Xgboost 1.1.1 error

I am using xgboost4spark and I have been trying out xgboost 1.1.1.
Some of my models’ predictions fail on the following error:
Check failed: weights_.Size() == num_row_ (15363 vs. 15362) : Size of weights must equal to number of rows.

The numbers in the error are of course not always the same (but the subtraction is always 1).
The same data/model works on xgboost 0.9.

Any one know what is wrong?


Basically it seems that since xgboost 1.0 xgboost is broken, at least xgboost4spark.
The same data/model that fails on 1.0 and above, always works on 0.9.
We also tried to retrain a new model using 1.0, but still some of those models will fail on prediction.
Even using the new “1.0” models that fail under 1.0 library, will always work under 0.9 library.

Without a fix, xgboost is unusable under Spark 3.0 - since 0.9 does not work under spark 3.

@ran.haim Your issue may be similar to the bug reported in #5848. I’ve submitted a pull request (#5929) that will fix it.