Rory Mitchell

copy_if integer overflow for sizes > 2^31

3

For sizes > 2^31 we get an integer overflow with copy_if using the device backend. This results in out of memory: ``` terminate called after throwing an instance of 'thrust::system::detail::bad_alloc'...

c++11 documentation

1

I noticed that my thrust::shuffle documentation would not build by default because of the preprocessor flag `#if THRUST_CPP_DIALECT >= 2011`. I want to make sure this documentation is available for...

Optimise histogram kernels

6

Performed loop unrolling and change compressed iterator to use byte aligned sizes, increasing global memory read throughput. max_depth=8 | dataset | master | hist | |---------|-------------|-------------| | airline | 89.51209751...

Cuda SASS code is compatible across major architectures, so we only need one from each, unless particular features are needed. As we don't use any special features (e.g. tensor cores)...

[BUG] No way to predict_proba with RandomForestClassifier trained on 64-bit data

3

```python import cuml import numpy as np X = np.array([[0.0], [1.0]]) y = np.array([0, 1]) model = cuml.ensemble.RandomForestClassifier().fit(X, y) pred = model.predict_proba(X) ``` ``` File "cuml/ensemble/randomforestclassifier.pyx", line 676, in cuml.ensemble.randomforestclassifier.RandomForestClassifier.predict_proba...

@trivialfis has a PR here https://github.com/dmlc/xgboost/pull/3548 implementing the span class from the cpp core guidelines and the c++20 draft. Ideally we would replace the dmlc-core array_view class with this but...

Use quantised gradients in gpu_hist histograms

7

Floating point addition and subtraction is causing small numerical errors in the learning process, sometimes causing the algorithm to create splits on 'phantom' gradients. We can solve this by consistently...

'scale_pos_weight' parameter

1

I want to use this issue to document some pitfalls of the 'scale_pos_weight' parameter and discuss if it needs to be changed or documented better. The following script illustrates some...

[BUG] Intermittent CI failures with xgboost

3

Sometimes we observe failures in CI of the following type: ``` cuml/tests/explainer/test_gpu_treeshap.py::test_with_hypothesis [85d2e7739a0d:2816 :0:3362] Caught signal 11 (Segmentation fault: address not mapped to object at address 0x55f54f237ba8) ==== backtrace (tid:...

Xgboost binary format compatibility

2

In #1215, changes to the xgboost binary format break integration with the shap library and lead users to downgrade xgboost or create other hack solutions. Xgboost makes no promises about...