diff --git a/unified-runtime/source/adapters/level_zero/helpers/kernel_helpers.hpp b/unified-runtime/source/adapters/level_zero/helpers/kernel_helpers.hpp
index 49345bb57e159..5dcf0c9123045 100644
--- a/unified-runtime/source/adapters/level_zero/helpers/kernel_helpers.hpp
+++ b/unified-runtime/source/adapters/level_zero/helpers/kernel_helpers.hpp
@@ -56,3 +56,18 @@ ur_result_t getSuggestedLocalWorkSize(ur_device_handle_t hDevice,
                                       ze_kernel_handle_t hZeKernel,
                                       size_t GlobalWorkSize3D[3],
                                       uint32_t SuggestedLocalWorkSize3D[3]);
+
+/**
+ * Handle uncommon conditions after kernel submission.
+ * Resets the offset to {0, 0, 0} if one was supplied.
+ * @param[in] hZeKernel The kernel handle.
+ * @param[in] pGlobalWorkOffset Pointer to offset array.
+ */
+inline void postSubmit(ze_kernel_handle_t hZeKernel,
+                       const size_t *pGlobalWorkOffset) {
+  // If this kernel was launched with an offset, clear it for the next launch.
+  // This slows down kernels with offsets but keeps the common case fast.
+  if (pGlobalWorkOffset != NULL) {
+    zeKernelSetGlobalOffsetExp(hZeKernel, 0, 0, 0);
+  }
+}
diff --git a/unified-runtime/source/adapters/level_zero/v2/command_list_manager.cpp b/unified-runtime/source/adapters/level_zero/v2/command_list_manager.cpp
index 397aa022bcee7..d6f865d80b5c3 100644
--- a/unified-runtime/source/adapters/level_zero/v2/command_list_manager.cpp
+++ b/unified-runtime/source/adapters/level_zero/v2/command_list_manager.cpp
@@ -219,6 +219,16 @@ ur_result_t ur_command_list_manager::appendKernelLaunch(
     waitListView.clear();
   };
 
+  // If the offset is {0, 0, 0}, pass NULL instead.
+  // This allows us to skip setting the offset.
+  bool hasOffset = false;
+  for (uint32_t i = 0; i < workDim; ++i) {
+    hasOffset |= pGlobalWorkOffset[i];
+  }
+  if (!hasOffset) {
+    pGlobalWorkOffset = NULL;
+  }
+
   UR_CALL(hKernel->prepareForSubmission(context, device, pGlobalWorkOffset,
                                         workDim, WG[0], WG[1], WG[2],
                                         memoryMigrate));
@@ -229,6 +239,8 @@ ur_result_t ur_command_list_manager::appendKernelLaunch(
              (zeCommandList.get(), hZeKernel, &zeThreadGroupDimensions,
               zeSignalEvent, waitListView.num, waitListView.handles));
 
+  postSubmit(hZeKernel, pGlobalWorkOffset);
+
   return UR_RESULT_SUCCESS;
 }
 
diff --git a/unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp b/unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp
index 0c98147b53582..286623277dc47 100644
--- a/unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp
+++ b/unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp
@@ -809,6 +809,16 @@ ur_result_t ur_queue_immediate_in_order_t::enqueueCooperativeKernelLaunchExp(
     waitListView.clear();
   };
 
+  // If the offset is {0, 0, 0}, pass NULL instead.
+  // This allows us to skip setting the offset.
+  bool hasOffset = false;
+  for (uint32_t i = 0; i < workDim; ++i) {
+    hasOffset |= pGlobalWorkOffset[i];
+  }
+  if (!hasOffset) {
+    pGlobalWorkOffset = NULL;
+  }
+
   UR_CALL(hKernel->prepareForSubmission(hContext, hDevice, pGlobalWorkOffset,
                                         workDim, WG[0], WG[1], WG[2],
                                         memoryMigrate));
@@ -822,6 +832,8 @@ ur_result_t ur_queue_immediate_in_order_t::enqueueCooperativeKernelLaunchExp(
 
   recordSubmittedKernel(hKernel);
 
+  postSubmit(hZeKernel, pGlobalWorkOffset);
+
   return UR_RESULT_SUCCESS;
 }