[Model] Apply #32631 for recent models (#33785)

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

vllm/model_executor/models/eagle2_5_vl.py

@@ -222,6 +222,7 @@ class Eagle2_5_VLForConditionalGeneration(
 
         self.select_layer = getattr(config, "select_layer", -1)
 
+        with self._mark_tower_model(vllm_config, "image"):
             # Vision encoder (SigLIP)
             self.vision_model = self._init_vision_model(
                 config,
@@ -229,6 +230,10 @@ class Eagle2_5_VLForConditionalGeneration(
                 prefix=maybe_prefix(prefix, "vision_model"),
             )
 
+            # MLP projection
+            self.mlp1 = self._init_mlp1(config)
+
+        with self._mark_language_model(vllm_config):
             # Language model (Qwen2)
             self.language_model = init_vllm_registered_model(
                 vllm_config=vllm_config,
@@ -236,9 +241,6 @@ class Eagle2_5_VLForConditionalGeneration(
                 prefix=maybe_prefix(prefix, "language_model"),
             )
 
-        # MLP projection
-        self.mlp1 = self._init_mlp1(config)
-
         self.img_context_token_id = None
 
         self.make_empty_intermediate_tensors = (
@@ -399,9 +401,6 @@ class Eagle2_5_VLForConditionalGeneration(
         ]
         return image_embeds.split(image_feature_sizes)
 
-    def get_language_model(self) -> torch.nn.Module:
-        return self.language_model
-
     def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
         """Embed multimodal inputs."""
         image_input = self._parse_and_validate_image_input(**kwargs)

vllm/model_executor/models/funaudiochat.py

@@ -820,9 +820,6 @@ class FunAudioChatForConditionalGeneration(nn.Module, SupportsMultiModal, Suppor
             self.language_model.make_empty_intermediate_tensors
         )
 
-    def get_language_model(self) -> torch.nn.Module:
-        return self.language_model
-
     def _get_continuous_audio_features(
         self,
         input_features: torch.Tensor,

vllm/model_executor/models/openpangu_vl.py

@@ -843,6 +843,8 @@ class OpenPanguVLForConditionalGeneration(
         self.config = config
         self.vllm_config = vllm_config
         quant_config = vllm_config.quant_config
+
+        with self._mark_tower_model(vllm_config, {"image", "video"}):
             self.visual = OpenPanguVisionTransformer(
                 vision_config=config.vision_config,
                 out_hidden_size=config.vision_config.out_hidden_size,
@@ -852,11 +854,13 @@ class OpenPanguVLForConditionalGeneration(
                 prefix=maybe_prefix(prefix, "visual"),
             )
 
+        with self._mark_language_model(vllm_config):
             self.language_model = init_vllm_registered_model(
                 vllm_config=vllm_config,
                 prefix=maybe_prefix("openpangu", "language_model"),
                 architectures=["PanguEmbeddedForCausalLM"],
             )
+
         self.make_empty_intermediate_tensors = (
             self.language_model.make_empty_intermediate_tensors
         )
@@ -1008,9 +1012,6 @@ class OpenPanguVLForConditionalGeneration(
                 )
         return mm_input_by_modality
 
-    def get_language_model(self) -> torch.nn.Module:
-        return self.language_model
-
     def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None:
         mm_input_by_modality = self._parse_and_validate_multimodal_inputs(**kwargs)
         if not mm_input_by_modality:

vllm/model_executor/models/qwen3_asr.py

@@ -296,13 +296,15 @@ class Qwen3ASRForConditionalGeneration(
         multimodal_config = vllm_config.model_config.multimodal_config
         self.config = thinker_config
         self.multimodal_config = multimodal_config
+        self.quant_config = quant_config
 
+        with self._mark_tower_model(vllm_config, "audio"):
             self.audio_tower = Qwen3OmniMoeAudioEncoder(
                 thinker_config.audio_config,
                 prefix=maybe_prefix(prefix, "audio_tower"),
             )
-        self.quant_config = quant_config
 
+        with self._mark_language_model(vllm_config):
             self.language_model = Qwen3ForCausalLM(
                 vllm_config=vllm_config.with_hf_config(
                     thinker_config.text_config, architectures=["Qwen3ForCausalLM"]
@@ -363,9 +365,6 @@ class Qwen3ASRForConditionalGeneration(
         )
         return audio_features.split(audio_output_lengths.tolist())
 
-    def get_language_model(self) -> torch.nn.Module:
-        return self.language_model
-
     def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None:
         mm_input_by_modality = self._parse_and_validate_multimodal_inputs(**kwargs)
         if not mm_input_by_modality: