diff --git a/trl/trainer/grpo_trainer.py b/trl/trainer/grpo_trainer.py
index bafd348b33e..5dd7571acd0 100644
--- a/trl/trainer/grpo_trainer.py
+++ b/trl/trainer/grpo_trainer.py
@@ -1424,7 +1424,7 @@ async def _run_async_tools(async_coros):
                 **self.chat_template_kwargs,
             )
             if self.use_vllm and self.vllm_mode == "colocate":
-                max_model_len = self.llm.llm_engine.model_config.max_model_len
+                max_model_len = self.vllm_generation.llm.llm_engine.model_config.max_model_len
             elif not self.use_vllm:
                 max_model_len = self.model.config.max_position_embeddings
             else: