fix

Author: Your Name

vllm/beam/filtering.py

@@ -58,6 +58,9 @@ async def get_n_valid_beams(self, create_completion: Callable,
         res = await asyncio.gather(*tasks)
         request.n = n
         beam_validator_res = self.validate(res)
+        if isinstance(beam_validator_res, ErrorResponse):
+            return beam_validator_res
+        
         filtered_res = [r for r, valid in zip(res, beam_validator_res) if valid]
         logger.debug("Filtered count: %d", len(filtered_res))
         if len(filtered_res) == 0:
@@ -68,6 +71,7 @@ async def get_n_valid_beams(self, create_completion: Callable,
     def validate(self, responses: list[AsyncGenerator],
                  debug_infos_G: list[BeamDebugInfo] = None):
         error_responses = [r for r in responses if isinstance(r, ErrorResponse)]
+        print(f"error_responses: {error_responses}")
         if len(error_responses) > 0:
             combined_message = "; ".join(er.message for er in error_responses)
             return ErrorResponse(

vllm/entrypoints/openai/serving_completion.py

@@ -106,6 +106,9 @@ async def _process_prefix(request: CompletionRequest):
             return res
 
         res = await _process_prefix(request)
+        if isinstance(res, ErrorResponse):
+            return res
+        
         input_str_len = len(res.choices[0].text)
 
         async def _should_stop(final):
@@ -121,6 +124,10 @@ async def _chunk_generator():
             while num_chunks < max_chunks and not should_stop:
                 num_chunks += 1
                 beams = await self.beam_validator.get_n_valid_beams(create_completion=self.create_completion, request=request, raw_request=raw_request)
+                if isinstance(beams, ErrorResponse):
+                    yield f"data: {beams.model_dump_json()}\n\n"
+                    break
+            
                 final = await self.beam_scorer.pick_best_beam(beams)
                 request.prompt = final.choices[0].text
                 should_stop = await _should_stop(final)