diff --git a/tensorflow/core/profiler/internal/gpu/cupti_collector.cc b/tensorflow/core/profiler/internal/gpu/cupti_collector.cc
index 840b2ee0ada..ab4061a3749 100644
--- a/tensorflow/core/profiler/internal/gpu/cupti_collector.cc
+++ b/tensorflow/core/profiler/internal/gpu/cupti_collector.cc
@@ -224,27 +224,25 @@ struct PerDeviceCollector {
 
     std::vector<Annotation> annotation_stack =
         ParseAnnotationStack(event.annotation);
-    // If multiple metadata have the same key name, show the values from the top
-    // of the stack (innermost annotation). Concatenate the values from
-    // "hlo_op".
-    absl::flat_hash_set<absl::string_view> key_set;
-    std::vector<absl::string_view> hlo_op_names;
-    for (auto annotation = annotation_stack.rbegin();
-         annotation != annotation_stack.rend(); ++annotation) {
-      for (const Annotation::Metadata& metadata : annotation->metadata) {
-        if (metadata.key == "tf_op") {
-          continue;  // ignored, obtained from HLO proto via DebugInfoMap
-        } else if (key_set.insert(metadata.key).second) {
-          xevent.ParseAndAddStatValue(
-              *plane->GetOrCreateStatMetadata(metadata.key), metadata.value);
-        }
-      }
-    }
     if (!annotation_stack.empty()) {
       xevent.AddStatValue(
           *plane->GetOrCreateStatMetadata(GetStatTypeStr(StatType::kTfOp)),
           *plane->GetOrCreateStatMetadata(annotation_stack.begin()->name));
     }
+    // If multiple metadata have the same key name, show the values from the top
+    // of the stack (innermost annotation). Concatenate the values from
+    // "hlo_op".
+    absl::flat_hash_set<absl::string_view> key_set;
+
+    for (auto annotation = annotation_stack.rbegin();
+         annotation != annotation_stack.rend(); ++annotation) {
+      for (const Annotation::Metadata& metadata : annotation->metadata) {
+        if (key_set.insert(metadata.key).second) {
+          xevent.ParseAndAddStatValue(
+              *plane->GetOrCreateStatMetadata(metadata.key), metadata.value);
+        }
+      }
+    }
   }
 
   absl::optional<int> GetDeviceAttribute(CUdevice device,
diff --git a/tensorflow/core/profiler/lib/traceme_encode.h b/tensorflow/core/profiler/lib/traceme_encode.h
index 4dcd6ea469b..1a97d8b0e19 100644
--- a/tensorflow/core/profiler/lib/traceme_encode.h
+++ b/tensorflow/core/profiler/lib/traceme_encode.h
@@ -133,6 +133,12 @@ TF_ATTRIBUTE_ALWAYS_INLINE inline std::string TraceMeOp(
     absl::string_view op_name, absl::string_view op_type) {
   return absl::StrCat(op_name, ":", op_type);
 }
+
+TF_ATTRIBUTE_ALWAYS_INLINE inline std::string TraceMeOp(const char* op_name,
+                                                        const char* op_type) {
+  return absl::StrCat(op_name, ":", op_type);
+}
+
 TF_ATTRIBUTE_ALWAYS_INLINE inline std::string TraceMeOp(
     std::string&& op_name, absl::string_view op_type) {
   absl::StrAppend(&op_name, ":", op_type);