Rory Mitchell

Here is the A100 benchmark. Everything looks good. | dataset | master | hist | |---------|-------------|-------------| | airline | 65.77564727 | 60.79124835 | | bosch | 13.05801762 | 13.36868745 |...

Unfortunately using aligned byte sizes in the compressed iterator increased the memory usage of the large sizes test by 1gb and I think it barely no longer fits on the...

I reverted the changes to compressed iterator. In the test for large sizes the bit packed version to is able to use 10 bits per symbol where the aligned version...

I think the criticism of this heuristic is quite valid, it makes the result unpredictable. The assumption that the hessian is uniformly distributed seems dubious. I think the only reliable...

After working on the row partitioning I realized that if we have count information it makes this step much faster. So if we counted rows in the histogram we lose...

In my opinion this PR should not be merged. It looks like unrelated pieces of code copy pasted from the original PR. It's definitely possible to implement the desired improvements...

Does just adding an additional ConstFloatPointer type work for you?

We can switch to doing some work on xgboost after you have dealt with the review issues until the continuous integration problem gets resolved.

Its always 1 in code. According to @tqchen it was intended to be used for multiclass trees (he can maybe confirm this). There is also something in metainfo to do...

So for my xgboost integration (https://github.com/dmlc/xgboost/pull/4473) I will try the approach of running the tracker on worker zero and assume the performance load of the tracker is negligible.