diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index c3536113e9bef..a04b89d296113 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -942,13 +942,20 @@ void AMDGPUTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
                 *this, Opt, ThinOrFullLTOPhase::FullLTOPostLink));
           }
         }
-        if (!NoKernelInfoEndLTO) {
-          FunctionPassManager FPM;
-          FPM.addPass(KernelInfoPrinter(this));
-          PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
-        }
       });
 
+  // Add kernel-info pass using OptimizerLastEPCallback to run during LTO
+  // while remark streamer is still active
+  if (!NoKernelInfoEndLTO) {
+    PB.registerOptimizerLastEPCallback([this](ModulePassManager &PM,
+                                              OptimizationLevel Level,
+                                              ThinOrFullLTOPhase Phase) {
+      FunctionPassManager FPM;
+      FPM.addPass(KernelInfoPrinter(this));
+      PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
+    });
+  }
+
   PB.registerRegClassFilterParsingCallback(
       [](StringRef FilterName) -> RegAllocFilterFunc {
         if (FilterName == "sgpr")
diff --git a/llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp b/llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp
index 0603994606d71..5f345369766cd 100644
--- a/llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp
@@ -261,12 +261,13 @@ void NVPTXTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
       });
 
   if (!NoKernelInfoEndLTO) {
-    PB.registerFullLinkTimeOptimizationLastEPCallback(
-        [this](ModulePassManager &PM, OptimizationLevel Level) {
-          FunctionPassManager FPM;
-          FPM.addPass(KernelInfoPrinter(this));
-          PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
-        });
+    PB.registerOptimizerLastEPCallback([this](ModulePassManager &PM,
+                                              OptimizationLevel Level,
+                                              ThinOrFullLTOPhase Phase) {
+      FunctionPassManager FPM;
+      FPM.addPass(KernelInfoPrinter(this));
+      PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
+    });
   }
 }