From c6490b6d5a0acc65a8abf4c5ff317faf3781f82e Mon Sep 17 00:00:00 2001
From: David Bayer <bayer.david99@gmail.com>
Date: Thu, 16 Oct 2025 09:24:44 +0200
Subject: [PATCH 1/2] Deprecate `<cuda/stream_ref>` header

---
 cub/cub/device/device_reduce.cuh                             | 2 +-
 cudax/examples/vector.cuh                                    | 2 +-
 .../experimental/__container/uninitialized_async_buffer.cuh  | 2 +-
 cudax/include/cuda/experimental/__launch/host_launch.cuh     | 2 +-
 cudax/include/cuda/experimental/__launch/launch.cuh          | 2 +-
 .../experimental/__memory_resource/device_memory_pool.cuh    | 2 +-
 .../experimental/__memory_resource/memory_resource_base.cuh  | 3 +--
 .../include/cuda/experimental/__stream/device_transform.cuh  | 2 +-
 cudax/include/cuda/experimental/__stream/stream_ref.cuh      | 2 +-
 cudax/test/common/utility.cuh                                | 2 +-
 cudax/test/containers/async_buffer/test_resources.h          | 2 +-
 cudax/test/containers/uninitialized_async_buffer.cu          | 2 +-
 cudax/test/memory_resource/device_memory_resource.cu         | 2 +-
 cudax/test/memory_resource/managed_memory_resource.cu        | 2 +-
 cudax/test/memory_resource/memory_pools.cu                   | 2 +-
 cudax/test/memory_resource/pinned_memory_resource.cu         | 2 +-
 docs/libcudacxx/extended_api/memory_resource.rst             | 2 +-
 examples/cudax/vector_add/vector.cuh                         | 2 +-
 libcudacxx/cmake/LibcudacxxBuildCompilerTargets.cmake        | 1 +
 .../include/cuda/__memory_resource/get_memory_resource.h     | 2 +-
 libcudacxx/include/cuda/__memory_resource/resource.h         | 2 +-
 libcudacxx/include/cuda/__memory_resource/resource_ref.h     | 2 +-
 libcudacxx/include/cuda/stream_ref                           | 5 +++++
 libcudacxx/test/libcudacxx/cuda/ccclrt/common/utility.cuh    | 2 +-
 .../memory_resource/async_resource_ref/equality.fail.cpp     | 2 +-
 .../memory_resource/async_resource_ref/equality.pass.cpp     | 2 +-
 .../memory_resource/async_resource_ref/properties.pass.cpp   | 2 +-
 .../test/libcudacxx/cuda/stream_ref/get_stream.pass.cpp      | 2 +-
 .../cuda/stream_ref/stream_ref.constructor.pass.cpp          | 2 +-
 .../libcudacxx/cuda/stream_ref/stream_ref.equality.pass.cpp  | 2 +-
 .../test/libcudacxx/cuda/stream_ref/stream_ref.get.pass.cpp  | 2 +-
 .../test/libcudacxx/cuda/stream_ref/stream_ref.query.cpp     | 2 +-
 .../libcudacxx/cuda/stream_ref/stream_ref.ready.pass.cpp     | 2 +-
 .../test/libcudacxx/cuda/stream_ref/stream_ref.sync.pass.cpp | 2 +-
 .../test/libcudacxx/cuda/stream_ref/stream_ref.wait.pass.cpp | 2 +-
 thrust/testing/cuda/stream_wrapper.cu                        | 2 +-
 36 files changed, 40 insertions(+), 35 deletions(-)

diff --git a/cub/cub/device/device_reduce.cuh b/cub/cub/device/device_reduce.cuh
index b1861d6ed2d..cd209b275ea 100644
--- a/cub/cub/device/device_reduce.cuh
+++ b/cub/cub/device/device_reduce.cuh
@@ -60,6 +60,7 @@
 #include <cuda/__iterator/tabulate_output_iterator.h>
 #include <cuda/__memory_resource/get_memory_resource.h>
 #include <cuda/__stream/get_stream.h>
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__execution/env.h>
 #include <cuda/std/__functional/identity.h>
 #include <cuda/std/__functional/invoke.h>
@@ -69,7 +70,6 @@
 #include <cuda/std/__type_traits/is_same.h>
 #include <cuda/std/cstdint>
 #include <cuda/std/limits>
-#include <cuda/stream_ref>
 
 CUB_NAMESPACE_BEGIN
 
diff --git a/cudax/examples/vector.cuh b/cudax/examples/vector.cuh
index 437e32aca88..d888043f916 100644
--- a/cudax/examples/vector.cuh
+++ b/cudax/examples/vector.cuh
@@ -24,9 +24,9 @@
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
 
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__type_traits/maybe_const.h>
 #include <cuda/std/span>
-#include <cuda/stream_ref>
 
 #include <cuda/experimental/__detail/utility.cuh>
 #include <cuda/experimental/__launch/param_kind.cuh>
diff --git a/cudax/include/cuda/experimental/__container/uninitialized_async_buffer.cuh b/cudax/include/cuda/experimental/__container/uninitialized_async_buffer.cuh
index effbe6fd355..207404c03a2 100644
--- a/cudax/include/cuda/experimental/__container/uninitialized_async_buffer.cuh
+++ b/cudax/include/cuda/experimental/__container/uninitialized_async_buffer.cuh
@@ -22,6 +22,7 @@
 #endif // no system header
 
 #include <cuda/__memory_resource/properties.h>
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__memory/addressof.h>
 #include <cuda/std/__memory/align.h>
 #include <cuda/std/__new/launder.h>
@@ -30,7 +31,6 @@
 #include <cuda/std/__utility/move.h>
 #include <cuda/std/__utility/swap.h>
 #include <cuda/std/span>
-#include <cuda/stream_ref>
 
 #include <cuda/experimental/__memory_resource/any_resource.cuh>
 
diff --git a/cudax/include/cuda/experimental/__launch/host_launch.cuh b/cudax/include/cuda/experimental/__launch/host_launch.cuh
index 744bd55a5ff..3932fe9ec0f 100644
--- a/cudax/include/cuda/experimental/__launch/host_launch.cuh
+++ b/cudax/include/cuda/experimental/__launch/host_launch.cuh
@@ -21,11 +21,11 @@
 #  pragma system_header
 #endif // no system header
 
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__functional/reference_wrapper.h>
 #include <cuda/std/__type_traits/decay.h>
 #include <cuda/std/__utility/forward.h>
 #include <cuda/std/tuple>
-#include <cuda/stream_ref>
 
 #include <cuda/std/__cccl/prologue.h>
 
diff --git a/cudax/include/cuda/experimental/__launch/launch.cuh b/cudax/include/cuda/experimental/__launch/launch.cuh
index 1386c930172..234598a272a 100644
--- a/cudax/include/cuda/experimental/__launch/launch.cuh
+++ b/cudax/include/cuda/experimental/__launch/launch.cuh
@@ -22,13 +22,13 @@
 #endif // no system header
 
 #include <cuda/__driver/driver_api.h>
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__exception/cuda_error.h>
 #include <cuda/std/__type_traits/is_function.h>
 #include <cuda/std/__type_traits/is_pointer.h>
 #include <cuda/std/__type_traits/type_identity.h>
 #include <cuda/std/__utility/forward.h>
 #include <cuda/std/__utility/pod_tuple.h>
-#include <cuda/stream_ref>
 
 #include <cuda/experimental/__execution/completion_signatures.cuh>
 #include <cuda/experimental/__execution/cpos.cuh>
diff --git a/cudax/include/cuda/experimental/__memory_resource/device_memory_pool.cuh b/cudax/include/cuda/experimental/__memory_resource/device_memory_pool.cuh
index 1eea0ea9596..26902a75226 100644
--- a/cudax/include/cuda/experimental/__memory_resource/device_memory_pool.cuh
+++ b/cudax/include/cuda/experimental/__memory_resource/device_memory_pool.cuh
@@ -28,10 +28,10 @@
 
 #include <cuda/__memory_resource/get_property.h>
 #include <cuda/__memory_resource/properties.h>
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__cuda/api_wrapper.h>
 #include <cuda/std/__new_>
 #include <cuda/std/span>
-#include <cuda/stream_ref>
 
 #include <cuda/experimental/__memory_resource/memory_pool_base.cuh>
 #include <cuda/experimental/__stream/stream.cuh>
diff --git a/cudax/include/cuda/experimental/__memory_resource/memory_resource_base.cuh b/cudax/include/cuda/experimental/__memory_resource/memory_resource_base.cuh
index ce3a92271e9..36b6d983d2d 100644
--- a/cudax/include/cuda/experimental/__memory_resource/memory_resource_base.cuh
+++ b/cudax/include/cuda/experimental/__memory_resource/memory_resource_base.cuh
@@ -27,11 +27,10 @@
 #endif // _CCCL_CUDA_COMPILER(CLANG)
 
 #include <cuda/__device/device_ref.h>
-#include <cuda/__memory_resource/properties.h>
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__concepts/concept_macros.h>
 #include <cuda/std/__cuda/api_wrapper.h>
 #include <cuda/std/cstddef>
-#include <cuda/stream_ref>
 
 #include <cuda/experimental/__memory_resource/any_resource.cuh>
 #include <cuda/experimental/__memory_resource/memory_pool_base.cuh>
diff --git a/cudax/include/cuda/experimental/__stream/device_transform.cuh b/cudax/include/cuda/experimental/__stream/device_transform.cuh
index 9d15d8571d3..95e2d7b2844 100644
--- a/cudax/include/cuda/experimental/__stream/device_transform.cuh
+++ b/cudax/include/cuda/experimental/__stream/device_transform.cuh
@@ -21,6 +21,7 @@
 #  pragma system_header
 #endif // no system header
 
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__memory/addressof.h>
 #include <cuda/std/__memory/construct_at.h>
 #include <cuda/std/__new/launder.h>
@@ -30,7 +31,6 @@
 #include <cuda/std/__utility/declval.h>
 #include <cuda/std/__utility/forward.h>
 #include <cuda/std/__utility/move.h>
-#include <cuda/stream_ref>
 
 #include <cuda/experimental/__detail/type_traits.cuh>
 #include <cuda/experimental/__detail/utility.cuh>
diff --git a/cudax/include/cuda/experimental/__stream/stream_ref.cuh b/cudax/include/cuda/experimental/__stream/stream_ref.cuh
index ae623c83b61..21d7cd27533 100644
--- a/cudax/include/cuda/experimental/__stream/stream_ref.cuh
+++ b/cudax/include/cuda/experimental/__stream/stream_ref.cuh
@@ -22,8 +22,8 @@
 
 #include <cuda/__device/all_devices.h>
 #include <cuda/__event/timed_event.h>
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__cuda/api_wrapper.h>
-#include <cuda/stream_ref>
 
 #include <cuda/experimental/__device/logical_device.cuh>
 #include <cuda/experimental/__execution/completion_behavior.cuh>
diff --git a/cudax/test/common/utility.cuh b/cudax/test/common/utility.cuh
index bad8cec90a0..78a3e3fda15 100644
--- a/cudax/test/common/utility.cuh
+++ b/cudax/test/common/utility.cuh
@@ -14,10 +14,10 @@
 #include <cuda_runtime_api.h>
 // cuda_runtime_api needs to come first
 
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/atomic>
 #include <cuda/std/__cuda/api_wrapper.h>
 #include <cuda/std/utility>
-#include <cuda/stream_ref>
 
 #include <cuda/experimental/launch.cuh>
 
diff --git a/cudax/test/containers/async_buffer/test_resources.h b/cudax/test/containers/async_buffer/test_resources.h
index cbb1e34cb0d..a8a0c22f3f2 100644
--- a/cudax/test/containers/async_buffer/test_resources.h
+++ b/cudax/test/containers/async_buffer/test_resources.h
@@ -11,9 +11,9 @@
 #ifndef CUDAX_TEST_CONTAINER_VECTOR_TEST_RESOURCES_H
 #define CUDAX_TEST_CONTAINER_VECTOR_TEST_RESOURCES_H
 
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/memory_resource>
 #include <cuda/std/type_traits>
-#include <cuda/stream_ref>
 
 #include <cuda/experimental/memory_resource.cuh>
 
diff --git a/cudax/test/containers/uninitialized_async_buffer.cu b/cudax/test/containers/uninitialized_async_buffer.cu
index 4e2786ed4de..cdf8732b61c 100644
--- a/cudax/test/containers/uninitialized_async_buffer.cu
+++ b/cudax/test/containers/uninitialized_async_buffer.cu
@@ -17,7 +17,7 @@
 #include <cuda/std/span>
 #include <cuda/std/type_traits>
 #include <cuda/std/utility>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include <cuda/experimental/container.cuh>
 #include <cuda/experimental/memory_resource.cuh>
diff --git a/cudax/test/memory_resource/device_memory_resource.cu b/cudax/test/memory_resource/device_memory_resource.cu
index 732dc414773..16c34b08875 100644
--- a/cudax/test/memory_resource/device_memory_resource.cu
+++ b/cudax/test/memory_resource/device_memory_resource.cu
@@ -10,7 +10,7 @@
 
 #include <cuda/std/cstdint>
 #include <cuda/std/type_traits>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include <cuda/experimental/memory_resource.cuh>
 
diff --git a/cudax/test/memory_resource/managed_memory_resource.cu b/cudax/test/memory_resource/managed_memory_resource.cu
index c1904ab59fd..da67c6e643e 100644
--- a/cudax/test/memory_resource/managed_memory_resource.cu
+++ b/cudax/test/memory_resource/managed_memory_resource.cu
@@ -10,7 +10,7 @@
 
 #include <cuda/std/cstdint>
 #include <cuda/std/type_traits>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include <cuda/experimental/memory_resource.cuh>
 #include <cuda/experimental/stream.cuh>
diff --git a/cudax/test/memory_resource/memory_pools.cu b/cudax/test/memory_resource/memory_pools.cu
index bffe71947c3..e848ad499c5 100644
--- a/cudax/test/memory_resource/memory_pools.cu
+++ b/cudax/test/memory_resource/memory_pools.cu
@@ -10,7 +10,7 @@
 
 #include <cuda/std/cstdint>
 #include <cuda/std/type_traits>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include <cuda/experimental/launch.cuh>
 #include <cuda/experimental/memory_resource.cuh>
diff --git a/cudax/test/memory_resource/pinned_memory_resource.cu b/cudax/test/memory_resource/pinned_memory_resource.cu
index aa3875baf1f..5d2ac5a495b 100644
--- a/cudax/test/memory_resource/pinned_memory_resource.cu
+++ b/cudax/test/memory_resource/pinned_memory_resource.cu
@@ -10,7 +10,7 @@
 
 #include <cuda/std/cstdint>
 #include <cuda/std/type_traits>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include <cuda/experimental/memory_resource.cuh>
 #include <cuda/experimental/stream.cuh>
diff --git a/docs/libcudacxx/extended_api/memory_resource.rst b/docs/libcudacxx/extended_api/memory_resource.rst
index 11a82ecb510..69edca07163 100644
--- a/docs/libcudacxx/extended_api/memory_resource.rst
+++ b/docs/libcudacxx/extended_api/memory_resource.rst
@@ -46,4 +46,4 @@ that are e.g. accessible on device, or whether it can utilize other allocation m
 
 To better support asynchronous CUDA `stream-ordered allocations <https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#stream-ordered-memory-allocator>`__
 libcu++ provides :ref:`cuda::stream_ref <libcudacxx-extended-api-streams-stream-ref>` as a wrapper around
-``cudaStream_t``. The definition of ``cuda::stream_ref`` can be found in the ``<cuda/stream_ref>`` header.
+``cudaStream_t``. The definition of ``cuda::stream_ref`` can be found in the ``<cuda/stream>`` header.
diff --git a/examples/cudax/vector_add/vector.cuh b/examples/cudax/vector_add/vector.cuh
index 533bc152631..2dc35c956b0 100644
--- a/examples/cudax/vector_add/vector.cuh
+++ b/examples/cudax/vector_add/vector.cuh
@@ -26,7 +26,7 @@
 
 #include <cuda/std/__type_traits/maybe_const.h>
 #include <cuda/std/span>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include <cuda/experimental/__detail/utility.cuh>
 #include <cuda/experimental/__launch/param_kind.cuh>
diff --git a/libcudacxx/cmake/LibcudacxxBuildCompilerTargets.cmake b/libcudacxx/cmake/LibcudacxxBuildCompilerTargets.cmake
index afbc958a925..b0825e08494 100644
--- a/libcudacxx/cmake/LibcudacxxBuildCompilerTargets.cmake
+++ b/libcudacxx/cmake/LibcudacxxBuildCompilerTargets.cmake
@@ -22,6 +22,7 @@ function(libcudacxx_build_compiler_targets)
   list(APPEND cxx_compile_definitions "CCCL_IGNORE_DEPRECATED_CPP_DIALECT")
   list(APPEND cxx_compile_definitions "CCCL_ENABLE_OPTIONAL_REF")
   list(APPEND cxx_compile_definitions "CCCL_IGNORE_DEPRECATED_DISCARD_MEMORY_HEADER")
+  list(APPEND cxx_compile_definitions "CCCL_IGNORE_DEPRECATED_STREAM_REF_HEADER")
 
   cccl_build_compiler_interface(libcudacxx.compiler_interface
     "${cuda_compile_options}"
diff --git a/libcudacxx/include/cuda/__memory_resource/get_memory_resource.h b/libcudacxx/include/cuda/__memory_resource/get_memory_resource.h
index a41d2f63fda..f9a2bd5c1ea 100644
--- a/libcudacxx/include/cuda/__memory_resource/get_memory_resource.h
+++ b/libcudacxx/include/cuda/__memory_resource/get_memory_resource.h
@@ -23,11 +23,11 @@
 
 #include <cuda/__memory_resource/properties.h>
 #include <cuda/__memory_resource/resource.h>
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__concepts/equality_comparable.h>
 #include <cuda/std/__execution/env.h>
 #include <cuda/std/__type_traits/is_same.h>
 #include <cuda/std/__type_traits/remove_cvref.h>
-#include <cuda/stream_ref>
 
 #include <cuda/std/__cccl/prologue.h>
 
diff --git a/libcudacxx/include/cuda/__memory_resource/resource.h b/libcudacxx/include/cuda/__memory_resource/resource.h
index 204ba26c4f8..9faba888a11 100644
--- a/libcudacxx/include/cuda/__memory_resource/resource.h
+++ b/libcudacxx/include/cuda/__memory_resource/resource.h
@@ -22,6 +22,7 @@
 #endif // no system header
 
 #include <cuda/__memory_resource/get_property.h>
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/std/__concepts/concept_macros.h>
 #include <cuda/std/__concepts/convertible_to.h>
 #include <cuda/std/__concepts/equality_comparable.h>
@@ -29,7 +30,6 @@
 #include <cuda/std/__tuple_dir/sfinae_helpers.h>
 #include <cuda/std/__type_traits/decay.h>
 #include <cuda/std/__type_traits/fold.h>
-#include <cuda/stream_ref>
 
 #include <cuda/std/__cccl/prologue.h>
 
diff --git a/libcudacxx/include/cuda/__memory_resource/resource_ref.h b/libcudacxx/include/cuda/__memory_resource/resource_ref.h
index 24ba8b6a27e..98b21244135 100644
--- a/libcudacxx/include/cuda/__memory_resource/resource_ref.h
+++ b/libcudacxx/include/cuda/__memory_resource/resource_ref.h
@@ -26,6 +26,7 @@
 #  include <cuda/__memory_resource/get_property.h>
 #  include <cuda/__memory_resource/properties.h>
 #  include <cuda/__memory_resource/resource.h>
+#  include <cuda/__stream/stream_ref.h>
 #  include <cuda/std/__concepts/concept_macros.h>
 #  include <cuda/std/__memory/addressof.h>
 #  include <cuda/std/__type_traits/is_base_of.h>
@@ -34,7 +35,6 @@
 #  include <cuda/std/__utility/exchange.h>
 #  include <cuda/std/__utility/move.h>
 #  include <cuda/std/cstddef>
-#  include <cuda/stream_ref>
 
 #  include <cuda/std/__cccl/prologue.h>
 
diff --git a/libcudacxx/include/cuda/stream_ref b/libcudacxx/include/cuda/stream_ref
index b2bf5f02d76..0229a84747b 100644
--- a/libcudacxx/include/cuda/stream_ref
+++ b/libcudacxx/include/cuda/stream_ref
@@ -48,6 +48,11 @@ private:
 #  pragma system_header
 #endif // no system header
 
+#if !defined(CCCL_IGNORE_DEPRECATED_STREAM_REF_HEADER)
+_CCCL_WARNING("The <cuda/stream_ref> header is deprecated and will be removed in the next major release. The \
+contents were moved to <cuda/stream>. To ignore this message, define CCCL_IGNORE_DEPRECATED_STREAM_REF_HEADER.")
+#endif // !CCCL_IGNORE_DEPRECATED_STREAM_REF_HEADER
+
 #include <cuda/__stream/get_stream.h>
 #include <cuda/__stream/stream_ref.h>
 
diff --git a/libcudacxx/test/libcudacxx/cuda/ccclrt/common/utility.cuh b/libcudacxx/test/libcudacxx/cuda/ccclrt/common/utility.cuh
index 84b56d18110..c21046db1db 100644
--- a/libcudacxx/test/libcudacxx/cuda/ccclrt/common/utility.cuh
+++ b/libcudacxx/test/libcudacxx/cuda/ccclrt/common/utility.cuh
@@ -15,10 +15,10 @@
 // cuda_runtime_api needs to come first
 
 #include <cuda/__runtime/ensure_current_context.h>
+#include <cuda/__stream/stream_ref.h>
 #include <cuda/atomic>
 #include <cuda/std/__cuda/api_wrapper.h>
 #include <cuda/std/utility>
-#include <cuda/stream_ref>
 
 #include <new> // IWYU pragma: keep (needed for placement new)
 
diff --git a/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/equality.fail.cpp b/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/equality.fail.cpp
index 3d82e201baa..4e32adf1ea2 100644
--- a/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/equality.fail.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/equality.fail.cpp
@@ -15,7 +15,7 @@
 #include <cuda/memory_resource>
 #include <cuda/std/cassert>
 #include <cuda/std/cstdint>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include "types.h"
 
diff --git a/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/equality.pass.cpp b/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/equality.pass.cpp
index ecb848a1f53..bfe4c0a27a8 100644
--- a/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/equality.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/equality.pass.cpp
@@ -15,7 +15,7 @@
 #include <cuda/memory_resource>
 #include <cuda/std/cassert>
 #include <cuda/std/cstdint>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include "types.h"
 
diff --git a/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/properties.pass.cpp b/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/properties.pass.cpp
index 3218189f1e0..756eb6bb695 100644
--- a/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/properties.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/memory_resource/async_resource_ref/properties.pass.cpp
@@ -15,7 +15,7 @@
 #include <cuda/memory_resource>
 #include <cuda/std/cassert>
 #include <cuda/std/cstdint>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include "test_macros.h"
 #include "types.h"
diff --git a/libcudacxx/test/libcudacxx/cuda/stream_ref/get_stream.pass.cpp b/libcudacxx/test/libcudacxx/cuda/stream_ref/get_stream.pass.cpp
index 64ce3cf40ce..bf5b8bf0f13 100644
--- a/libcudacxx/test/libcudacxx/cuda/stream_ref/get_stream.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/stream_ref/get_stream.pass.cpp
@@ -10,7 +10,7 @@
 // UNSUPPORTED: nvrtc
 
 #include <cuda/std/type_traits>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 __host__ __device__ void test()
 {
diff --git a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.constructor.pass.cpp b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.constructor.pass.cpp
index d12a604ed09..7c131466a39 100644
--- a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.constructor.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.constructor.pass.cpp
@@ -11,7 +11,7 @@
 
 #include <cuda/std/cassert>
 #include <cuda/std/type_traits>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 static_assert(cuda::std::is_default_constructible<cuda::stream_ref>::value, "");
 static_assert(!cuda::std::is_constructible<cuda::stream_ref, int>::value, "");
diff --git a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.equality.pass.cpp b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.equality.pass.cpp
index 1f38523ed84..35f1e3a1d32 100644
--- a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.equality.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.equality.pass.cpp
@@ -11,7 +11,7 @@
 // UNSUPPORTED: nvrtc
 
 #include <cuda/std/cassert>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 __host__ __device__ void test()
 {
diff --git a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.get.pass.cpp b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.get.pass.cpp
index 4a2254e5d28..81a8c81b3ec 100644
--- a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.get.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.get.pass.cpp
@@ -11,7 +11,7 @@
 // UNSUPPORTED: nvrtc
 
 #include <cuda/std/cassert>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 __host__ __device__ void test()
 {
diff --git a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.query.cpp b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.query.cpp
index e598278ee5e..13f3fec86d4 100644
--- a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.query.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.query.cpp
@@ -11,7 +11,7 @@
 // UNSUPPORTED: nvrtc
 
 #include <cuda/std/cassert>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 __host__ __device__ void test()
 {
diff --git a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.ready.pass.cpp b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.ready.pass.cpp
index 139a68a8819..5bcad3f1191 100644
--- a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.ready.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.ready.pass.cpp
@@ -11,7 +11,7 @@
 // UNSUPPORTED: nvrtc
 
 #include <cuda/std/cassert>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include "test_macros.h"
 
diff --git a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.sync.pass.cpp b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.sync.pass.cpp
index 7818e04d026..89889cf5e87 100644
--- a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.sync.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.sync.pass.cpp
@@ -11,7 +11,7 @@
 // UNSUPPORTED: nvrtc
 
 #include <cuda/std/cassert>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include <atomic>
 #include <chrono>
diff --git a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.wait.pass.cpp b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.wait.pass.cpp
index dfd6ef1b4b4..c4eb77710f6 100644
--- a/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.wait.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/stream_ref/stream_ref.wait.pass.cpp
@@ -11,7 +11,7 @@
 // UNSUPPORTED: nvrtc
 
 #include <cuda/std/cassert>
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include <atomic>
 #include <chrono>
diff --git a/thrust/testing/cuda/stream_wrapper.cu b/thrust/testing/cuda/stream_wrapper.cu
index 24d0138cb7d..8fc0ea32df4 100644
--- a/thrust/testing/cuda/stream_wrapper.cu
+++ b/thrust/testing/cuda/stream_wrapper.cu
@@ -2,7 +2,7 @@
 #include <thrust/iterator/constant_iterator.h>
 #include <thrust/reduce.h>
 
-#include <cuda/stream_ref>
+#include <cuda/stream>
 
 #include <unittest/unittest.h>
 

From d8d89ce2deba87daf1925c8286ef92f713b50157 Mon Sep 17 00:00:00 2001
From: David Bayer <bayer.david99@gmail.com>
Date: Thu, 16 Oct 2025 09:41:17 +0200
Subject: [PATCH 2/2] fix namespace

---
 cudax/test/execution/test_stream_context.cu | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/cudax/test/execution/test_stream_context.cu b/cudax/test/execution/test_stream_context.cu
index af1f25fd179..467e3d6f55e 100644
--- a/cudax/test/execution/test_stream_context.cu
+++ b/cudax/test/execution/test_stream_context.cu
@@ -135,7 +135,7 @@ void bulk_on_stream_scheduler()
 
   using _env_t = cudax::env_t<cuda::mr::device_accessible>;
   auto mr      = cudax::device_memory_resource{_dev};
-  auto mr2     = cudax::any_resource<cudax::device_accessible>(mr);
+  auto mr2     = cudax::any_resource<cuda::device_accessible>(mr);
   _env_t env{mr, cuda::get_stream(sch), ex::par_unseq};
   auto buf = cudax::make_async_buffer<int>(sctx, mr2, 10, 40, env); // a device buffer of 10 integers, initialized to 40
   cuda::std::span data{buf};