[maleadt]

Device-only code in the global method table gets compiled for the host by sysimage builds: Julia 1.12+ compiles every method with a compileable signature, and 1.11 keeps native code instances (e.g. those left by Core.Compiler.return_type(initialize_rng_state) in finish_module!). The image then refers to device symbols like julia.gpu.state_getter, and since sysimages are loaded with RTLD_NOW, it fails to load. Same fix as JuliaGPU/KernelAbstractions.jl#930 and JuliaGPU/OpenCL.jl#598, using the shared definitions from GPUToolbox 3.5.

CUDA already marked its runtime in #3249 and host_leaks(CUDACore) was empty, so this is mostly for coherence with the other back ends:

• @device_override, @device_function and @device_functions are one-liners around GPUToolbox.Overlays.device_override / device_function (which accepts blocks); ExprTools is no longer needed.
• finish_module! looks up initialize_rng_state through the job's method tables and calls it as a void function, instead of inferring its return type with the native interpreter, which on 1.11 left native code instances of it behind. initialize_rng_state is now a device function.
• A host leaks test next to the method table tests checks that GPUToolbox.Overlays.host_leaks(CUDACore) stays empty.
• The compile-all test no longer sees the CPU stubs, which GPUToolbox now emits, so its sanity bound drops from 100 to 50 compiled methods (99 remain on 1.10).

Tested locally on an RTX 5080 with Julia 1.13 (device, exceptions, codegen, execution, KernelAbstractions and KernelInterface tests) and 1.10.

[github-actions]

Benchmark results

Compared with 4eb6468cd, the merge base.
The environment differs from the baseline's: BFloat16sExt 10.4.2 → 10.4.3, LLVM 10.4.2 → 10.4.3.

Times are the minimum time of every benchmark. A change is unusual when it exceeds both three times the benchmark's typical change between consecutive runs on the default branch and a minimum per category (performance 3%, latency 5%, applications 3%, other 3%). A benchmark that starts or stops allocating is listed however few the allocations. A single run can still be a fluke, so rerun before drawing conclusions.

	Geometric mean	Unusual: slower	Unusual: faster	Compared
Performance	+1.2%	1	0	79
Latency	−0.2%	0	0	3
Applications	+0.0%	0	0	6

Benchmark	Baseline	This PR	Change
Performance
cuda/synchronization/context/nonblocking	4.86 µs	7.35 µs	+51.3%	🔴 slower

All benchmarks

Benchmark	Baseline	This PR	Change
Performance
array/accumulate/Float32/1d	40.2 µs	39.8 µs	−0.8%
array/accumulate/Float32/dims=1	28.2 µs	28.1 µs	−0.4%
array/accumulate/Float32/dims=1L	82.9 µs	82.5 µs	−0.4%
array/accumulate/Float32/dims=2	46.8 µs	47.1 µs	+0.6%
array/accumulate/Float32/dims=2L	366 µs	364 µs	−0.5%
array/accumulate/Int64/1d	66.7 µs	66.5 µs	−0.3%
array/accumulate/Int64/dims=1	45.8 µs	45.8 µs	+0.1%
array/accumulate/Int64/dims=1L	141 µs	141 µs	+0.0%
array/accumulate/Int64/dims=2	67.0 µs	67.2 µs	+0.3%
array/accumulate/Int64/dims=2L	760 µs	761 µs	+0.0%
array/broadcast	9.24 µs	9.34 µs	+1.0%
array/broadcast launch	5.60 µs	5.69 µs	+1.6%
array/construct	1.38 µs	1.32 µs	−4.4%
array/copy	17.2 µs	17.0 µs	−1.1%
array/copyto!/cpu_to_gpu	207 µs	207 µs	−0.1%
array/copyto!/gpu_to_cpu	240 µs	240 µs	−0.0%
array/copyto!/gpu_to_gpu	8.90 µs	8.74 µs	−1.8%
array/iteration/findall/bool	46.3 µs	47.9 µs	+3.3%
array/iteration/findall/int	58.1 µs	58.4 µs	+0.5%
array/iteration/findfirst/bool	31.2 µs	32.0 µs	+2.5%
array/iteration/findfirst/int	31.5 µs	32.0 µs	+1.7%
array/iteration/findmin/1d	48.7 µs	49.0 µs	+0.7%
array/iteration/findmin/2d	37.8 µs	37.8 µs	+0.0%
array/iteration/logical	46.9 µs	48.4 µs	+3.3%
array/iteration/scalar	60.4 µs	60.1 µs	−0.6%
array/permutedims/2d	45.4 µs	45.3 µs	−0.3%
array/permutedims/3d	43.0 µs	43.1 µs	+0.2%
array/permutedims/4d	36.1 µs	36.2 µs	+0.4%
array/random/rand!/Float32	8.20 µs	8.23 µs	+0.4%
array/random/rand!/Int64	14.3 µs	14.3 µs	+0.3%
array/random/rand/Float32	11.3 µs	11.7 µs	+3.0%
array/random/rand/Int64	18.7 µs	18.7 µs	−0.1%
array/random/randn!/Float32	24.7 µs	24.7 µs	−0.2%
array/random/randn/Float32	33.2 µs	33.1 µs	−0.2%
array/reductions/mapreduce/Float32/1d	29.5 µs	30.0 µs	+1.6%
array/reductions/mapreduce/Float32/dims=1	16.2 µs	16.7 µs	+3.0%
array/reductions/mapreduce/Float32/dims=1L	54.0 µs	54.0 µs	−0.0%
array/reductions/mapreduce/Float32/dims=2	23.8 µs	24.1 µs	+1.2%
array/reductions/mapreduce/Float32/dims=2L	54.8 µs	54.4 µs	−0.6%
array/reductions/mapreduce/Int64/1d	29.7 µs	29.5 µs	−0.7%
array/reductions/mapreduce/Int64/dims=1	18.3 µs	18.7 µs	+2.4%
array/reductions/mapreduce/Int64/dims=1L	93.1 µs	93.0 µs	−0.1%
array/reductions/mapreduce/Int64/dims=2	24.7 µs	24.9 µs	+0.9%
array/reductions/mapreduce/Int64/dims=2L	79.7 µs	79.7 µs	+0.0%
array/reductions/reduce/Float32/1d	29.5 µs	29.8 µs	+0.9%
array/reductions/reduce/Float32/dims=1	16.3 µs	16.6 µs	+1.5%
array/reductions/reduce/Float32/dims=1L	53.9 µs	54.0 µs	+0.1%
array/reductions/reduce/Float32/dims=2	24.0 µs	24.3 µs	+1.5%
array/reductions/reduce/Float32/dims=2L	54.4 µs	54.4 µs	−0.1%
array/reductions/reduce/Int64/1d	29.4 µs	30.2 µs	+2.8%
array/reductions/reduce/Int64/dims=1	18.0 µs	18.3 µs	+1.7%
array/reductions/reduce/Int64/dims=1L	92.8 µs	92.9 µs	+0.1%
array/reductions/reduce/Int64/dims=2	24.7 µs	25.1 µs	+1.7%
array/reductions/reduce/Int64/dims=2L	79.8 µs	79.5 µs	−0.4%
array/reverse/1d	18.2 µs	18.7 µs	+2.6%
array/reverse/1d_inplace	10.8 µs	11.1 µs	+3.0%
array/reverse/1dL	72.0 µs	72.7 µs	+1.0%
array/reverse/1dL_inplace	69.1 µs	69.3 µs	+0.3%
array/reverse/2d	20.6 µs	20.9 µs	+1.3%
array/reverse/2d_inplace	13.4 µs	13.6 µs	+1.7%
array/reverse/2dL	79.0 µs	79.4 µs	+0.5%
array/reverse/2dL_inplace	69.4 µs	69.7 µs	+0.5%
array/sorting/1d	656 µs	657 µs	+0.1%
array/sorting/2d	197 µs	197 µs	+0.2%
array/sorting/by	802 µs	804 µs	+0.2%
cuda/graph/capture	24.9 µs	25.1 µs	+0.8%
cuda/graph/eager	24.6 µs	24.9 µs	+1.5%
cuda/graph/launch	13.9 µs	14.0 µs	+0.6%
cuda/synchronization/context/auto	2.08 µs	2.19 µs	+5.3%
cuda/synchronization/context/blocking	919 ns	906 ns	−1.5%
cuda/synchronization/context/nonblocking	4.86 µs	7.35 µs	+51.3%	🔴 slower
cuda/synchronization/stream/auto	846 ns	839 ns	−0.9%
cuda/synchronization/stream/blocking	786 ns	786 ns	+0.0%
cuda/synchronization/stream/nonblocking	7.26 µs	7.31 µs	+0.7%
kernel/indexing	10.3 µs	10.5 µs	+2.1%
kernel/indexing_checked	10.8 µs	10.9 µs	+0.5%
kernel/launch	1.33 µs	1.34 µs	+0.8%
kernel/occupancy	30.0 ns	30.0 ns	+0.0%
kernel/rand	13.8 µs	14.1 µs	+2.7%
Latency
import	3.55 s	3.50 s	−1.4%
precompile	5.43 s	5.44 s	+0.3%
ttfp	4.01 s	4.03 s	+0.5%
Applications
byval/reference	147 µs	147 µs	+0.1%
byval/slices=1	147 µs	147 µs	−0.0%
byval/slices=2	289 µs	289 µs	+0.0%
byval/slices=3	431 µs	431 µs	−0.0%
cudadevrt	104 µs	104 µs	+0.1%
volumerhs	9.68 ms	9.67 ms	−0.1%

History: https://juliagpu.org/perf/?pkg=cuda

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.71%. Comparing base (4eb6468) to head (8a12729).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3502      +/-   ##
==========================================
- Coverage   88.51%   85.71%   -2.80%     
==========================================
  Files         198      197       -1     
  Lines       22021    21734     -287     
==========================================
- Hits        19491    18629     -862     
- Misses       2530     3105     +575     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.