Jiaming Yuan
0715ab3c10
Use dlopen to load NCCL. ( #9796 )
...
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.
2023-11-22 19:27:31 +08:00
Jiaming Yuan
ada377c57e
[coll] Reduce the scope of lock in the event loop. ( #9784 )
2023-11-15 14:16:19 +08:00
Jiaming Yuan
6fd4a30667
[coll] Increase timeout for allgather test. ( #9777 )
2023-11-09 05:26:40 +08:00
Jiaming Yuan
44099f585d
[coll] Add C API for the tracker. ( #9773 )
2023-11-08 18:17:14 +08:00
Jiaming Yuan
6c0a190f6d
[coll] Add comm group. ( #9759 )
...
- Implement `CommGroup` for double dispatching.
- Small cleanup to tracker for handling abort.
2023-11-07 11:12:31 +08:00
Jiaming Yuan
4da4e092b5
[coll] Improvements and fixes for tracker and allreduce. ( #9745 )
...
- Allow the tracker to wait.
- Fix allreduce type cast
- Return args from the federated tracker.
2023-11-02 04:06:46 +08:00
Jiaming Yuan
6755179e77
[coll] Add nccl. ( #9726 )
2023-10-28 16:33:58 +08:00
Jiaming Yuan
7a02facc9d
Serialize expand entry for allgather. ( #9702 )
2023-10-24 14:33:28 +08:00
Philip Hyunsu Cho
5e6cb63a56
[CI] Set up CI for Mac M1 ( #9699 )
2023-10-22 23:33:19 -07:00
Jiaming Yuan
b771f58453
[coll] Define interface for bridging. ( #9695 )
...
* Define the basic interface that will shared by nccl, federated and native.
2023-10-20 16:20:48 +08:00
Jiaming Yuan
5d1bcde719
[coll] allgatherv. ( #9688 )
2023-10-19 03:13:50 +08:00
Jiaming Yuan
4c0e4422d0
[coll] allgather. ( #9681 )
2023-10-18 10:22:18 +08:00
Jiaming Yuan
48ac9b6cbe
[coll] Allreduce. ( #9679 )
2023-10-17 13:57:14 +08:00
Jiaming Yuan
53049b16b8
[coll] Broadcast. ( #9659 )
2023-10-14 09:34:37 +08:00
Rong Ou
e164d51c43
Improve allgather functions ( #9649 )
2023-10-12 23:31:43 +08:00
Jiaming Yuan
946ae1c440
[coll] Implement a new tracker and a communicator. ( #9650 )
...
* [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.
2023-10-12 12:49:16 +08:00
Jiaming Yuan
b14e535e78
[Coll] Implement get host address in libxgboost. ( #9644 )
...
- Port `xgboost.tracker.get_host_ip` in C++.
2023-10-10 10:01:14 +08:00
Jiaming Yuan
8c676c889d
Remove internal use of gpu_id. ( #9568 )
2023-09-20 23:29:51 +08:00
Jiaming Yuan
38ac52dd87
Build a simple event loop for collective. ( #9593 )
2023-09-20 02:09:07 +08:00
Jiaming Yuan
b438d684d2
Utilities and cleanups for socket. ( #9576 )
...
- Use c++-17 nodiscard and nested ns.
- Add bind method to socket.
- Remove rabit parameters.
2023-09-14 01:41:42 +08:00
Jiaming Yuan
ccfc90e4c6
[rabit] Improved connection handling. ( #9531 )
...
- Enable timeout.
- Report connection error from the system.
- Handle retry for both tracker connection and peer connection.
2023-08-30 13:00:04 +08:00
Rong Ou
c2b85ab68a
Clean up MGPU C++ tests ( #9430 )
2023-08-02 14:31:18 +08:00
Rong Ou
15ca12a77e
Fix NCCL test hang ( #9367 )
2023-07-07 11:21:35 +08:00
Rong Ou
f90771eec6
Fix device communicator dependency ( #9346 )
2023-06-29 10:34:30 +08:00
Rong Ou
d8beb517ed
Support bitwise allreduce in NCCL communicator ( #9300 )
2023-06-17 01:56:50 +08:00
Rong Ou
e70810be8a
Refactor device communicator to make allreduce more flexible ( #9295 )
2023-06-14 03:53:03 +08:00
Rong Ou
52311dcec9
Fix multi-threaded gtests ( #9148 )
2023-05-10 19:15:32 +08:00
Jiaming Yuan
ea04d4c46c
[doc] [dask] Troubleshooting NCCL errors. ( #8943 )
2023-03-22 22:17:26 +08:00
Rong Ou
cbf98cb9c6
Add Allgather to collective communicator ( #8765 )
...
* Add Allgather to collective communicator
2023-02-09 11:31:22 +08:00
Rong Ou
78396f8a6e
Initial support for column-split cpu predictor ( #8676 )
2023-01-18 06:33:13 +08:00
Rong Ou
77b069c25d
Support bitwise allreduce operations in the communicator ( #8623 )
2022-12-25 06:40:05 +08:00
Rong Ou
a8255ea678
Add an in-memory collective communicator ( #8494 )
2022-12-01 00:24:12 +08:00
Jiaming Yuan
b791446623
Initial support for IPv6 ( #8225 )
...
- Merge rabit socket into XGBoost.
- Dask interface support.
- Add test to the socket.
2022-09-21 18:06:50 +08:00
Rong Ou
a2686543a9
Common interface for collective communication ( #8057 )
...
* implement broadcast for federated communicator
* implement allreduce
* add communicator factory
* add device adapter
* add device communicator to factory
* add rabit communicator
* add rabit communicator to the factory
* add nccl device communicator
* add synchronize to device communicator
* add back print and getprocessorname
* add python wrapper and c api
* clean up types
* fix non-gpu build
* try to fix ci
* fix std::size_t
* portable string compare ignore case
* c style size_t
* fix lint errors
* cross platform setenv
* fix memory leak
* fix lint errors
* address review feedback
* add python test for rabit communicator
* fix failing gtest
* use json to configure communicators
* fix lint error
* get rid of factories
* fix cpu build
* fix include
* fix python import
* don't export collective.py yet
* skip collective communicator pytest on windows
* add review feedback
* update documentation
* remove mpi communicator type
* fix tests
* shutdown the communicator separately
Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu>
2022-09-12 15:21:12 -07:00
