Rong Ou
def77870f3
Test categorical features with column-split gpu quantile ( #9595 )
2023-09-23 09:55:09 +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
Rong Ou
d8c3cc92ae
More support for column split in gpu predictor ( #9562 )
2023-09-14 08:13:13 +08:00
Jiaming Yuan
300f9ace06
Fix default metric configuration. ( #9575 )
2023-09-13 13:05:47 -07:00
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
Rong Ou
66a0832778
Add tests for gpu_approx ( #9553 )
2023-09-07 17:21:58 +08:00
Jiaming Yuan
adea842c83
Fix inplace predict with fallback when base margin is used. ( #9536 )
...
- Copy meta info from proxy DMatrix.
- Use `std::call_once` to emit less warnings.
2023-09-05 01:04:24 +08:00
Rong Ou
c928dd4ff5
Support vertical federated learning with gpu_hist ( #9539 )
2023-09-03 11:37:11 +08:00
Rong Ou
9bab06cbca
Support column split in gpu hist updater ( #9384 )
2023-08-31 18:09:35 +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
Jiaming Yuan
ddf2e68821
Use the new DeviceOrd in the linalg module. ( #9527 )
2023-08-29 13:37:29 +08:00
Jiaming Yuan
972730cde0
Use matrix for gradient. ( #9508 )
...
- 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)
2023-08-24 05:29:52 +08:00
Rong Ou
6103dca0bb
Support column split in GPU evaluate splits ( #9511 )
2023-08-23 16:33:43 +08:00
Jiaming Yuan
3c09399f29
Fix device dispatch for linear updater. ( #9507 )
2023-08-23 00:17:35 +08:00
Jiaming Yuan
044fea1281
Drop support for loading remote files. ( #9504 )
2023-08-21 23:34:05 +08:00
Jiaming Yuan
1caa93221a
Use realloc for histogram cache and expose the cache limit. ( #9455 )
2023-08-10 14:05:27 +08:00
Jiaming Yuan
f05a23b41c
Use weakref instead of id for DataIter cache. ( #9445 )
...
- Fix case where Python reuses id from freed objects.
- Small optimization to column matrix with QDM by using `realloc` instead of copying data.
2023-08-10 00:40:06 +08:00
Philip Hyunsu Cho
7ce090e775
Handle UTF-8 paths correctly on Windows platform ( #9443 )
...
* Fix round-trip serialization with UTF-8 paths
* Add compiler version check
* Add comment to C API functions
* Add Python tests
* [CI] Updatre MacOS deployment target
* Use std::filesystem instead of dmlc::TemporaryDirectory
2023-08-07 23:27:25 -07:00
Jiaming Yuan
54029a59af
Bound the size of the histogram cache. ( #9440 )
...
- A new histogram collection with a limit in size.
- Unify histogram building logic between hist, multi-hist, and approx.
2023-08-08 03:21:26 +08:00
Rong Ou
bde1ebc209
Switch back to the GPUIDX macro ( #9438 )
2023-08-04 15:14:31 +08:00
Jiaming Yuan
1332ff787f
Unify the code path between local and distributed training. ( #9433 )
...
This removes the need for a local histogram space during distributed training, which cuts the cache size by half.
2023-08-03 21:46:36 +08:00
Jiaming Yuan
e93a274823
Small cleanup for histogram routines. ( #9427 )
...
* Small cleanup for histogram routines.
- Extract hist train param from GPU hist.
- Make histogram const after construction.
- Unify parameter names.
2023-08-02 18:28:26 +08:00
Rong Ou
c2b85ab68a
Clean up MGPU C++ tests ( #9430 )
2023-08-02 14:31:18 +08:00
Jiaming Yuan
912e341d57
Initial GPU support for the approx tree method. ( #9414 )
2023-07-31 15:50:28 +08:00
Rong Ou
7579905e18
Retry switching to per-thread default stream ( #9416 )
2023-07-26 07:09:12 +08:00
Jiaming Yuan
3a9996173e
Revert "Switch to per-thread default stream ( #9396 )" ( #9413 )
...
This reverts commit f7f673b00c15458fb4dd74a2a0d2ba80369c5faf.
2023-07-24 12:03:28 -07:00
Jiaming Yuan
a196443a07
Implement sketching with Hessian on GPU. ( #9399 )
...
- Prepare for implementing approx on GPU.
- Unify the code path between weighted and uniform sketching on DMatrix.
2023-07-24 15:43:03 +08:00
Jiaming Yuan
22b0a55a04
Remove hist builder class. ( #9400 )
...
* Remove hist build class.
* Cleanup this stateless class.
* Add comment to thread block.
2023-07-22 10:43:12 +08:00
Jiaming Yuan
0de7c47495
Fix metric serialization. ( #9405 )
2023-07-22 08:39:21 +08:00
Rong Ou
f7f673b00c
Switch to per-thread default stream ( #9396 )
2023-07-20 08:21:00 +08:00
Jiaming Yuan
04aff3af8e
Define the new device parameter. ( #9362 )
2023-07-13 19:30:25 +08:00
Rong Ou
3632242e0b
Support column split with GPU quantile ( #9370 )
2023-07-11 12:15:56 +08:00
Jiaming Yuan
97ed944209
Unify the hist tree method for different devices. ( #9363 )
2023-07-11 10:04:39 +08:00
Jiaming Yuan
20c52f07d2
Support exporting cut values ( #9356 )
2023-07-08 15:32:41 +08:00
Rong Ou
15ca12a77e
Fix NCCL test hang ( #9367 )
2023-07-07 11:21:35 +08:00
Jiaming Yuan
41c6813496
Preserve order of saved updaters config. ( #9355 )
...
- Save the updater sequence as an array instead of object.
- Warn only once.
The compatibility is kept, but we should be able to break it as the config is not loaded
in pickle model and it's declared to be not stable.
2023-07-05 20:20:07 +08:00
Jiaming Yuan
645037e376
Improve test coverage with predictor configuration. ( #9354 )
...
* Improve test coverage with predictor configuration.
- Test with ext memory.
- Test with QDM.
- Test with dart.
2023-07-05 15:17:22 +08:00
Jiaming Yuan
d0916849a6
Remove unused weight from buffer for cat features. ( #9341 )
2023-07-04 01:07:09 +08:00
Jiaming Yuan
39390cc2ee
[breaking] Remove the predictor param, allow fallback to prediction using DMatrix. ( #9129 )
...
- 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.
2023-07-03 19:23:54 +08:00
Rong Ou
3a0f787703
Support column split in GPU predictor ( #9343 )
2023-07-03 04:05:34 +08:00
Rong Ou
f90771eec6
Fix device communicator dependency ( #9346 )
2023-06-29 10:34:30 +08:00
Jiaming Yuan
f4798718c7
Use hist as the default tree method. ( #9320 )
2023-06-27 23:04:24 +08:00
Jiaming Yuan
bc267dd729
Use ptr from mmap for GHistIndexMatrix and ColumnMatrix. ( #9315 )
...
* Use ptr from mmap for `GHistIndexMatrix` and `ColumnMatrix`.
- Define a resource for holding various types of memory pointers.
- Define ref vector for holding resources.
- Swap the underlying resources for GHist and ColumnM.
- Add documentation for current status.
- s390x support is removed. It should work if you can compile XGBoost, all the old workaround code does is to get GCC to compile.
2023-06-27 19:05:46 +08:00
Jiaming Yuan
54da4b3185
Cleanup to prepare for using mmap pointer in external memory. ( #9317 )
...
- Update SparseDMatrix comment.
- Use a pointer in the bitfield. We will replace the `std::vector<bool>` in `ColumnMatrix` with bitfield.
- Clean up the page source. The timer is removed as it's inaccurate once we swap the mmap pointer into the page.
2023-06-22 06:43:11 +08:00
Jiaming Yuan
ee6809e642
Use mmap for external memory. ( #9282 )
...
- Have basic infrastructure for mmap.
- Release file write handle.
2023-06-19 18:52:55 +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
Jiaming Yuan
152e2fb072
Unify test helpers for creating ctx. ( #9274 )
2023-06-10 03:35:22 +08:00
Rong Ou
ff122d61ff
More tests for cpu predictor with column split ( #9270 )
2023-06-08 22:47:19 +08:00
