From aaed49e8d48212dd1f4db508715543b9b1ed6e63 Mon Sep 17 00:00:00 2001
From: Mark Harris <783069+harrism@users.noreply.github.com>
Date: Wed, 7 Feb 2024 09:12:05 +0000
Subject: [PATCH 1/4] Automate header grouping and ordering with clang-format

---
 .clang-format | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/.clang-format b/.clang-format
index 26b9a5bf4..b36983d2e 100644
--- a/.clang-format
+++ b/.clang-format
@@ -71,8 +71,18 @@ ForEachMacros:
   - foreach
   - Q_FOREACH
   - BOOST_FOREACH
-IncludeBlocks: Preserve
-IncludeIsMainRegex: '([-_](test|unittest))?$'
+IncludeBlocks: Regroup
+IncludeCategories:
+  - Regex:           '^"' # quoted includes
+    Priority:        1
+  - Regex:           '^<rmm/' # RMM includes
+    Priority:        2
+  - Regex:           '^<cuda/' # CUDA includes
+    Priority:        3
+  - Regex:           '^<.*\..*' # other system includes (e.g. with a '.')
+    Priority:        4
+  - Regex:           '^<[^.]+' # STL includes (no '.')
+    Priority:        5
 IndentCaseLabels: true
 IndentPPDirectives: None
 IndentWidth:     2

From 10a2918cb9c9c0caa58c6f8f74ba02f097f50f26 Mon Sep 17 00:00:00 2001
From: Mark Harris <783069+harrism@users.noreply.github.com>
Date: Wed, 7 Feb 2024 09:15:27 +0000
Subject: [PATCH 2/4] Run new clang-format rules to change header grouping.

---
 .../cuda_stream_pool/cuda_stream_pool_bench.cpp     |  3 +--
 benchmarks/device_uvector/device_uvector_bench.cu   |  2 --
 .../multi_stream_allocations_bench.cu               |  6 ++----
 .../random_allocations/random_allocations.cpp       |  3 +--
 benchmarks/replay/replay.cpp                        | 13 +++++--------
 benchmarks/utilities/log_parser.hpp                 |  6 +++---
 include/rmm/cuda_stream.hpp                         |  2 +-
 include/rmm/cuda_stream_view.hpp                    |  4 ++--
 include/rmm/detail/dynamic_load_runtime.hpp         |  1 -
 include/rmm/detail/logging_assert.hpp               |  3 ++-
 include/rmm/device_buffer.hpp                       |  4 ++--
 include/rmm/device_uvector.hpp                      |  4 ++--
 include/rmm/device_vector.hpp                       |  2 +-
 include/rmm/exec_policy.hpp                         |  2 +-
 include/rmm/mr/device/arena_memory_resource.hpp     |  1 -
 .../rmm/mr/device/cuda_async_memory_resource.hpp    |  5 ++---
 .../mr/device/cuda_async_view_memory_resource.hpp   |  5 ++---
 include/rmm/mr/device/cuda_memory_resource.hpp      |  3 +--
 include/rmm/mr/device/detail/arena.hpp              |  1 -
 .../detail/stream_ordered_memory_resource.hpp       |  1 -
 .../rmm/mr/device/fixed_size_memory_resource.hpp    |  5 ++---
 include/rmm/mr/device/managed_memory_resource.hpp   |  3 +--
 include/rmm/mr/device/pool_memory_resource.hpp      |  8 +++-----
 include/rmm/mr/device/thrust_allocator_adaptor.hpp  |  6 +++---
 include/rmm/mr/host/new_delete_resource.hpp         |  3 +--
 include/rmm/thrust_rmm_allocator.h                  |  2 +-
 python/rmm/_lib/_torch_allocator.cpp                |  4 ++--
 tests/cuda_stream_pool_tests.cpp                    |  3 +--
 tests/cuda_stream_tests.cpp                         |  4 ++--
 tests/device_buffer_tests.cu                        |  3 +--
 tests/device_scalar_tests.cpp                       |  7 +++----
 tests/logger_tests.cpp                              |  2 +-
 tests/mr/device/arena_mr_tests.cpp                  |  3 +--
 tests/mr/device/callback_mr_tests.cpp               |  6 +++---
 tests/mr/device/mr_multithreaded_tests.cpp          |  4 ++--
 tests/mr/device/mr_ref_multithreaded_tests.cpp      |  4 ++--
 tests/mr/device/mr_ref_test.hpp                     |  4 ++--
 tests/mr/device/polymorphic_allocator_tests.cpp     |  7 ++++---
 tests/mr/device/thrust_allocator_tests.cu           |  1 -
 tests/mr/device/tracking_mr_tests.cpp               |  1 -
 tests/mr/host/mr_ref_tests.cpp                      |  3 +--
 tests/mr/host/mr_tests.cpp                          |  3 +--
 42 files changed, 65 insertions(+), 92 deletions(-)

diff --git a/benchmarks/cuda_stream_pool/cuda_stream_pool_bench.cpp b/benchmarks/cuda_stream_pool/cuda_stream_pool_bench.cpp
index 6710ffe50..5adf23354 100644
--- a/benchmarks/cuda_stream_pool/cuda_stream_pool_bench.cpp
+++ b/benchmarks/cuda_stream_pool/cuda_stream_pool_bench.cpp
@@ -17,9 +17,8 @@
 #include <rmm/cuda_stream_pool.hpp>
 #include <rmm/detail/error.hpp>
 
-#include <cuda_runtime_api.h>
-
 #include <benchmark/benchmark.h>
+#include <cuda_runtime_api.h>
 
 #include <stdexcept>
 
diff --git a/benchmarks/device_uvector/device_uvector_bench.cu b/benchmarks/device_uvector/device_uvector_bench.cu
index 8b7f9a5ba..7181de325 100644
--- a/benchmarks/device_uvector/device_uvector_bench.cu
+++ b/benchmarks/device_uvector/device_uvector_bench.cu
@@ -27,9 +27,7 @@
 #include <rmm/mr/device/pool_memory_resource.hpp>
 
 #include <benchmark/benchmark.h>
-
 #include <cuda_runtime_api.h>
-
 #include <thrust/device_vector.h>
 #include <thrust/memory.h>
 
diff --git a/benchmarks/multi_stream_allocations/multi_stream_allocations_bench.cu b/benchmarks/multi_stream_allocations/multi_stream_allocations_bench.cu
index 4943e507f..e407476b0 100644
--- a/benchmarks/multi_stream_allocations/multi_stream_allocations_bench.cu
+++ b/benchmarks/multi_stream_allocations/multi_stream_allocations_bench.cu
@@ -14,8 +14,6 @@
  * limitations under the License.
  */
 
-#include <benchmarks/utilities/cxxopts.hpp>
-
 #include <rmm/cuda_device.hpp>
 #include <rmm/cuda_stream.hpp>
 #include <rmm/cuda_stream_pool.hpp>
@@ -28,9 +26,9 @@
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <rmm/mr/device/pool_memory_resource.hpp>
 
-#include <cuda_runtime_api.h>
-
 #include <benchmark/benchmark.h>
+#include <benchmarks/utilities/cxxopts.hpp>
+#include <cuda_runtime_api.h>
 
 #include <cstddef>
 
diff --git a/benchmarks/random_allocations/random_allocations.cpp b/benchmarks/random_allocations/random_allocations.cpp
index 2856cd323..57116743b 100644
--- a/benchmarks/random_allocations/random_allocations.cpp
+++ b/benchmarks/random_allocations/random_allocations.cpp
@@ -14,8 +14,6 @@
  * limitations under the License.
  */
 
-#include <benchmarks/utilities/cxxopts.hpp>
-
 #include <rmm/cuda_device.hpp>
 #include <rmm/mr/device/arena_memory_resource.hpp>
 #include <rmm/mr/device/binning_memory_resource.hpp>
@@ -26,6 +24,7 @@
 #include <rmm/mr/device/pool_memory_resource.hpp>
 
 #include <benchmark/benchmark.h>
+#include <benchmarks/utilities/cxxopts.hpp>
 
 #include <array>
 #include <cstddef>
diff --git a/benchmarks/replay/replay.cpp b/benchmarks/replay/replay.cpp
index 253708ace..096adf37d 100644
--- a/benchmarks/replay/replay.cpp
+++ b/benchmarks/replay/replay.cpp
@@ -14,10 +14,6 @@
  * limitations under the License.
  */
 
-#include <benchmarks/utilities/cxxopts.hpp>
-#include <benchmarks/utilities/log_parser.hpp>
-#include <benchmarks/utilities/simulated_memory_resource.hpp>
-
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/detail/error.hpp>
 #include <rmm/mr/device/arena_memory_resource.hpp>
@@ -27,15 +23,16 @@
 #include <rmm/mr/device/owning_wrapper.hpp>
 #include <rmm/mr/device/pool_memory_resource.hpp>
 
+#include <benchmark/benchmark.h>
+#include <benchmarks/utilities/cxxopts.hpp>
+#include <benchmarks/utilities/log_parser.hpp>
+#include <benchmarks/utilities/simulated_memory_resource.hpp>
+#include <spdlog/common.h>
 #include <thrust/execution_policy.h>
 #include <thrust/iterator/constant_iterator.h>
 #include <thrust/iterator/discard_iterator.h>
 #include <thrust/reduce.h>
 
-#include <benchmark/benchmark.h>
-
-#include <spdlog/common.h>
-
 #include <atomic>
 #include <chrono>
 #include <iterator>
diff --git a/benchmarks/utilities/log_parser.hpp b/benchmarks/utilities/log_parser.hpp
index bc1aee47c..2283ace93 100644
--- a/benchmarks/utilities/log_parser.hpp
+++ b/benchmarks/utilities/log_parser.hpp
@@ -16,12 +16,12 @@
 
 #pragma once
 
-#include <chrono>
+#include "rapidcsv.h"
+
 #include <rmm/detail/error.hpp>
 #include <rmm/mr/device/device_memory_resource.hpp>
 
-#include "rapidcsv.h"
-
+#include <chrono>
 #include <cstdint>
 #include <iomanip>
 #include <limits>
diff --git a/include/rmm/cuda_stream.hpp b/include/rmm/cuda_stream.hpp
index cb5654ac3..197095c39 100644
--- a/include/rmm/cuda_stream.hpp
+++ b/include/rmm/cuda_stream.hpp
@@ -16,13 +16,13 @@
 
 #pragma once
 
-#include <functional>
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/detail/error.hpp>
 #include <rmm/detail/logging_assert.hpp>
 
 #include <cuda_runtime_api.h>
 
+#include <functional>
 #include <memory>
 
 namespace rmm {
diff --git a/include/rmm/cuda_stream_view.hpp b/include/rmm/cuda_stream_view.hpp
index f8564b16b..51ac0880d 100644
--- a/include/rmm/cuda_stream_view.hpp
+++ b/include/rmm/cuda_stream_view.hpp
@@ -18,10 +18,10 @@
 
 #include <rmm/detail/error.hpp>
 
-#include <cuda_runtime_api.h>
-
 #include <cuda/stream_ref>
 
+#include <cuda_runtime_api.h>
+
 #include <atomic>
 #include <cstddef>
 #include <cstdint>
diff --git a/include/rmm/detail/dynamic_load_runtime.hpp b/include/rmm/detail/dynamic_load_runtime.hpp
index 43626144a..9e67f9bbc 100644
--- a/include/rmm/detail/dynamic_load_runtime.hpp
+++ b/include/rmm/detail/dynamic_load_runtime.hpp
@@ -18,7 +18,6 @@
 #include <rmm/cuda_device.hpp>
 
 #include <cuda_runtime_api.h>
-
 #include <dlfcn.h>
 
 #include <memory>
diff --git a/include/rmm/detail/logging_assert.hpp b/include/rmm/detail/logging_assert.hpp
index 321bddab8..7eb667211 100644
--- a/include/rmm/detail/logging_assert.hpp
+++ b/include/rmm/detail/logging_assert.hpp
@@ -20,9 +20,10 @@
 // logger can be extremely expensive to compile, so we want to avoid including
 // it.
 #if !defined(NDEBUG)
-#include <cassert>
 #include <rmm/detail/error.hpp>
 #include <rmm/logger.hpp>
+
+#include <cassert>
 #endif
 
 /**
diff --git a/include/rmm/device_buffer.hpp b/include/rmm/device_buffer.hpp
index 167d07c4b..7dc7a30e0 100644
--- a/include/rmm/device_buffer.hpp
+++ b/include/rmm/device_buffer.hpp
@@ -21,6 +21,8 @@
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <rmm/resource_ref.hpp>
 
+#include <cuda/memory_resource>
+
 #include <cuda_runtime_api.h>
 
 #include <cassert>
@@ -28,8 +30,6 @@
 #include <stdexcept>
 #include <utility>
 
-#include <cuda/memory_resource>
-
 namespace rmm {
 /**
  * @addtogroup data_containers
diff --git a/include/rmm/device_uvector.hpp b/include/rmm/device_uvector.hpp
index 0750ef796..ff6a7d837 100644
--- a/include/rmm/device_uvector.hpp
+++ b/include/rmm/device_uvector.hpp
@@ -23,11 +23,11 @@
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <rmm/resource_ref.hpp>
 
+#include <cuda/memory_resource>
+
 #include <cstddef>
 #include <vector>
 
-#include <cuda/memory_resource>
-
 namespace rmm {
 /**
  * @addtogroup data_containers
diff --git a/include/rmm/device_vector.hpp b/include/rmm/device_vector.hpp
index 7caf5845b..64fd8b993 100644
--- a/include/rmm/device_vector.hpp
+++ b/include/rmm/device_vector.hpp
@@ -16,9 +16,9 @@
 
 #pragma once
 
+#include <rmm/detail/thrust_namespace.h>
 #include <rmm/mr/device/thrust_allocator_adaptor.hpp>
 
-#include <rmm/detail/thrust_namespace.h>
 #include <thrust/device_vector.h>
 
 namespace rmm {
diff --git a/include/rmm/exec_policy.hpp b/include/rmm/exec_policy.hpp
index 5acd062e3..a343afb8b 100644
--- a/include/rmm/exec_policy.hpp
+++ b/include/rmm/exec_policy.hpp
@@ -22,10 +22,10 @@
 #pragma once
 
 #include <rmm/cuda_stream_view.hpp>
+#include <rmm/detail/thrust_namespace.h>
 #include <rmm/mr/device/thrust_allocator_adaptor.hpp>
 #include <rmm/resource_ref.hpp>
 
-#include <rmm/detail/thrust_namespace.h>
 #include <thrust/system/cuda/execution_policy.h>
 #include <thrust/version.h>
 
diff --git a/include/rmm/mr/device/arena_memory_resource.hpp b/include/rmm/mr/device/arena_memory_resource.hpp
index b274e0c18..680a7fdec 100644
--- a/include/rmm/mr/device/arena_memory_resource.hpp
+++ b/include/rmm/mr/device/arena_memory_resource.hpp
@@ -23,7 +23,6 @@
 #include <rmm/mr/device/device_memory_resource.hpp>
 
 #include <cuda_runtime_api.h>
-
 #include <spdlog/common.h>
 
 #include <cstddef>
diff --git a/include/rmm/mr/device/cuda_async_memory_resource.hpp b/include/rmm/mr/device/cuda_async_memory_resource.hpp
index 69e23624d..1bd761eb2 100644
--- a/include/rmm/mr/device/cuda_async_memory_resource.hpp
+++ b/include/rmm/mr/device/cuda_async_memory_resource.hpp
@@ -19,13 +19,12 @@
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/detail/dynamic_load_runtime.hpp>
 #include <rmm/detail/error.hpp>
+#include <rmm/detail/thrust_namespace.h>
 #include <rmm/mr/device/cuda_async_view_memory_resource.hpp>
 #include <rmm/mr/device/device_memory_resource.hpp>
 
-#include <rmm/detail/thrust_namespace.h>
-#include <thrust/optional.h>
-
 #include <cuda_runtime_api.h>
+#include <thrust/optional.h>
 
 #include <cstddef>
 #include <limits>
diff --git a/include/rmm/mr/device/cuda_async_view_memory_resource.hpp b/include/rmm/mr/device/cuda_async_view_memory_resource.hpp
index 6c24cccc3..c07730f70 100644
--- a/include/rmm/mr/device/cuda_async_view_memory_resource.hpp
+++ b/include/rmm/mr/device/cuda_async_view_memory_resource.hpp
@@ -19,12 +19,11 @@
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/detail/dynamic_load_runtime.hpp>
 #include <rmm/detail/error.hpp>
-#include <rmm/mr/device/device_memory_resource.hpp>
-
 #include <rmm/detail/thrust_namespace.h>
-#include <thrust/optional.h>
+#include <rmm/mr/device/device_memory_resource.hpp>
 
 #include <cuda_runtime_api.h>
+#include <thrust/optional.h>
 
 #include <cstddef>
 #include <limits>
diff --git a/include/rmm/mr/device/cuda_memory_resource.hpp b/include/rmm/mr/device/cuda_memory_resource.hpp
index 0652d6360..16100b06a 100644
--- a/include/rmm/mr/device/cuda_memory_resource.hpp
+++ b/include/rmm/mr/device/cuda_memory_resource.hpp
@@ -15,10 +15,9 @@
  */
 #pragma once
 
-#include <rmm/mr/device/device_memory_resource.hpp>
-
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/detail/error.hpp>
+#include <rmm/mr/device/device_memory_resource.hpp>
 
 #include <cstddef>
 
diff --git a/include/rmm/mr/device/detail/arena.hpp b/include/rmm/mr/device/detail/arena.hpp
index c7965ca34..edc637616 100644
--- a/include/rmm/mr/device/detail/arena.hpp
+++ b/include/rmm/mr/device/detail/arena.hpp
@@ -24,7 +24,6 @@
 #include <rmm/logger.hpp>
 
 #include <cuda_runtime_api.h>
-
 #include <fmt/core.h>
 #include <spdlog/common.h>
 
diff --git a/include/rmm/mr/device/detail/stream_ordered_memory_resource.hpp b/include/rmm/mr/device/detail/stream_ordered_memory_resource.hpp
index c7c8d9178..a3a8b7581 100644
--- a/include/rmm/mr/device/detail/stream_ordered_memory_resource.hpp
+++ b/include/rmm/mr/device/detail/stream_ordered_memory_resource.hpp
@@ -22,7 +22,6 @@
 #include <rmm/mr/device/device_memory_resource.hpp>
 
 #include <cuda_runtime_api.h>
-
 #include <fmt/core.h>
 
 #include <cstddef>
diff --git a/include/rmm/mr/device/fixed_size_memory_resource.hpp b/include/rmm/mr/device/fixed_size_memory_resource.hpp
index 296ef103c..d34ad8d7e 100644
--- a/include/rmm/mr/device/fixed_size_memory_resource.hpp
+++ b/include/rmm/mr/device/fixed_size_memory_resource.hpp
@@ -19,15 +19,14 @@
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/detail/error.hpp>
 #include <rmm/detail/logging_assert.hpp>
+#include <rmm/detail/thrust_namespace.h>
 #include <rmm/mr/device/detail/fixed_size_free_list.hpp>
 #include <rmm/mr/device/detail/stream_ordered_memory_resource.hpp>
 
-#include <rmm/detail/thrust_namespace.h>
+#include <cuda_runtime_api.h>
 #include <thrust/iterator/counting_iterator.h>
 #include <thrust/iterator/transform_iterator.h>
 
-#include <cuda_runtime_api.h>
-
 #include <algorithm>
 #include <cstddef>
 #include <list>
diff --git a/include/rmm/mr/device/managed_memory_resource.hpp b/include/rmm/mr/device/managed_memory_resource.hpp
index 708adad8e..ea43315de 100644
--- a/include/rmm/mr/device/managed_memory_resource.hpp
+++ b/include/rmm/mr/device/managed_memory_resource.hpp
@@ -15,10 +15,9 @@
  */
 #pragma once
 
-#include <rmm/mr/device/device_memory_resource.hpp>
-
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/detail/error.hpp>
+#include <rmm/mr/device/device_memory_resource.hpp>
 
 #include <cstddef>
 
diff --git a/include/rmm/mr/device/pool_memory_resource.hpp b/include/rmm/mr/device/pool_memory_resource.hpp
index 5f70e73d1..d37982d00 100644
--- a/include/rmm/mr/device/pool_memory_resource.hpp
+++ b/include/rmm/mr/device/pool_memory_resource.hpp
@@ -19,20 +19,18 @@
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/detail/error.hpp>
 #include <rmm/detail/logging_assert.hpp>
+#include <rmm/detail/thrust_namespace.h>
 #include <rmm/logger.hpp>
 #include <rmm/mr/device/detail/coalescing_free_list.hpp>
 #include <rmm/mr/device/detail/stream_ordered_memory_resource.hpp>
 #include <rmm/mr/device/device_memory_resource.hpp>
 
-#include <rmm/detail/thrust_namespace.h>
+#include <cuda_runtime_api.h>
+#include <fmt/core.h>
 #include <thrust/iterator/counting_iterator.h>
 #include <thrust/iterator/transform_iterator.h>
 #include <thrust/optional.h>
 
-#include <fmt/core.h>
-
-#include <cuda_runtime_api.h>
-
 #include <algorithm>
 #include <cstddef>
 #include <iostream>
diff --git a/include/rmm/mr/device/thrust_allocator_adaptor.hpp b/include/rmm/mr/device/thrust_allocator_adaptor.hpp
index ece495c37..783f3c41f 100644
--- a/include/rmm/mr/device/thrust_allocator_adaptor.hpp
+++ b/include/rmm/mr/device/thrust_allocator_adaptor.hpp
@@ -16,16 +16,16 @@
 
 #pragma once
 
+#include <rmm/detail/thrust_namespace.h>
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <rmm/resource_ref.hpp>
 
-#include <rmm/detail/thrust_namespace.h>
+#include <cuda/memory_resource>
+
 #include <thrust/device_malloc_allocator.h>
 #include <thrust/device_ptr.h>
 #include <thrust/memory.h>
 
-#include <cuda/memory_resource>
-
 namespace rmm::mr {
 /**
  * @addtogroup device_resource_adaptors
diff --git a/include/rmm/mr/host/new_delete_resource.hpp b/include/rmm/mr/host/new_delete_resource.hpp
index ccb294d21..a48f61447 100644
--- a/include/rmm/mr/host/new_delete_resource.hpp
+++ b/include/rmm/mr/host/new_delete_resource.hpp
@@ -15,10 +15,9 @@
  */
 #pragma once
 
-#include <rmm/mr/host/host_memory_resource.hpp>
-
 #include <rmm/aligned.hpp>
 #include <rmm/detail/aligned.hpp>
+#include <rmm/mr/host/host_memory_resource.hpp>
 
 #include <cstddef>
 #include <utility>
diff --git a/include/rmm/thrust_rmm_allocator.h b/include/rmm/thrust_rmm_allocator.h
index 0b78d5497..ad71e107a 100644
--- a/include/rmm/thrust_rmm_allocator.h
+++ b/include/rmm/thrust_rmm_allocator.h
@@ -17,10 +17,10 @@
 #pragma once
 
 #include <rmm/cuda_stream_view.hpp>
+#include <rmm/detail/thrust_namespace.h>
 #include <rmm/device_vector.hpp>
 #include <rmm/mr/device/thrust_allocator_adaptor.hpp>
 
-#include <rmm/detail/thrust_namespace.h>
 #include <thrust/execution_policy.h>
 
 namespace rmm {
diff --git a/python/rmm/_lib/_torch_allocator.cpp b/python/rmm/_lib/_torch_allocator.cpp
index 5e5b2dcb4..dc92e4639 100644
--- a/python/rmm/_lib/_torch_allocator.cpp
+++ b/python/rmm/_lib/_torch_allocator.cpp
@@ -14,12 +14,12 @@
  * limitations under the License.
  */
 
-#include <cuda_runtime_api.h>
-
 #include <rmm/cuda_device.hpp>
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/mr/device/per_device_resource.hpp>
 
+#include <cuda_runtime_api.h>
+
 // These signatures must match those required by CUDAPluggableAllocator in
 // github.com/pytorch/pytorch/blob/main/torch/csrc/cuda/CUDAPluggableAllocator.h
 // Since the loading is done at runtime via dlopen, no error checking
diff --git a/tests/cuda_stream_pool_tests.cpp b/tests/cuda_stream_pool_tests.cpp
index de17f8c3c..3943c56df 100644
--- a/tests/cuda_stream_pool_tests.cpp
+++ b/tests/cuda_stream_pool_tests.cpp
@@ -18,9 +18,8 @@
 #include <rmm/detail/error.hpp>
 #include <rmm/device_uvector.hpp>
 
-#include <gtest/gtest.h>
-
 #include <cuda_runtime_api.h>
+#include <gtest/gtest.h>
 
 struct CudaStreamPoolTest : public ::testing::Test {
   rmm::cuda_stream_pool pool{};
diff --git a/tests/cuda_stream_tests.cpp b/tests/cuda_stream_tests.cpp
index 465e5d4c5..c9aaa6a87 100644
--- a/tests/cuda_stream_tests.cpp
+++ b/tests/cuda_stream_tests.cpp
@@ -17,13 +17,13 @@
 #include <rmm/cuda_stream.hpp>
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/device_buffer.hpp>
-#include <sstream>
 
 #include <cuda_runtime_api.h>
-
 #include <gtest/gtest-death-test.h>
 #include <gtest/gtest.h>
 
+#include <sstream>
+
 struct CudaStreamTest : public ::testing::Test {};
 
 TEST_F(CudaStreamTest, Equality)
diff --git a/tests/device_buffer_tests.cu b/tests/device_buffer_tests.cu
index c095eecf8..fb3439819 100644
--- a/tests/device_buffer_tests.cu
+++ b/tests/device_buffer_tests.cu
@@ -26,11 +26,10 @@
 #include <rmm/mr/device/pool_memory_resource.hpp>
 #include <rmm/resource_ref.hpp>
 
+#include <gtest/gtest.h>
 #include <thrust/equal.h>
 #include <thrust/sequence.h>
 
-#include <gtest/gtest.h>
-
 namespace testing {
 namespace thrust = THRUST_NS_QUALIFIER;
 }  // namespace testing
diff --git a/tests/device_scalar_tests.cpp b/tests/device_scalar_tests.cpp
index 5a7825533..e529d5c05 100644
--- a/tests/device_scalar_tests.cpp
+++ b/tests/device_scalar_tests.cpp
@@ -14,18 +14,17 @@
  * limitations under the License.
  */
 
-#include <gtest/gtest.h>
-
 #include <rmm/cuda_stream.hpp>
 #include <rmm/device_scalar.hpp>
 #include <rmm/mr/device/device_memory_resource.hpp>
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <rmm/resource_ref.hpp>
 
-#include <cuda_runtime_api.h>
-
 #include <cuda/memory_resource>
 
+#include <cuda_runtime_api.h>
+#include <gtest/gtest.h>
+
 #include <chrono>
 #include <cstddef>
 #include <random>
diff --git a/tests/logger_tests.cpp b/tests/logger_tests.cpp
index f6531bc59..5e3b2f899 100644
--- a/tests/logger_tests.cpp
+++ b/tests/logger_tests.cpp
@@ -15,11 +15,11 @@
  */
 
 #include "./byte_literals.hpp"
-#include <benchmarks/utilities/log_parser.hpp>
 
 #include <rmm/mr/device/cuda_memory_resource.hpp>
 #include <rmm/mr/device/logging_resource_adaptor.hpp>
 
+#include <benchmarks/utilities/log_parser.hpp>
 #include <gmock/gmock.h>
 #include <gtest/gtest.h>
 
diff --git a/tests/mr/device/arena_mr_tests.cpp b/tests/mr/device/arena_mr_tests.cpp
index c8c60f7fb..1eb38888e 100644
--- a/tests/mr/device/arena_mr_tests.cpp
+++ b/tests/mr/device/arena_mr_tests.cpp
@@ -26,10 +26,9 @@
 
 #include <gmock/gmock.h>
 #include <gtest/gtest.h>
-
-#include <memory>
 #include <sys/stat.h>
 
+#include <memory>
 #include <thread>
 #include <vector>
 
diff --git a/tests/mr/device/callback_mr_tests.cpp b/tests/mr/device/callback_mr_tests.cpp
index 95dac93ec..34a2cc8cc 100644
--- a/tests/mr/device/callback_mr_tests.cpp
+++ b/tests/mr/device/callback_mr_tests.cpp
@@ -16,19 +16,19 @@
 
 #include "../../byte_literals.hpp"
 #include "../../mock_resource.hpp"
+
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/detail/error.hpp>
 #include <rmm/device_buffer.hpp>
 #include <rmm/mr/device/callback_memory_resource.hpp>
 #include <rmm/mr/device/device_memory_resource.hpp>
 
-#include <cstddef>
-
 #include <fmt/core.h>
-
 #include <gmock/gmock.h>
 #include <gtest/gtest.h>
 
+#include <cstddef>
+
 namespace rmm::test {
 namespace {
 
diff --git a/tests/mr/device/mr_multithreaded_tests.cpp b/tests/mr/device/mr_multithreaded_tests.cpp
index 113b59e8f..704fec0de 100644
--- a/tests/mr/device/mr_multithreaded_tests.cpp
+++ b/tests/mr/device/mr_multithreaded_tests.cpp
@@ -16,8 +16,6 @@
 
 #include "mr_test.hpp"
 
-#include <gtest/gtest.h>
-
 #include <rmm/cuda_stream.hpp>
 #include <rmm/mr/device/arena_memory_resource.hpp>
 #include <rmm/mr/device/cuda_memory_resource.hpp>
@@ -25,6 +23,8 @@
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <rmm/mr/device/pool_memory_resource.hpp>
 
+#include <gtest/gtest.h>
+
 #include <thread>
 #include <vector>
 
diff --git a/tests/mr/device/mr_ref_multithreaded_tests.cpp b/tests/mr/device/mr_ref_multithreaded_tests.cpp
index 48d642a32..b310766ab 100644
--- a/tests/mr/device/mr_ref_multithreaded_tests.cpp
+++ b/tests/mr/device/mr_ref_multithreaded_tests.cpp
@@ -16,8 +16,6 @@
 
 #include "mr_ref_test.hpp"
 
-#include <gtest/gtest.h>
-
 #include <rmm/cuda_stream.hpp>
 #include <rmm/mr/device/arena_memory_resource.hpp>
 #include <rmm/mr/device/cuda_memory_resource.hpp>
@@ -27,6 +25,8 @@
 
 #include <cuda/memory_resource>
 
+#include <gtest/gtest.h>
+
 #include <thread>
 #include <vector>
 
diff --git a/tests/mr/device/mr_ref_test.hpp b/tests/mr/device/mr_ref_test.hpp
index f999e08f4..3a7719ef0 100644
--- a/tests/mr/device/mr_ref_test.hpp
+++ b/tests/mr/device/mr_ref_test.hpp
@@ -35,10 +35,10 @@
 #include <rmm/mr/device/pool_memory_resource.hpp>
 #include <rmm/resource_ref.hpp>
 
-#include <gtest/gtest.h>
-
 #include <cuda/memory_resource>
 
+#include <gtest/gtest.h>
+
 #include <cstddef>
 #include <cstdint>
 #include <functional>
diff --git a/tests/mr/device/polymorphic_allocator_tests.cpp b/tests/mr/device/polymorphic_allocator_tests.cpp
index 4c8346730..3b73d4a49 100644
--- a/tests/mr/device/polymorphic_allocator_tests.cpp
+++ b/tests/mr/device/polymorphic_allocator_tests.cpp
@@ -14,15 +14,16 @@
  * limitations under the License.
  */
 
-#include <memory>
-
-#include <gtest/gtest.h>
 #include <rmm/cuda_stream.hpp>
 #include <rmm/mr/device/cuda_memory_resource.hpp>
 #include <rmm/mr/device/managed_memory_resource.hpp>
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <rmm/mr/device/polymorphic_allocator.hpp>
 
+#include <gtest/gtest.h>
+
+#include <memory>
+
 namespace {
 
 struct allocator_test : public ::testing::Test {
diff --git a/tests/mr/device/thrust_allocator_tests.cu b/tests/mr/device/thrust_allocator_tests.cu
index 038f4b664..e25a2eece 100644
--- a/tests/mr/device/thrust_allocator_tests.cu
+++ b/tests/mr/device/thrust_allocator_tests.cu
@@ -23,7 +23,6 @@
 #include <rmm/resource_ref.hpp>
 
 #include <gtest/gtest.h>
-
 #include <thrust/reduce.h>
 
 // explicit instantiation for test coverage purposes
diff --git a/tests/mr/device/tracking_mr_tests.cpp b/tests/mr/device/tracking_mr_tests.cpp
index 2af720d3f..65d2f955c 100644
--- a/tests/mr/device/tracking_mr_tests.cpp
+++ b/tests/mr/device/tracking_mr_tests.cpp
@@ -23,7 +23,6 @@
 #include <rmm/mr/device/tracking_resource_adaptor.hpp>
 
 #include <gtest/gtest.h>
-
 #include <spdlog/sinks/ostream_sink.h>
 
 namespace rmm::test {
diff --git a/tests/mr/host/mr_ref_tests.cpp b/tests/mr/host/mr_ref_tests.cpp
index 416641f18..0f11dc164 100644
--- a/tests/mr/host/mr_ref_tests.cpp
+++ b/tests/mr/host/mr_ref_tests.cpp
@@ -21,10 +21,9 @@
 #include <rmm/mr/host/new_delete_resource.hpp>
 #include <rmm/mr/host/pinned_memory_resource.hpp>
 
-#include <cuda_runtime_api.h>
-
 #include <cuda/memory_resource>
 
+#include <cuda_runtime_api.h>
 #include <gtest/gtest.h>
 
 #include <cstddef>
diff --git a/tests/mr/host/mr_tests.cpp b/tests/mr/host/mr_tests.cpp
index e0078c920..a06f800d9 100644
--- a/tests/mr/host/mr_tests.cpp
+++ b/tests/mr/host/mr_tests.cpp
@@ -21,10 +21,9 @@
 #include <rmm/mr/host/new_delete_resource.hpp>
 #include <rmm/mr/host/pinned_memory_resource.hpp>
 
-#include <cuda_runtime_api.h>
-
 #include <cuda/memory_resource>
 
+#include <cuda_runtime_api.h>
 #include <gtest/gtest.h>
 
 #include <cstddef>

From 384d8c8d036d318da0e5b645e0533b56967d256c Mon Sep 17 00:00:00 2001
From: Mark Harris <783069+harrism@users.noreply.github.com>
Date: Wed, 7 Feb 2024 23:15:58 +0000
Subject: [PATCH 3/4] Separate group for CCCL (thrust/cub/cuda) and CUDART
 includes

---
 .clang-format                                            | 4 +++-
 benchmarks/cuda_stream_pool/cuda_stream_pool_bench.cpp   | 3 ++-
 benchmarks/device_uvector/device_uvector_bench.cu        | 3 ++-
 .../multi_stream_allocations_bench.cu                    | 3 ++-
 benchmarks/replay/replay.cpp                             | 9 +++++----
 benchmarks/synchronization/synchronization.hpp           | 3 ++-
 include/rmm/cuda_stream_view.hpp                         | 1 -
 include/rmm/detail/dynamic_load_runtime.hpp              | 1 +
 include/rmm/device_buffer.hpp                            | 1 -
 include/rmm/mr/device/arena_memory_resource.hpp          | 1 +
 include/rmm/mr/device/detail/arena.hpp                   | 1 +
 .../mr/device/detail/stream_ordered_memory_resource.hpp  | 1 +
 include/rmm/mr/device/pool_memory_resource.hpp           | 3 ++-
 include/rmm/mr/device/thrust_allocator_adaptor.hpp       | 1 -
 include/rmm/mr/pinned_host_memory_resource.hpp           | 1 -
 tests/cuda_stream_pool_tests.cpp                         | 1 +
 tests/cuda_stream_tests.cpp                              | 1 +
 tests/device_buffer_tests.cu                             | 3 ++-
 tests/device_scalar_tests.cpp                            | 2 +-
 tests/mr/device/thrust_allocator_tests.cu                | 3 ++-
 tests/mr/host/mr_ref_tests.cpp                           | 2 +-
 tests/mr/host/mr_tests.cpp                               | 2 +-
 22 files changed, 31 insertions(+), 19 deletions(-)

diff --git a/.clang-format b/.clang-format
index b36983d2e..74c37f464 100644
--- a/.clang-format
+++ b/.clang-format
@@ -77,7 +77,9 @@ IncludeCategories:
     Priority:        1
   - Regex:           '^<rmm/' # RMM includes
     Priority:        2
-  - Regex:           '^<cuda/' # CUDA includes
+  - Regex:           '^<(thrust|cub|cuda)/' # CCCL includes
+    Priority:        3
+  - Regex:           '^<(cuda_runtime_api.h|device_types.h)' # CUDART includes
     Priority:        3
   - Regex:           '^<.*\..*' # other system includes (e.g. with a '.')
     Priority:        4
diff --git a/benchmarks/cuda_stream_pool/cuda_stream_pool_bench.cpp b/benchmarks/cuda_stream_pool/cuda_stream_pool_bench.cpp
index 5adf23354..6710ffe50 100644
--- a/benchmarks/cuda_stream_pool/cuda_stream_pool_bench.cpp
+++ b/benchmarks/cuda_stream_pool/cuda_stream_pool_bench.cpp
@@ -17,9 +17,10 @@
 #include <rmm/cuda_stream_pool.hpp>
 #include <rmm/detail/error.hpp>
 
-#include <benchmark/benchmark.h>
 #include <cuda_runtime_api.h>
 
+#include <benchmark/benchmark.h>
+
 #include <stdexcept>
 
 static void BM_StreamPoolGetStream(benchmark::State& state)
diff --git a/benchmarks/device_uvector/device_uvector_bench.cu b/benchmarks/device_uvector/device_uvector_bench.cu
index 7181de325..36c9183f9 100644
--- a/benchmarks/device_uvector/device_uvector_bench.cu
+++ b/benchmarks/device_uvector/device_uvector_bench.cu
@@ -26,11 +26,12 @@
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <rmm/mr/device/pool_memory_resource.hpp>
 
-#include <benchmark/benchmark.h>
 #include <cuda_runtime_api.h>
 #include <thrust/device_vector.h>
 #include <thrust/memory.h>
 
+#include <benchmark/benchmark.h>
+
 #include <cstdio>
 #include <type_traits>
 
diff --git a/benchmarks/multi_stream_allocations/multi_stream_allocations_bench.cu b/benchmarks/multi_stream_allocations/multi_stream_allocations_bench.cu
index e407476b0..d38e89a3b 100644
--- a/benchmarks/multi_stream_allocations/multi_stream_allocations_bench.cu
+++ b/benchmarks/multi_stream_allocations/multi_stream_allocations_bench.cu
@@ -26,9 +26,10 @@
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <rmm/mr/device/pool_memory_resource.hpp>
 
+#include <cuda_runtime_api.h>
+
 #include <benchmark/benchmark.h>
 #include <benchmarks/utilities/cxxopts.hpp>
-#include <cuda_runtime_api.h>
 
 #include <cstddef>
 
diff --git a/benchmarks/replay/replay.cpp b/benchmarks/replay/replay.cpp
index 096adf37d..5afed036a 100644
--- a/benchmarks/replay/replay.cpp
+++ b/benchmarks/replay/replay.cpp
@@ -23,15 +23,16 @@
 #include <rmm/mr/device/owning_wrapper.hpp>
 #include <rmm/mr/device/pool_memory_resource.hpp>
 
+#include <thrust/execution_policy.h>
+#include <thrust/iterator/constant_iterator.h>
+#include <thrust/iterator/discard_iterator.h>
+#include <thrust/reduce.h>
+
 #include <benchmark/benchmark.h>
 #include <benchmarks/utilities/cxxopts.hpp>
 #include <benchmarks/utilities/log_parser.hpp>
 #include <benchmarks/utilities/simulated_memory_resource.hpp>
 #include <spdlog/common.h>
-#include <thrust/execution_policy.h>
-#include <thrust/iterator/constant_iterator.h>
-#include <thrust/iterator/discard_iterator.h>
-#include <thrust/reduce.h>
 
 #include <atomic>
 #include <chrono>
diff --git a/benchmarks/synchronization/synchronization.hpp b/benchmarks/synchronization/synchronization.hpp
index 1f26beb37..596b3684b 100644
--- a/benchmarks/synchronization/synchronization.hpp
+++ b/benchmarks/synchronization/synchronization.hpp
@@ -61,9 +61,10 @@
 #include <rmm/cuda_stream_view.hpp>
 
 // Google Benchmark library
-#include <benchmark/benchmark.h>
 #include <cuda_runtime_api.h>
 
+#include <benchmark/benchmark.h>
+
 class cuda_event_timer {
  public:
   /**
diff --git a/include/rmm/cuda_stream_view.hpp b/include/rmm/cuda_stream_view.hpp
index 51ac0880d..a34897df6 100644
--- a/include/rmm/cuda_stream_view.hpp
+++ b/include/rmm/cuda_stream_view.hpp
@@ -19,7 +19,6 @@
 #include <rmm/detail/error.hpp>
 
 #include <cuda/stream_ref>
-
 #include <cuda_runtime_api.h>
 
 #include <atomic>
diff --git a/include/rmm/detail/dynamic_load_runtime.hpp b/include/rmm/detail/dynamic_load_runtime.hpp
index 9e67f9bbc..43626144a 100644
--- a/include/rmm/detail/dynamic_load_runtime.hpp
+++ b/include/rmm/detail/dynamic_load_runtime.hpp
@@ -18,6 +18,7 @@
 #include <rmm/cuda_device.hpp>
 
 #include <cuda_runtime_api.h>
+
 #include <dlfcn.h>
 
 #include <memory>
diff --git a/include/rmm/device_buffer.hpp b/include/rmm/device_buffer.hpp
index 7dc7a30e0..4663b5733 100644
--- a/include/rmm/device_buffer.hpp
+++ b/include/rmm/device_buffer.hpp
@@ -22,7 +22,6 @@
 #include <rmm/resource_ref.hpp>
 
 #include <cuda/memory_resource>
-
 #include <cuda_runtime_api.h>
 
 #include <cassert>
diff --git a/include/rmm/mr/device/arena_memory_resource.hpp b/include/rmm/mr/device/arena_memory_resource.hpp
index 680a7fdec..b274e0c18 100644
--- a/include/rmm/mr/device/arena_memory_resource.hpp
+++ b/include/rmm/mr/device/arena_memory_resource.hpp
@@ -23,6 +23,7 @@
 #include <rmm/mr/device/device_memory_resource.hpp>
 
 #include <cuda_runtime_api.h>
+
 #include <spdlog/common.h>
 
 #include <cstddef>
diff --git a/include/rmm/mr/device/detail/arena.hpp b/include/rmm/mr/device/detail/arena.hpp
index edc637616..c7965ca34 100644
--- a/include/rmm/mr/device/detail/arena.hpp
+++ b/include/rmm/mr/device/detail/arena.hpp
@@ -24,6 +24,7 @@
 #include <rmm/logger.hpp>
 
 #include <cuda_runtime_api.h>
+
 #include <fmt/core.h>
 #include <spdlog/common.h>
 
diff --git a/include/rmm/mr/device/detail/stream_ordered_memory_resource.hpp b/include/rmm/mr/device/detail/stream_ordered_memory_resource.hpp
index a3a8b7581..c7c8d9178 100644
--- a/include/rmm/mr/device/detail/stream_ordered_memory_resource.hpp
+++ b/include/rmm/mr/device/detail/stream_ordered_memory_resource.hpp
@@ -22,6 +22,7 @@
 #include <rmm/mr/device/device_memory_resource.hpp>
 
 #include <cuda_runtime_api.h>
+
 #include <fmt/core.h>
 
 #include <cstddef>
diff --git a/include/rmm/mr/device/pool_memory_resource.hpp b/include/rmm/mr/device/pool_memory_resource.hpp
index d37982d00..4f2d46a7c 100644
--- a/include/rmm/mr/device/pool_memory_resource.hpp
+++ b/include/rmm/mr/device/pool_memory_resource.hpp
@@ -26,11 +26,12 @@
 #include <rmm/mr/device/device_memory_resource.hpp>
 
 #include <cuda_runtime_api.h>
-#include <fmt/core.h>
 #include <thrust/iterator/counting_iterator.h>
 #include <thrust/iterator/transform_iterator.h>
 #include <thrust/optional.h>
 
+#include <fmt/core.h>
+
 #include <algorithm>
 #include <cstddef>
 #include <iostream>
diff --git a/include/rmm/mr/device/thrust_allocator_adaptor.hpp b/include/rmm/mr/device/thrust_allocator_adaptor.hpp
index 783f3c41f..f2b1a6027 100644
--- a/include/rmm/mr/device/thrust_allocator_adaptor.hpp
+++ b/include/rmm/mr/device/thrust_allocator_adaptor.hpp
@@ -21,7 +21,6 @@
 #include <rmm/resource_ref.hpp>
 
 #include <cuda/memory_resource>
-
 #include <thrust/device_malloc_allocator.h>
 #include <thrust/device_ptr.h>
 #include <thrust/memory.h>
diff --git a/include/rmm/mr/pinned_host_memory_resource.hpp b/include/rmm/mr/pinned_host_memory_resource.hpp
index e92b2985e..ee409dbac 100644
--- a/include/rmm/mr/pinned_host_memory_resource.hpp
+++ b/include/rmm/mr/pinned_host_memory_resource.hpp
@@ -21,7 +21,6 @@
 
 #include <cuda/memory_resource>
 #include <cuda/stream_ref>
-
 #include <cuda_runtime_api.h>
 
 #include <cstddef>
diff --git a/tests/cuda_stream_pool_tests.cpp b/tests/cuda_stream_pool_tests.cpp
index 3943c56df..ad1d421d7 100644
--- a/tests/cuda_stream_pool_tests.cpp
+++ b/tests/cuda_stream_pool_tests.cpp
@@ -19,6 +19,7 @@
 #include <rmm/device_uvector.hpp>
 
 #include <cuda_runtime_api.h>
+
 #include <gtest/gtest.h>
 
 struct CudaStreamPoolTest : public ::testing::Test {
diff --git a/tests/cuda_stream_tests.cpp b/tests/cuda_stream_tests.cpp
index c9aaa6a87..1cc068434 100644
--- a/tests/cuda_stream_tests.cpp
+++ b/tests/cuda_stream_tests.cpp
@@ -19,6 +19,7 @@
 #include <rmm/device_buffer.hpp>
 
 #include <cuda_runtime_api.h>
+
 #include <gtest/gtest-death-test.h>
 #include <gtest/gtest.h>
 
diff --git a/tests/device_buffer_tests.cu b/tests/device_buffer_tests.cu
index fb3439819..c095eecf8 100644
--- a/tests/device_buffer_tests.cu
+++ b/tests/device_buffer_tests.cu
@@ -26,10 +26,11 @@
 #include <rmm/mr/device/pool_memory_resource.hpp>
 #include <rmm/resource_ref.hpp>
 
-#include <gtest/gtest.h>
 #include <thrust/equal.h>
 #include <thrust/sequence.h>
 
+#include <gtest/gtest.h>
+
 namespace testing {
 namespace thrust = THRUST_NS_QUALIFIER;
 }  // namespace testing
diff --git a/tests/device_scalar_tests.cpp b/tests/device_scalar_tests.cpp
index e529d5c05..6f80a5de1 100644
--- a/tests/device_scalar_tests.cpp
+++ b/tests/device_scalar_tests.cpp
@@ -21,8 +21,8 @@
 #include <rmm/resource_ref.hpp>
 
 #include <cuda/memory_resource>
-
 #include <cuda_runtime_api.h>
+
 #include <gtest/gtest.h>
 
 #include <chrono>
diff --git a/tests/mr/device/thrust_allocator_tests.cu b/tests/mr/device/thrust_allocator_tests.cu
index e25a2eece..372d5c327 100644
--- a/tests/mr/device/thrust_allocator_tests.cu
+++ b/tests/mr/device/thrust_allocator_tests.cu
@@ -22,9 +22,10 @@
 #include <rmm/mr/device/thrust_allocator_adaptor.hpp>
 #include <rmm/resource_ref.hpp>
 
-#include <gtest/gtest.h>
 #include <thrust/reduce.h>
 
+#include <gtest/gtest.h>
+
 // explicit instantiation for test coverage purposes
 template class rmm::mr::thrust_allocator<int>;
 
diff --git a/tests/mr/host/mr_ref_tests.cpp b/tests/mr/host/mr_ref_tests.cpp
index 0f11dc164..49b5bf061 100644
--- a/tests/mr/host/mr_ref_tests.cpp
+++ b/tests/mr/host/mr_ref_tests.cpp
@@ -22,8 +22,8 @@
 #include <rmm/mr/host/pinned_memory_resource.hpp>
 
 #include <cuda/memory_resource>
-
 #include <cuda_runtime_api.h>
+
 #include <gtest/gtest.h>
 
 #include <cstddef>
diff --git a/tests/mr/host/mr_tests.cpp b/tests/mr/host/mr_tests.cpp
index a06f800d9..4daa80721 100644
--- a/tests/mr/host/mr_tests.cpp
+++ b/tests/mr/host/mr_tests.cpp
@@ -22,8 +22,8 @@
 #include <rmm/mr/host/pinned_memory_resource.hpp>
 
 #include <cuda/memory_resource>
-
 #include <cuda_runtime_api.h>
+
 #include <gtest/gtest.h>
 
 #include <cstddef>

From bea71191d540de66150678ad0856085c42f19d5a Mon Sep 17 00:00:00 2001
From: Mark Harris <mharris@nvidia.com>
Date: Tue, 20 Feb 2024 07:48:32 +0000
Subject: [PATCH 4/4] Match libcudf's groupings closer.

---
 .clang-format | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.clang-format b/.clang-format
index 74c37f464..35ebdbfae 100644
--- a/.clang-format
+++ b/.clang-format
@@ -79,7 +79,7 @@ IncludeCategories:
     Priority:        2
   - Regex:           '^<(thrust|cub|cuda)/' # CCCL includes
     Priority:        3
-  - Regex:           '^<(cuda_runtime_api.h|device_types.h)' # CUDART includes
+  - Regex:           '^<(cooperative_groups|cuco|cuda.h|cuda_runtime|device_types|math_constants|nvtx3)' # CUDA includes
     Priority:        3
   - Regex:           '^<.*\..*' # other system includes (e.g. with a '.')
     Priority:        4