Index: clang/lib/Sema/SemaOverload.cpp
===================================================================
--- clang/lib/Sema/SemaOverload.cpp
+++ clang/lib/Sema/SemaOverload.cpp
@@ -12750,6 +12750,13 @@
   DeclAccessPair DAP;
   SmallVector<FunctionDecl *, 2> AmbiguousDecls;
 
+  // Return positive for better, negative for worse, 0 for equal preference.
+  auto CheckCUDAPreference = [&](FunctionDecl *FD1, FunctionDecl *FD2) {
+    FunctionDecl *Caller = getCurFunctionDecl(/*AllowLambda=*/true);
+    return static_cast<int>(IdentifyCUDAPreference(Caller, FD1)) -
+           static_cast<int>(IdentifyCUDAPreference(Caller, FD2));
+  };
+
   auto CheckMoreConstrained = [&](FunctionDecl *FD1,
                                   FunctionDecl *FD2) -> std::optional<bool> {
     if (FunctionDecl *MF = FD1->getInstantiatedFromMemberFunction())
@@ -12780,9 +12787,31 @@
     if (!checkAddressOfFunctionIsAvailable(FD))
       continue;
 
+    // If we found a better result, update Result.
+    auto FoundBetter = [&]() {
+      IsResultAmbiguous = false;
+      DAP = I.getPair();
+      Result = FD;
+    };
+
     // We have more than one result - see if it is more constrained than the
     // previous one.
     if (Result) {
+      // Check CUDA preference first. If the candidates have differennt CUDA
+      // preference, choose the one with higher CUDA preference. Otherwise,
+      // choose the one with more constraints.
+      if (getLangOpts().CUDA) {
+        int PreferenceByCUDA = CheckCUDAPreference(FD, Result);
+        // FD has different preference than Result.
+        if (PreferenceByCUDA != 0) {
+          // FD is more preferable than Result.
+          if (PreferenceByCUDA > 0)
+            FoundBetter();
+          continue;
+        }
+      }
+      // FD has the same CUDA prefernece than Result. Continue check
+      // constraints.
       std::optional<bool> MoreConstrainedThanPrevious =
           CheckMoreConstrained(FD, Result);
       if (!MoreConstrainedThanPrevious) {
@@ -12794,9 +12823,7 @@
         continue;
       // FD is more constrained - replace Result with it.
     }
-    IsResultAmbiguous = false;
-    DAP = I.getPair();
-    Result = FD;
+    FoundBetter();
   }
 
   if (IsResultAmbiguous)
@@ -12806,9 +12833,15 @@
     SmallVector<const Expr *, 1> ResultAC;
     // We skipped over some ambiguous declarations which might be ambiguous with
     // the selected result.
-    for (FunctionDecl *Skipped : AmbiguousDecls)
+    for (FunctionDecl *Skipped : AmbiguousDecls) {
+      // If skipped candidate has different CUDA preference than the result,
+      // there is no ambiguity. Otherwise check whether they have different
+      // constraints.
+      if (getLangOpts().CUDA && CheckCUDAPreference(Skipped, Result) != 0)
+        continue;
       if (!CheckMoreConstrained(Skipped, Result))
         return nullptr;
+    }
     Pair = DAP;
   }
   return Result;
Index: clang/test/SemaCUDA/template-arg-deduction.cu
===================================================================
--- /dev/null
+++ clang/test/SemaCUDA/template-arg-deduction.cu
@@ -0,0 +1,27 @@
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fsyntax-only -verify %s
+// RUN: %clang_cc1 -triple nvptx64-nvidia-cuda -fsyntax-only -fcuda-is-device -verify %s
+
+// expected-no-diagnostics
+
+#include "Inputs/cuda.h"
+
+void foo();
+__device__ void foo();
+
+template<class F>
+void host_temp(F f);
+
+template<class F>
+__device__ void device_temp(F f);
+
+void host_caller() {
+  host_temp(foo);
+}
+
+__global__ void kernel_caller() {
+  device_temp(foo);
+}
+
+__device__ void device_caller() {
+  device_temp(foo);
+}