Merge branch 'main' into feature/sparse-page-io-hybrid

Author: mhaseeb123

.github/workflows/pr.yaml

@@ -422,6 +422,8 @@ jobs:
     with:
       build_type: pull-request
       script: ci/test_cpp.sh
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   conda-python-build:
     needs: conda-cpp-build
     permissions:
@@ -465,6 +467,8 @@ jobs:
     with:
       build_type: pull-request
       script: "ci/test_python_cudf.sh"
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   conda-python-other-tests:
     # Tests for dask_cudf, cudf_polars, custreamz, cudf_kafka are separated for CI parallelism
     needs: [conda-python-build-noarch, changed-files]
@@ -482,6 +486,8 @@ jobs:
       # https://github.com/rapidsai/cudf/pull/22381/changes#r3196736965
       container-options: "--cap-add CAP_SYS_PTRACE --shm-size=8g --ulimit=nofile=1000000:1000000"
       script: "ci/test_python_other.sh"
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   conda-java-tests:
     needs: [conda-cpp-build, changed-files]
     permissions:
@@ -601,6 +607,8 @@ jobs:
     with:
       build_type: pull-request
       script: ci/test_wheel_cudf_streaming.sh
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   wheel-build-pylibcudf:
     needs: [checks, wheel-build-libcudf]
     permissions:
@@ -651,6 +659,8 @@ jobs:
     with:
       build_type: pull-request
       script: ci/test_wheel_cudf.sh
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   wheel-build-cudf-polars:
     needs: wheel-build-pylibcudf
     permissions:
@@ -775,7 +785,8 @@ jobs:
     if: (fromJSON(needs.changed-files.outputs.changed_file_groups).test_python_wheels || fromJSON(needs.changed-files.outputs.changed_file_groups).test_cudf_pandas) && fromJSON(needs.changed-files.outputs.changed_file_groups).neither_cudf_polars_nor_dask_cudf
     with:
       # This selects the latest supported Python + CUDA minor versions for each ARCH/CUDA major version combo
-      matrix_filter: group_by([(.ARCH), (.CUDA_VER|split(".")|map(tonumber)|.[0])]) | map(max_by([(.PY_VER|split(".")|map(tonumber)), (.CUDA_VER|split(".")|map(tonumber))]))
+      # Filter out GB300 due to https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: group_by([(.ARCH), (.CUDA_VER|split(".")|map(tonumber)|.[0])]) | map(max_by([(.PY_VER|split(".")|map(tonumber)), (.CUDA_VER|split(".")|map(tonumber))])) | map(select(.GPU != "gb300" and .GPU != "gh200"))
       build_type: pull-request
       script: ci/cudf_pandas_scripts/run_tests.sh
   third-party-integration-tests-cudf-pandas:

.github/workflows/test.yaml

@@ -54,6 +54,8 @@ jobs:
       date: ${{ inputs.date }}
       script: ci/test_cpp.sh
       sha: ${{ inputs.sha }}
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   conda-cpp-benchmark-tests:
     permissions:
       actions: read
@@ -121,6 +123,8 @@ jobs:
       date: ${{ inputs.date }}
       sha: ${{ inputs.sha }}
       script: "ci/test_python_cudf.sh"
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   conda-python-other-tests:
     # Tests for dask_cudf, custreamz, cudf_kafka are separated for CI parallelism
     permissions:
@@ -137,6 +141,8 @@ jobs:
       date: ${{ inputs.date }}
       sha: ${{ inputs.sha }}
       script: "ci/test_python_other.sh"
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   conda-java-tests:
     permissions:
       actions: read
@@ -218,6 +224,8 @@ jobs:
       date: ${{ inputs.date }}
       sha: ${{ inputs.sha }}
       script: ci/cudf_pandas_scripts/run_tests.sh
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   third-party-integration-tests-cudf-pandas:
     permissions:
       actions: read
@@ -271,6 +279,8 @@ jobs:
       date: ${{ inputs.date }}
       sha: ${{ inputs.sha }}
       script: "ci/test_wheel_cudf_polars.sh"
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   cudf-polars-polars-tests:
     permissions:
       actions: read
@@ -286,6 +296,8 @@ jobs:
       date: ${{ inputs.date }}
       sha: ${{ inputs.sha }}
       script: "ci/test_cudf_polars_polars_tests.sh"
+      # https://github.com/rapidsai/cudf/issues/23498
+      matrix_filter: map(select(.GPU != "gb300" and .GPU != "gh200"))
   narwhals-tests:
     permissions:
       actions: read

cpp/doxygen/developer_guide/DEVELOPER_GUIDE.md

@@ -1578,7 +1578,7 @@ the null masks of both struct fields.
 ## Dictionary columns
 
 Dictionaries provide an efficient way to represent low-cardinality data by storing a single copy
-of each value. A dictionary comprises a column of distinct keys and a column containing an index into
+of each value. A dictionary comprises a column of keys and a column containing an index into
 the keys column for each row of the parent column. The keys column may have any fixed-width data_type
 or STRING data_type. The indices represent the corresponding positions of each
 element's value in the keys. The indices child column can have any signed integer type
@@ -1589,8 +1589,11 @@ input column will produce equivalent dictionary columns but the keys may be in a
 and therefore the indices will not match as well. Using `cudf::dictionary::decode()` on both dictionary
 columns should produce the same result.
 
-Although `cudf::make_dictionary_column()` expects distinct keys, the API does not enforce this constraint.
-Using a dictionary column with non-distinct keys in libcudf APIs may result in undefined behavior.
+The libcudf APIs also accept dictionary columns with non-unique keys.
+However, output dictionary columns will generally contain unique keys in an unspecified order.
+The exceptions are `cudf::make_dictionary_column()`, which accepts keys and indices without
+changing them, and `cudf::dictionary::set_keys()`, which strictly honors the given keys
+(both order and duplicates).
 
 ## Nested column challenges
 

cpp/include/cudf/column/column_device_view.cuh

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2019-2026, NVIDIA CORPORATION.
+ * SPDX-FileCopyrightText: Copyright (c) 2019-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * SPDX-License-Identifier: Apache-2.0
  */
 #pragma once
@@ -11,10 +11,12 @@
 #include <cudf/strings/strings_column_view.hpp>
 #include <cudf/structs/struct_view.hpp>
 #include <cudf/utilities/default_stream.hpp>
+#include <cudf/utilities/memory_resource.hpp>
 #include <cudf/utilities/span.hpp>
 #include <cudf/utilities/type_dispatcher.hpp>
 
 #include <rmm/cuda_stream_view.hpp>
+#include <rmm/resource_ref.hpp>
 
 #include <cuda/iterator>
 #include <cuda/std/utility>
@@ -498,11 +500,14 @@ class alignas(16) column_device_view : public column_device_view_core {
    *
    * @param source_view The `column_view` to make usable in device code
    * @param stream CUDA stream used for device memory operations for children columns.
+   * @param mr Device memory resource used to allocate the returned child-view storage
    * @return A `unique_ptr` to a `column_device_view` that makes the data from
    *`source_view` available in device memory.
    */
   static std::unique_ptr<column_device_view, std::function<void(column_device_view*)>> create(
-    column_view source_view, rmm::cuda_stream_view stream = cudf::get_default_stream());
+    column_view source_view,
+    rmm::cuda_stream_view stream      = cudf::get_default_stream(),
+    rmm::device_async_resource_ref mr = cudf::get_current_device_resource_ref());
 
   /**
    * @brief Destroy the `column_device_view` object.
@@ -645,13 +650,15 @@ class alignas(16) mutable_column_device_view : public mutable_column_device_view
    *
    * @param source_view The `column_view` to make usable in device code
    * @param stream CUDA stream used for device memory operations for children columns.
+   * @param mr Device memory resource used to allocate the returned child-view storage
    * @return A `unique_ptr` to a `mutable_column_device_view` that makes the
    * data from `source_view` available in device memory.
    */
   static std::unique_ptr<mutable_column_device_view,
                          std::function<void(mutable_column_device_view*)>>
   create(mutable_column_view source_view,
-         rmm::cuda_stream_view stream = cudf::get_default_stream());
+         rmm::cuda_stream_view stream      = cudf::get_default_stream(),
+         rmm::device_async_resource_ref mr = cudf::get_current_device_resource_ref());
 
   /**
    * @brief Returns reference to element at the specified index.

cpp/include/cudf/detail/row_operator/lexicographic.cuh

@@ -677,8 +677,10 @@ struct less_equivalent_comparator
  *
  */
 struct preprocessed_table {
-  using table_device_view_owner =
-    std::invoke_result_t<decltype(table_device_view::create), table_view, rmm::cuda_stream_view>;
+  using table_device_view_owner = std::invoke_result_t<decltype(table_device_view::create),
+                                                       table_view,
+                                                       rmm::cuda_stream_view,
+                                                       rmm::device_async_resource_ref>;
 
   /**
    * @brief Preprocess table for use with lexicographical comparison

cpp/include/cudf/detail/row_operator/preprocessed_table.cuh

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION.
+ * SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * SPDX-License-Identifier: Apache-2.0
  */
 
@@ -68,8 +68,10 @@ struct preprocessed_table {
   template <template <typename> class Hash>
   friend class ::cudf::detail::row::primitive::row_hasher;
 
-  using table_device_view_owner =
-    std::invoke_result_t<decltype(table_device_view::create), table_view, rmm::cuda_stream_view>;
+  using table_device_view_owner = std::invoke_result_t<decltype(table_device_view::create),
+                                                       table_view,
+                                                       rmm::cuda_stream_view,
+                                                       rmm::device_async_resource_ref>;
 
   preprocessed_table(table_device_view_owner&& table,
                      std::vector<rmm::device_buffer>&& null_buffers,

cpp/include/cudf/dictionary/dictionary_column_view.hpp

@@ -22,7 +22,8 @@ namespace CUDF_EXPORT cudf {
  * @brief A wrapper class for operations on a dictionary column.
  *
  * A dictionary column contains a set of keys and a column of indices.
- * The keys are a sorted set of unique values for the column.
+ * The keys are not required to be unique.
+ *
  * The indices represent the corresponding positions of each element's
  * value in the keys.
  */

cpp/include/cudf/dictionary/dictionary_factories.hpp

@@ -26,10 +26,7 @@ namespace CUDF_EXPORT cudf {
  * @brief Construct a dictionary column by copying the provided `keys`
  * and `indices`.
  *
- * It is expected that `keys_column.has_nulls() == false`.
- * It is assumed the elements in `keys_column` are unique and
- * are in a strict, total order. Meaning, `keys_column[i]` is ordered before
- * `keys_column[i+1]` for all `i in [0,n-1)` where `n` is the number of keys.
+ * The keys_column must contain no nulls.
  *
  * The indices values must be in the range [0,keys_column.size()).
  *
@@ -49,7 +46,7 @@ namespace CUDF_EXPORT cudf {
  * @throw std::invalid_argument if keys_column contains nulls
  * @throw std::invalid_argument if indices_column type is not a signed integer
  *
- * @param keys_column Column of unique, ordered values to use as the new dictionary column's keys.
+ * @param keys_column Column of non-null values to use as the new dictionary column's keys.
  * @param indices_column Indices to use for the new dictionary column.
  * @param stream CUDA stream used for device memory operations and kernel launches.
  * @param mr Device memory resource used to allocate the returned column's device memory.
@@ -66,16 +63,12 @@ std::unique_ptr<column> make_dictionary_column(
  * and indices columns.
  *
  * The keys_column and indices columns must contain no nulls.
- * It is assumed the elements in `keys_column` are unique and
- * are in a strict, total order. Meaning, `keys_column[i]` is ordered before
- * `keys_column[i+1]` for all `i in [0,n-1)` where `n` is the number of keys.
- *
  * The indices values must be in the range [0,keys_column.size()).
  *
  * @throw std::invalid_argument if keys_column or indices_column contains nulls
  * @throw std::invalid_argument if indices_column type is not a signed integer type
  *
- * @param keys_column Column of unique, ordered values to use as the new dictionary column's keys.
+ * @param keys_column Column of non-null values to use as the new dictionary column's keys.
  * @param indices_column Indices to use for the new dictionary column.
  * @param null_mask Null mask for the output column.
  * @param null_count Number of nulls for the output column.
@@ -101,7 +94,7 @@ std::unique_ptr<column> make_dictionary_column(std::unique_ptr<column> keys_colu
  *
  * @throw cudf::logic_error if keys_column contains nulls
  *
- * @param keys_column Column of unique, ordered values to use as the new dictionary column's keys.
+ * @param keys_column Column of non-null values to use as the new dictionary column's keys.
  * @param indices_column Indices values and null-mask to use for the new dictionary column.
  * @param stream CUDA stream used for device memory operations and kernel launches.
  * @param mr Device memory resource used to allocate the returned column's device memory.

cpp/include/cudf/dictionary/update_keys.hpp

@@ -121,6 +121,9 @@ std::unique_ptr<column> remove_unused_keys(
  * The indices are updated to reflect the position values of the new keys.
  * Any indices pointing to removed keys sets those rows to null.
  *
+ * Although duplicate keys are allowed, indices in the returned dictionary may
+ * only reference one of the duplicates.
+ *
  * @code{.pseudo}
  * d1 = {keys=["a", "b", "c"], indices=[2, 0, 1, 2, 1]}
  * d2 = set_keys(existing_dict, ["b","c","d"])

cpp/include/cudf/reduction/detail/reduction_functions.hpp

@@ -107,11 +107,14 @@ std::unique_ptr<scalar> max(column_view const& col,
  * If all elements in input column are null, output scalar is null.
  *
  * @param col input column to compute reduction
+ * @param dispatch_type The type to dispatch on. For dictionary columns this must be the keys type;
+ *                      for all other columns it must equal `col.type()`.
  * @param stream CUDA stream used for device memory operations and kernel launches
  * @param mr Device memory resource used to allocate the returned scalar's device memory
  * @return Index of the minimum element as scalar of type `output_dtype`
  */
 std::unique_ptr<scalar> argmin(column_view const& col,
+                               data_type dispatch_type,
                                rmm::cuda_stream_view stream,
                                rmm::device_async_resource_ref mr);
 
@@ -121,11 +124,14 @@ std::unique_ptr<scalar> argmin(column_view const& col,
  * If all elements in input column are null, output scalar is null.
  *
  * @param col input column to compute reduction
+ * @param dispatch_type The type to dispatch on. For dictionary columns this must be the keys type;
+ *                      for all other columns it must equal `col.type()`.
  * @param stream CUDA stream used for device memory operations and kernel launches
  * @param mr Device memory resource used to allocate the returned scalar's device memory
  * @return Index of the maximum element as scalar of type `output_dtype`
  */
 std::unique_ptr<scalar> argmax(column_view const& col,
+                               data_type dispatch_type,
                                rmm::cuda_stream_view stream,
                                rmm::device_async_resource_ref mr);