xgboost failing with julia 1.8.5

[naveenpalli]

XGBoost version 2.2.5
Julia Version 1.8.5
Commit 17cfb8e65e (2023-01-08 06:45 UTC)
Platform Info:
OS: Windows (x86_64-w64-mingw32)
CPU: 8 × Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-13.0.1 (ORCJIT, skylake)
Threads: 1 on 8 virtual cores

--- sample code below crashes at the last line

using XGBoost
using DataFrames

(X,y) = (DataFrame(randn(100,3), [:l, :r, :u]), randn(100))

b = xgboost((X,y), max_depth=4)

[ExpandingMan]

Can you show the full stack trace? I can't reproduce any error.

[ablaom]

Yes, please post stack trace, but my guess is JuliaLang/julia#48187 . Only workaround I know of is downgrade julia to < 1.8.4 or use non-windows machine

[naveenpalli]

How do I generate the full stack trace? Running on Windows 11

[ExpandingMan]

Oh, I didn't notice this was windows, yeah it's probably that. Also Julia 1.9 just came out today so that might work (and it's best to upgrade anyway).

How do I generate the full stack trace? Running on Windows 11

What appears when you see the error?

[jeremiedb]

Unfortunately the problem hasn't been resolved with Julia v1.9.
AFAIK, the most active thread discussing the issue is: https://discourse.julialang.org/t/issue-with-xgboost-jl-and-libsvm-jl-when-julia-1-8-4/92396/46

[naveenpalli]

Yeah, didn't fix the issue when I tried in 1.9.0 as well. Seems to be ongoing since 1.8.4.
No errors are generated it silently crashes. The REPL dies when you run the four lines above and if you put in a script nothing prints.

[ExpandingMan]

I'm getting the sense that there are a lack of windows people around to debug the issue. I don't think any of us who have made recent commits to this package are capable of debugging the issue. Probably something JLL related but 🤷

[jd-foster]

As discussed in the linked discourse thread, it seems to be an incompatibility arising after an update to building julia under gcc 12 since v1.8.4, and the libgomp (OpenMP) library. A possible work-around is to use your own system's version of libgomp, but not a good long-term solution. And, yes, it needs someone who knows how to debug windows crashes.

[ExpandingMan]

Do we know what library that xgboost relies on has libgomp as a build dependency? I'm wondering if it could be as simple as updating a jll build script, but I don't even know which one my specify this as a dependency.

[jd-foster]

The julia stdlib CompilerSupportLibraries and the CompilerSupportLibraries_jll supply a specific libgomp:

https://github.com/JuliaPackaging/Yggdrasil/blob/543e52c046761a2ad39c1d93dbbe108b79a48b22/X/XGBoost/build_tarballs.jl#LL86C1-L86C1

See also my halting attempts: https://discourse.julialang.org/t/issue-with-xgboost-jl-and-libsvm-jl-when-julia-1-8-4/92396/28?u=jd-foster

[naveenpalli]

Ok, quick and dirty fix to at least avoid crashes. As noted above, offending library seems to be libgomp-1.dll that is installed in the bin directory. Replacing with the 1.8.3 version seems to fix the crashes.

Mode LastWriteTime Length Name

---

-a--- 11/14/2022 1:07 PM 1761059 libgomp-1.dll
-a--- 5/8/2023 6:28 AM 1511051 libgomp-1.dll.installed_1_9_0

the 1_9_0 version (second one) comes with 1.8.5 and 1.9.0, the first one is from 1.8.3. For now, just replace this one if you need to use XGBoost. Of course, someone needs to figure out why libgomp-1.dll.installed_1_9_0 crashes.

[devmotion]

Seems like a duplicate of #153?

[giordano]

Fixed in Julia master by JuliaLang/julia#50135 (and yes, this is a duplicate)

[ExpandingMan]

Thanks so much for your efforts on this @giordano .

[giordano]

I merely executed the solution found by @jd-foster 🙂

[ExpandingMan]

Ok, well thanks @jd-foster then!

[giordano]

I reported the fact that this issue has been fixed (at least on Julia master) because maybe you want to close this ticket (and #153) 🙂

[ExpandingMan]

I'll wait until 1.9.2 is out and we see CI/CD pass to close them.

[giordano]

For CI you can check it works on nightly, like JuliaApproximation/FastTransforms.jl#222

[naveenpalli]

Issue fixed in latest nightly 👍 . Great job @jd-foster and @giordano.

julia> b = xgboost((X,y), max_depth=4)
[ Info: XGBoost: starting training.
[ Info: [1] train-rmse:0.92113782176854342
[ Info: [2] train-rmse:0.86453193144589746
[ Info: [3] train-rmse:0.81174178860290347
[ Info: [4] train-rmse:0.77007250879661970
[ Info: [5] train-rmse:0.70126246007321902
[ Info: [6] train-rmse:0.64168739063396085
[ Info: [7] train-rmse:0.61086880841385705
[ Info: [8] train-rmse:0.56906845909755754
[ Info: [9] train-rmse:0.54034384738000574
[ Info: [10] train-rmse:0.51089554296834572
[ Info: Training rounds complete.

[tylerjthomas9]

We should be good to go. All the tests are now green thanks to Julia v1.9.2

#184

[ExpandingMan]

As with #153, I believe this is now resolved on 1.9.2. Please re-open if this is not the case.