From 32601ef7a1ce8aaa6d777778435ec499248906fb Mon Sep 17 00:00:00 2001 From: waizuichougou <105572299+waizuichougou@users.noreply.github.com> Date: Sat, 5 Sep 2026 14:39:32 +0800 Subject: [PATCH] [Perf][Multimodal] Avoid duplicate text embedding in Qwen2.5-Omni (#55415) Signed-off-by: waizuichougou <2082431897@qq.com> --- vllm/model_executor/models/qwen2_5_omni_thinker.py | 12 +++++------- 1 file changed, 5 insertions(+), 7 deletions(-) diff --git a/vllm/model_executor/models/qwen2_5_omni_thinker.py b/vllm/model_executor/models/qwen2_5_omni_thinker.py index b09cc504bd..26b69bd6dd 100644 --- a/vllm/model_executor/models/qwen2_5_omni_thinker.py +++ b/vllm/model_executor/models/qwen2_5_omni_thinker.py @@ -1473,14 +1473,12 @@ class Qwen2_5OmniThinkerForConditionalGeneration( if multimodal_embeddings is None or is_multimodal is None: return super().embed_input_ids(input_ids) - inputs_embeds = self._embed_text_input_ids( - input_ids, - self.get_language_model().embed_input_ids, - is_multimodal=is_multimodal, - ) - if len(multimodal_embeddings) == 0: - return inputs_embeds + return super().embed_input_ids( + input_ids, + multimodal_embeddings=multimodal_embeddings, + is_multimodal=is_multimodal, + ) # Check for audio-in-video: interleaved video and audio tokens # in the multimodal region. Only use the interleaved path when