[Bugfix] Standardize merging multimodal embeddings (#26771)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/ernie45_vl.py

@@ -1645,12 +1645,12 @@ class Ernie4_5_VLMoeForConditionalGeneration(
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                vision_embeddings = self._process_image_input(image_input)
+                image_embeddings = self._process_image_input(image_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
                 video_embeddings = self._process_video_input(video_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 

vllm/model_executor/models/glm4_1v.py

@@ -1608,11 +1608,11 @@ class Glm4vForConditionalGeneration(
         for modality in mm_input_by_modality:
             multimodal_input = mm_input_by_modality[modality]
             if modality == "image":
-                vision_embeddings = self._process_image_input(multimodal_input)
+                image_embeddings = self._process_image_input(multimodal_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "video":
                 video_embeddings = self._process_video_input(multimodal_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
         return multimodal_embeddings
 
     def forward(

vllm/model_executor/models/hyperclovax_vision.py

@@ -749,12 +749,12 @@ class HCXVisionForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                vision_embeddings = self._process_image_input(image_input)
+                image_embeddings = self._process_image_input(image_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
                 video_embeddings = self._process_video_input(video_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 

vllm/model_executor/models/interns1.py

@@ -753,12 +753,12 @@ class InternS1ForConditionalGeneration(
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                vision_embeddings = self._process_vision_input(image_input)
+                image_embeddings = self._process_vision_input(image_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
                 video_embeddings = self._process_vision_input(video_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 

vllm/model_executor/models/internvl.py

@@ -1358,12 +1358,12 @@ class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA)
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                vision_embeddings = self._process_vision_input(image_input)
+                image_embeddings = self._process_vision_input(image_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
                 video_embeddings = self._process_vision_input(video_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 

vllm/model_executor/models/keye.py

@@ -1459,12 +1459,12 @@ class BaseKeyeModule(nn.Module):
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                vision_embeddings = self._process_image_input(image_input)
+                image_embeddings = self._process_image_input(image_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
                 video_embeddings = self._process_video_input(video_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
         return multimodal_embeddings
 
     def forward(

vllm/model_executor/models/llava_onevision.py

@@ -881,8 +881,8 @@ class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
         for modality in mm_input_by_modality:
             multimodal_input = mm_input_by_modality[modality]
             if modality == "image":
-                vision_embeddings = self._process_image_input(multimodal_input)
+                image_embeddings = self._process_image_input(multimodal_input)
-                multimodal_embeddings += tuple(vision_embeddings)
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "video":
                 video_embeddings = self._process_video_pixels(multimodal_input)
                 multimodal_embeddings += tuple(video_embeddings)

vllm/model_executor/models/minicpmo.py

@@ -762,7 +762,7 @@ class MiniCPMO(MiniCPMV2_6):
         for modality in modalities:
             if modality == "audios":
                 audio_input = modalities["audios"]
-                audio_features = self._process_audio_input(audio_input)
+                audio_embeddings = self._process_audio_input(audio_input)
-                multimodal_embeddings += tuple(audio_features)
+                multimodal_embeddings += tuple(audio_embeddings)
 
         return multimodal_embeddings

vllm/model_executor/models/minicpmv.py

@@ -1129,12 +1129,12 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                image_features = self._process_vision_input(image_input)
+                image_embeddings = self._process_vision_input(image_input)
-                multimodal_embeddings += tuple(image_features)
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
-                video_features = self._process_vision_input(video_input)
+                video_embeddings = self._process_vision_input(video_input)
-                multimodal_embeddings += tuple(video_features)
+                multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 

vllm/model_executor/models/nano_nemotron_vl.py

@@ -1263,12 +1263,12 @@ class NemotronH_Nano_VL_V2(
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                vision_embeddings = self._process_image_input(image_input)
+                image_embeddings = self._process_image_input(image_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
                 video_embeddings = self._process_video_input(video_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 

vllm/model_executor/models/nemotron_vl.py

@@ -575,8 +575,8 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, Suppor
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                vision_embeddings = self._process_image_input(image_input)
+                image_embeddings = self._process_image_input(image_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
 
         return multimodal_embeddings
 

vllm/model_executor/models/ovis2_5.py

@@ -616,12 +616,12 @@ class Ovis2_5(nn.Module, SupportsMultiModal, SupportsPP):
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                vision_embeddings = self._process_visual_input(image_input)
+                image_embeddings = self._process_visual_input(image_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
                 video_embeddings = self._process_visual_input(video_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 

vllm/model_executor/models/phi4_multimodal.py

@@ -1430,8 +1430,8 @@ class Phi4MultimodalForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
             if modality == "images":
                 audio_projection_mode = "vision"
                 image_input = modalities["images"]
-                vision_embeddings = self._process_image_input(image_input)
+                image_embeddings = self._process_image_input(image_input)
-                multimodal_embeddings += tuple(vision_embeddings)
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "audios":
                 audio_input = modalities["audios"]
                 audio_embeddings = self._process_audio_input(

vllm/model_executor/models/phi4mm.py

@@ -1248,8 +1248,8 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
             if modality == "images":
                 audio_projection_mode = "vision"
                 image_input = modalities["images"]
-                vision_embeddings = self._process_image_input(image_input)
+                image_embeddings = self._process_image_input(image_input)
-                multimodal_embeddings += tuple(vision_embeddings)
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "audios":
                 audio_input = modalities["audios"]
                 audio_embeddings = self._process_audio_input(

vllm/model_executor/models/qwen2_5_omni_thinker.py

@@ -1210,14 +1210,14 @@ class Qwen2_5OmniThinkerForConditionalGeneration(
         for modality in mm_input_by_modality:
             multimodal_input = mm_input_by_modality[modality]
             if modality == "image":
-                vision_embeddings = self._process_image_input(multimodal_input)
+                image_embeddings = self._process_image_input(multimodal_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "video":
                 video_embeddings = self._process_video_input(multimodal_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
             if modality == "audio":
                 audio_embeddings = self._process_audio_input(multimodal_input)
-                multimodal_embeddings += audio_embeddings
+                multimodal_embeddings += tuple(audio_embeddings)
         return multimodal_embeddings
 
     # TODO (ywang96): support overlapping modality embeddings so that

vllm/model_executor/models/qwen2_5_vl.py

@@ -1586,19 +1586,19 @@ class Qwen2_5_VLForConditionalGeneration(
         for modality in mm_input_by_modality:
             multimodal_input = mm_input_by_modality[modality]
             if modality == "image":
-                vision_embeddings = self._process_image_input(multimodal_input)
+                image_embeddings = self._process_image_input(multimodal_input)
                 if self.is_multimodal_pruning_enabled:
-                    vision_embeddings = self._postprocess_image_embeds_evs(
+                    image_embeddings = self._postprocess_image_embeds_evs(
-                        vision_embeddings, multimodal_input
+                        image_embeddings, multimodal_input
                     )
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "video":
                 video_embeddings = self._process_video_input(multimodal_input)
                 if self.is_multimodal_pruning_enabled:
                     video_embeddings = self._postprocess_video_embeds_evs(
                         video_embeddings, multimodal_input
                     )
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
         return multimodal_embeddings
 
     def forward(

vllm/model_executor/models/qwen2_vl.py

@@ -1561,12 +1561,12 @@ class Qwen2VLForConditionalGeneration(
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                vision_embeddings = self._process_image_input(image_input)
+                image_embeddings = self._process_image_input(image_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
                 video_embeddings = self._process_video_input(video_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 

vllm/model_executor/models/qwen3_omni_moe_thinker.py

@@ -1260,14 +1260,14 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
         for modality in mm_input_by_modality:
             multimodal_input = mm_input_by_modality[modality]
             if modality == "image":
-                vision_embeddings = self._process_image_input(multimodal_input)
+                image_embeddings = self._process_image_input(multimodal_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "video":
                 video_embeddings = self._process_video_input(multimodal_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
             if modality == "audio":
                 audio_embeddings = self._process_audio_input(multimodal_input)
-                multimodal_embeddings += audio_embeddings
+                multimodal_embeddings += tuple(audio_embeddings)
         return multimodal_embeddings
 
     def get_input_embeddings(

vllm/model_executor/models/qwen3_vl.py

@@ -1601,11 +1601,11 @@ class Qwen3VLForConditionalGeneration(
         for modality in mm_input_by_modality:
             multimodal_input = mm_input_by_modality[modality]
             if modality == "image":
-                vision_embeddings = self._process_image_input(multimodal_input)
+                image_embeddings = self._process_image_input(multimodal_input)
-                multimodal_embeddings += vision_embeddings
+                multimodal_embeddings += tuple(image_embeddings)
             if modality == "video":
                 video_embeddings = self._process_video_input(multimodal_input)
-                multimodal_embeddings += video_embeddings
+                multimodal_embeddings += tuple(video_embeddings)
         return multimodal_embeddings
 
     def _compute_deepstack_embeds(