New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
1.5.0 Release Candidate #7260
Comments
This time R-hub has Apple M1 instances. Let's wait for the test result. ;-) |
R-Hub
|
Hi all, the RC packages are released to public repositories now. Please help test the new release and provide feedbacks. Enjoy! |
Everything looks good. I will start preparing the stable release tomorrow. |
Need to resolve #7294 first. |
Hi @hetong007 @hcho3 I have made the 1.5 stable release and uploaded all Python packages to pypi. Could you please help upload the R package and JVM packages? |
@trivialfis I've uploaded the JVM package to Maven Central. |
I'm quite occupied these two days but I'll start submitting the R package later this week. Hope it'll go into cran smoothly. |
The release branch consistently passes However, the pre-test via Win-Builder failed with R-devel, R-release and R-oldrelease consistently, due to too-large training errors from several checks in
Full check logs (might be removed in 72 hours):
The error seems concerning to me as it indicates the model was not trained as expected in our test on Windows. Do we expect such behavior or have we observed similar issues? Also, please correct me if I made any mistakes in the pre-test. |
Interesting, thanks for running the tests! I will look into it. Oddly, I ran the tests with R-hub #7260 (comment) and it worked fine there. |
Might be an issue with the new prediction interface. |
@hetong007 Do you know how to reproduce that particular environment locally? (or on platforms like AWS) |
@trivialfis Thanks for the effort! According to win-builder, the configuration is:
AFAIK, windows server 2008 is only available via "Community AMIs", but I'm not sure if that's critical to reproduce. I can also try later at home on my Windows 10 PC. If that doesn't reproduce, then maybe try an official Windows Server 2012 (the closest environment)? Unfortunately, based on a few previous submissions, passing the win-builder check system has been a hard requirement to get into CRAN now. Thus we need to fix it eventually. |
Yes. It's likely a bug in the R package. Just I need to reproduce it in order to fix it. So far neither the tests on GitHub action nor the one on R-hub show the error so I'm wondering where can I build that environment. On R-hub I tested |
Interestingly, I did the following on my Ubuntu 16.04 and observed a test failure:
And the message is:
Hope that may help our debugging. But I'm not sure how it passes the local check though... |
Unable to reproduce on Ubuntu 20.04. Let me spin up a 16.04 machine. |
Apologies! I messed up my environment thus didn't test with the latest release for the above results :( Will still test on my Windows 10 though, and ping here about the results. |
Just a quick note, I have tested with |
Hmm, I keep getting the error about |
After replacing |
The evaluation log and feature importance are also different so it's not just the prediction function rewrite. |
@hetong007 Did you reproduce it? I tried to change the |
I couldn't reproduce it on my windows 10... I tried to either install the zip from our github release, or directly download the compiled zip from win-builder. They both passed the manual execution on |
@hetong007 Thanks for the help! I will test the package on win builder with the maintainer field changed to my email later today. |
Hi @hetong007 @hcho3 I ran the test on the win builder. Here is the installation log: https://win-builder.r-project.org/Jujkkd7LoWF8/00install.out (also attached below). I noticed one weird thing, from this section:
It's using rtools40, according to https://cran.microsoft.com/snapshot/2020-07-16/bin/windows/Rtools/ , it should be using gcc toolchains 8.3.0. But in the log somehow |
Reproduced the error locally. |
Thanks, that's fantastic! Would you mind to leave a pointer to the exact commit? There seems to be two ways to handle this:
|
@hetong007 I opened a PR for the fix: #7390 . |
@hetong007 Hi, I ran some tests on 1.5 branch a3d195e . Oracle Solaris 10, x86, 32 bit, R-release Windows Server 2008 R2 SP1, R-release, 32/64 bit Apple Silicon (M1), macOS 11.6 Big Sur, R-release Debian Linux, R-release, GCC win builder |
Thanks a lot! I've submitted the latest release branch to CRAN. It passes the auto-check, however seems to break some dependencies. I'll need to look into these packages, and probably communicate with the maintainers before we can have it in CRAN. Details are below:
|
It shouldn't have breaking change (but has new deprecation) I will look into it. |
From the above log, the error from |
Building that package now. Might take some time to debug, I'm working on other things at the moment. Will keep you posted. |
663136a is the commit to blame. |
Okay, I get the issue now. I re-implemented the global feature importance in C++ to eliminate the text parsing. But there's a subtle difference between the original Python and R implmentation. For Python there were "total_gain" and "gain", the latter is just the former divided by split count (or "frequency" in R). But for R this should not occur. |
Opened a PR for the fix #7394 . Sorry about the inconvenience and thanks for all the help! |
Ran tests on R-hub and win-builder Passing:
Failing:
The windows instance on R-hub failed during "download and unpack files", which doesn't seem to be related to XGBoost:
|
2 failures caused by timeout. The tests have passed, but timeout at building doc. This on centos is concerning:
Other tests don't have this note. |
Doesn't make sense, I checked the binary locally, there's only |
@trivialfis Thanks a ton for digging into this! It has successfully passed the reverse dependency test, and on its way to CRAN. Let's hope it'll pass the real tests across all platforms :) |
@hetong007 Do you have the bandwidth to review R PRs? My experience with R is very limited and would be great to have an expert at supervising. ;-) |
No problem. I can still handle the R part, but not familiar with things that touch low-level source now. |
Excellent! |
There might be some warnings we need to fix:
|
Opened a PR: #7404 I did not reproduce the warning using the development version of clang though. |
Running tests on R-hub and win-builder. I will update this comment accordingly:
** DONE EIX |
@hetong007 Should I bump the version to 1.5.0.2 or 1.5.1? We are preparing the 1.5.1 for Python and JVM and it would be great to have some consistency. But I'm not entirely sure about the CRAN policy. Right now, the other 2 packages use |
Do we plan 1.5.1 to include ONLY the incremental fixes for CRAN submission? If so, we can bump to 1.5.1, otherwise I think it is better to use 1.5.0.2. |
Got it, I will keep it as 1.5.0.2. Opened a PR for incrementing the version number: #7452 . |
All tests have passed. |
Closing as all 1.5.0 packages are submitted. Thank you all! |
Roadmap #6846 . Draft of release note: #7271 .
We are about to release version 1.5.0 of XGBoost. In the next two weeks, we invite everyone to try out the release candidate (RC).
Feedback period: until the end of October 13, 2021. No new feature will be added to the release; only critical bug fixes will be added.
@dmlc/xgboost-committer
Available packages:
Linux x86_64: https://github.com/dmlc/xgboost/releases/download/v1.5.0rc1/xgboost_r_gpu_linux.tar.gz
Windows x86_64: https://github.com/dmlc/xgboost/releases/download/v1.5.0rc1/xgboost_r_gpu_win64.tar.gz
Show instructions (Maven/SBT)
Maven
SBT
Starting from 1.2.0, XGBoost4J-Spark supports training with NVIDIA GPUs. To enable this capability, download artifacts suffixed with
-gpu
, as follows:Show instructions (Maven/SBT)
Maven
SBT
TO-DOs
PRs to be backported
verbose_eval
in Pythoncv
function. (Fix cv print period #7291)The text was updated successfully, but these errors were encountered: