fixes

Author: Your Name

vllm/beam/beam.py

@@ -32,7 +32,7 @@ async def pick_best_beam(self, responses: list[
             ranking_scores = self.ranking_computer.compute(
                 heads_tensor, debug_info
             )
-            scores *= ranking_scores
+            scores += ranking_scores
 
         for i in range(len(responses)):
             debug_info[i].final_score = scores[i]

vllm/entrypoints/openai/serving_completion.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import asyncio
+import math
 import time
 from collections.abc import AsyncGenerator, AsyncIterator
 from collections.abc import Sequence as GenericSequence
@@ -13,7 +14,7 @@
 from typing_extensions import assert_never
 
 from vllm.beam.beam import BeamScorer
-from vllm.beam.filtering import BeamValidator
+from vllm.beam.filtering import _CHUNK_SIZE, BeamValidator
 from vllm.beam.metrics import report_metrics
 from vllm.beam.penalty import MEOW_CLASSI_IDX, PenaltyComputer
 from vllm.config import ModelConfig
@@ -102,13 +103,14 @@ async def _process_prefix(request: CompletionRequest):
         async def _should_stop(final):
             return final.choices[0].finish_reason == "stop" or final.choices[0].is_filtered
 
+        max_chunks = math.ceil(request.max_tokens / _CHUNK_SIZE)
         async def _chunk_generator():
             num_chunks = 0
             should_stop = False
             output = None
 
             # TODO(@tanuj): calc created tokens
-            while num_chunks < 4 and not should_stop:
+            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)
                 final = await self.beam_scorer.pick_best_beam(beams)