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.
- Add a test for blocking calls.
- Do not require the queue to be empty after waking up; this frees up the thread to answer blocking calls.
- Handle EOF in read.
- Improve the error message in the result. Allow concatenation of multiple results.
* Update collective implementation.
- Cleanup resource during `Finalize` to avoid handling threads in destructor.
- Calculate the size for allgather automatically.
- Use simple allgather for small (smaller than the number of worker) allreduce.
- Use the array interface internally.
- Deprecate `XGDMatrixSetDenseInfo`.
- Deprecate `XGDMatrixSetUIntInfo`.
- Move the handling of `DataType` into the deprecated C function.
---------
Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu>
- 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.
* Cleanup code for distributed training.
- Merge `GetNcclResult` into nccl stub.
- Split up utilities from the main dask module.
- Let Channel return `Result` to accommodate nccl channel.
- Remove old `use_label_encoder` parameter.
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.
- 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.
- Define a new data type, the proto file is copied for now.
- Merge client and communicator into `FederatedColl`.
- Define CUDA variant.
- Migrate tests for CPU, add tests for CUDA.
* [coll] Implement a new tracker and a communicator.
The new tracker and communicators communicate through the use of JSON documents. Along
with which, communicators are aware of each other.
Jiaming Yuan
Rong OuandJiaming Yuan