Merge branch-25.06 into branch-25.08

Author: Matt711

ci/test_narwhals.sh

@@ -26,9 +26,12 @@ rapids-logger "Check narwhals versions"
 python -c "import narwhals; print(narwhals.show_versions())"
 
 rapids-logger "Run narwhals tests for cuDF"
-python -m pytest \
+PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest \
     --cache-clear \
     --junitxml="${RAPIDS_TESTS_DIR}/junit-cudf-narwhals.xml" \
+    -p xdist \
+    -p env \
+    -p no:pytest_benchmark \
     -p cudf.testing.narwhals_test_plugin \
     --numprocesses=8 \
     --dist=worksteal \
@@ -44,9 +47,12 @@ test_nan \
 "
 
 rapids-logger "Run narwhals tests for cuDF Polars"
-NARWHALS_POLARS_GPU=1 python -m pytest \
+PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 NARWHALS_POLARS_GPU=1 python -m pytest \
     --cache-clear \
     --junitxml="${RAPIDS_TESTS_DIR}/junit-cudf-polars-narwhals.xml" \
+    -p xdist \
+    -p env \
+    -p no:pytest_benchmark \
     -k "not ( \
         ${TESTS_THAT_NEED_NARWHALS_FIX_FOR_CUDF_POLARS} \
     )" \
@@ -84,10 +90,13 @@ TESTS_THAT_NEED_NARWHALS_FIX_FOR_CUDF_PANDAS=" \
 test_dtypes \
 "
 
-NARWHALS_DEFAULT_CONSTRUCTORS=pandas python -m pytest \
+PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 NARWHALS_DEFAULT_CONSTRUCTORS=pandas python -m pytest \
     -p cudf.pandas \
     --cache-clear \
     --junitxml="${RAPIDS_TESTS_DIR}/junit-cudf-pandas-narwhals.xml" \
+    -p xdist \
+    -p env \
+    -p no:pytest_benchmark \
     -k "not ( \
         ${TESTS_THAT_NEED_CUDF_FIX} or \
         ${TESTS_TO_ALWAYS_SKIP} or \

cpp/src/rolling/detail/rolling.cuh

@@ -126,9 +126,16 @@ struct DeviceRolling {
 
     bool output_is_valid = (count >= min_periods);
 
-    // store the output value, one per thread
-    cudf::detail::rolling_store_output_functor<OutputType, op == aggregation::MEAN>{}(
-      output.element<OutputType>(current_index), val, count);
+    if (output_is_valid) {
+      // store the output value, one per thread, but only if the
+      // output is valid. min_periods is required to be >= 1, and so
+      // here, count must be nonzero. We need to avoid storing if
+      // count is zero since this could cause UB in some aggregations,
+      // which may cause the compiler to deduce nonsense about the loop
+      // that increments count.
+      cudf::detail::rolling_store_output_functor<OutputType, op == aggregation::MEAN>{}(
+        output.element<OutputType>(current_index), val, count);
+    }
 
     return output_is_valid;
   }

cpp/src/rolling/detail/rolling.hpp

@@ -35,26 +35,23 @@ struct rolling_store_output_functor {
 // Specialization for MEAN
 template <typename _T>
 struct rolling_store_output_functor<_T, true> {
-  // SFINAE for non-bool types
+  // Don't store the output if count is zero since integral division by zero is
+  // undefined behaviour. The caller must ensure that the relevant row is
+  // marked as invalid with a null.
+
+  // SFINAE for non-bool, non-timestamp types
   template <typename T                                                             = _T,
             std::enable_if_t<!(cudf::is_boolean<T>() || cudf::is_timestamp<T>())>* = nullptr>
   CUDF_HOST_DEVICE inline void operator()(T& out, T& val, size_type count)
   {
-    out = val / count;
-  }
-
-  // SFINAE for bool type
-  template <typename T = _T, std::enable_if_t<cudf::is_boolean<T>()>* = nullptr>
-  CUDF_HOST_DEVICE inline void operator()(T& out, T& val, size_type count)
-  {
-    out = static_cast<int32_t>(val) / count;
+    if (count > 0) { out = val / count; }
   }
 
   // SFINAE for timestamp types
   template <typename T = _T, std::enable_if_t<cudf::is_timestamp<T>()>* = nullptr>
   CUDF_HOST_DEVICE inline void operator()(T& out, T& val, size_type count)
   {
-    out = static_cast<T>(val.time_since_epoch() / count);
+    if (count > 0) { out = static_cast<T>(val.time_since_epoch() / count); }
   }
 };