v0.9.34

NNlib v0.9.34

Diff since v0.9.33

Merged pull requests:

• CompatHelper: bump compat for CUDA in [weakdeps] to 6, (keep existing compat) (#675) (@github-actions[bot])
• CompatHelper: bump compat for cuDNN in [weakdeps] to 6, (keep existing compat) (#676) (@github-actions[bot])

v0.9.33

NNlib v0.9.33

Diff since v0.9.32

Merged pull requests:

• Extending scatter! to work with CUDA sparse arrays (#648) (@alonsoC1s)
• Add Metal queue specification for Buildkite agents (#667) (@CarloLucibello)
• fix: gather and scatter on empty arrays (#668) (@lenianiva)
• revert 648 (extend scatter to CUDA sparse) (#670) (@CarloLucibello)

Closed issues:

• gather is not friendly with matrix of size 0 on GPU (#411)
• Specialized scatter! dispatch for CUSPARSE.CuSparseMatrix (#647)
• cuda extension broken due to missing AbstractCuSparseArray (#669)

v0.9.1

NNlib v0.9.1

Diff since v0.9.0

This release has been identified as a backport.
Automated changelogs for backports tend to be wildly incorrect.
Therefore, the list of issues and pull requests is hidden.

v0.9.32

NNlib v0.9.32

Diff since v0.9.31

Merged pull requests:

• Fix cudnnConvolutionDescriptor deprecation warning in Julia 1.12 (#649) (@Copilot)
• Add GitHub Copilot instructions for NNlib.jl (#651) (@Copilot)
• Generalize batched_vec to support N-D batches (#652) (@Copilot)
• Refactor gelu_tanh to use tanh for compiler pattern matching (#653) (@Copilot)
• Bump julia-actions/cache from 1 to 2 (#656) (@dependabot[bot])
• Bump codecov/codecov-action from 3 to 5 (#657) (@dependabot[bot])
• Bump actions/checkout from 3 to 5 (#658) (@dependabot[bot])
• Bump thollander/actions-comment-pull-request from 1.0.1 to 3.0.1 (#659) (@dependabot[bot])
• Bump julia-actions/setup-julia from 1 to 2 (#660) (@dependabot[bot])
• update min julia version to 1.10 (#661) (@CarloLucibello)
• Bump actions/checkout from 5 to 6 (#663) (@dependabot[bot])
• Bump tibdex/github-app-token from 1 to 2 (#664) (@dependabot[bot])
• fix CI (#665) (@CarloLucibello)
• Metal Extension with fix for tanh_fast (#666) (@CarloLucibello)

Closed issues:

• gelu_tanh should actually use tanh (#640)
• batched_vec for N-D batches (#645)
• Deprecation warning: cudnnConvolutionDescriptor constructor extension during NNlibCUDACUDNNExt precompilation (Julia 1.12) (#646)
• ✨ Set up Copilot instructions (#650)

v0.9.31

NNlib v0.9.31

Diff since v0.9.30

Merged pull requests:

• Bump AMDGPU compat to 2.0 (#643) (@pxl-th)

Closed issues:

• Precise order of operations for convolution (#642)

v0.9.30

NNlib v0.9.30

Diff since v0.9.29

Merged pull requests:

• Use the correct nthreads (#636) (@IanButterworth)
• CompatHelper: bump compat for ForwardDiff in [weakdeps] to 1, (keep existing compat) (#637) (@github-actions[bot])

Closed issues:

• Wrong number of threads used if interactive pool exists. (#634)

v0.9.29

NNlib v0.9.29

Diff since v0.9.28

Merged pull requests:

• CompatHelper: add new compat entry for SpecialFunctions in [weakdeps] at version 2, (keep existing compat) (#630) (@github-actions[bot])
• Add type alias for gelu -> gelu_tanh (#632) (@ToucheSir)
• Don't spawn if: only one iteration or no threads or threading disabled (#633) (@IanButterworth)

Closed issues:

• Issue with change to gelu (#631)

v0.9.28

NNlib v0.9.28

Diff since v0.9.27

Merged pull requests:

• Add Grid Sampling for 3D images. (#627) (@gRox167)
• Adds full gelu without approximation (#629) (@se-schmitt)

Closed issues:

• "Full" gelu without approximation (#628)

v0.9.27

NNlib v0.9.27

Diff since v0.9.26

Merged pull requests:

• Fix NaN gradient in spectrogram (#617) (@pxl-th)
• Use return_type not _return_type to fix an error on nightly (#618) (@mcabbott)
• Add unthunk in conv.jl (#620) (@mcabbott)

Closed issues:

• Test failures with Zygote 0.7 (#619)

v0.9.26

NNlib v0.9.26

Diff since v0.9.25

Merged pull requests:

• Add Atomix.jl v1 compat (#615) (@christiangnrd)