diff --git a/vllm/v1/worker/gpu/model_runner.py b/vllm/v1/worker/gpu/model_runner.py
index 5b067c87e0..7a82032313 100644
--- a/vllm/v1/worker/gpu/model_runner.py
+++ b/vllm/v1/worker/gpu/model_runner.py
@@ -1555,6 +1555,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         if hasattr(self, "kv_cache_config"):
             del self.kv_cache_config
         free_before_shutdown(self.vllm_config)
+        if hasattr(self, "model_state"):
+            del self.model_state
+        if getattr(self, "speculator", None) is not None:
+            self.speculator = None
         if hasattr(self, "model"):
             del self.model