[Frontend]Flatten beam-search beams with itertools.chain instead of sum (#48829)
Signed-off-by: Wang Xingda <wangxingda1993@126.com> Co-authored-by: 王兴达 <wangxingda@360itdeMacBook-Pro.local> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Wang Xingda committed
Jul 18, 2026 at 06:09 UTC
fae543015cd49dfb4db3afb3a2e96d2eccb3e1db
1 file changed
+3
-1
vllm/entrypoints/generate/beam_search/offline.py
+3
-1
@@ -207,7 +207,9 @@ class BeamSearchOfflineMixin(OfflineInferenceMixin):
207
Returns True if all beams are exhausted and search should stop.
208
"""
209
all_beams: list[BeamSearchSequence] = list(
210
- sum((instance.beams for instance in instances_batch), [])
210
+ itertools.chain.from_iterable(
211
+ instance.beams for instance in instances_batch
212
+ )
213
)
214
pos = [0] + list(
215
itertools.accumulate(len(instance.beams) for instance in instances_batch)