diff --git a/libkineto/src/Config.cpp b/libkineto/src/Config.cpp
index 18103440b..83d57ad42 100644
--- a/libkineto/src/Config.cpp
+++ b/libkineto/src/Config.cpp
@@ -15,13 +15,9 @@
 #include <fmt/ranges.h>
 #include <time.h>
 #include <chrono>
-#include <fstream>
 #include <functional>
-#include <iomanip>
-#include <istream>
 #include <mutex>
 #include <ostream>
-#include <sstream>
 
 #include "Logger.h"
 #include "ThreadUtil.h"
diff --git a/libkineto/src/ConfigLoader.h b/libkineto/src/ConfigLoader.h
index 84366495c..31b3dc727 100644
--- a/libkineto/src/ConfigLoader.h
+++ b/libkineto/src/ConfigLoader.h
@@ -19,10 +19,6 @@
 
 #include "Config.h"
 
-// TODO(T90238193)
-// @lint-ignore-every CLANGTIDY facebook-hte-RelativeInclude
-#include "ILoggerObserver.h"
-
 namespace libkineto {
 class LibkinetoApi;
 }
@@ -95,7 +91,7 @@ class ConfigLoader {
   }
 
   bool hasNewConfig(const Config& oldConfig);
-  int contextCountForGpu(uint32_t gpu);
+  int contextCountForGpu(uint32_t device);
 
   void handleOnDemandSignal();
 
diff --git a/libkineto/src/CuptiActivityApi.cpp b/libkineto/src/CuptiActivityApi.cpp
index d9e4f56df..425be9bf2 100644
--- a/libkineto/src/CuptiActivityApi.cpp
+++ b/libkineto/src/CuptiActivityApi.cpp
@@ -8,12 +8,10 @@
 
 #include "CuptiActivityApi.h"
 
-#include <assert.h>
 #include <chrono>
 #include <mutex>
 #include <thread>
 
-#include "Config.h"
 #include "DeviceUtil.h"
 #include "Logger.h"
 
@@ -214,7 +212,7 @@ std::unique_ptr<CuptiActivityBufferMap> CuptiActivityApi::activityBuffers() {
 int CuptiActivityApi::processActivitiesForBuffer(
     uint8_t* buf,
     size_t validSize,
-    std::function<void(const CUpti_Activity*)> handler) {
+    const std::function<void(const CUpti_Activity*)>& handler) {
   int count = 0;
   if (buf && validSize) {
     CUpti_Activity* record{nullptr};
@@ -229,7 +227,7 @@ int CuptiActivityApi::processActivitiesForBuffer(
 
 const std::pair<int, size_t> CuptiActivityApi::processActivities(
     CuptiActivityBufferMap& buffers,
-    std::function<void(const CUpti_Activity*)> handler) {
+    const std::function<void(const CUpti_Activity*)>& handler) {
   std::pair<int, size_t> res{0, 0};
 #ifdef HAS_CUPTI
   for (auto& pair : buffers) {
diff --git a/libkineto/src/CuptiActivityApi.h b/libkineto/src/CuptiActivityApi.h
index e9ee6d56c..448cc5415 100644
--- a/libkineto/src/CuptiActivityApi.h
+++ b/libkineto/src/CuptiActivityApi.h
@@ -11,7 +11,6 @@
 #include <atomic>
 #include <condition_variable>
 #include <functional>
-#include <list>
 #include <memory>
 #include <mutex>
 #include <set>
@@ -48,7 +47,7 @@ class CuptiActivityApi {
   CuptiActivityApi(const CuptiActivityApi&) = delete;
   CuptiActivityApi& operator=(const CuptiActivityApi&) = delete;
 
-  virtual ~CuptiActivityApi() {}
+  virtual ~CuptiActivityApi() = default;
 
   static CuptiActivityApi& singleton();
 
@@ -65,7 +64,7 @@ class CuptiActivityApi {
 
   virtual const std::pair<int, size_t> processActivities(
       CuptiActivityBufferMap&,
-      std::function<void(const CUpti_Activity*)> handler);
+      const std::function<void(const CUpti_Activity*)>& handler);
 
   void setMaxBufferSize(int size);
   void setDeviceBufferSize(size_t size);
@@ -91,7 +90,7 @@ class CuptiActivityApi {
   int processActivitiesForBuffer(
       uint8_t* buf,
       size_t validSize,
-      std::function<void(const CUpti_Activity*)> handler);
+      const std::function<void(const CUpti_Activity*)>& handler);
   static void CUPTIAPI bufferRequestedTrampoline(
       uint8_t** buffer,
       size_t* size,
diff --git a/libkineto/src/CuptiActivityProfiler.h b/libkineto/src/CuptiActivityProfiler.h
index 9328ad144..30d2203d4 100644
--- a/libkineto/src/CuptiActivityProfiler.h
+++ b/libkineto/src/CuptiActivityProfiler.h
@@ -10,7 +10,6 @@
 
 #include <atomic>
 #include <chrono>
-#include <condition_variable>
 #include <deque>
 #include <list>
 #include <map>
@@ -149,11 +148,11 @@ class CuptiActivityProfiler {
     gpuActivityPresent_ = val;
   }
 
-  inline bool gpuActivityPresent() {
+  inline bool gpuActivityPresent() const {
     return gpuActivityPresent_;
   }
 
-  inline bool traceNonEmpty() {
+  inline bool traceNonEmpty() const {
     return cpuActivityPresent_ || gpuActivityPresent_;
   }
 
@@ -172,7 +171,7 @@ class CuptiActivityProfiler {
 
   // Collect CPU and GPU traces
   void collectTrace(
-      bool collectionDone,
+      bool collection_done,
       const std::chrono::time_point<std::chrono::system_clock>& now);
 
   // Ensure collectTrace is done
diff --git a/libkineto/src/RoctracerActivityApi.cpp b/libkineto/src/RoctracerActivityApi.cpp
index 34fa97aa0..4d512e3ae 100644
--- a/libkineto/src/RoctracerActivityApi.cpp
+++ b/libkineto/src/RoctracerActivityApi.cpp
@@ -63,7 +63,8 @@ inline bool inRange(int64_t start, int64_t end, int64_t stamp) {
   return ((stamp > start) && (stamp < end));
 }
 
-inline bool RoctracerActivityApi::isLogged(libkineto::ActivityType atype) {
+inline bool RoctracerActivityApi::isLogged(
+    libkineto::ActivityType atype) const {
   return activityMaskSnapshot_ & (1 << static_cast<uint32_t>(atype));
 }
 
diff --git a/libkineto/src/RoctracerActivityApi.h b/libkineto/src/RoctracerActivityApi.h
index 491f13f7e..b72c1be62 100644
--- a/libkineto/src/RoctracerActivityApi.h
+++ b/libkineto/src/RoctracerActivityApi.h
@@ -7,17 +7,14 @@
  */
 
 #pragma once
+#ifdef HAS_ROCTRACER
 
 #include <atomic>
 #include <functional>
-#include <map>
 #include <set>
-#include <vector>
 
-#ifdef HAS_ROCTRACER
 #include <roctracer.h>
 #include "RoctracerLogger.h"
-#endif
 
 #include "ActivityType.h"
 #include "GenericTraceActivity.h"
@@ -66,9 +63,10 @@ class RoctracerActivityApi {
   // Enabled Activity Filters
   uint32_t activityMask_{0};
   uint32_t activityMaskSnapshot_{0};
-  bool isLogged(libkineto::ActivityType atype);
+  bool isLogged(libkineto::ActivityType atype) const;
 
   RoctracerLogger* d;
 };
 
 } // namespace KINETO_NAMESPACE
+#endif
diff --git a/libkineto/src/output_json.cpp b/libkineto/src/output_json.cpp
index 119c399a4..efcc05c19 100644
--- a/libkineto/src/output_json.cpp
+++ b/libkineto/src/output_json.cpp
@@ -11,9 +11,7 @@
 #include <fmt/format.h>
 #include <time.h>
 #include <fstream>
-#include <map>
 #include "Config.h"
-#include "DeviceProperties.h"
 #include "TraceSpan.h"
 
 #include "Logger.h"
diff --git a/libkineto/src/plugin/xpupti/XpuptiActivityProfiler.cpp b/libkineto/src/plugin/xpupti/XpuptiActivityProfiler.cpp
index e722e49ae..b2a56c6fa 100644
--- a/libkineto/src/plugin/xpupti/XpuptiActivityProfiler.cpp
+++ b/libkineto/src/plugin/xpupti/XpuptiActivityProfiler.cpp
@@ -10,14 +10,21 @@ std::vector<std::array<unsigned char, 16>>
     XpuptiActivityProfilerSession::deviceUUIDs_ = {};
 std::vector<std::string> XpuptiActivityProfilerSession::correlateRuntimeOps_ = {
     "piextUSMEnqueueFill",
+    "urEnqueueUSMFill",
     "piextUSMEnqueueFill2D",
+    "urEnqueueUSMFill2D",
     "piextUSMEnqueueMemcpy",
+    "urEnqueueUSMMemcpy",
     "piextUSMEnqueueMemset",
     "piextUSMEnqueueMemcpy2D",
+    "urEnqueueUSMMemcpy2D",
     "piextUSMEnqueueMemset2D",
     "piEnqueueKernelLaunch",
+    "urEnqueueKernelLaunch",
     "piextEnqueueKernelLaunchCustom",
-    "piextEnqueueCooperativeKernelLaunch"};
+    "urEnqueueKernelLaunchCustomExp",
+    "piextEnqueueCooperativeKernelLaunch",
+    "urEnqueueCooperativeKernelLaunchExp"};
 
 // =========== Session Constructor ============= //
 XpuptiActivityProfilerSession::XpuptiActivityProfilerSession(
diff --git a/libkineto/test/CuptiActivityProfilerTest.cpp b/libkineto/test/CuptiActivityProfilerTest.cpp
index 713a8b83b..4a406401a 100644
--- a/libkineto/test/CuptiActivityProfilerTest.cpp
+++ b/libkineto/test/CuptiActivityProfilerTest.cpp
@@ -219,7 +219,7 @@ class MockCuptiActivities : public CuptiActivityApi {
  public:
   virtual const std::pair<int, size_t> processActivities(
       CuptiActivityBufferMap&, /*unused*/
-      std::function<void(const CUpti_Activity*)> handler) override {
+      const std::function<void(const CUpti_Activity*)>& handler) override {
     for (CUpti_Activity* act : activityBuffer->activities) {
       handler(act);
     }