diff --git a/.github/workflows/rocprofiler-sdk-continuous_integration.yml b/.github/workflows/rocprofiler-sdk-continuous_integration.yml
index 573395f021c..32c0be11a32 100644
--- a/.github/workflows/rocprofiler-sdk-continuous_integration.yml
+++ b/.github/workflows/rocprofiler-sdk-continuous_integration.yml
@@ -325,6 +325,8 @@ jobs:
             ${{ matrix.system.ci-flags }} -- \
             -DROCPROFILER_DEP_ROCMCORE=ON \
             -DROCPROFILER_BUILD_DOCS=OFF \
+            -DROCPROFILER_BUILD_FMT=OFF \
+            -DROCPROFILER_INTERNAL_RCCL_API_TRACE=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.system.build-type }} \
             -DCMAKE_INSTALL_PREFIX=/opt/rocprofiler-sdk \
             -DCPACK_GENERATOR='DEB;RPM;TGZ' \
@@ -632,6 +634,8 @@ jobs:
             -- \
             -DROCPROFILER_DEP_ROCMCORE=ON \
             -DROCPROFILER_BUILD_DOCS=OFF \
+            -DROCPROFILER_BUILD_FMT=OFF \
+            -DROCPROFILER_INTERNAL_RCCL_API_TRACE=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.system.build-type }} \
             -DCMAKE_PREFIX_PATH='${{ env.ROCM_PATH }};${{ env.ROCM_PATH }}/llvm' \
             -DPython3_EXECUTABLE=$(which python3) \
@@ -846,6 +850,8 @@ jobs:
           --memcheck ${{ matrix.system.sanitizer }}
           --run-attempt ${{ github.run_attempt }}
           --
+          -DROCPROFILER_BUILD_FMT=OFF \
+          -DROCPROFILER_INTERNAL_RCCL_API_TRACE=ON \
           -DCMAKE_BUILD_TYPE=${{ matrix.system.build-type }}
           -DCMAKE_INSTALL_PREFIX="${{ env.ROCM_PATH }}"
           -DCMAKE_PREFIX_PATH='${{ env.ROCM_PATH }};${{ env.ROCM_PATH }}/llvm'
diff --git a/projects/rocprofiler-sdk/cmake/rocprofiler_linting.cmake b/projects/rocprofiler-sdk/cmake/rocprofiler_linting.cmake
index 36a7e1c802e..234ca66a2a9 100644
--- a/projects/rocprofiler-sdk/cmake/rocprofiler_linting.cmake
+++ b/projects/rocprofiler-sdk/cmake/rocprofiler_linting.cmake
@@ -78,7 +78,9 @@ macro(ROCPROFILER_ACTIVATE_CLANG_TIDY)
                                 "path to clang-tidy executable")
 
         set(CMAKE_CXX_CLANG_TIDY
-            ${ROCPROFILER_CLANG_TIDY_EXE} -header-filter=${PROJECT_SOURCE_DIR}/source/.*
+            ${ROCPROFILER_CLANG_TIDY_EXE}
+            -header-filter=${PROJECT_SOURCE_DIR}/source/.*
+            --exclude-header-filter=${PROJECT_SOURCE_DIR}/source/include/rocprofiler-sdk/rccl/details/.*
             --warnings-as-errors=*,-misc-header-include-cycle)
 
         # Create a preprocessor definition that depends on .clang-tidy content so the
diff --git a/projects/rocprofiler-sdk/source/lib/common/logging.cpp b/projects/rocprofiler-sdk/source/lib/common/logging.cpp
index 6c07322b065..24aecbc23db 100644
--- a/projects/rocprofiler-sdk/source/lib/common/logging.cpp
+++ b/projects/rocprofiler-sdk/source/lib/common/logging.cpp
@@ -25,6 +25,7 @@
 #include "lib/common/filesystem.hpp"
 
 #include <fmt/format.h>
+#include <fmt/ranges.h>
 #include <glog/logging.h>
 #include <glog/vlog_is_on.h>
 
diff --git a/projects/rocprofiler-sdk/source/lib/output/generateRocpd.cpp b/projects/rocprofiler-sdk/source/lib/output/generateRocpd.cpp
index 564a9bd45f0..65d9a5d7cf7 100644
--- a/projects/rocprofiler-sdk/source/lib/output/generateRocpd.cpp
+++ b/projects/rocprofiler-sdk/source/lib/output/generateRocpd.cpp
@@ -56,6 +56,7 @@
 #include <rocprofiler-sdk-rocpd/types.h>
 
 #include <fmt/format.h>
+#include <fmt/ranges.h>
 #include <sqlite3.h>
 #include <unistd.h>
 
diff --git a/projects/rocprofiler-sdk/source/lib/output/generateStats.cpp b/projects/rocprofiler-sdk/source/lib/output/generateStats.cpp
index 22c1a7f4259..3860be129b1 100644
--- a/projects/rocprofiler-sdk/source/lib/output/generateStats.cpp
+++ b/projects/rocprofiler-sdk/source/lib/output/generateStats.cpp
@@ -33,6 +33,7 @@
 #include <rocprofiler-sdk/marker/api_id.h>
 
 #include <fmt/format.h>
+#include <fmt/ranges.h>
 
 #include <unistd.h>
 #include <cstdint>
diff --git a/projects/rocprofiler-sdk/source/lib/output/sql/common.cpp b/projects/rocprofiler-sdk/source/lib/output/sql/common.cpp
index 32587752da2..c442a4f9e53 100644
--- a/projects/rocprofiler-sdk/source/lib/output/sql/common.cpp
+++ b/projects/rocprofiler-sdk/source/lib/output/sql/common.cpp
@@ -31,6 +31,7 @@
 #include <rocprofiler-sdk/cxx/operators.hpp>
 
 #include <fmt/format.h>
+#include <fmt/ranges.h>
 #include <sqlite3.h>
 
 #include <iomanip>
diff --git a/projects/rocprofiler-sdk/source/lib/python/rocpd/libpyrocpd.cpp b/projects/rocprofiler-sdk/source/lib/python/rocpd/libpyrocpd.cpp
index 3bf5b4cb68a..343e7bb9f4b 100644
--- a/projects/rocprofiler-sdk/source/lib/python/rocpd/libpyrocpd.cpp
+++ b/projects/rocprofiler-sdk/source/lib/python/rocpd/libpyrocpd.cpp
@@ -54,6 +54,7 @@
 #include <rocprofiler-sdk-rocpd/sql.h>
 
 #include <fmt/format.h>
+#include <fmt/ranges.h>
 #include <gotcha/gotcha.h>
 #include <pybind11/detail/common.h>
 #include <pybind11/pybind11.h>
diff --git a/projects/rocprofiler-sdk/source/lib/python/rocpd/source/serialization/sql.cpp b/projects/rocprofiler-sdk/source/lib/python/rocpd/source/serialization/sql.cpp
index d10e878be5d..9192abc45bc 100644
--- a/projects/rocprofiler-sdk/source/lib/python/rocpd/source/serialization/sql.cpp
+++ b/projects/rocprofiler-sdk/source/lib/python/rocpd/source/serialization/sql.cpp
@@ -23,6 +23,9 @@
 #include "lib/python/rocpd/source/serialization/sql.hpp"
 #include "lib/output/sql/common.hpp"
 
+#include <fmt/format.h>
+#include <fmt/ranges.h>
+
 namespace cereal
 {
 SQLite3InputArchive::SQLite3InputArchive(sqlite3*         conn,
diff --git a/projects/rocprofiler-sdk/source/lib/python/rocpd/source/sql_generator.hpp b/projects/rocprofiler-sdk/source/lib/python/rocpd/source/sql_generator.hpp
index 6284857afbf..8563734cda3 100644
--- a/projects/rocprofiler-sdk/source/lib/python/rocpd/source/sql_generator.hpp
+++ b/projects/rocprofiler-sdk/source/lib/python/rocpd/source/sql_generator.hpp
@@ -31,6 +31,8 @@
 #include "lib/output/generator.hpp"
 #include "lib/output/sql/common.hpp"
 
+#include <fmt/format.h>
+#include <fmt/ranges.h>
 #include <sqlite3.h>
 
 #include <chrono>
diff --git a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk-tool/tool.cpp b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk-tool/tool.cpp
index 7b03fa3eb86..e36582550d9 100644
--- a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk-tool/tool.cpp
+++ b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk-tool/tool.cpp
@@ -79,6 +79,7 @@
 #include <rocprofiler-sdk/cxx/operators.hpp>
 
 #include <fmt/core.h>
+#include <fmt/ranges.h>
 
 #include <time.h>
 #include <unistd.h>
diff --git a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/metrics.hpp b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/metrics.hpp
index bc9525612ea..d3186c332cd 100644
--- a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/metrics.hpp
+++ b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/metrics.hpp
@@ -30,6 +30,7 @@
 #include <hsa/hsa_ven_amd_aqlprofile.h>
 
 #include <cstdint>
+#include <memory>
 #include <optional>
 #include <string>
 #include <unordered_map>
diff --git a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/device_counting.cpp b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/device_counting.cpp
index 9587b219f3b..d82a6ec40be 100644
--- a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/device_counting.cpp
+++ b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/device_counting.cpp
@@ -39,6 +39,7 @@
 #include <rocprofiler-sdk/rocprofiler.h>
 
 #include <fmt/core.h>
+#include <fmt/ranges.h>
 #include <gtest/gtest.h>
 #include <hsa/hsa.h>
 #include <hsa/hsa_api_trace.h>
diff --git a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/dimension.cpp b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/dimension.cpp
index 387c167bd05..22fcda9bf7c 100644
--- a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/dimension.cpp
+++ b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/dimension.cpp
@@ -39,6 +39,7 @@
 #include <rocprofiler-sdk/rocprofiler.h>
 
 #include <fmt/core.h>
+#include <fmt/ranges.h>
 #include <gtest/gtest.h>
 #include <hsa/hsa.h>
 #include <hsa/hsa_api_trace.h>
diff --git a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hip/details/format.hpp b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hip/details/format.hpp
index f65cf016a63..46f734d65df 100644
--- a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hip/details/format.hpp
+++ b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hip/details/format.hpp
@@ -31,8 +31,8 @@
 #include <hip/hip_deprecated.h>
 #include <hip/hip_version.h>
 
-#include "fmt/core.h"
-#include "fmt/ranges.h"
+#include <fmt/core.h>
+#include <fmt/ranges.h>
 
 #define ROCP_SDK_HIP_FORMATTER(TYPE, ...)                                                          \
     template <>                                                                                    \
diff --git a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hsa/details/fmt.hpp b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hsa/details/fmt.hpp
index c136a2739c8..f0d7d9e94f0 100644
--- a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hsa/details/fmt.hpp
+++ b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hsa/details/fmt.hpp
@@ -22,6 +22,9 @@
 
 #include "lib/rocprofiler-sdk/hsa/queue.hpp"
 
+#include <fmt/format.h>
+#include <fmt/ranges.h>
+
 namespace fmt
 {
 template <>
diff --git a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hsa/queue.cpp b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hsa/queue.cpp
index 82955da324c..c6b7a5420ae 100644
--- a/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hsa/queue.cpp
+++ b/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hsa/queue.cpp
@@ -39,6 +39,8 @@
 #include <rocprofiler-sdk/external_correlation.h>
 #include <rocprofiler-sdk/fwd.h>
 
+#include <fmt/format.h>
+#include <fmt/ranges.h>
 #include <hsa/hsa.h>
 #include <hsa/hsa_ext_amd.h>
 
diff --git a/projects/rocprofiler-sdk/source/lib/tests/common/parse.cpp b/projects/rocprofiler-sdk/source/lib/tests/common/parse.cpp
index dc6995c28fa..abe0fe5c07e 100644
--- a/projects/rocprofiler-sdk/source/lib/tests/common/parse.cpp
+++ b/projects/rocprofiler-sdk/source/lib/tests/common/parse.cpp
@@ -23,6 +23,7 @@
 #include <rocprofiler-sdk/cxx/details/tokenize.hpp>
 
 #include <fmt/format.h>
+#include <fmt/ranges.h>
 #include <gtest/gtest.h>
 
 #include <cstddef>