[Minor] Fix duplication of ignored seq group in engine step (vllm-pro…

…ject#1666)

tests/test_regression.py

@@ -0,0 +1,27 @@
+"""Containing tests that check for regressions in vLLM's behavior.
+
+It should include tests that are reported by users and making sure they
+will never happen again.
+
+"""
+from vllm import LLM, SamplingParams
+
+
+def test_duplicated_ignored_sequence_group():
+    """https://github.com/vllm-project/vllm/issues/1655"""
+
+    sampling_params = SamplingParams(temperature=0.01,
+                                     top_p=0.1,
+                                     max_tokens=256)
+    llm = LLM(model="facebook/opt-125m",
+              max_num_batched_tokens=4096,
+              tensor_parallel_size=1)
+    prompts = ["This is a short prompt", "This is a very long prompt " * 1000]
+    outputs = llm.generate(prompts, sampling_params=sampling_params)
+
+    assert len(prompts) == len(outputs)
+
+
+if __name__ == "__main__":
+    import pytest
+    pytest.main([__file__])

vllm/engine/llm_engine.py

@@ -567,7 +567,7 @@ def step(self) -> List[RequestOutput]:
             blocks_to_copy=scheduler_outputs.blocks_to_copy,
         )
 
-        return self._process_model_outputs(output, scheduler_outputs) + ignored
+        return self._process_model_outputs(output, scheduler_outputs)
 
     def _log_system_stats(
         self,