This PR adds optional support for loading nccl with `dlopen` as an alternative of compile time linking. This is to address the size bloat issue with the PyPI binary release.
- Add CMake option to load `nccl` at runtime.
- Add an NCCL stub.
After this, `nccl` will be fetched from PyPI when using pip to install XGBoost, either by a user or by `pyproject.toml`. Others who want to link the nccl at compile time can continue to do so without any change.
At the moment, this is Linux only since we only support MNMG on Linux.
- Use the `linalg::Matrix` for storing gradients.
- New API for the custom objective.
- Custom objective for multi-class/multi-target is now required to return the correct shape.
- Custom objective for Python can accept arrays with any strides. (row-major, column-major)
* Define `best_iteration` only if early stopping is used.
This is the behavior specified by the document but not honored in the actual code.
- Don't set the attributes if there's no early stopping.
- Clean up the code for callbacks, and replace assertions with proper exceptions.
- Assign the attributes when early stopping `save_best` is used.
- Turn the attributes into Python properties.
---------
Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu>
* Handle the new `device` parameter in dask and demos.
- Check no ordinal is specified in the dask interface.
- Update demos.
- Update dask doc.
- Update the condition for QDM.
- A `DeviceOrd` struct is implemented to indicate the device. It will eventually replace the `gpu_id` parameter.
- The `predictor` parameter is removed.
- Fallback to `DMatrix` when `inplace_predict` is not available.
- The heuristic for choosing a predictor is only used during training.
- Fix prediction range.
- Support prediction cache in mt-hist.
- Support model slicing.
- Make the booster a Python iterable by defining `__iter__`.
- Cleanup removed/deprecated parameters.
- A new field in the output model `iteration_indptr` for pointing to the ranges of trees for each iteration.
- Remove parameter serialization in the scikit-learn interface.
The scikit-lear interface `save_model` will save only the model and discard all
hyper-parameters. This is to align with the native XGBoost interface, which distinguishes
the hyper-parameter and model parameters.
With the scikit-learn interface, model parameters are attributes of the estimator. For
instance, `n_features_in_`, `n_classes_` are always accessible with
`estimator.n_features_in_` and `estimator.n_classes_`, but not with the
`estimator.get_params`.
- Define a `load_model` method for classifier to load its own attributes.
- Set n_estimators to None by default.
* Support sklearn cross validation for ranker.
- Add a convention for X to include a special `qid` column.
sklearn utilities consider only `X`, `y` and `sample_weight` for supervised learning
algorithms, but we need an additional qid array for ranking.
It's important to be able to support the cross validation function in sklearn since all
other tuning functions like grid search are based on cross validation.
* Use array interface for CSC matrix.
Use array interface for CSC matrix and align the interface with CSR and dense.
- Fix nthread issue in the R package DMatrix.
- Unify the behavior of handling `missing` with other inputs.
- Unify the behavior of handling `missing` around R, Python, Java, and Scala DMatrix.
- Expose `num_non_missing` to the JVM interface.
- Deprecate old CSR and CSC constructors.
- Group C API.
- Add C API sphinx doc.
- Consistent use of `OptionalArg` and the parameter name `config`.
- Remove call to deprecated functions in demo.
- Fix some formatting errors.
- Add links to c examples in the document (only visible with doxygen pages)
- Fix arrow.
