diff --git a/vllm/model_executor/models/eagle2_5_vl.py b/vllm/model_executor/models/eagle2_5_vl.py
index 3ce9b9c4d..19d21de5b 100644
--- a/vllm/model_executor/models/eagle2_5_vl.py
+++ b/vllm/model_executor/models/eagle2_5_vl.py
@@ -222,22 +222,24 @@ class Eagle2_5_VLForConditionalGeneration(
 
         self.select_layer = getattr(config, "select_layer", -1)
 
-        # Vision encoder (SigLIP)
-        self.vision_model = self._init_vision_model(
-            config,
-            quant_config=quant_config,
-            prefix=maybe_prefix(prefix, "vision_model"),
-        )
+        with self._mark_tower_model(vllm_config, "image"):
+            # Vision encoder (SigLIP)
+            self.vision_model = self._init_vision_model(
+                config,
+                quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "vision_model"),
+            )
 
-        # Language model (Qwen2)
-        self.language_model = init_vllm_registered_model(
-            vllm_config=vllm_config,
-            hf_config=config.text_config,
-            prefix=maybe_prefix(prefix, "language_model"),
-        )
+            # MLP projection
+            self.mlp1 = self._init_mlp1(config)
 
-        # 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,
+                hf_config=config.text_config,
+                prefix=maybe_prefix(prefix, "language_model"),
+            )
 
         self.img_context_token_id = None
 
@@ -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)
diff --git a/vllm/model_executor/models/funaudiochat.py b/vllm/model_executor/models/funaudiochat.py
index 16afec3cf..b7b8659a4 100644
--- a/vllm/model_executor/models/funaudiochat.py
+++ b/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,
diff --git a/vllm/model_executor/models/openpangu_vl.py b/vllm/model_executor/models/openpangu_vl.py
index d7df2cbb4..e9288e6dd 100644
--- a/vllm/model_executor/models/openpangu_vl.py
+++ b/vllm/model_executor/models/openpangu_vl.py
@@ -843,20 +843,24 @@ class OpenPanguVLForConditionalGeneration(
         self.config = config
         self.vllm_config = vllm_config
         quant_config = vllm_config.quant_config
-        self.visual = OpenPanguVisionTransformer(
-            vision_config=config.vision_config,
-            out_hidden_size=config.vision_config.out_hidden_size,
-            hidden_size=config.hidden_size,
-            norm_eps=getattr(config.vision_config, "rms_norm_eps", 1e-6),
-            quant_config=self._maybe_ignore_quant_config(quant_config),
-            prefix=maybe_prefix(prefix, "visual"),
-        )
 
-        self.language_model = init_vllm_registered_model(
-            vllm_config=vllm_config,
-            prefix=maybe_prefix("openpangu", "language_model"),
-            architectures=["PanguEmbeddedForCausalLM"],
-        )
+        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,
+                hidden_size=config.hidden_size,
+                norm_eps=getattr(config.vision_config, "rms_norm_eps", 1e-6),
+                quant_config=self._maybe_ignore_quant_config(quant_config),
+                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:
diff --git a/vllm/model_executor/models/qwen3_asr.py b/vllm/model_executor/models/qwen3_asr.py
index e63e03e23..9dac8d75b 100644
--- a/vllm/model_executor/models/qwen3_asr.py
+++ b/vllm/model_executor/models/qwen3_asr.py
@@ -296,19 +296,21 @@ class Qwen3ASRForConditionalGeneration(
         multimodal_config = vllm_config.model_config.multimodal_config
         self.config = thinker_config
         self.multimodal_config = multimodal_config
-
-        self.audio_tower = Qwen3OmniMoeAudioEncoder(
-            thinker_config.audio_config,
-            prefix=maybe_prefix(prefix, "audio_tower"),
-        )
         self.quant_config = quant_config
 
-        self.language_model = Qwen3ForCausalLM(
-            vllm_config=vllm_config.with_hf_config(
-                thinker_config.text_config, architectures=["Qwen3ForCausalLM"]
-            ),
-            prefix=maybe_prefix(prefix, "language_model"),
-        )
+        with self._mark_tower_model(vllm_config, "audio"):
+            self.audio_tower = Qwen3OmniMoeAudioEncoder(
+                thinker_config.audio_config,
+                prefix=maybe_prefix(prefix, "audio_tower"),
+            )
+
+        with self._mark_language_model(vllm_config):
+            self.language_model = Qwen3ForCausalLM(
+                vllm_config=vllm_config.with_hf_config(
+                    thinker_config.text_config, architectures=["Qwen3ForCausalLM"]
+                ),
+                prefix=maybe_prefix(prefix, "language_model"),
+            )
 
         self.make_empty_intermediate_tensors = (
             self.language_model.make_empty_intermediate_tensors
@@ -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: