Memory leak with unified memory?

[Tortar]

Describe the bug

When using a CuArray with unified memory, running a loop many times increases memory until the program is killed. This seems related with GC not being too smart on this.

To reproduce

using CUDA

N = 10_000_000

data_cpu = rand(Float32, N)
uni_arr = cu(data_cpu; unified=true)

function f(a)
    CUDA.@sync a = sin.(a)
end

for _ in 1:1000 @time f(uni_arr) end # executing this causes an OOM error in CPU RAM

Manifest.toml

Status `~/.julia/environments/v1.10/Project.toml`
⌃ [46ada45e] Agents v6.0.9
  [56664e29] Ark v0.3.0 `~/.julia/dev/Ark`
  [6e4b80f9] BenchmarkTools v1.6.3
  [336ed68f] CSV v0.10.15
  [052768ef] CUDA v5.9.6
  [324d7699] CategoricalArrays v1.0.2
  [0ca39b1e] Chairmarks v1.3.1
  [944b1d66] CodecZlib v0.7.8
  [a93c6f00] DataFrames v1.8.1
⌅ [864edb3b] DataStructures v0.18.22
  [85a47980] Dictionaries v0.4.6
⌃ [a0c0ee7d] DifferentiationInterface v0.7.12
⌃ [31c24e10] Distributions v0.25.122
  [e30172f5] Documenter v1.16.1
⌃ [7da242da] Enzyme v0.13.109
  [ff5a1669] FieldViews v0.3.3
  [41ab1584] InvertedIndices v1.3.1
  [da2b9cff] Mooncake v0.4.193 `~/.julia/dev/Mooncake`
  [d96e819e] Parameters v0.12.3
⌃ [91a5bcdd] Plots v1.41.2
  [21216c6a] Preferences v1.5.1
⌃ [295af30f] Revise v3.12.3
  [ac92255e] Speculator v0.3.0
  [90137ffa] StaticArrays v1.9.16
⌃ [2913bbd2] StatsBase v0.34.8
  [ff63dad9] StreamSampling v0.7.6 `~/.julia/dev/StreamSampling`
  [09ab397b] StructArrays v0.7.2
  [2fbcfb34] UniqueVectors v1.2.0
  [02edb123] WeightVectors v0.1.0
Info Packages marked with ⌃ and ⌅ have new versions available. Those with ⌃ may be upgradable, but those with ⌅ are restricted by compatibility constraints from upgrading. To see why use `status --outdated`

Expected behavior

No memory increase as when using unified=false

Version info

Details on Julia:

Julia Version 1.10.10
Commit 95f30e51f41 (2025-06-27 09:51 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 12 × AMD Ryzen 5 5600H with Radeon Graphics
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, znver3)
Threads: 1 default, 0 interactive, 1 GC (on 12 virtual cores)

Details on CUDA:

CUDA toolchain: 
- runtime 13.0, artifact installation
- driver 580.95.5 for 13.0
- compiler 13.1

CUDA libraries: 
- CUBLAS: 13.1.0
- CURAND: 10.4.0
- CUFFT: 12.0.0
- CUSOLVER: 12.0.4
- CUSPARSE: 12.6.3
- CUPTI: 2025.3.1 (API 13.0.1)
- NVML: 13.0.0+580.95.5

Julia packages: 
- CUDA: 5.9.6
- GPUArrays: 11.3.2
- GPUCompiler: 1.7.5
- KernelAbstractions: 0.9.39
- CUDA_Driver_jll: 13.1.0+0
- CUDA_Compiler_jll: 0.3.0+1
- CUDA_Runtime_jll: 0.19.2+0

Toolchain:
- Julia: 1.10.10
- LLVM: 15.0.7

1 device:
  0: NVIDIA GeForce GTX 1650 (sm_75, 1.743 GiB / 4.000 GiB available)

[Tortar]

Seems like this work:

for i in 1:1000 
    if i % 100 == 0 GC.gc() end
    @time f(uni_arr)
end

though it would be much more robust not to manually need to invoke GC