#254)

Version 0.2.2 (January 22, 2026)

Added

• Added support for nn.Dropout (#239)
• Added LayerNorm predictor (#240)
• Added GCNConv and TAGConv predictors (#244)

Other

• Updated printing in docstrings for JuMP@1.29.4 (#241)
• Added checklists to docs (#242)

Version 0.2.1 (January 15, 2026)

Added

• Added support for custom layers in PyTorch (#237)

Version 0.2.0 (January 14, 2026)

This release contains a number of breaking changes. See below for details on how to upgrade from v0.1.19.

Breaking

• Remove ::Matrix method that maps over columns (#223).

  This change means you cannot do MathOptAI.add_predictor(model, predictor, x) where x::Matrix and have it return a y::Matrix corresponding to mapping the predictor over the columns of x.

  To upgrade, replace:

  y, formulation = MathOptAI.add_predictor(model, predictor, x::Matrix)

  with:

  ret = map(1:size(x, 2)) do i    return MathOptAI.add_predictor(model, predictor, x[:, i])endformulation = last.(ret)y = reduce(hcat, first.(ret))

  and write additional code as needed to handle the formulation objects.

• Replace GrayBox by VectorNonlinearOracle implementation (#230)

  We have removed the old GrayBox predictor and replaced it by the VectorNonlinearOracle predictor, which has now been renamed to GrayBox. In addition, the ; vector_nonlinear_oracle = true keyword argument has been removed, and the new ; gray_box = true is equivalent to the old ; vector_nonlinear_oracle = true.

  To upgrade, replace:

  MathOptAI.add_predictor(model, predictor, x; vector_nonlinear_oracle = true)

  with:

  MathOptAI.add_predictor(model, predictor, x; gray_box = true)

• Change config dictionary to have functions as values (#233)

  We have changed how the ; config = Dict() keyword works for the Flux, Lux, and PyTorch extensions. Where previously the values were instantiated AbstractPredictor objects, they must now be constructors that, when called, return an AbstractPredictor object.

  As an example of upgrading, replace:

  ; config = Dict(Flux.relu => MathOptAI.ReLU())

  with:

  ; config = Dict(Flux.relu => MathOptAI.ReLU)# or, alternatively; config = Dict(Flux.relu => () -> MathOptAI.ReLU())

Added

• Add LeakyReLU predictor (#218)
• Add AvgPool2d, Conv2d, and MaxPool2d predictors (#220), (#222), (#224)
• Add support for custom models in Flux (#227)
• Add MaxPool2dBigM predictor (#231), (#235)

Fixed

• Simplify test to fix flakey TestPythonCallExt.test_gelu (#228)

Other

• Remove comment that vector_nonlinear_oracle is experimental (#219)
• Added this changelog (#234), (#236)

Version 0.1.19 (December 5, 2025)

Added

• Add replace_weights_with_variables (#215)

Version 0.1.18 (October 24, 2025)

Added

• Propagate start values through layers (#210)

Other

• Update to JuliaFormatter@2 (#211)
• Make some predictors callable (#212)

Version 0.1.17 (October 23, 2025)

Fixed

• Iterate over PyTorch Sequential module itself, rather than its children (#207)

Version 0.1.16 (October 22, 2025)

Added

• Add GELU predictor (#202)
• Add support for MOI.VectorNonlinearOracle (#204)

Other

• Clarify the problem class of each predictor (#205)

Version 0.1.15 (September 30, 2025)

Added

• Add extension hooks for InfiniteOpt (#200)

Other

• Refactor different ReLU into separate files (#199)

Version 0.1.14 (September 8, 2025)

Fixed

• Fix binary decision tree by making the tolerance default to 1e-6 (#196)
• Fix docs because of a change in Lux (#197)

Other

• Update README.md with MOAI paper (#194)

Version 0.1.13 (July 8, 2025)

Fixed

• Add DimensionMismatch checks for Affine and Scale (#192)

Version 0.1.12 (June 3, 2025)

Added

• Add a reduced_space fallback for extensions (#187)

Other

• Update test for change in MOI@1.40.2 (#186)
• Add Softmax to the PyTorch manual page (#188)

Version 0.1.11 (June 2, 2025)

Added

• Add LinearCombination predictor (#181)
• Add EvoTrees.jl extension (#182)

Other

• Fix docstrings in MathOptAIEvoTreesExt.jl (#183)
• Update to Lux@1 (#184)

Version 0.1.10 (April 14, 2025)

Added

• Allow relaxation of ReLUQuadratic (#178)

Fixed

• Fix duplicate test name (#176)

Version 0.1.9 (April 1, 2025)

Added

• Add VectorNonlinearOracle predictor (#172)

Version 0.1.8 (March 31, 2025)

Added

• Set device in output_size (#170)
• Support SoftMax in PyTorchModel (#173)

Other

• Add paper to README (#169)

Version 0.1.7 (January 17, 2025)

Fixed

• Evaluate torch model to get output dimension (#166)

Version 0.1.6 (November 7, 2024)

Other

• Clarify instructions to load PythonCall (#163)

Version 0.1.5 (November 6, 2024)

Added

• Add gray_box_device option for PyTorchModel (#159)

Fixed

• Fix flakey tests (#161)

Version 0.1.4 (October 25, 2024)

Fixed

• Add bound constraints to the Formulation object (#155)

Other

• Improve the docstrings (#156)
• Fix various typos (#157)

Version 0.1.3 (October 24, 2024)

Other

• Delete .github/workflows/documentation-deploy.yml (#150)
• Improve manual for NN extensions (#152)
• Fix documentation link in README.md (#153)

Version 0.1.2 (October 23, 2024)

Other

• Add DOCUMENTER_KEY (#148)

Version 0.1.1 (October 23, 2024)

Other

• Update installation instructions for registered package (#141)
• Create documentation-deploy.yml (#146)

Version 0.1.0 (October 22, 2024)

Initial release. Too many pull requests to reference.