Bump the all-julia-packages group across 1 directory with 5 updates

[dependabot]

Updates the requirements on CUDA, ChainRulesCore, Mooncake, AMDGPU and Enzyme to permit the latest version.
Updates CUDA to 6.2.1

Release notes

Sourced from CUDA's releases.

v6.2.1

CUDA v6.2.1

Diff since v6.2.0

Merged pull requests:

• Update CI (#3175) (@​maleadt)
• Ignore patch in llvm and ptxas compat helpers (#3176) (@​AntonOresten)
• Bump GPUToolbox and remove the ability to reset devices (#3178) (@​maleadt)
• Add narrow-precision floats to cudaDataType (#3180) (@​AntonOresten)
• Bump actions/checkout from 6 to 7 (#3182) (@​dependabot[bot])
• Update to GPUToolbox v3. (#3183) (@​maleadt)
• Skip device-to-device copies of zero-byte arrays. (#3184) (@​maleadt)
• don't promote to i64 unnecessarily in unsafe_cached_load (#3186) (@​simeonschaub)
• Fix cluster barriers silently trapping on Julia <= 1.11 (#3188) (@​AntonOresten)
• Implementing specialized coo_type for CuSparseCOO (#3190) (@​alonsoC1s)

Closed issues:

• rCuFFTPlan not defined (#2509)
• Julia 1.12: Switch to Once (#2544)
• Norm of Hermitian matrices trigger scalar indexing error (#2965)
• LLVM 18 / Julia 1.12: Invalid vectorization leads to ptxas error (#3036)
• LLVM 20 / Julia 1.13: Back-end crash during texture tests (#3037)
• CUDACore - warning during precompilation (#3177)
• Error on CUDA.rand(RGB{Float32}, ...) in CUDA.jl 6.1 and 6.2 (#3179)
• deepcopy fails for CuArray with zero-size element type (#3181)
• Calling findnz(A) with a COO-formatted argument causes a stack overflow (#3189)

Commits

• 2816631 Bump version.
• 1b62bc2 Fix cluster barriers silently trapping on Julia <= 1.11 (#3188)
• 5886f9d Implementing specialized coo_type for CuSparseCOO (#3190)
• 8c6af3b Add narrow-precision floats to cudaDataType (#3180)
• ae647ae Don't promote to i64 unnecessarily in unsafe_cached_load (#3186)
• fbb9098 Skip device-to-device copies of zero-byte arrays. (#3184)
• 9dc5657 Update to GPUToolbox v3. (#3183)
• f4a33a2 Bump actions/checkout from 6 to 7 (#3182)
• bc81d40 Bump GPUToolbox and remove the ability to reset devices (#3178)
• f36a98c Ignore patch versions in LLVM and ptxas compatibility queries (#3176)
• Additional commits viewable in compare view

Updates ChainRulesCore to 1.26.1

Release notes

Sourced from ChainRulesCore's releases.

v1.26.1

What's Changed

• Bump actions/checkout from 4 to 5 by @​dependabot[bot] in JuliaDiff/ChainRulesCore.jl#699
• Bump actions/checkout from 5 to 6 by @​dependabot[bot] in JuliaDiff/ChainRulesCore.jl#700
• Bump codecov/codecov-action from 5 to 6 by @​dependabot[bot] in JuliaDiff/ChainRulesCore.jl#705
• Fix some tests and bump version by @​JamesWrigley in JuliaDiff/ChainRulesCore.jl#706

New Contributors

• @​JamesWrigley made their first contribution in JuliaDiff/ChainRulesCore.jl#706

Full Changelog: JuliaDiff/ChainRulesCore.jl@v1.26.0...v1.26.1

Commits

• 4253bff Fix some tests and bump version (#706)
• dad0d24 Bump codecov/codecov-action from 5 to 6 (#705)
• 004e43e Bump actions/checkout from 5 to 6 (#700)
• 115c282 Bump actions/checkout from 4 to 5 (#699)
• 414b660 Update Project.toml
• 124e0cb Merge pull request #698 from ChrisRackauckas-Claude/update-compat-bounds-2025
• 1d5c651 Update Downgrade.yml
• 6facda3 Create Downgrade.yml
• 632c0dc Update compat bounds to exclude releases >1 year old
• f396d36 Merge pull request #697 from ChrisRackauckas-Claude/bump-julia-v1.10
• Additional commits viewable in compare view

Updates Mooncake to 0.5.45

Release notes

Sourced from Mooncake's releases.

v0.5.45

Mooncake v0.5.45

Diff since v0.5.44

Merged pull requests:

• Unpin JET, Revise, and LoweredCodeUtils; allow JET 0.12 (#1271) (@​sunxd3)

Changelog

Sourced from Mooncake's changelog.

0.5.32

• Fix forward-over-reverse Hessian-vector products on closures that capture a Ref wrapped in a NoTangent-typed aggregate, which previously threw UndefRefError. prepare_hvp_cache now eagerly compiles the inner rrule!! together with its forward-mode dual callables and routes the outer forward pass through a new DerivedFoRRule, so the inner IdDict constructor is no longer inlined past Mooncake's rule (#1193, #1202).
• Fix the gradient of copysign(x, y) with respect to x: the derivative is sign(x) * sign(y), and the missing sign(x) factor previously gave the wrong gradient sign when x < 0 (#1196).
• Handle mutation of non-const globals (setglobal!) in forward mode on Julia 1.12+ (#1194).
• Version-bound the Core._call_latest(CoreLogging.handle_message, ...) rule (and its keyword-argument variant) to Julia below 1.12 (#1200).

0.5.31

• Guard codual_type / fcodual_type against unbound TypeVars (#1191, #1192).

0.5.30

• Bump the LogExpFunctions compat bound to 1 (#1186).

0.5.29

• Add a max_fd_step keyword to TestUtils.test_rule that caps the finite-difference step sizes, keeping perturbations of domain-restricted functions (log, sqrt, cholesky) inside their domains (#1173).
• Pre-allocate and reuse the Hessian, gradient, and basis-direction buffers in the Hessian cache so that repeated value_gradient_and_hessian!! calls avoid allocation (#1178).
• Add consistency checks for rule reuse (#1172).
• Mark CUDACore.cudaError_enum as having no tangent (#1175).

0.5.28

• Throw a clear UnhandledLanguageFeatureException for try / catch blocks in reverse-mode AD instead of an opaque IR-verification failure (#1161).
• Import the ChainRules svd rule via @from_rrule (#1163, closes #670).
• Guard the friendly_tangent_cache array branch against NoTangent element types (e.g. SparseMatrixCSC{Int}) (#1150).

0.5.27

• Add a cached value_and_jacobian!! interface for both forward-mode (chunked) and reverse-mode (row-by-row) caches (#1153).
• Fix tangent_type for Union{NoRData, RData{...}} (#1133).
• Add foreigncall zero-derivative rules for jl_get_world_counter and jl_matching_methods, supporting forward-over-reverse over those calls (#1143).
• Handle Ptr in zero_tangent by delegating to uninit_tangent (#1139).
• Update the CUDA extension for CUDA + cuDNN 6 (#1148).

0.5.26

• Add Config(empty_cache=true) to free internal caches before rebuilding rules.

config = Mooncake.Config(empty_cache=true)
cache = Mooncake.prepare_gradient_cache(sin, 1.0; config)

0.5.25

• Add nfwd: a new N-wide forward-mode implementation built around NDual, with Nfwd / NfwdMooncake internals and broad tests for scalar, array, and rule-building paths.
• Expand Mooncake's forward-mode interface and caching around nfwd, including prepared derivative/gradient cache improvements and lower-allocation hot paths for repeated calls.
• Route a broader scalar-math set through nfwd-backed direct primitive frule!! / rrule!! wrappers, reducing dependence on imported ChainRules rules for these cases.

... (truncated)

Commits

• 1ca59a9 Unpin JET, Revise, and LoweredCodeUtils; allow JET 0.12 (#1271)
• 1839665 Update Project.toml
• 28e0eee Bugfix for broadcasts on in MooncakeCUDA extension. (#1270)
• 3e95d05 README: explain why some rules must be written by hand (#1269)
• cd15108 Fix three tangent-construction crashes hit while debugging SciMLSensitivity #...
• 1da3bd6 Some fixes for rules and typos. (#1268)
• 7e305a9 More concise comments (#1267)
• 57e8150 Better rules for nnlib dropout and scatter (#1264)
• f46746d Restrict nnlib rules' scoping (#1265)
• c5f254f Improve nnlib activation rules (#1263)
• Additional commits viewable in compare view

Updates AMDGPU to 2.7.1

Release notes

Sourced from AMDGPU's releases.

v2.7.1

AMDGPU v2.7.1

Diff since v2.7.0

Merged pull requests:

• Switch to CompilerCaching.jl. (#961) (@​maleadt)
• Group all Julia Github Actions PRs into a single one (#985) (@​luraess)
• Allow Enzyme test to fail (#993) (@​luraess)
• Increase evals from 1 to 4 in the benchmark suite (#999) (@​luraess)
• Support ROCm 7.14+ versioned core-* library layout (#1000) (@​luraess)
• Make versioninfo() robust to version-probe crashes (#1001) (@​luraess)
• don't pick up device libs in current dir during precompilation (#1003) (@​simeonschaub)
• don't return reshaped array from mapreducedim! (#1004) (@​simeonschaub)
• use AMDGPU_LLVM_Backend_jll instead of in-tree backend (#1005) (@​simeonschaub)
• Reduce per-kernel allocations from 12 down to 1 (#1009) (@​pxl-th)
• support DocumenterVitepress 0.3 (#1010) (@​luraess)

Closed issues:

• Trying install AMDGPU on ubuntu 26.04 with no success. (#920)
• Discovery will fail on ROCm 7.14 — libraries move to /opt/rocm/core-<ver>/lib (#986)

Commits

• 81e79cb Bump patch
• 6ee5c7f Merge pull request #1010
• 16bd7d0 Reduce per-kernel allocations from 12 down to 1 (#1009)
• 892611b support DocumenterVitepress 0.3
• a364664 use AMDGPU_LLVM_Backend_jll instead of in-tree backend (#1005)
• e3e5b25 Merge pull request #1001
• 2b21e7a Merge pull request #1000
• 724d15f Merge pull request #1004
• 74626e8 don't return ReshapedArray from mapreducedim!
• f39609f Merge pull request #1003
• Additional commits viewable in compare view

Updates Enzyme to 0.13.198

Release notes

Sourced from Enzyme's releases.

v0.13.198

Enzyme v0.13.198

Diff since v0.13.197

Merged pull requests:

• CompatHelper: bump compat for CMake_jll to 4 for package deps, (keep existing compat) (#3421) (@​github-actions[bot])
• Fix return activity (#3426) (@​wsmoses)

Commits

• 59156d0 Update version to 0.13.198
• 4da56af Fix return activity (#3426)
• 30103be CompatHelper: bump compat for CMake_jll to 4 for package deps, (keep existing...
• 2b1e15a Simplify ref condition in CI workflow
• 47a71dd Bump version from 0.13.196 to 0.13.197
• 2779a94 Mark fieldnames as inactive (#3424)
• 7075a21 Handle poison incomings in nodecayed_phis! all_args prefilter (#3422)
• 8b4526e Bump version from 0.13.195 to 0.13.196
• 9c450ac Handle batched any return type (#3419)
• 18cd701 Mark norm_recursive_check inactive (#3416)
• Additional commits viewable in compare view