diff --git a/source/adapters/cuda/enqueue.cpp b/source/adapters/cuda/enqueue.cpp
index 906fd49d1d..1c074025a9 100644
--- a/source/adapters/cuda/enqueue.cpp
+++ b/source/adapters/cuda/enqueue.cpp
@@ -530,6 +530,21 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunchCustomExp(
   }
 
   std::vector<CUlaunchAttribute> launch_attribute(numPropsInLaunchPropList);
+
+  // Early exit for zero size kernel
+  if (*pGlobalWorkSize == 0) {
+    return urEnqueueEventsWaitWithBarrier(hQueue, numEventsInWaitList,
+                                          phEventWaitList, phEvent);
+  }
+
+  // Set the number of threads per block to the number of threads per warp
+  // by default unless user has provided a better number
+  size_t ThreadsPerBlock[3] = {32u, 1u, 1u};
+  size_t BlocksPerGrid[3] = {1u, 1u, 1u};
+
+  uint32_t LocalSize = hKernel->getLocalSize();
+  CUfunction CuFunc = hKernel->get();
+
   for (uint32_t i = 0; i < numPropsInLaunchPropList; i++) {
     switch (launchPropList[i].id) {
     case UR_EXP_LAUNCH_PROPERTY_ID_IGNORE: {
@@ -540,12 +555,32 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunchCustomExp(
 
       launch_attribute[i].id = CU_LAUNCH_ATTRIBUTE_CLUSTER_DIMENSION;
       // Note that cuda orders from right to left wrt SYCL dimensional order.
-      launch_attribute[i].value.clusterDim.x =
-          launchPropList[i].value.clusterDim[2];
-      launch_attribute[i].value.clusterDim.y =
-          launchPropList[i].value.clusterDim[1];
-      launch_attribute[i].value.clusterDim.z =
-          launchPropList[i].value.clusterDim[0];
+      if (workDim == 3) {
+        launch_attribute[i].value.clusterDim.x =
+            launchPropList[i].value.clusterDim[2];
+        launch_attribute[i].value.clusterDim.y =
+            launchPropList[i].value.clusterDim[1];
+        launch_attribute[i].value.clusterDim.z =
+            launchPropList[i].value.clusterDim[0];
+      } else if (workDim == 2) {
+        launch_attribute[i].value.clusterDim.x =
+            launchPropList[i].value.clusterDim[1];
+        launch_attribute[i].value.clusterDim.y =
+            launchPropList[i].value.clusterDim[0];
+        launch_attribute[i].value.clusterDim.z =
+            launchPropList[i].value.clusterDim[2];
+      } else {
+        launch_attribute[i].value.clusterDim.x =
+            launchPropList[i].value.clusterDim[0];
+        launch_attribute[i].value.clusterDim.y =
+            launchPropList[i].value.clusterDim[1];
+        launch_attribute[i].value.clusterDim.z =
+            launchPropList[i].value.clusterDim[2];
+      }
+
+      UR_CHECK_ERROR(cuFuncSetAttribute(
+          CuFunc, CU_FUNC_ATTRIBUTE_NON_PORTABLE_CLUSTER_SIZE_ALLOWED, 1));
+
       break;
     }
     case UR_EXP_LAUNCH_PROPERTY_ID_COOPERATIVE: {
@@ -560,20 +595,6 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunchCustomExp(
     }
   }
 
-  // Early exit for zero size kernel
-  if (*pGlobalWorkSize == 0) {
-    return urEnqueueEventsWaitWithBarrier(hQueue, numEventsInWaitList,
-                                          phEventWaitList, phEvent);
-  }
-
-  // Set the number of threads per block to the number of threads per warp
-  // by default unless user has provided a better number
-  size_t ThreadsPerBlock[3] = {32u, 1u, 1u};
-  size_t BlocksPerGrid[3] = {1u, 1u, 1u};
-
-  uint32_t LocalSize = hKernel->getLocalSize();
-  CUfunction CuFunc = hKernel->get();
-
   // This might return UR_RESULT_ERROR_ADAPTER_SPECIFIC, which cannot be handled
   // using the standard UR_CHECK_ERROR
   if (ur_result_t Ret =
diff --git a/test/conformance/exp_launch_properties/launch_properties.cpp b/test/conformance/exp_launch_properties/launch_properties.cpp
index bc252392eb..4cc0dcfe22 100644
--- a/test/conformance/exp_launch_properties/launch_properties.cpp
+++ b/test/conformance/exp_launch_properties/launch_properties.cpp
@@ -78,7 +78,7 @@ TEST_P(urEnqueueKernelLaunchCustomTest, Success) {
         if (compute_capability >= 9.0) {
             ur_exp_launch_property_t cluster_dims_prop;
             cluster_dims_prop.id = UR_EXP_LAUNCH_PROPERTY_ID_CLUSTER_DIMENSION;
-            cluster_dims_prop.value.clusterDim[0] = 1;
+            cluster_dims_prop.value.clusterDim[0] = 16;
             cluster_dims_prop.value.clusterDim[1] = 1;
             cluster_dims_prop.value.clusterDim[2] = 1;