Segmentation fault with categorical variables

[simsurace]

After encountering a problem when fitting with categorical features, I found the following minimal reproducing example for triggering a segfault:

julia> versioninfo()
Julia Version 1.11.9
Commit 53a02c0720c (2026-02-06 00:27 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: macOS (arm64-apple-darwin24.0.0)
  CPU: 12 × Apple M3 Pro
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, apple-m2)
Threads: 1 default, 0 interactive, 1 GC (on 6 virtual cores)
Environment:
  JULIA_EDITOR = code

julia> using CategoricalArrays

julia> using DataFrames

julia> using EvoTrees

julia> for max_value in 1:255
           values = 1:max_value

           df = DataFrame()
           df.cat_feature = map(1:10^3) do _
               rand(values)
           end |> categorical 
           df.y = rand(10^3)

           config = EvoTreeRegressor()

           EvoTrees.fit(
               config,
               df;
               target_name = "y",
               feature_names = ["cat_feature"],
           )
           @info "Successfull for max_value = $max_value"
       end
[ Info: Successfull for max_value = 1
... # (left out intermediate log messages for simplicity)
[ Info: Successfull for max_value = 67

[98526] signal 11 (2): Segmentation fault: 11
in expression starting at REPL[4]:1
dataids at ./abstractarray.jl:1561 [inlined]
dataids at ./abstractarray.jl:1562 [inlined]
mightalias at ./multidimensional.jl:1065 [inlined]
unalias at ./abstractarray.jl:1500 [inlined]
broadcast_unalias at ./broadcast.jl:946 [inlined]
preprocess at ./broadcast.jl:953 [inlined]
preprocess_args at ./broadcast.jl:956 [inlined]
preprocess_args at ./broadcast.jl:955 [inlined]
preprocess at ./broadcast.jl:952 [inlined]
copyto! at ./broadcast.jl:969 [inlined]
copyto! at ./broadcast.jl:925 [inlined]
materialize! at ./broadcast.jl:883 [inlined]
materialize! at ./broadcast.jl:880 [inlined]
get_best_split at /Users/simone/.julia/packages/EvoTrees/LRLAF/src/fit-utils.jl:429
macro expansion at /Users/simone/.julia/packages/EvoTrees/LRLAF/src/fit.jl:94 [inlined]
#980#threadsfor_fun#149 at ./threadingconstructs.jl:253
#980#threadsfor_fun at ./threadingconstructs.jl:220 [inlined]
#1 at ./threadingconstructs.jl:154
unknown function (ip: 0x120f20053)
jl_apply at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XC9YQX9HH2.0/build/default-honeycrisp-XC9YQX9HH2-0/julialang/julia-release-1-dot-11/src/./julia.h:2157 [inlined]
start_task at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XC9YQX9HH2.0/build/default-honeycrisp-XC9YQX9HH2-0/julialang/julia-release-1-dot-11/src/task.c:1204
Allocations: 38616310 (Pool: 38614614; Big: 1696); GC: 28
[1]    98526 segmentation fault  julia --project=.

When running Julia with bounds checks, I get:

ERROR: TaskFailedException

    nested task error: TaskFailedException
    
        nested task error: BoundsError: attempt to access 3×64×1 Array{Float64, 3} at index [1, 65, 1]
        Stacktrace:
          [1] throw_boundserror(A::Array{Float64, 3}, I::Tuple{Int64, Int64, Int64})
            @ Base ./essentials.jl:14
          [2] checkbounds
            @ ./abstractarray.jl:699 [inlined]
          [3] getindex
            @ ./array.jl:929 [inlined]
          [4] getindex
            @ ./abstractarray.jl:1315 [inlined]
          [5] macro expansion
            @ ~/.julia/packages/EvoTrees/LRLAF/src/fit-utils.jl:329 [inlined]
          [6] macro expansion
            @ ./simdloop.jl:77 [inlined]
          [7] macro expansion
            @ ~/.julia/packages/EvoTrees/LRLAF/src/fit-utils.jl:327 [inlined]
          [8] (::EvoTrees.var"#899#threadsfor_fun#139"{EvoTrees.var"#899#threadsfor_fun#138#140"{…}})(tid::Int64; onethread::Bool)
            @ EvoTrees ./threadingconstructs.jl:253
          [9] #899#threadsfor_fun
            @ ./threadingconstructs.jl:220 [inlined]
         [10] (::Base.Threads.var"#1#2"{EvoTrees.var"#899#threadsfor_fun#139"{EvoTrees.var"#899#threadsfor_fun#138#140"{…}}, Int64})()
            @ Base.Threads ./threadingconstructs.jl:154
    Stacktrace:
     [1] threading_run(fun::EvoTrees.var"#899#threadsfor_fun#139"{EvoTrees.var"#899#threadsfor_fun#138#140"{…}}, static::Bool)
       @ Base.Threads ./threadingconstructs.jl:173
     [2] macro expansion
       @ ./threadingconstructs.jl:190 [inlined]
     [3] update_hist!
       @ ~/.julia/packages/EvoTrees/LRLAF/src/fit-utils.jl:326 [inlined]
     [4] macro expansion
       @ ~/.julia/packages/EvoTrees/LRLAF/src/fit.jl:82 [inlined]
     [5] (::EvoTrees.var"#952#threadsfor_fun#150"{EvoTrees.var"#952#threadsfor_fun#147#151"{…}})(tid::Int64; onethread::Bool)
       @ EvoTrees ./threadingconstructs.jl:253
     [6] #952#threadsfor_fun
       @ ./threadingconstructs.jl:220 [inlined]
     [7] (::Base.Threads.var"#1#2"{EvoTrees.var"#952#threadsfor_fun#150"{EvoTrees.var"#952#threadsfor_fun#147#151"{…}}, Int64})()
       @ Base.Threads ./threadingconstructs.jl:154
Stacktrace:
 [1] threading_run(fun::EvoTrees.var"#952#threadsfor_fun#150"{EvoTrees.var"#952#threadsfor_fun#147#151"{…}}, static::Bool)
   @ Base.Threads ./threadingconstructs.jl:173
 [2] macro expansion
   @ ./threadingconstructs.jl:190 [inlined]
 [3] grow_tree!(tree::EvoTrees.Tree{…}, nodes::Vector{…}, params::EvoTreeRegressor, ∇::Matrix{…}, js::Vector{…}, is::Vector{…}, left::Vector{…}, right::Vector{…}, x_bin::Matrix{…}, feattypes::Vector{…}, monotone_constraints::Vector{…})
   @ EvoTrees ~/.julia/packages/EvoTrees/LRLAF/src/fit.jl:81
 [4] grow_evotree!(m::EvoTree{…}, cache::EvoTrees.CacheBaseCPU{…}, params::EvoTreeRegressor)
   @ EvoTrees ~/.julia/packages/EvoTrees/LRLAF/src/fit.jl:21
 [5] fit(params::EvoTreeRegressor, dtrain::DataFrame; target_name::String, feature_names::Vector{…}, weight_name::Nothing, offset_name::Nothing, deval::Nothing, print_every_n::Int64, verbosity::Int64)
   @ EvoTrees ~/.julia/packages/EvoTrees/LRLAF/src/fit.jl:339
 [6] top-level scope
   @ REPL[4]:12
Some type information was truncated. Use `show(err)` to see complete types.

[jeremiedb]

@simsurace I think this issue is likely stemming from a limitation where categorical vector can have a number of levels <= nbins.
This should however not be a limitation for cpu implementation, and I suspect only a minimal patch would solve this issue.
However, I think that the GPU implementation would be a little more tricky, and as such, I think that I'd favor enforcing the check that number of cat levels <= nbins.

[simsurace]

Sure, showing a helpful error message would be fine in my view, rather than hitting an out-of-bounds error or even a segfault.

[jeremiedb]

Will be fixed with EvoTrees v0.18.5 (see PR #323) - An informative error will now be returned.
Ideally, I'd like to be have each feature encoded to a personnalized number of bins; especially useful for categorical of handling of large number of binary vars without exploding memory footprint.

[simsurace]

Thanks, awesome! The per-feature encoding would be slick, but not really urgent for me to have this.