[Bugfix][LoRA] Use an explicit capability flag for tower connector LoRA (#53092)

Signed-off-by: linitra24 <Joy25810@foxmail.com>

vllm/lora/model_manager.py

@@ -201,8 +201,8 @@ class LoRAModelManager:
         self.punica_wrapper_mapping[lm_prefix] = llm_punica_wrapper
 
         # First, determine if the model supports tower connector LoRA.
-        self.supports_tower_connector_lora = self.supports_mm and hasattr(
-            self.model, "get_num_mm_encoder_tokens"
+        self.supports_tower_connector_lora = (
+            self.supports_mm and self.model.supports_tower_connector_lora
         )
 
         # Then, handle the case where the feature is disabled in the config.

vllm/model_executor/models/blip2.py

@@ -525,6 +525,8 @@ class Blip2MultiModalProcessor(BaseMultiModalProcessor[Blip2ProcessingInfo]):
 class Blip2ForConditionalGeneration(
     nn.Module, SupportsLoRA, SupportsMultiModal, SupportsPP, SupportsQuant
 ):
+    supports_tower_connector_lora = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/dots_ocr.py

@@ -618,6 +618,7 @@ class DotsOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
         "fc13": ["fc1", "fc3"],
     }
     supports_encoder_tp_data = True
+    supports_tower_connector_lora = True
 
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:

vllm/model_executor/models/gemma3_mm.py

@@ -524,6 +524,8 @@ class Gemma3ForConditionalGeneration(
         }
     )
 
+    supports_tower_connector_lora = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/gemma4_mm.py

@@ -978,6 +978,7 @@ class Gemma4ForConditionalGeneration(
     # window instead of dropping them (which would make image attention
     # causal-only for images larger than the sliding window).
     mm_prefix_clamp_sliding_window: bool = True
+    supports_tower_connector_lora = True
 
     packed_modules_mapping = {
         "qkv_proj": [

vllm/model_executor/models/glm4_1v.py

@@ -1755,6 +1755,7 @@ class Glm4vForConditionalGeneration(
     )
 
     supports_encoder_tp_data = True
+    supports_tower_connector_lora = True
 
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:

vllm/model_executor/models/idefics3.py

@@ -503,6 +503,8 @@ class Idefics3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsLo
         ],
     }
 
+    supports_tower_connector_lora = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/interfaces.py

@@ -167,6 +167,12 @@ class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol):
     `multimodal_config.mm_device_do_normalize`.
     """
 
+    supports_tower_connector_lora: ClassVar[bool] = False
+    """
+    A flag that indicates whether this model supports
+    `lora_config.enable_tower_connector_lora`.
+    """
+
     requires_raw_input_tokens: ClassVar[bool] = False
     """
     A flag that indicates this model processes input id tokens

vllm/model_executor/models/internvl.py

@@ -553,6 +553,7 @@ class InternVLChatModel(
     SupportsEncoderCudaGraph,
 ):
     supports_encoder_tp_data = True
+    supports_tower_connector_lora = True
 
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix=dict.fromkeys(

vllm/model_executor/models/keye.py

@@ -1235,6 +1235,8 @@ class BaseKeyeModule(nn.Module, SupportsMultiModal):
         }
     )
 
+    supports_tower_connector_lora = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/llava.py

@@ -520,6 +520,8 @@ class LlavaForConditionalGeneration(
         }
     )
 
+    supports_tower_connector_lora = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/llava_next.py

@@ -231,6 +231,7 @@ class LlavaNextMultiModalProcessor(BaseLlavaNextMultiModalProcessor[_I]):
 class LlavaNextForConditionalGeneration(
     nn.Module, SupportsLoRA, SupportsMultiModal, SupportsPP
 ):
+    supports_tower_connector_lora = True
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],

vllm/model_executor/models/llava_next_video.py

@@ -323,6 +323,8 @@ class LlavaNextVideoForConditionalGeneration(
         }
     )
 
+    supports_tower_connector_lora = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("video"):

vllm/model_executor/models/mllama4.py

@@ -740,6 +740,7 @@ class Llama4ForConditionalGeneration(
     }
 
     supports_encoder_tp_data = True
+    supports_tower_connector_lora = True
 
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:

vllm/model_executor/models/paligemma.py

@@ -274,6 +274,8 @@ class PaliGemmaForConditionalGeneration(
         }
     )
 
+    supports_tower_connector_lora = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/pixtral.py

@@ -309,6 +309,8 @@ class PixtralForConditionalGeneration(
         "gate_up_proj": ["gate_proj", "up_proj"],
     }
 
+    supports_tower_connector_lora = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):

vllm/model_executor/models/qwen2_5_vl.py

@@ -1256,6 +1256,7 @@ class Qwen2_5_VLForConditionalGeneration(
 
     supports_encoder_tp_data = True
     supports_mm_device_do_normalize = True
+    supports_tower_connector_lora = True
 
     def iter_mm_grid_thw(
         self, mm_features: list[MultiModalFeatureSpec]

vllm/model_executor/models/qwen2_vl.py

@@ -1194,6 +1194,7 @@ class Qwen2VLForConditionalGeneration(
 
     supports_encoder_tp_data = True
     supports_mm_device_do_normalize = True
+    supports_tower_connector_lora = True
 
     def iter_mm_grid_thw(
         self, mm_features: list[MultiModalFeatureSpec]

vllm/model_executor/models/qwen3_asr.py

@@ -361,6 +361,7 @@ class Qwen3ASRForConditionalGeneration(
     }
 
     supported_languages = ISO639_1_SUPPORTED_LANGS
+    supports_tower_connector_lora = True
 
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={

vllm/model_executor/models/qwen3_vl.py

@@ -1753,6 +1753,7 @@ class Qwen3VLForConditionalGeneration(
     }
 
     supports_encoder_tp_data = True
+    supports_tower_connector_lora = True
 
     supported_video_pruning_methods = ("evs", "vidcom2")
 

vllm/model_executor/models/ultravox.py

@@ -610,6 +610,8 @@ class UltravoxModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
         }
     )
 
+    supports_tower_connector_lora = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("audio"):