v6.5.1

CUDA v6.5.1

Diff since v6.5.0

Merged pull requests:

• Print every thread's backtrace when a gated test stalls (#3484) (@maleadt)
• Record more benchmark data, and publish it to juliagpu.org/perf (#3485) (@maleadt)
• Comment on pull requests from the JuliaGPU benchmark history (#3487) (@maleadt)
• Redirect the benchmark page to juliagpu.org/perf (#3488) (@maleadt)
• Let the benchmark agent pick benchmark jobs first (#3489) (@maleadt)
• Precompile host code for KernelAbstractions launches (#3490) (@maleadt)
• Keep check_parameter_memory from being invalidated by AcceleratedKernels' CUDA extension (#3491) (@maleadt)
• Bump benchmark-action, to flag benchmarks that start allocating (#3492) (@maleadt)
• Define the KernelAbstractions methods in CUDACore rather than an extension (#3493) (@maleadt)
• Run the toolkit and special CI steps on Julia 1.13 (#3494) (@maleadt)
• Avoid boxed captured variables (#3496) (@maleadt)
• Stop using GPUCompiler's internal type_tag in the relocation test (#3500) (@maleadt)
• Keep CUPTI teardown failures from hiding why a PC sampling session failed (#3501) (@maleadt)
• Define device-only functions as overlays, so sysimages load (#3502) (@maleadt)
• cuDNN: copy views that cuDNN can't address (#3503) (@maleadt)
• cuDNN: add ceil_mode to pooling (#3504) (@maleadt)
• Return memory freed outside of a pool to the driver (#3507) (@maleadt)
• Precompile displaying arrays, and link kernels without a closure (#3508) (@maleadt)
• Handle contention in profiler error and deadlock tests (#3511) (@maleadt)
• Check that cuBLAS can address every matrix argument (#3513) (@kshyatt)
• Fixes and updates for quantum wrappers (#3514) (@kshyatt)
• Adapt to ParallelTestRunner 2.10 (#3515) (@christiangnrd)
• Fix bugs in the cuStateVec and cuTensorNet wrappers (#3517) (@maleadt)

Closed issues:

• CUDA.jl as weak dependency compatibility: cuFFT issue with Atomix.jl (#3486)
• cuDNN: ∇maxpool! errors or returns wrong gradients when x and dx (or y and dy) have different strides (#3497)
• cuDNN: views with negative strides throw instead of being copied (#3498)
• cuDNN: pooling has no ceil_mode (#3499)
• CUDA.jl v6.5.0 OOM during model training (#3512)