Fix hanging when prompt exceeds limit (#1029)

vllm/core/scheduler.py

@@ -175,7 +175,7 @@ class Scheduler:
                 num_curr_seqs += num_new_seqs
                 scheduled.append(seq_group)
 
-            if scheduled:
+            if scheduled or ignored_seq_groups:
                 scheduler_outputs = SchedulerOutputs(
                     scheduled_seq_groups=scheduled,
                     prompt_run=True,

vllm/engine/llm_engine.py

@@ -294,14 +294,12 @@ class LLMEngine:
     def _schedule(
         self
     ) -> Tuple[List[SequenceGroupMetadata], SchedulerOutputs,
-               Optional[List[RequestOutput]]]:
+               List[RequestOutput]]:
         seq_group_metadata_list, scheduler_outputs = self.scheduler.schedule()
-        if scheduler_outputs.is_empty():
-            return seq_group_metadata_list, scheduler_outputs, [
-                RequestOutput.from_seq_group(seq_group)
-                for seq_group in scheduler_outputs.ignored_seq_groups
-            ]
-        return seq_group_metadata_list, scheduler_outputs, None
+        return seq_group_metadata_list, scheduler_outputs, [
+            RequestOutput.from_seq_group(seq_group)
+            for seq_group in scheduler_outputs.ignored_seq_groups
+        ]
 
     def _check_beam_search_early_stopping(
         self,
@@ -545,10 +543,9 @@ class LLMEngine:
         and updates the scheduler with the model outputs. Finally, it decodes
         the sequences and returns the newly generated results.
         """
-        (seq_group_metadata_list, scheduler_outputs,
-         early_return) = self._schedule()
-        if early_return is not None:
-            return early_return
+        seq_group_metadata_list, scheduler_outputs, ignored = self._schedule()
+        if scheduler_outputs.is_empty():
+            return ignored
 
         # Execute the model.
         output = self._run_workers(
@@ -559,7 +556,7 @@ class LLMEngine:
             blocks_to_copy=scheduler_outputs.blocks_to_copy,
         )
 
-        return self._process_model_outputs(output, scheduler_outputs)
+        return self._process_model_outputs(output, scheduler_outputs) + ignored
 
     def _log_system_stats(
         self,