FluxML/Flux.jl v0.16.13

on GitHub

Flux v0.16.13

Diff since v0.16.12

Merged pull requests:

• Fix three DDP correctness issues: PMI guardrails, data padding, unuse… (#2694) (@edopica)
• [tests] add enzyme gradient tests (#2721) (@hardikxk)
• Update SparseArrays requirement to 1.10 (#2727) (@dependabot[bot])
• Update LinearAlgebra requirement to 1.10 (#2728) (@dependabot[bot])
• Update Random requirement to 1.10 (#2729) (@dependabot[bot])
• Bump actions/checkout from 4 to 7 (#2731) (@dependabot[bot])
• Bump julia-actions/setup-julia from 2 to 3 (#2732) (@dependabot[bot])
• Fix AMDGPU Chain(Conv) gradient test for flipped conv weights (#2734) (@Dhruba531)
• Bump JuliaRegistries/TagBot from 1.25.11 to 1.25.13 (#2735) (@dependabot[bot])
• [tests] check Enzyme gradients by default on CPU (#2736) (@CarloLucibello)
• Fix gpu falling back to the CPU on Julia 1.13, and the Reactant GPU tests (#2737) (@CarloLucibello)
• Reactant on CUDA: separate Buildkite job, backend check, wider tests (#2738) (@CarloLucibello)
• Add Attention layer on NNlib's fused attention, deprecate MultiHeadAttention (#2739) (@CarloLucibello)
• Update GPUArrays requirement from 11.2 to 11.2, 12.1 (#2741) (@dependabot[bot])
• Define expm1 for Nil, so outputsize works with elu, celu, selu (#2742) (@CarloLucibello)
• Make recurrent layers type stable (#2743) (@CarloLucibello)
• Reactant: don't reuse a stale compiled training step (#2745) (@CarloLucibello)
• Reactant: docs fixes, benchmark env vars, Metal CI job (#2746) (@CarloLucibello)
• Reactant: fresh Dropout mask at every trainstep! (#2747) (@CarloLucibello)
• Reactant: reject non-Enzyme AD backends, skip the update on a non-finite loss (#2748) (@CarloLucibello)
• Resume a saved optimiser state on a Reactant device (#2749) (@CarloLucibello)
• [tests] Expand Reactant test coverage: layers, losses, optimisers, precision (#2750) (@CarloLucibello)
• Require NNlib 0.9.50 and Optimisers 0.4.10, fix the tests they affect (#2751) (@CarloLucibello)
• CI: fix macOS stalls, cut test and precompile time (#2752) (@CarloLucibello)
• CI: run the Reactant tests in their own job (#2753) (@CarloLucibello)

Closed issues:

• test Enzyme gradient for loss functions (#2444)
• broken AMDGPU support (#2658)
• Audit Open Issues - August 2026 (#2715)
• Please return the old Flux.train!. (#2730)
• outputsize fails with elu, celu, selu on NNlib master: Nil has no expm1 (#2740)
• Reactant support audit: stale compile cache, Dropout, optimiser gaps, test coverage (#2744)