Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
7,086 Commits 32 Branches 59 Tags
Commit Graph

790 Commits

Author SHA1 Message Date
Jiaming Yuan
96bbf80457
[EM] Suport quantile objectives for GPU-based external memory. (#10820) 
- Improved error message for memory usage.
- Support quantile-based objectives for GPU external memory.
 2024-09-17 13:27:02 +08:00
Jiaming Yuan
d94f6679fc
[EM] Avoid synchronous calls and unnecessary ATS access. (#10811) 
- Pass context into various functions.
- Factor out some CUDA algorithms.
- Use ATS only for update position.
 2024-09-10 14:33:14 +08:00
Jiaming Yuan
ed5f33df16
[EM] Multi-level quantile sketching for GPU. (#10813) 2024-09-10 13:08:34 +08:00
Jiaming Yuan
5f7f31d464
[EM] Refactor ellpack construction. (#10810) 
- Remove the calculation of n_symbols in the accessor.
- Pack initialization steps into the parameter list.
- Pass the context into various ctors.
- Specialization for dense data to prepare for further compression.
 2024-09-09 14:10:10 +08:00
Jiaming Yuan
e1a2c1bbb3
[EM] Merge GPU partitioning with histogram building. (#10766) 
- Stop concatenating pages if there's no subsampling.
- Use a single iteration for histogram build and partitioning.
 2024-08-31 03:25:37 +08:00
Jiaming Yuan
98ac153265
Avoid warning from NVCC. (#10757) 2024-08-30 16:11:31 +08:00
Jiaming Yuan
34d4ab455e
[EM] Avoid stream sync in quantile sketching. (#10765) 
.
 2024-08-30 12:33:24 +08:00
Jiaming Yuan
61dd854a52
[EM] Refactor GPU histogram builder. (#10764) 
- Expose the maximum number of cached nodes to be consistent with the CPU implementation. Also easier for testing.
- Extract the subtraction trick for easier testing.
- Split up the `GradientQuantiser` to avoid circular dependency.
 2024-08-30 02:39:14 +08:00
Jiaming Yuan
4fe67f10b4
[EM] Have one partitioner for each batch. (#10760) 
- Initialize one partitioner for each batch.
- Collect partition size during initialization.
- Support base ridx in the finalization.
 2024-08-29 01:35:17 +08:00
Jiaming Yuan
64afe9873b
Increase timeout in C++ tests from 1 to 5 seconds. (#10756) 
To avoid CI failures on FreeBSD.
 2024-08-28 02:27:14 +08:00
Jiaming Yuan
bde1265caf
[EM] Return a full DMatrix instead of a Ellpack from the GPU sampler. (#10753) 2024-08-28 01:05:11 +08:00
Jiaming Yuan
d6ebcfb032
[EM] Support CPU quantile objective for external memory. (#10751) 2024-08-27 04:16:57 +08:00
Jiaming Yuan
25966e4ba8
[EM] Pass batch parameter into extmem format. (#10736) 
- Allow customization for format reading.
- Customize the number of pre-fetch batches.
 2024-08-27 02:37:50 +08:00
Jiaming Yuan
fd0138c91c
[coll] Improve column split tests with named threads. (#10735) 2024-08-24 12:43:47 +08:00
Jiaming Yuan
55aef8f546
[EM] Avoid resizing host cache. (#10734) 
* [EM] Avoid resizing host cache.

- Add SAM allocator and resource.
- Use page-based cache instead of stream-based cache.
 2024-08-23 06:34:01 +08:00
Jiaming Yuan
142bdc73ec
[EM] Support SHAP contribution with QDM. (#10724) 
- Add GPU support.
- Add external memory support.
- Update the GPU tree shap.
 2024-08-22 05:25:10 +08:00
Jiaming Yuan
cb54374550
Update clang-tidy. (#10730) 
- Install cmake using pip.
- Fix compile command generation.
- Clean up the tidy script and remove the need to load the yaml file.
- Fix modernized type traits.
- Fix span class. Polymorphism support is dropped
 2024-08-22 04:12:18 +08:00
Dmitry Razdoburdin
24d225c1ab
[SYCL] Implement UpdatePredictionCache and connect updater with leraner. (#10701) 
---------

Co-authored-by: Dmitry Razdoburdin <>
 2024-08-22 02:07:44 +08:00
Jiaming Yuan
402e7837fb
Fix potential race in feature constraint. (#10719) 2024-08-21 16:50:31 +08:00
Jiaming Yuan
508ac13243
Check cub errors. (#10721) 
- Make sure cuda error returned by cub scan is caught.
- Avoid temporary buffer allocation in thrust device vector.
 2024-08-21 02:50:26 +08:00
Jiaming Yuan
ec3f327c20
Add managed memory allocator. (#10711) 2024-08-17 03:02:34 +08:00
Jiaming Yuan
8d7fe262d9
[EM] Enable access to the number of batches. (#10691) 
- Expose `NumBatches` in `DMatrix`.
- Small cleanup for removing legacy CUDA stream and ~force CUDA context initialization~.
- Purge old external memory data generation code.
 2024-08-17 02:59:45 +08:00
Jiaming Yuan
abe65e3769
Reduce thread contention in column split histogram test. (#10708) 2024-08-17 01:00:32 +08:00
Jiaming Yuan
582ea104b5
[EM] Enable prediction cache for GPU. (#10707) 
- 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.
 2024-08-15 21:41:59 +08:00
Dmitry Razdoburdin
773ded684b
[sycl] Add depth-wise policy (#10690) 
Co-authored-by: Dmitry Razdoburdin <>
 2024-08-13 18:12:35 +08:00
Jiaming Yuan
2ecc85ffad
[EM] Support ExtMemQdm in the GPU predictor. (#10694) 2024-08-13 12:21:11 +08:00
Jiaming Yuan
43704549a2
[coll] Reduce the amount of open files (socket). (#10693) 
Reduce the chance of hitting `Failed to call `socket`: Too many open files`.
 2024-08-13 05:23:49 +08:00
Jiaming Yuan
d414fdf2e7
[EM] Add GPU version of the external memory QDM. (#10689) 2024-08-10 10:49:43 +08:00
Jiaming Yuan
7bccc1ea2c
[EM] CPU implementation for external memory QDM. (#10682) 
- A new DMatrix type.
- Extract common code into a new QDM base class.

Not yet working:
- Not exposed to the interface yet, will wait for the GPU implementation.
- ~No meta info yet, still working on the source.~
- Exporting data to CSR is not supported yet.
 2024-08-09 09:38:02 +08:00
Dmitry Razdoburdin
e555a238bc
[SYCL]. Add implementation for loss-guided policy (#10681) 
---------

Co-authored-by: Dmitry Razdoburdin <>
 2024-08-09 09:04:46 +08:00
Jiaming Yuan
cc3b56fc37
Cleanup GPU Hist tests. (#10677) 
* Cleanup GPU Hist tests.

- Remove GPU Hist gradient sampling test. The same properties are tested in the gradient
  sampler test suite.
- Move basic histogram tests into the histogram test suite.
- Remove the header inclusion of the `updater_gpu_hist.cu` in tests.
 2024-08-06 11:50:44 +08:00
Jiaming Yuan
77c844cef7
Reduce thread contention in column split tests. (#10658) 
---------

Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu>
 2024-08-01 18:36:46 +08:00
Dmitry Razdoburdin
7720272870
[sycl] add split applications and tests (#10636) 
Co-authored-by: Dmitry Razdoburdin <>
 2024-07-26 15:25:49 +08:00
Jiaming Yuan
a19bbc9be5
Avoid caching allocator for large allocations. (#10582) 2024-07-23 03:48:03 +08:00
Dmitry Razdoburdin
f6cae4da85
[SYCL] Add splits evaluation (#10605) 
---------

Co-authored-by: Dmitry Razdoburdin <>
 2024-07-22 18:14:06 +08:00
Jiaming Yuan
6d9fcb771e
Move device histogram storage into histogram.cuh. (#10608) 2024-07-21 14:10:13 +08:00
Jiaming Yuan
292bb677e5
[EM] Support mmap backed ellpack. (#10602) 
- Support resource view in ellpack.
- Define the CUDA version of MMAP resource.
- Define the CUDA version of malloc resource.
- Refactor cuda runtime API wrappers, and add memory access related wrappers.
- gather windows macros into a single header.
 2024-07-18 08:20:21 +08:00
Jiaming Yuan
e9fbce9791
Refactor DeviceUVector. (#10595) 
Create a wrapper instead of using inheritance to avoid inconsistent interface of the class.
 2024-07-18 03:33:01 +08:00
Jiaming Yuan
a6a8a55ffa
Merge approx tests. (#10583) 2024-07-16 19:03:48 +08:00
Jiaming Yuan
6c403187ec
Fix column split race condition. (#10572) 2024-07-12 01:07:12 +08:00
Jiaming Yuan
1ca4bfd20e
Avoid thrust vector initialization. (#10544) 
* Avoid thrust vector initialization.

- Add a wrapper for rmm device uvector.
- Split up the `Resize` method for HDV.
 2024-07-11 17:29:27 +08:00
Jiaming Yuan
89da9f9741
[fed] Split up federated test CMake file. (#10566) 
- Collect all federated test files into the same directory.
- Independently list the files.
 2024-07-11 13:09:18 +08:00
Jiaming Yuan
5f910cd4ff
[EM] Handle base idx in GPU histogram. (#10549) 2024-07-11 03:26:30 +08:00
Jiaming Yuan
34b154c284
Avoid the use of size_t in the partitioner. (#10541) 
- 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.
 2024-07-11 00:43:08 +08:00
Jiaming Yuan
620b2b155a
Cache GPU histogram kernel configuration. (#10538) 2024-07-04 15:38:59 +08:00
Jiaming Yuan
628411a654
Enhance the threadpool implementation. (#10531) 
- Accept an initialization function.
- Support void return tasks.
 2024-07-03 12:13:27 +08:00
Jiaming Yuan
9cb4c938da
[EM] Move prefetch in reset into the end of the iteration. (#10529) 2024-07-03 03:48:18 +08:00
Philip Hyunsu Cho
09d32f1f2b
Fix build and C++ tests for FreeBSD (#10480) 2024-06-28 01:47:55 -07:00
Jiaming Yuan
e8a962575a
[EM] Allow staging ellpack on host for GPU external memory. (#10488) 
- New parameter `on_host`.
- Abstract format creation and stream creation into policy classes.
 2024-06-28 04:42:18 +08:00
Jiaming Yuan
26eb68859f
Consistently report error in tests. (#10453) 2024-06-21 14:35:22 +08:00