- Expose `NumBatches` in `DMatrix`.
- Small cleanup for removing legacy CUDA stream and ~force CUDA context initialization~.
- Purge old external memory data generation code.
- Use `UpdatePosition` for all nodes and skip `FinalizePosition` when external memory is used.
- Create `encode/decode` for node position, this is just as a refactor.
- Reuse code between update position and finalization.
- Avoid the use of size_t in the partitioner.
- Use `Span` instead of `Elem` where `node_id` is not needed.
- Remove the `const_cast`.
- Make sure the constness is not removed in the `Elem` by making it reference only.
size_t is implementation-defined, which causes issue when we want to pass pointer or span.
This PR replaces the original RABIT implementation with a new one, which has already been partially merged into XGBoost. The new one features:
- Federated learning for both CPU and GPU.
- NCCL.
- More data types.
- A unified interface for all the underlying implementations.
- Improved timeout handling for both tracker and workers.
- Exhausted tests with metrics (fixed a couple of bugs along the way).
- A reusable tracker for Python and JVM packages.
- Use std::uint64_t instead of size_t to avoid implementation-defined type.
- Rename to bst_idx_t, to account for other types of indexing.
- Small cleanup to the base header.
- CUDA implementation.
- Extract the broadcasting logic, we will need the context parameter after revamping the collective implementation.
- Some changes to the event loop for fixing a deadlock in CI.
- Move argsort into algorithms.cuh, add support for cuda stream.
* [coll] Pass context to various functions.
In the future, the `Context` object would be required for collective operations, this PR
passes the context object to some required functions to prepare for swapping out the
implementation.
- Pass context from booster to DMatrix.
- Use context instead of integer for `n_threads`.
- Check the consistency configuration for `max_bin`.
- Test for all combinations of initialization options.
* Pass sparse page as adapter, which prepares for quantile dmatrix.
* Remove old external memory code like `rbegin` and extra `Init` function.
* Simplify type dispatch.
* Generate column matrix from gHistIndex.
* Avoid synchronization with the sparse page once the cache is written.
* Cleanups: Remove member variables/functions, change the update routine to look like approx and gpu_hist.
* Remove pruner.
* Extract partitioner from hist.
* Implement categorical data support by passing the gradient index directly into the partitioner.
* Organize/update document.
* Remove code for negative hessian.
* Implement `MaxCategory` in quantile.
* Implement partition-based split for GPU evaluation. Currently, it's based on the existing evaluation function.
* Extract an evaluator from GPU Hist to store the needed states.
* Added some CUDA stream/event utilities.
* Update document with references.
* Fixed a bug in approx evaluator where the number of data points is less than the number of categories.
This PR prepares the GHistIndexMatrix to host the column matrix which is used by the hist tree method by accepting sparse_threshold parameter.
Some cleanups are made to ensure the correct batch param is being passed into DMatrix along with some additional tests for correctness of SimpleDMatrix.
Other than modularizing the split evaluation function, this PR also removes some more functions including `InitNewNodes` and `BuildNodeStats` among some other unused variables. Also, scattered code like setting leaf weights is grouped into the split evaluator and `NodeEntry` is simplified and made private. Another subtle difference with the original implementation is that the modified code doesn't call `tree[nidx].Parent()` to traversal upward.
