Deadlock during extension loading (FluxCUDAcuDNNExt) on Julia 1.12.2 — works on 1.11.7 #60307
Description
Upgrading from Julia 1.11.7 to 1.12.2 causes a consistent extension-loading failure when importing Flux with CUDA/cuDNN enabled. The exact same environment (same Project.toml, same CUDA toolkit, same GPU) works without issue on 1.11.7.
On Julia 1.12.2, loading Flux triggers a loader deadlock:
Error during loading of extension FluxCUDAcuDNNExt of Flux
ConcurrencyViolationError("deadlock detected in loading FluxCUDAExt using FluxCUDAExt (while loading FluxCUDAcuDNNExt)")
This appears to originate entirely from Julia's extension loading system, not CUDA.jl or Flux itself.
Error output
┌ Error: Error during loading of extension FluxCUDAcuDNNExt of Flux, use `Base.retry_load_extensions()` to retry.
│ exception =
│ 1-element ExceptionStack:
│ ConcurrencyViolationError("deadlock detected in loading FluxCUDAExt using FluxCUDAExt (while loading FluxCUDAcuDNNExt)")
│ Stacktrace:
│ [1] canstart_loading(modkey::Base.PkgId, build_id::UInt128, stalecheck::Bool)
│ @ Base .\loading.jl:2207
│ [2] _require_search_from_serialized(pkg::Base.PkgId, sourcepath::String, build_id::UInt128, stalecheck::Bool; reasons::Dict{String, Int64}, DEPOT_PATH::Vector{String})
│ @ Base .\loading.jl:2041
│ [3] __require_prelocked(pkg::Base.PkgId, env::Nothing)
│ @ Base .\loading.jl:2624
│ [4] _require_prelocked(uuidkey::Base.PkgId, env::Nothing)
│ @ Base .\loading.jl:2490
│ [5] _require_prelocked(uuidkey::Base.PkgId)
│ @ Base .\loading.jl:2484
│ [6] run_extension_callbacks(extid::Base.ExtensionId)
│ @ Base .\loading.jl:1604
│ [7] run_extension_callbacks(pkgid::Base.PkgId)
│ @ Base .\loading.jl:1641
│ [8] run_package_callbacks(modkey::Base.PkgId)
│ @ Base .\loading.jl:1457
│ ...
└ @ Base loading.jl:1614
(The full trace is longer; happy to attach it if helpful.)
Reproduction
Minimal example:
julia> using CUDA
julia> using Flux # or reverse order, either triggers extension loading deadlockThis immediately throws the ConcurrencyViolationError only on Julia 1.12.x.
Environment
Julia
Julia Version 1.12.2
LLVM: 18.1.7
OS: Windows 11
Hardware
GPU: NVIDIA RTX 2080 Ti (sm_75)
NVIDIA driver: 577.0
CUDA libraries (from CUDA.jl diagnostics)
CUDA runtime: 12.9
CUBLAS: 12.9.1
CUFFT: 11.4.1
CUDNN: 8.x via cuDNN_jll (FluxCUDAcuDNNExt)
Julia packages
CUDA.jl: 5.8.5 and 5.9.5 tested (same result)
Flux.jl: 0.16.5
cuDNN.jl: 1.4.x
Notes
-
This does not occur on Julia 1.11.7 with the same project and same GPU configuration.
-
Setting
ENV["JULIA_CUDA_USE_CUDNN"] = "false"
does not prevent the deadlock.
-
Precompilation succeeds; the crash occurs during extension loading at runtime.
-
The issue appears related to concurrency behaviour introduced in Julia 1.12’s extension system.
Request
Could the core team review whether this is a regression in the extension loader?
If additional logs, a minimal test environment, or a reproducible package set would assist, I’m happy to provide them.