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)
• Some recent cuda commit broke the enzyme.jl tests? (#3172)
• CUDA 6.2 regression from lost FMA contraction (#3203)
• Buildkite upload step quite slow (#3223)
• Regression on Julia 1.12 over Julia 1.10 for DifferentialRiccatiEquations.jl (#3236)