[Bugfix] Fix a bug in RequestOutput.finished (vllm-project#202)

Author: WoosukKwon

examples/llm_engine_example.py

@@ -30,7 +30,7 @@ def main(args: argparse.Namespace):
 
         request_outputs = engine.step()
         for request_output in request_outputs:
-            if request_output.finished():
+            if request_output.finished:
                 print(request_output)
 
         if not (engine.has_unfinished_requests() or test_prompts):

vllm/engine/async_llm_engine.py

@@ -154,7 +154,7 @@ async def generate(
             yield request_output
 
             # Once finished, release the resources of the sequence group.
-            if request_output.finished():
+            if request_output.finished:
                 if self.log_requests:
                     logger.info(f"Finished request {request_id}.")
 

vllm/entrypoints/llm.py

@@ -133,7 +133,7 @@ def _run_engine(self, use_tqdm: bool) -> List[RequestOutput]:
         while self.llm_engine.has_unfinished_requests():
             step_outputs = self.llm_engine.step()
             for output in step_outputs:
-                if output.finished():
+                if output.finished:
                     outputs.append(output)
                     if use_tqdm:
                         pbar.update(1)

vllm/outputs.py

@@ -60,11 +60,13 @@ def __init__(
         prompt: str,
         prompt_token_ids: List[int],
         outputs: List[CompletionOutput],
+        finished: bool,
     ) -> None:
         self.request_id = request_id
         self.prompt = prompt
         self.prompt_token_ids = prompt_token_ids
         self.outputs = outputs
+        self.finished = finished
 
     @classmethod
     def from_seq_group(cls, seq_group: SequenceGroup) -> "RequestOutput":
@@ -95,13 +97,13 @@ def from_seq_group(cls, seq_group: SequenceGroup) -> "RequestOutput":
         # Every sequence in the sequence group should have the same prompt.
         prompt = top_n_seqs[0].prompt
         prompt_token_ids = top_n_seqs[0].data.prompt_token_ids
-        return cls(seq_group.request_id, prompt, prompt_token_ids, outputs)
+        finished = seq_group.is_finished()
+        return cls(seq_group.request_id, prompt, prompt_token_ids, outputs,
+                   finished)
 
     def __repr__(self) -> str:
         return (f"RequestOutput(request_id={self.request_id}, "
                 f"prompt={self.prompt!r}, "
                 f"prompt_token_ids={self.prompt_token_ids}, "
-                f"outputs={self.outputs})")
-
-    def finished(self) -> bool:
-        return all(output.finished() for output in self.outputs)
+                f"outputs={self.outputs}, "
+                f"finished={self.finished})")