diff --git a/python/sglang/srt/configs/model_config.py b/python/sglang/srt/configs/model_config.py
index bdf7ca22a..b51d19d26 100644
--- a/python/sglang/srt/configs/model_config.py
+++ b/python/sglang/srt/configs/model_config.py
@@ -180,6 +180,14 @@ class ModelConfig:
         self.is_audio_model = enable_multimodal and is_audio_model(
             self.hf_config.architectures
         )
+        # TODO: requires further polishing
+        self.is_image_understandable_model = enable_multimodal and hasattr(
+            self.hf_config, "vision_config"
+        )
+        self.is_audio_understandable_model = enable_multimodal and hasattr(
+            self.hf_config, "audio_config"
+        )
+
         self.is_multimodal_chunked_prefill_supported = (
             enable_multimodal
             and is_multimodal_chunked_prefill_supported(self.hf_config.architectures)
diff --git a/python/sglang/srt/entrypoints/http_server.py b/python/sglang/srt/entrypoints/http_server.py
index 289db654c..76e999498 100644
--- a/python/sglang/srt/entrypoints/http_server.py
+++ b/python/sglang/srt/entrypoints/http_server.py
@@ -483,6 +483,8 @@ async def get_model_info():
         "is_generation": _global_state.tokenizer_manager.is_generation,
         "preferred_sampling_params": _global_state.tokenizer_manager.server_args.preferred_sampling_params,
         "weight_version": _global_state.tokenizer_manager.server_args.weight_version,
+        "has_image_understanding": _global_state.tokenizer_manager.model_config.is_image_understandable_model,
+        "has_audio_understanding": _global_state.tokenizer_manager.model_config.is_audio_understandable_model,
     }
     return result
 
@@ -1429,6 +1431,10 @@ def launch_server(
             _global_state.tokenizer_manager.socket_mapping.clear_all_sockets()
 
 
+# Minimal 2x2 black PNG (base64, 1x1 image would cause ambiguous in image channel dimension)
+MINIMUM_PNG_PICTURE_BASE64 = "iVBORw0KGgoAAAANSUhEUgAAAAIAAAACCAYAAABytg0kAAAAEElEQVQImWNgYGD4z8DAAAQYAAH7u8WFAAAAAElFTkSuQmCC"
+
+
 def _execute_server_warmup(
     server_args: ServerArgs,
     pipe_finish_writer: Optional[multiprocessing.connection.Connection],
@@ -1460,8 +1466,16 @@ def _execute_server_warmup(
 
     model_info = res.json()
 
+    is_vlm = bool(model_info.get("has_image_understanding", False))
+
     # Send a warmup request
-    request_name = "/generate" if model_info["is_generation"] else "/encode"
+    if model_info["is_generation"]:
+        if is_vlm and not server_args.skip_tokenizer_init:
+            request_name = "/v1/chat/completions"
+        else:
+            request_name = "/generate"
+    else:
+        request_name = "/encode"
     max_new_tokens = 8 if model_info["is_generation"] else 1
     json_data = {
         "sampling_params": {
@@ -1474,6 +1488,31 @@ def _execute_server_warmup(
         # TODO Workaround the bug that embedding errors for list of size 1
         if server_args.dp_size == 1:
             json_data["input_ids"] = json_data["input_ids"][0]
+    elif is_vlm and server_args.disaggregation_mode == "null":
+        # TODO: ChatCompletionRequest does not have bootstrap info required by disaggregation mode, disable image-warmup for now
+        json_data = {
+            "model": _global_state.tokenizer_manager.served_model_name,
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image_url",
+                            "image_url": {
+                                "url": f"data:image/png;base64,{MINIMUM_PNG_PICTURE_BASE64}"
+                            },
+                        },
+                        {
+                            "type": "text",
+                            "text": "Describe the image.",
+                        },
+                    ],
+                }
+            ],
+            "max_tokens": max_new_tokens,
+            "stream": False,
+            "temperature": 0.0,
+        }
     else:
         json_data["text"] = ["The capital city of France is"] * server_args.dp_size
         # TODO Workaround the bug that embedding errors for list of size 1
diff --git a/python/sglang/srt/models/qwen2_5_vl.py b/python/sglang/srt/models/qwen2_5_vl.py
index 29a9f37b0..b6abf722b 100644
--- a/python/sglang/srt/models/qwen2_5_vl.py
+++ b/python/sglang/srt/models/qwen2_5_vl.py
@@ -23,7 +23,7 @@
 # limitations under the License.
 """Inference-only Qwen2-VL model compatible with HuggingFace weights."""
 import logging
-from functools import lru_cache, partial
+from functools import partial
 from typing import Iterable, List, Optional, Tuple, Type
 
 import torch
@@ -61,7 +61,6 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2 import Qwen2Model
 from sglang.srt.models.utils import permute_inv
 from sglang.srt.utils import add_prefix
-from sglang.srt.utils.hf_transformers_utils import get_processor
 
 logger = logging.getLogger(__name__)
 
@@ -117,8 +116,6 @@ class Qwen2_5_VisionBlock(nn.Module):
         rms_norm_eps: float = 1e-6,
     ) -> None:
         super().__init__()
-        if norm_layer is None:
-            norm_layer = partial(nn.LayerNorm, eps=1e-6)
         self.norm1 = RMSNorm(dim, eps=rms_norm_eps)
         self.norm2 = RMSNorm(dim, eps=rms_norm_eps)
 
@@ -458,9 +455,6 @@ class Qwen2_5_VisionTransformer(nn.Module):
         return x
 
 
-cached_get_processor = lru_cache(get_processor)
-
-
 class Qwen2_5_VLForConditionalGeneration(nn.Module):
     # BitandBytes specific attributes
     default_bitsandbytes_target_modules = [