Squash 6587

vllm/engine/async_llm_engine.py

@@ -130,7 +130,10 @@ def process_request_output(self,
         """Process a request output from the engine."""
         request_id = request_output.request_id
 
-        self._request_streams[request_id].put(request_output)
+        # Avoid a KeyError which can occur if the request was aborted while the
+        # outputs were generated
+        if request_id in self._request_streams:
+            self._request_streams[request_id].put(request_output)
         if request_output.finished:
             if verbose:
                 logger.info("Finished request %s.", request_id)

vllm/engine/output_processor/single_step.py

@@ -90,7 +90,10 @@ def _process_sequence_group_outputs(self, seq_group: SequenceGroup,
             for parent_seq in parent_seqs
         }
         for sample in samples:
-            parent_child_dict[sample.parent_seq_id].append(sample)
+            # Avoid a KeyError which can occur if the request was aborted while
+            # the outputs were generated
+            if sample.parent_seq_id in parent_child_dict:
+                parent_child_dict[sample.parent_seq_id].append(sample)
         # List of (child, parent)
         child_seqs: List[Tuple[Sequence, Sequence]] = []