Merge remote-tracking branch 'upstream/master' into redesign-jvm

.github/workflows/main.yml

@@ -180,7 +180,7 @@ jobs:
     - uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29 # v4.1.6
       with:
         submodules: 'true'
-    - uses: actions/setup-python@39cd14951b08e74b54015e9e001cdefcf80e669f # v5.1.1
+    - uses: actions/setup-python@f677139bbe7f9c59b41e40162b753c062f5d49a3 # v5.2.0
       with:
         python-version: "3.10"
         architecture: 'x64'

.github/workflows/python_tests.yml

@@ -319,7 +319,7 @@ jobs:
           submodules: 'true'
 
       - name: Set up Python 3.10
-        uses: actions/setup-python@39cd14951b08e74b54015e9e001cdefcf80e669f # v5.1.1
+        uses: actions/setup-python@f677139bbe7f9c59b41e40162b753c062f5d49a3 # v5.2.0
         with:
           python-version: "3.10"
 

.github/workflows/r_tests.yml

@@ -84,7 +84,7 @@ jobs:
         key: ${{ runner.os }}-r-${{ matrix.config.r }}-7-${{ hashFiles('R-package/DESCRIPTION') }}
         restore-keys: ${{ runner.os }}-r-${{ matrix.config.r }}-7-${{ hashFiles('R-package/DESCRIPTION') }}
 
-    - uses: actions/setup-python@39cd14951b08e74b54015e9e001cdefcf80e669f # v5.1.1
+    - uses: actions/setup-python@f677139bbe7f9c59b41e40162b753c062f5d49a3 # v5.2.0
       with:
         python-version: "3.10"
         architecture: 'x64'

python-package/xgboost/dask/__init__.py

@@ -1568,14 +1568,15 @@ def inplace_predict(  # pylint: disable=unused-argument
 
 async def _async_wrap_evaluation_matrices(
     client: Optional["distributed.Client"],
+    device: Optional[str],
     tree_method: Optional[str],
     max_bin: Optional[int],
     **kwargs: Any,
 ) -> Tuple[DaskDMatrix, Optional[List[Tuple[DaskDMatrix, str]]]]:
     """A switch function for async environment."""
 
     def _dispatch(ref: Optional[DaskDMatrix], **kwargs: Any) -> DaskDMatrix:
-        if _can_use_qdm(tree_method):
+        if _can_use_qdm(tree_method, device):
             return DaskQuantileDMatrix(
                 client=client, ref=ref, max_bin=max_bin, **kwargs
             )
@@ -1776,6 +1777,7 @@ async def _fit_async(
         params = self.get_xgb_params()
         dtrain, evals = await _async_wrap_evaluation_matrices(
             client=self.client,
+            device=self.device,
             tree_method=self.tree_method,
             max_bin=self.max_bin,
             X=X,
@@ -1865,6 +1867,7 @@ async def _fit_async(
         params = self.get_xgb_params()
         dtrain, evals = await _async_wrap_evaluation_matrices(
             self.client,
+            device=self.device,
             tree_method=self.tree_method,
             max_bin=self.max_bin,
             X=X,
@@ -2067,6 +2070,7 @@ async def _fit_async(
         params = self.get_xgb_params()
         dtrain, evals = await _async_wrap_evaluation_matrices(
             self.client,
+            device=self.device,
             tree_method=self.tree_method,
             max_bin=self.max_bin,
             X=X,

python-package/xgboost/sklearn.py

@@ -67,8 +67,9 @@ def _check_rf_callback(
         )
 
 
-def _can_use_qdm(tree_method: Optional[str]) -> bool:
-    return tree_method in ("hist", "gpu_hist", None, "auto")
+def _can_use_qdm(tree_method: Optional[str], device: Optional[str]) -> bool:
+    not_sycl = (device is None) or (not device.startswith("sycl"))
+    return tree_method in ("hist", "gpu_hist", None, "auto") and not_sycl
 
 
 class _SklObjWProto(Protocol):  # pylint: disable=too-few-public-methods
@@ -1031,7 +1032,7 @@ def _duplicated(parameter: str) -> None:
 
     def _create_dmatrix(self, ref: Optional[DMatrix], **kwargs: Any) -> DMatrix:
         # Use `QuantileDMatrix` to save memory.
-        if _can_use_qdm(self.tree_method) and self.booster != "gblinear":
+        if _can_use_qdm(self.tree_method, self.device) and self.booster != "gblinear":
             try:
                 return QuantileDMatrix(
                     **kwargs, ref=ref, nthread=self.n_jobs, max_bin=self.max_bin

python-package/xgboost/spark/core.py

@@ -1028,7 +1028,10 @@ def _train_booster(
             context = BarrierTaskContext.get()
 
             dev_ordinal = None
-            use_qdm = _can_use_qdm(booster_params.get("tree_method", None))
+            use_qdm = _can_use_qdm(
+                booster_params.get("tree_method", None),
+                booster_params.get("device", None),
+            )
             verbosity = booster_params.get("verbosity", 1)
             msg = "Training on CPUs"
             if run_on_gpu:

src/common/compressed_iterator.h

@@ -77,13 +77,11 @@ class CompressedBufferWriter {
   static size_t CalculateBufferSize(size_t num_elements, size_t num_symbols) {
     constexpr int kBitsPerByte = 8;
     size_t compressed_size = static_cast<size_t>(std::ceil(
-        static_cast<double>(detail::SymbolBits(num_symbols) * num_elements) /
-        kBitsPerByte));
+        static_cast<double>(detail::SymbolBits(num_symbols) * num_elements) / kBitsPerByte));
     // Handle atomicOr where input must be unsigned int, hence 4 bytes aligned.
-    size_t ret =
-        std::ceil(static_cast<double>(compressed_size + detail::kPadding) /
-                  static_cast<double>(sizeof(unsigned int))) *
-        sizeof(unsigned int);
+    size_t ret = std::ceil(static_cast<double>(compressed_size + detail::kPadding) /
+                           static_cast<double>(sizeof(std::uint32_t))) *
+                 sizeof(std::uint32_t);
     return ret;
   }