Cannot load CUDA.jl with Julia 1.13

[eschnett]

I am using the master branch of CUDA.jl with Julia 1.13-beta1 on an H100 system. using CUDA leads to this error:

$ julia
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.13.0-beta1 (2026-01-12)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org release
|__/                   |

(@v1.13) pkg> status
Status `~/.julia/environments/v1.13/Project.toml`
  [052768ef] CUDA v5.9.6
  [295af30f] Revise v3.13.1

(@v1.13) pkg> add https://github.com/JuliaGPU/CUDA.jl#master
    Updating git-repo `https://github.com/JuliaGPU/CUDA.jl`
   Resolving package versions...
    Updating `~/.julia/environments/v1.13/Project.toml`
  [052768ef] ~ CUDA v5.9.6 ⇒ v5.9.6 `https://github.com/JuliaGPU/CUDA.jl#master`
    Updating `~/.julia/environments/v1.13/Manifest.toml`
  [052768ef] ~ CUDA v5.9.6 ⇒ v5.9.6 `https://github.com/JuliaGPU/CUDA.jl#master`
ERROR: LoadError: ArgumentError: Function definition contains invalid or missing arguments
Expr
  head: Symbol =
  args: Array{Any}((2,))
    1: Symbol #2354#val
    2: Expr
      head: Symbol =
      args: Array{Any}((2,))
        1: Expr
          head: Symbol call
          args: Array{Any}((1,))
            1: Symbol gridDim
        2: Expr
          head: Symbol block
          args: Array{Any}((3,))
            1: Expr
              head: Symbol meta
              args: Array{Any}((1,))
                1: Symbol inline
            2: LineNumberNode
              line: Int64 71
              file: Symbol /home/eschnett/.julia/packages/CUDA/jXARD/src/device/intrinsics/indexing.jl
            3: Expr
              head: Symbol tuple
              args: Array{Any}((3,))
                1: Expr
                2: Expr
                3: Expr

Stacktrace:
  [1] invalid_def
    @ ~/.julia/packages/ExprTools/MDS6z/src/function.jl:30 [inlined]
  [2] splitdef(ex::Expr; throw::Bool)
    @ ExprTools ~/.julia/packages/ExprTools/MDS6z/src/function.jl:121
  [3] splitdef(ex::Expr)
    @ ExprTools ~/.julia/packages/ExprTools/MDS6z/src/function.jl:23
  [4] var"@device_function"(__source__::LineNumberNode, __module__::Module, ex::Any)
    @ CUDA ~/.julia/packages/CUDA/jXARD/src/device/utils.jl:29
  [5] fl_lower(ex::Expr, mod::Module, filename::String, lineno::UInt64, world::UInt64, warn::Bool)
    @ Base ./flfrontend.jl:24
  [6] include(mapexpr::Function, mod::Module, _path::String)
    @ Base ./Base.jl:310
  [7] top-level scope
    @ ~/.julia/packages/CUDA/jXARD/src/device/intrinsics.jl:8
  [8] include(mapexpr::Function, mod::Module, _path::String)
    @ Base ./Base.jl:310
  [9] top-level scope
    @ ~/.julia/packages/CUDA/jXARD/src/CUDA.jl:78
 [10] include(mod::Module, _path::String)
    @ Base ./Base.jl:309
 [11] include_package_for_output(pkg::Base.PkgId, input::String, depot_path::Vector{String}, dl_load_path::Vector{String}, load_path::Vector{String}, concrete_deps::Vector{Pair{Base.PkgId, UInt128}}, source::Nothing)
    @ Base ./loading.jl:3142
 [12] top-level scope
    @ stdin:5
 [13] eval(m::Module, e::Any)
    @ Core ./boot.jl:489
 [14] include_string(mapexpr::typeof(identity), mod::Module, code::String, filename::String)
    @ Base ./loading.jl:2988
 [15] include_string
    @ ./loading.jl:2998 [inlined]
 [16] exec_options(opts::Base.JLOptions)
    @ Base ./client.jl:342
 [17] _start()
    @ Base ./client.jl:577
in expression starting at /home/eschnett/.julia/packages/CUDA/jXARD/src/device/utils.jl:56
in expression starting at /home/eschnett/.julia/packages/CUDA/jXARD/src/device/intrinsics/indexing.jl:64
in expression starting at /home/eschnett/.julia/packages/CUDA/jXARD/src/device/intrinsics.jl:8
in expression starting at /home/eschnett/.julia/packages/CUDA/jXARD/src/CUDA.jl:1
in expression starting at stdin:5
  ✗ CUDA
  ✗ Atomix → AtomixCUDAExt
Precompiling CUDA finished.
  0 dependencies successfully precompiled in 26 seconds. 102 already precompiled.
  2 dependencies errored.
  For a report of the errors see `julia> err`. To retry use `pkg> precompile`

Things work with Julia 1.12.

[eschnett]

The problem is the macro @device_functions. Function definitions apparently look different in Julia 1.13. In 1.12, the defintion

           """
           f
           """
           @inline f(x) = (x, x)

leads to

    #= REPL[6]:2 =#
    begin
        f(x) = begin
                $(Expr(:meta, :inline))
                #= REPL[6]:5 =#
                (x, x)
            end
        (Base.Docs.doc!)(Main, (Base.Docs.Binding)(Main, :f), (Base.Docs.docstr)((Core.svec)("f\n"), (Dict{Symbol, Any})(:path => "REPL[6]", :linenumber => 2, :module => Main)), Union{Tuple{Any}})
    end

whereas 1.13 produces

    #= REPL[1]:2 =#
    begin
        var"#3#val" = (f(x) = begin
                    $(Expr(:meta, :inline))
                    #= REPL[1]:5 =#
                    (x, x)
                end)
        (Base.Docs.doc!)(Main, (Base.Docs.Binding)(Main, :f), (Base.Docs.docstr)((Core.svec)("f\n"), (Dict{Symbol, Any})(:path => "REPL[1]", :linenumber => 2, :module => Main)), Union{Tuple{Any}})
        var"#3#val"
    end

That is, a local symbol "#3#val is invented, assigned, and its value is used later on.

Also, removing the @inline declaration from the definition of gridDim avoids the problem.

[eschnett]

The macro @device_functions is used in a single file for 10 small functions. We might as well remove it.

[maleadt]

#2998 adds it in more places, so ideally we fix thisl

[eschnett]

I don't know how to fix the macro, sorry.