[dependabot]

Updates the requirements on cuBLAS to permit the latest version.

Release notes

Sourced from cuBLAS's releases.

v6.3.0

CUDA v6.3.0

Diff since v6.2.1

Merged pull requests:

• Switch to CompilerCaching.jl. (#3185) (@​maleadt)
• Modernize cuDNN wrappers around the backend graph API (#3191) (@​maleadt)
• Restore testing of libraries + fix regressions (#3192) (@​maleadt)
• Make kernel launches allocation-free (on Julia 1.12+) (#3196) (@​maleadt)
• Improve/add low-level API for conversion-free launches (#3197) (@​maleadt)
• Fix PTX selection during precompilation (#3198) (@​maleadt)
• Fix 3-arg QR ldiv! for multi-column right-hand sides (#3199) (@​hexaeder)
• docs: fix some typos (#3201) (@​jeis4wpi)
• Enforce PTX intrinsic target requirements (#3202) (@​maleadt)
• Remove ineffective command-line LLVM argument. (#3204) (@​maleadt)
• Fix enzyme tests (#3211) (@​kshyatt)
• Restrict Crayons to 4.1.x to unblock CI (#3212) (@​kshyatt)
• Harden intrinsic compile-time validation (#3213) (@​maleadt)
• Fix Crayons compatibility. (#3215) (@​maleadt)
• Dump the CUBIN on load failure. (#3216) (@​maleadt)
• Prepare for CUDA 13.4 (#3217) (@​maleadt)
• Don't let ^ accuracy depend on the exponent's width. (#3218) (@​maleadt)
• don't return reshaped array from mapreducedim! (#3219) (@​simeonschaub)
• Delete unnecessary axes() method (#3221) (@​JamesWrigley)
• Enable GPUArrays sparse tests (#3222) (@​christiangnrd)
• Add support for CUDA programmatic dependent launch (PDL). (#3225) (@​maleadt)
• CUDA: Select compiler artifacts independently (#3227) (@​maleadt)
• Fix cuDNN tests importing CUDA instead of CUDACore (#3228) (@​AntonOresten)
• Thread safety fixes (#3229) (@​maleadt)
• cuDNN: support block-scaled (MXFP8/NVFP4) graph ops (#3230) (@​AntonOresten)
• cuDNN: set default pointwise alpha2 to 1 (#3231) (@​AntonOresten)
• cuDNN: support layer and RMS norm modes (#3232) (@​AntonOresten)
• Fix bad typing in Enzyme ext (#3233) (@​kshyatt)
• Restore testing old deps (#3235) (@​kshyatt)
• Reduce kernel launch and first-use latency (#3237) (@​maleadt)
• cuDNN: implement the composite SDPA backward (#3238) (@​AntonOresten)
• Adapt to cuTENSOR 2.6 and 2.7 (#3239) (@​maleadt)
• Narrow the trsm data operand to CUDA arrays. (#3240) (@​maleadt)
• cuDNN: accept F8_128x4 scale tiles on any blocked dimension (#3241) (@​AntonOresten)

Closed issues:

• cuDNN: Provide wrappers for the declarative API (#2266)
• Issue when differentiating a CUDA kernel ( only without KA ) (#2947)
• Re-instate old dependencies test (#2971)
• CUBLAS.getrs_strided_batched! throws MethodError (#3033)
• Scalar indexing when mul! into SubArray (#3041)
• Scalar indexing on mul!(::SubArray, ::Number, ::UniformScaling) (#3044)
• Wrong matrix subtyping in cuSPARSE (#3104)

... (truncated)

Changelog

Sourced from cuBLAS's changelog.

CUDA.jl release history

This document lists the noteworthy changes in every release of CUDA.jl, newest first. It is a hand-written summary aimed at users; for the complete list of merged pull requests, see the GitHub releases.

CUDA.jl follows semantic versioning, so intentional breaking changes are reserved for major releases. Some minor releases nevertheless contain changes that are technically breaking, such as stricter validation of previously invalid uses, corrections to unsupported behavior, or changes to support floors; those are labeled separately below. Patch releases are listed as subsections of the minor release they belong to.

v6.3 (August 2026)

Compiled kernels are now cached by Julia's compiler. Adapting to GPUCompiler 2, which builds on CompilerCaching.jl, CUDA.jl dropped its own per-context dictionary of compiled kernels: the compilation artifacts (the cubin image and the entry-point name) are attached to the CodeInstance that Julia already caches, keyed on the method instance, world age and compiler configuration, while the context-specific CuFunction handles remain session-local and are never serialized. On Julia 1.11 and later that cache is Julia's integrated code cache, so a kernel compiled while a package is being precompiled can travel in that package's image instead of being recompiled in every fresh session; on Julia 1.10 the cache stays session-local (#3185).

Technically breaking changes:

• The CUDA compiler artifacts (ptxas and friends) are now selected independently from the CUDA runtime. CUDA.set_runtime_version! and CUDA.reset_runtime_version! write the version and local preferences for both CUDA_Runtime_jll and CUDA_Compiler_jll, and both UUIDs have to be listed in the project's [extras] section for those preferences to be picked up. The preferences can also be set separately, which is useful to test a compiler release against a fixed runtime; keep both within the same CUDA major version. The local_compiler preference is gone: a local compiler is requested through CUDA_Compiler_jll's local preference, and is discovered with CUDA_Runtime_Discovery. Requires CUDA_Compiler_jll 0.5 and CUDA_Runtime_jll 0.24 (#3227).
• Device intrinsics that need a minimum compute capability now fail at compile time with an explanatory error, instead of generating PTX that ptxas rejects. This affects 16-bit atomic_cas!, dp4a, nanosleep, map_shared_rank and the WMMA operations. The run-time compute-capability checks that used to select a compare-and-swap fallback for Float16, Float64 and BFloat16 atomic addition were removed; the back-end expands those on devices without native support

... (truncated)

Commits

• 2b2b8ae Bump version to v6.3.0.
• b0f48d3 cuDNN: accept F8_128x4 scale tiles on any blocked dimension (#3241)
• 1ee19d2 Harden intrinsic compile-time validation (#3213)
• 039fd82 Add release history.
• 4fa9999 Adapt to cuTENSOR 2.6 and 2.7 (#3239)
• afa2910 Narrow the trsm data operand to CUDA arrays. (#3240)
• 0695881 cuDNN: implement the composite SDPA backward (#3238)
• 12db784 Merge pull request #3237 from JuliaGPU/tb/latency
• a50cf54 Record the CUDA benchmark environment
• ddfdd71 Precompile GPU-dependent initialization and reflection
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

[dependabot]

Superseded by #29.