diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index 08dbd614fdcf..d11c8ad19303 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -5638,6 +5638,9 @@ def profile_cudagraph_memory(self) -> int:
         for instance in list(CUDAGraphWrapper._all_instances):
             if id(instance) in original_pools:
                 instance.graph_pool = original_pools[id(instance)]
+        for key_set in self.cudagraph_dispatcher.cudagraph_keys.values():
+            key_set.clear()
+        self.cudagraph_dispatcher.keys_initialized = False
         self.maybe_remove_all_loras(self.lora_config)
         self._cleanup_profiling_kv_cache()
         compilation_counter.num_cudagraph_captured = saved_num_cudagraph_captured