Prepare librtcx for standalone usage

cpp/librtcx/CMakeLists.txt

@@ -11,6 +11,8 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
   include(cmake/rapids_config.cmake)
   include(rapids-cmake)
   include(rapids-cpm)
+  include(rapids-export)
+  include(rapids-find)
   rapids_cpm_init()
 endif()
 
@@ -23,7 +25,11 @@ project(
 option(RTCX_STATIC_LINK_NVRTC "Use static linking for NVRTC" OFF)
 option(RTCX_STATIC_LINK_NVJITLINK "Use static linking for nvJitLink" OFF)
 
-find_package(CUDAToolkit REQUIRED)
+rapids_find_package(
+  CUDAToolkit REQUIRED
+  BUILD_EXPORT_SET rtcx-exports
+  INSTALL_EXPORT_SET rtcx-exports
+)
 
 if(NOT TARGET zstd)
   set(CPM_DOWNLOAD_zstd ON)
@@ -87,3 +93,74 @@ if(RTCX_STATIC_LINK_NVJITLINK)
 else()
   target_compile_definitions(rtcx PRIVATE RTCX_STATIC_LINK_LIBNVJITLINK=0)
 endif()
+
+# =============================================================================
+# Install / Export
+# =============================================================================
+include(GNUInstallDirs)
+include(${rapids-cmake-dir}/cmake/install_lib_dir.cmake)
+rapids_cmake_install_lib_dir(lib_dir)
+
+# Option to control install (default ON when built standalone, OFF when used as subdirectory)
+option(RTCX_INSTALL "Enable installation of rtcx targets" ${PROJECT_IS_TOP_LEVEL})
+
+set(rtcx_install_code_string
+    [=[
+# Embed functions (add_embed, embed_includes, embed_blob, embed)
+# are included automatically so consumers don't need explicit include().
+include("${CMAKE_CURRENT_LIST_DIR}/embed.cmake")
+
+# Set rtcx_LIBCXX_DIR for consumers using embed_includes with libcxx headers.
+set(rtcx_LIBCXX_DIR "${PACKAGE_PREFIX_DIR}/share/rtcx/libcxx")
+]=]
+)
+
+string(
+  CONFIGURE
+    [=[
+# Embed functions (add_embed, embed_includes, embed_blob, embed)
+# are included automatically so consumers don't need explicit include().
+include("@CMAKE_CURRENT_SOURCE_DIR@/embed.cmake")
+
+# Set rtcx_LIBCXX_DIR for consumers using embed_includes with libcxx headers.
+set(rtcx_LIBCXX_DIR "@CMAKE_CURRENT_SOURCE_DIR@/libcxx")
+]=]
+    rtcx_build_code_string
+  @ONLY
+)
+
+if(NOT RTCX_INSTALL)
+  set(rtcx_exclude_from_install EXCLUDE_FROM_ALL)
+endif()
+
+install(
+  TARGETS rtcx
+  EXPORT rtcx-exports
+  ARCHIVE DESTINATION ${lib_dir}
+  INCLUDES
+  DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
+  ${rtcx_exclude_from_install}
+)
+
+if(RTCX_INSTALL)
+  install(FILES rtcx.hpp sha256.hpp embed.hpp DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/rtcx)
+  install(FILES embed.cmake embed.in.cpp DESTINATION ${lib_dir}/cmake/rtcx)
+  install(DIRECTORY libcxx/ DESTINATION ${CMAKE_INSTALL_DATADIR}/rtcx/libcxx)
+
+  rapids_export(
+    INSTALL rtcx
+    EXPORT_SET rtcx-exports
+    GLOBAL_TARGETS rtcx
+    NAMESPACE rtcx::
+    FINAL_CODE_BLOCK rtcx_install_code_string
+  )
+endif()
+
+# Build-tree export (always, so CPM consumers can find_package from the build tree)
+rapids_export(
+  BUILD rtcx
+  EXPORT_SET rtcx-exports
+  GLOBAL_TARGETS rtcx
+  NAMESPACE rtcx::
+  FINAL_CODE_BLOCK rtcx_build_code_string
+)

cpp/librtcx/RAPIDS_BRANCH

@@ -0,0 +1 @@
+../../RAPIDS_BRANCH

cpp/librtcx/VERSION

@@ -0,0 +1 @@
+../../VERSION

cpp/librtcx/cmake/rapids_config.cmake

@@ -4,7 +4,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # cmake-format: on
 # =============================================================================
-file(READ "${CMAKE_CURRENT_LIST_DIR}/../../../VERSION" _rapids_version)
+file(READ "${CMAKE_CURRENT_LIST_DIR}/../VERSION" _rapids_version)
 if(_rapids_version MATCHES [[^([0-9][0-9])\.([0-9][0-9])\.([0-9][0-9])]])
   set(RAPIDS_VERSION_MAJOR "${CMAKE_MATCH_1}")
   set(RAPIDS_VERSION_MINOR "${CMAKE_MATCH_2}")
@@ -19,11 +19,11 @@ else()
   )
 endif()
 
-file(STRINGS "${CMAKE_CURRENT_LIST_DIR}/../../../RAPIDS_BRANCH" RAPIDS_BRANCH)
+file(STRINGS "${CMAKE_CURRENT_LIST_DIR}/../RAPIDS_BRANCH" RAPIDS_BRANCH)
 if(NOT RAPIDS_BRANCH)
   message(
     FATAL_ERROR
-      "Could not determine branch name to use for checking out rapids-cmake. The file \"${CMAKE_CURRENT_LIST_DIR}/../../../RAPIDS_BRANCH\" is missing."
+      "Could not determine branch name to use for checking out rapids-cmake. The file \"${CMAKE_CURRENT_LIST_DIR}/../RAPIDS_BRANCH\" is missing."
   )
 endif()
 

cpp/librtcx/embed.cmake

@@ -5,12 +5,6 @@
 # cmake-format: on
 # =============================================================================
 
-if(NOT TARGET zstd)
-  message(
-    FATAL_ERROR "zstd library is required for JIT embedding. Please ensure it is found by CMake."
-  )
-endif()
-
 # This function initializes a target for JIT embedding. It must be called before any calls to
 # embed_includes() or embed_blob() for the target. It creates a dedicated INTERFACE library target
 # that is used to track registered files and dependencies via target properties. The TARGET argument
@@ -274,8 +268,14 @@ function(embed TARGET)
 
   set(RUNNER "${TARGET}__jit_embed_run")
   add_executable(${RUNNER} EXCLUDE_FROM_ALL "${EMBED_SCRIPT}")
-  target_include_directories(${RUNNER} PRIVATE ${ZSTD_INCLUDE_DIR})
-  target_link_libraries(${RUNNER} PRIVATE ${CMAKE_DL_LIBS} zstd)
+  target_link_libraries(${RUNNER} PRIVATE ${CMAKE_DL_LIBS})
+  if(NOT ARG_COMPRESSION STREQUAL "none")
+    if(NOT TARGET zstd)
+      message(FATAL_ERROR "embed(): zstd target is required when COMPRESSION is not none.")
+    endif()
+    target_include_directories(${RUNNER} PRIVATE ${ZSTD_INCLUDE_DIR})
+    target_link_libraries(${RUNNER} PRIVATE zstd)
+  endif()
   set_target_properties(${RUNNER} PROPERTIES CXX_STANDARD 20 CXX_STANDARD_REQUIRED YES)
   target_include_directories(${RUNNER} PRIVATE ${CMAKE_CURRENT_FUNCTION_LIST_DIR})
 

python/cudf_streaming/cudf_streaming/examples/ray_shuffle_example.py

@@ -16,11 +16,11 @@
     partition_and_pack,
     unpack_and_concat,
 )
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.integrations.ray import RapidsMPFActor, setup_ray_ucxx_cluster
 from rapidsmpf.memory.buffer_resource import BufferResource
 from rapidsmpf.memory.spill import unspill_partitions
 from rapidsmpf.shuffler import Shuffler
-from rapidsmpf.testing import assert_eq
 
 
 class ShufflingActor(RapidsMPFActor):

python/cudf_streaming/cudf_streaming/testing.py

@@ -0,0 +1,66 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Testing utilities for cudf_streaming."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import pylibcudf
+
+from rmm.pylibrmm.stream import DEFAULT_STREAM
+
+if TYPE_CHECKING:
+    from rmm.pylibrmm.stream import Stream
+
+
+def assert_eq(
+    left: pylibcudf.Table,
+    right: pylibcudf.Table,
+    *,
+    sort_rows: int | None = None,
+    stream: Stream | None = None,
+) -> None:
+    """
+    Assert that two tables are equivalent using pylibcudf.
+
+    Parameters
+    ----------
+    left
+        plc.Table to compare.
+    right
+        plc.Table to compare.
+    sort_rows
+        If not None, sort both tables by this column before comparing.
+        An ``int`` is treated as a column index.
+    stream
+        CUDA stream to use for the comparison.
+
+    Raises
+    ------
+    AssertionError
+        If the two tables do not compare equal.
+    """
+    if stream is None:
+        stream = DEFAULT_STREAM
+
+    if sort_rows is not None:
+        column_order = [pylibcudf.types.Order.ASCENDING]
+        null_precedence = [pylibcudf.types.NullOrder.BEFORE]
+        left = pylibcudf.sorting.stable_sort_by_key(
+            left,
+            pylibcudf.Table([left.columns()[sort_rows]]),
+            column_order,
+            null_precedence,
+            stream=stream,
+        )
+        right = pylibcudf.sorting.stable_sort_by_key(
+            right,
+            pylibcudf.Table([right.columns()[sort_rows]]),
+            column_order,
+            null_precedence,
+            stream=stream,
+        )
+    if not pylibcudf.table_equality.tables_equal(left, right, stream=stream):
+        raise AssertionError(f"Table are not equal with {sort_rows=}")

python/cudf_streaming/cudf_streaming/tests/test_allgather.py

@@ -16,9 +16,9 @@
     packed_data_from_cudf_packed_columns,
     unpack_and_concat,
 )
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.coll import AllGather
 from rapidsmpf.memory.buffer_resource import BufferResource
-from rapidsmpf.testing import assert_eq
 
 if TYPE_CHECKING:
     import rmm.mr

python/cudf_streaming/cudf_streaming/tests/test_bloom_filter.py

@@ -13,13 +13,13 @@
 from cudf_streaming.streaming import ChannelMetadata
 from cudf_streaming.streaming.bloom_filter import BloomFilter
 from cudf_streaming.streaming.table_chunk import TableChunk
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.streaming.core.actor import define_actor, run_actor_network
 from rapidsmpf.streaming.core.leaf_actor import (
     pull_from_channel,
     push_to_channel,
 )
 from rapidsmpf.streaming.core.message import Message
-from rapidsmpf.testing import assert_eq
 
 if TYPE_CHECKING:
     from collections.abc import Awaitable

python/cudf_streaming/cudf_streaming/tests/test_integration_partition.py

@@ -13,9 +13,9 @@
     split_and_pack,
     unpack_and_concat,
 )
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.memory.buffer_resource import BufferResource
 from rapidsmpf.memory.spill import spill_partitions, unspill_partitions
-from rapidsmpf.testing import assert_eq
 from rmm.pylibrmm.stream import DEFAULT_STREAM
 
 if TYPE_CHECKING:

python/cudf_streaming/cudf_streaming/tests/test_partition.py

@@ -13,13 +13,13 @@
     unpack_and_concat,
 )
 from cudf_streaming.streaming.table_chunk import TableChunk
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.streaming.core.actor import run_actor_network
 from rapidsmpf.streaming.core.leaf_actor import (
     pull_from_channel,
     push_to_channel,
 )
 from rapidsmpf.streaming.core.message import Message
-from rapidsmpf.testing import assert_eq
 
 if TYPE_CHECKING:
     from rapidsmpf.streaming.chunks.partition import PartitionMapChunk

python/cudf_streaming/cudf_streaming/tests/test_shuffler.py

@@ -13,12 +13,12 @@
     partition_and_pack,
     unpack_and_concat,
 )
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.memory.buffer_resource import BufferResource
 from rapidsmpf.memory.spill import unspill_partitions
 from rapidsmpf.shuffler import (
     Shuffler,
 )
-from rapidsmpf.testing import assert_eq
 
 if TYPE_CHECKING:
     import rmm.mr

python/cudf_streaming/cudf_streaming/tests/test_sparse_alltoall.py

@@ -14,9 +14,9 @@
     packed_data_from_cudf_packed_columns,
     unpack_and_concat,
 )
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.coll.sparse_alltoall import SparseAlltoall
 from rapidsmpf.memory.buffer_resource import BufferResource
-from rapidsmpf.testing import assert_eq
 
 if TYPE_CHECKING:
     import rmm.mr

python/cudf_streaming/cudf_streaming/tests/test_streaming_allgather.py

@@ -15,6 +15,7 @@
     unpack_and_concat,
 )
 from cudf_streaming.streaming.table_chunk import TableChunk
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.streaming.chunks.packed_data import PackedDataChunk
 from rapidsmpf.streaming.coll.allgather import AllGather, allgather
 from rapidsmpf.streaming.core.actor import define_actor, run_actor_network
@@ -23,7 +24,6 @@
     push_to_channel,
 )
 from rapidsmpf.streaming.core.message import Message
-from rapidsmpf.testing import assert_eq
 
 if TYPE_CHECKING:
     from collections.abc import Awaitable

python/cudf_streaming/cudf_streaming/tests/test_streaming_define_actor.py

@@ -9,14 +9,14 @@
 import pytest
 
 from cudf_streaming.streaming.table_chunk import TableChunk
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.streaming.chunks.arbitrary import ArbitraryChunk
 from rapidsmpf.streaming.core.actor import define_actor, run_actor_network
 from rapidsmpf.streaming.core.leaf_actor import (
     pull_from_channel,
     push_to_channel,
 )
 from rapidsmpf.streaming.core.message import Message
-from rapidsmpf.testing import assert_eq
 
 
 @pytest.fixture

python/cudf_streaming/cudf_streaming/tests/test_streaming_fanout.py

@@ -11,14 +11,14 @@
 import pytest
 
 from cudf_streaming.streaming.table_chunk import TableChunk
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.streaming.core.actor import run_actor_network
 from rapidsmpf.streaming.core.fanout import FanoutPolicy, fanout
 from rapidsmpf.streaming.core.leaf_actor import (
     pull_from_channel,
     push_to_channel,
 )
 from rapidsmpf.streaming.core.message import Message
-from rapidsmpf.testing import assert_eq
 
 _INT64 = plc.DataType(plc.TypeId.INT64)
 

python/cudf_streaming/cudf_streaming/tests/test_streaming_leaf_actor.py

@@ -8,13 +8,13 @@
 import pylibcudf as plc
 
 from cudf_streaming.streaming.table_chunk import TableChunk
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.streaming.core.actor import run_actor_network
 from rapidsmpf.streaming.core.leaf_actor import (
     pull_from_channel,
     push_to_channel,
 )
 from rapidsmpf.streaming.core.message import Message
-from rapidsmpf.testing import assert_eq
 
 if TYPE_CHECKING:
     from rapidsmpf.streaming.core.channel import Channel

python/cudf_streaming/cudf_streaming/tests/test_streaming_shuffler.py

@@ -19,6 +19,7 @@
     unpack_and_concat as streaming_unpack_and_concat,
 )
 from cudf_streaming.streaming.table_chunk import TableChunk
+from cudf_streaming.testing import assert_eq
 from rapidsmpf.shuffler import PartitionAssignment
 from rapidsmpf.streaming.coll.shuffler import (
     ShufflerAsync,
@@ -30,7 +31,6 @@
     push_to_channel,
 )
 from rapidsmpf.streaming.core.message import Message
-from rapidsmpf.testing import assert_eq
 
 if TYPE_CHECKING:
     from collections.abc import Awaitable