[Perf][Multimodal] Avoid duplicate text embedding in Qwen2.5-Omni (#55415)
Signed-off-by: waizuichougou <2082431897@qq.com>
waizuichougou committed
Sep 5, 2026 at 14:39 UTC
32601ef7a1ce8aaa6d777778435ec499248906fb
1 file changed
+5
-7
vllm/model_executor/models/qwen2_5_omni_thinker.py
+5
-7
@@ -1473,14 +1473,12 @@ class Qwen2_5OmniThinkerForConditionalGeneration(
1473
if multimodal_embeddings is None or is_multimodal is None:
1474
return super().embed_input_ids(input_ids)
1475
1476
- inputs_embeds = self._embed_text_input_ids(
1477
- input_ids,
1478
- self.get_language_model().embed_input_ids,
1479
- is_multimodal=is_multimodal,
1480
- )
1481
-
1476
if len(multimodal_embeddings) == 0:
1483
- return inputs_embeds
1477
+ return super().embed_input_ids(
1478
+ input_ids,
1479
+ multimodal_embeddings=multimodal_embeddings,
1480
+ is_multimodal=is_multimodal,
1481
+ )
1482
1483
# Check for audio-in-video: interleaved video and audio tokens
1484
# in the multimodal region. Only use the interleaved path when