Adapt for julia 1.13

[XLin0mu]

Fixes #2817

[Antipath1]

Can you add inside of the CI.yml to the version "~1.13.0-0" so the CI runs against the 1.13 Beta ?

[Antipath1]

Thank you so much for putting this PR together! I really appreciate you taking the time to address 1.13 compatibility while it's still in Beta.

I've left a couple of inline suggestions to make sure we don't accidentally break compatibility with older versions while fixing 1.13. Let me know if you have any questions about them.

@maintainers: Could someone please add a 1.13 label to this PR and like in general as there will probably be more problems with 1.13 in the future I would suspect?

[Antipath1]

I am confused what the reasoning for removing this line entirely is for shouldn't this be wrapped in an @static if Version < 1.13 block instead?

[Antipath1]

Suggested change

	if VERSION < v"1.13-"
	if v"1.11-" <= VERSION < v"1.13-"

Memoryref wass only introduced in 1.11 which causes the CI to fail on 1.10.

[XLin0mu]

It's a copy-pasted mistake, logic about Base.memoryref for 1.11 still exists in later context, thanks for remind!

[XLin0mu]

For test/integration/random

Problem happend in:

# integration/random.jl: line 60-67
    fn2() = begin
        # TODO(#2253): MersenneTwister without seed leads to illegal instructions
        rng = MersenneTwister(0)
        x = zeros(Float64, 2, 3)
        Random.rand!(rng, x)
        return x
    end
    hlo = @code_hlo fn2()

Trace from test log:

Error in testset Random.jl Overlay:
Error During Test at /__w/Reactant.jl/Reactant.jl/test/integration/random.jl:5
  Got exception outside of a @test
  ArgumentError: cannot reinterpret `Reactant.TracedRNumber{UInt32}` as `UInt32`, type `Reactant.TracedRNumber{UInt32}` is not a bits type
  Stacktrace:
    [1] throwbits
      @ ./reinterpretarray.jl:16
    [2] call_with_reactant
      @ ./none:-1 [inlined]
    [3] call_with_reactant(::Reactant.EnsureReturnType{Union{}}, ::Base.var"#throwbits#throwbits##0", ::Type{Reactant.TracedRNumber{UInt32}}, ::Type{UInt32}, ::Type{Reactant.TracedRNumber{UInt32}})
      @ Reactant /__w/Reactant.jl/Reactant.jl/src/utils.jl:0
    [4] call_with_reactant
      @ ./none:-1 [inlined]
    [5] call_with_reactant(::Reactant.EnsureReturnType{Union{}}, ::typeof(reinterpret), ::Type{UInt32}, ::Reactant.TracedRArray{UInt32, 1})
      @ Reactant /__w/Reactant.jl/Reactant.jl/src/utils.jl:0
    [6] initstate!
      @ /__w/_tool/julia/1.13.0-beta3/x64/share/julia/stdlib/v1.13/Random/src/MersenneTwister.jl:250
    [7] call_with_reactant
      @ ./none:-1 [inlined]
    [8] call_with_reactant(::Reactant.EnsureReturnType{Union{}}, ::typeof(Random.initstate!), ::Random.MersenneTwister, ::Reactant.TracedRArray{UInt32, 1}, ::Int64)
      @ Reactant /__w/Reactant.jl/Reactant.jl/src/utils.jl:0
    [9] call_with_reactant
      @ ./none:-1 [inlined]
   [10] call_with_reactant(::Reactant.EnsureReturnType{Union{}}, ::typeof(Random.seed!), ::Random.MersenneTwister, ::Int64)
      @ Reactant /__w/Reactant.jl/Reactant.jl/src/utils.jl:0
   [11] MersenneTwister
      @ /__w/_tool/julia/1.13.0-beta3/x64/share/julia/stdlib/v1.13/Random/src/MersenneTwister.jl:75

Which directly casued by this change:

in julia 1.12:

# stdlib/Random/src/RNGs.jl: line 378
seed!(r::MersenneTwister, seed) = initstate!(r, hash_seed(seed), seed)

in julia 1.13:

# stdlib\Random\src\MersenneTwister.jl: line 265
seed!(r::MersenneTwister, seed) = initstate!(r, rand(SeedHasher(seed), UInt32, 8), seed)

This exception seems relative to implementation of Reactant and more changes in 1.13 stdlib/Random.

For test/integration/optimisers

Problem happend in:

# integration/optimisers.jl: line 16-28
@testset "Correct Aliasing" begin
    ps = Reactant.to_rarray((a=ones(4), b=ones(2), c=ones(4)))
    opt = Reactant.to_rarray(Descent(0.001f0); track_numbers=true)

    st_opt = @jit Optimisers.setup(opt, ps)

    @test st_opt.a.rule.eta.data === st_opt.b.rule.eta.data

    gs = Reactant.to_rarray((a=ones(4), b=ones(2), c=ones(4)))

    hlo = @code_hlo Optimisers.update(st_opt, ps, gs) # HERE
    @test length(findall("stablehlo.broadcast_in_dim", repr(hlo))) == 2
end

Trace when execute the line manually:

julia> t′ = Optimisers.fmap(copy, tree; exclude=Optimisers.maywrite)  # walks inside Leaf
ERROR: could not load symbol "jl_expand":
/root/julia/bin/../lib/julia/libjulia-internal.so.1.13: undefined symbol: jl_expand
Stacktrace:
  [1] __cacheget_generator__(world::UInt64, source::Method, self::Type, cache::Type, x::Type, args::Tuple{DataType, DataType})
    @ Functors ~/.julia/packages/Functors/LbNAu/src/cache.jl:24
  [2] (::Functors.CachedWalk{…})(::Function, ::ConcretePJRTNumber{…})
    @ Functors ~/.julia/packages/Functors/LbNAu/src/walks.jl:175
  [3] recurse
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:54 [inlined]
  [4] map
    @ ./tuple.jl:354 [inlined]
  [5] map
    @ ./namedtuple.jl:263 [inlined]
  [6] _map
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:3 [inlined]
  [7] DefaultWalk
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:73 [inlined]
  [8] ExcludeWalk
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:126 [inlined]
  [9] CachedWalk
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:177 [inlined]
 [10] recurse
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:54 [inlined]
 [11] map(f::Functors.var"#recurse#execute##0"{Functors.CachedWalk{…}}, t::Tuple{Descent{…}, Nothing, Bool})
    @ Base ./tuple.jl:356
 [12] map(::Function, ::@NamedTuple{rule::Descent{ConcretePJRTNumber{Float32, 1}}, state::Nothing, frozen::Bool})
    @ Base ./namedtuple.jl:263
 [13] _map
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:3 [inlined]
 [14] (::Functors.DefaultWalk)(::Function, ::Optimisers.Leaf{Descent{ConcretePJRTNumber{Float32, 1}}, Nothing})
    @ Functors ~/.julia/packages/Functors/LbNAu/src/walks.jl:73
 [15] ExcludeWalk
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:126 [inlined]
 [16] (::Functors.CachedWalk{…})(::Function, ::Optimisers.Leaf{…})
    @ Functors ~/.julia/packages/Functors/LbNAu/src/walks.jl:177
 [17] recurse
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:54 [inlined]
 [18] map
    @ ./tuple.jl:356 [inlined]
 [19] map
    @ ./namedtuple.jl:263 [inlined]
 [20] _map
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:3 [inlined]
 [21] DefaultWalk
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:73 [inlined]
 [22] ExcludeWalk
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:126 [inlined]
 [23] CachedWalk
    @ ~/.julia/packages/Functors/LbNAu/src/walks.jl:177 [inlined]
 [24] execute(::Functors.CachedWalk{…}, ::@NamedTuple{…})
    @ Functors ~/.julia/packages/Functors/LbNAu/src/walks.jl:55
 [25] fmap(::Function, ::@NamedTuple{…}; exclude::Function, walk::Functors.DefaultWalk, cache::IdDict{…}, prune::Functors.NoKeyword)
    @ Functors ~/.julia/packages/Functors/LbNAu/src/maps.jl:11
 [26] kwcall(::@NamedTuple{…}, ::typeof(Functors.fmap), ::Function, ::@NamedTuple{…})
    @ Functors ~/.julia/packages/Functors/LbNAu/src/maps.jl:3
 [27] top-level scope
    @ REPL[20]:1
Some type information was truncated. Use `show(err)` to see complete types.

Which caused by the things about :jl_expand changing from:

# 1.12 Base/meta.jl:line 163
lower(m::Module, @nospecialize(x)) = ccall(:jl_expand, Any, (Any, Any), x, m)

# 1.12  Base/expr.jl:line 1667-1668
function generated_body_to_codeinfo(ex::Expr, defmod::Module, isva::Bool)
    ci = ccall(:jl_expand, Any, (Any, Any), ex, defmod)

to:

# 1.13 Base/meta.jl:line 275
lower(m::Module, @nospecialize(x)) = Core._lower(x, m, "none", 0, typemax(Csize_t), false)[1]

# 1.13  Base/expr.jl:line 1800-1802
function generated_body_to_codeinfo(ex::Expr, defmod::Module, isva::Bool)
    ci = ccall(:jl_fl_lower, Any, (Any, Any, Ptr{UInt8}, Csize_t, Csize_t, Cint),
               ex, defmod, "none", 0, typemax(Csize_t), 0)[1]

Where the :jl_expand disapeard in julia 1.13