Model: Support Qwen 2.5 vl (#3258)

docs/references/supported_models.md

@@ -4,7 +4,7 @@
 - Llama / Llama 2 / Llama 3 / Llama 3.1 / Llama 3.2
 - Mistral / Mixtral / Mistral NeMo / Mistral Small 3
 - Gemma / Gemma 2
-- Qwen / Qwen 2 / Qwen 2 MoE / Qwen 2 VL
+- Qwen / Qwen 2 / Qwen 2 MoE / Qwen 2 VL / Qwen 2.5 VL
 - DeepSeek / DeepSeek 2 / [DeepSeek 3](https://github.com/sgl-project/sglang/tree/main/benchmark/deepseek_v3)
 - OLMoE
 - [LLaVA-OneVision](https://llava-vl.github.io/blog/2024-08-05-llava-onevision/)
@@ -54,7 +54,7 @@ To support a new model in SGLang, you only need to add a single file under [SGLa
 You can learn from existing model implementations and create new files for the new models.
 For most models, you should be able to find a similar model to start with (e.g., starting from Llama).
 
-## How to Support a New vision LLM
+## How to Support a New vLM
 
 To support a new vision-language model (vLM) in SGLang, there are several key components in addition to the standard LLM.
 

python/sglang/lang/chat_template.py

@@ -427,6 +427,8 @@ def match_chat_ml(model_path: str):
     if "tinyllama" in model_path:
         return get_chat_template("chatml")
     # Now the suffix for qwen2 chat model is "instruct"
+    if "qwen" in model_path and "vl" in model_path:
+        return get_chat_template("qwen2-vl")
     if "qwen" in model_path:
         if "vl" in model_path:
             return get_chat_template("qwen2-vl")
@@ -443,6 +445,12 @@ def match_chat_ml(model_path: str):
         return get_chat_template("chatml-llava")
 
 
+@register_chat_template_matching_function
+def match_chat_minicpm(model_path: str):
+    if "minicpm" in model_path:
+        return get_chat_template("minicpmv")
+
+
 @register_chat_template_matching_function
 def match_chat_yi(model_path: str):
     model_path = model_path.lower()

python/sglang/srt/configs/__init__.py

 from sglang.srt.configs.chatglm import ChatGLMConfig
 from sglang.srt.configs.dbrx import DbrxConfig
 from sglang.srt.configs.exaone import ExaoneConfig
-from sglang.srt.configs.qwen2vl import Qwen2VLConfig, Qwen2VLVisionConfig
+from sglang.srt.configs.qwen2_5_vl_config import (
+    Qwen2_5_VLConfig,
+    Qwen2_5_VLVisionConfig,
+)
 
 __all__ = [
     "ExaoneConfig",
-    "Qwen2VLConfig",
-    "Qwen2VLVisionConfig",
     "ChatGLMConfig",
     "DbrxConfig",
+    "Qwen2_5_VLConfig",
+    "Qwen2_5_VLVisionConfig",
 ]

python/sglang/srt/configs/model_config.py

@@ -403,6 +403,7 @@ def is_multimodal_model(model_architectures: List[str]):
         or "LlavaVidForCausalLM" in model_architectures
         or "MllamaForConditionalGeneration" in model_architectures
         or "Qwen2VLForConditionalGeneration" in model_architectures
+        or "Qwen2_5_VLForConditionalGeneration" in model_architectures
         or "MiniCPMV" in model_architectures
     ):
         return True

python/sglang/srt/configs/qwen2vl.py

-# coding=utf-8
-# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team.
-# All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Qwen2VL model configuration"""
-
-import os
-from typing import Union
-
-from transformers import PretrainedConfig
-
-
-class Qwen2VLVisionConfig(PretrainedConfig):
-    model_type = "qwen2_vl"
-
-    def __init__(
-        self,
-        depth=32,
-        embed_dim=1280,
-        hidden_size=3584,
-        hidden_act="quick_gelu",
-        mlp_ratio=4,
-        num_heads=16,
-        in_channels=3,
-        patch_size=14,
-        spatial_merge_size=2,
-        temporal_patch_size=2,
-        **kwargs,
-    ):
-        super().__init__(**kwargs)
-
-        self.depth = depth
-        self.embed_dim = embed_dim
-        self.hidden_size = hidden_size
-        self.hidden_act = hidden_act
-        self.mlp_ratio = mlp_ratio
-        self.num_heads = num_heads
-        self.in_channels = in_channels
-        self.patch_size = patch_size
-        self.spatial_merge_size = spatial_merge_size
-        self.temporal_patch_size = temporal_patch_size
-
-    @classmethod
-    def from_pretrained(
-        cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
-    ) -> "PretrainedConfig":
-        cls._set_token_in_kwargs(kwargs)
-
-        config_dict, kwargs = cls.get_config_dict(
-            pretrained_model_name_or_path, **kwargs
-        )
-
-        if config_dict.get("model_type") == "qwen2_vl":
-            config_dict = config_dict["vision_config"]
-
-        return cls.from_dict(config_dict, **kwargs)
-
-
-class Qwen2VLConfig(PretrainedConfig):
-    model_type = "qwen2_vl"
-
-    def __init__(
-        self,
-        vocab_size=152064,
-        hidden_size=8192,
-        intermediate_size=29568,
-        num_hidden_layers=80,
-        num_attention_heads=64,
-        num_key_value_heads=8,
-        hidden_act="silu",
-        max_position_embeddings=32768,
-        initializer_range=0.02,
-        rms_norm_eps=1e-05,
-        use_cache=True,
-        tie_word_embeddings=False,
-        rope_theta=1000000.0,
-        use_sliding_window=False,
-        sliding_window=4096,
-        max_window_layers=80,
-        attention_dropout=0.0,
-        vision_config=None,
-        rope_scaling=None,
-        **kwargs,
-    ):
-        if isinstance(vision_config, dict):
-            self.vision_config = Qwen2VLVisionConfig(**vision_config)
-        elif vision_config is None:
-            self.vision_config = Qwen2VLVisionConfig()
-
-        self.vocab_size = vocab_size
-        self.max_position_embeddings = max_position_embeddings
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.use_sliding_window = use_sliding_window
-        self.sliding_window = sliding_window
-        self.max_window_layers = max_window_layers
-
-        # for backward compatibility
-        if num_key_value_heads is None:
-            num_key_value_heads = num_attention_heads
-
-        self.num_key_value_heads = num_key_value_heads
-        self.hidden_act = hidden_act
-        self.initializer_range = initializer_range
-        self.rms_norm_eps = rms_norm_eps
-        self.use_cache = use_cache
-        self.rope_theta = rope_theta
-        self.attention_dropout = attention_dropout
-        self.rope_scaling = rope_scaling
-
-        # NOTE(HandH1998): This is necessary for configuring the `rope_type`` of qwen2vl models after removing dependencies on vllm.
-        if self.rope_scaling is not None and "type" in self.rope_scaling:
-            if self.rope_scaling["type"] == "mrope":
-                self.rope_scaling["type"] = "default"
-            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
-
-        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

python/sglang/srt/hf_transformers_utils.py

@@ -30,16 +30,15 @@ from transformers import (
 )
 from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
 
-from sglang.srt.configs import ChatGLMConfig, DbrxConfig, ExaoneConfig, Qwen2VLConfig
+from sglang.srt.configs import ChatGLMConfig, DbrxConfig, ExaoneConfig, Qwen2_5_VLConfig
 
 _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
     ChatGLMConfig.model_type: ChatGLMConfig,
     DbrxConfig.model_type: DbrxConfig,
     ExaoneConfig.model_type: ExaoneConfig,
-    Qwen2VLConfig.model_type: Qwen2VLConfig,
+    Qwen2_5_VLConfig.model_type: Qwen2_5_VLConfig,
 }
 
-
 for name, cls in _CONFIG_REGISTRY.items():
     with contextlib.suppress(ValueError):
         AutoConfig.register(name, cls)

python/sglang/srt/managers/image_processor.py

 # TODO: also move pad_input_ids into this module
 import asyncio
 import concurrent.futures
+import dataclasses
 import logging
 import multiprocessing as mp
 import os
@@ -8,6 +9,7 @@ from abc import ABC, abstractmethod
 from typing import List, Optional, Union
 
 import numpy as np
+import PIL
 import transformers
 from decord import VideoReader, cpu
 from PIL import Image
@@ -34,11 +36,22 @@ def init_global_processor(server_args: ServerArgs):
     )
 
 
+@dataclasses.dataclass
+class BaseImageProcessorOutput:
+    image_hashes: list[int]
+    image_sizes: list[int]
+    all_frames: [PIL.Image]
+    # input_text, with each frame of video/image represented with a image_token
+    input_text: str
+
+
 class BaseImageProcessor(ABC):
     def __init__(self, hf_config, server_args, _processor):
         self.hf_config = hf_config
         self._processor = _processor
         self.server_args = server_args
+        # FIXME: not accurate, model and image specific
+        self.NUM_TOKEN_PER_FRAME = 330
 
         self.executor = concurrent.futures.ProcessPoolExecutor(
             initializer=init_global_processor,
@@ -48,9 +61,128 @@ class BaseImageProcessor(ABC):
         )
 
     @abstractmethod
-    async def process_images_async(self, image_data, input_text, **kwargs):
+    async def process_images_async(
+        self, image_data, input_text, max_req_input_len, **kwargs
+    ):
         pass
 
+    def get_estimated_frames_list(self, image_data):
+        """
+        estimate the total frame count from all visual input
+        """
+        # Before processing inputs
+        estimated_frames_list = []
+        for image in image_data:
+            if isinstance(image, str) and image.startswith("video:"):
+                path = image[len("video:") :]
+                # Estimate frames for the video
+                vr = VideoReader(path, ctx=cpu(0))
+                num_frames = len(vr)
+            else:
+                # For images, each contributes one frame
+                num_frames = 1
+            estimated_frames_list.append(num_frames)
+
+        return estimated_frames_list
+
+    def encode_video(self, video_path, frame_count_limit=None):
+        if not os.path.exists(video_path):
+            logger.error(f"Video {video_path} does not exist")
+            return []
+
+        if frame_count_limit == 0:
+            return []
+
+        def uniform_sample(l, n):
+            gap = len(l) / n
+            idxs = [int(i * gap + gap / 2) for i in range(n)]
+            return [l[i] for i in idxs]
+
+        vr = VideoReader(video_path, ctx=cpu(0))
+        sample_fps = round(vr.get_avg_fps() / 1)  # FPS
+        frame_idx = [i for i in range(0, len(vr), sample_fps)]
+        if frame_count_limit is not None and len(frame_idx) > frame_count_limit:
+            frame_idx = uniform_sample(frame_idx, frame_count_limit)
+        frames = vr.get_batch(frame_idx).asnumpy()
+        frames = [Image.fromarray(v.astype("uint8")) for v in frames]
+        return frames
+
+    def load_images(
+        self,
+        max_req_input_len: int,
+        input_ids: list,
+        image_data,
+        image_token: str,
+    ) -> BaseImageProcessorOutput:
+        """
+        Each frame of video/image will be replaced by a single image token
+        """
+        image_hashes, image_sizes = [], []
+        all_frames = []
+        new_text_parts = []
+
+        if isinstance(input_ids, list):
+            assert len(input_ids) and isinstance(input_ids[0], int)
+            input_text = self._processor.tokenizer.decode(input_ids)
+        else:
+            input_text = input_ids
+
+        text_parts = input_text.split(image_token)
+
+        # roughly calculate the max number of frames under the max_req_input_len limit
+        def calculate_max_num_frames() -> int:
+            ret = (max_req_input_len - len(input_ids)) // self.NUM_TOKEN_PER_FRAME
+            return min(ret, 100)
+
+        MAX_NUM_FRAMES = calculate_max_num_frames()
+        estimated_frames_list = self.get_estimated_frames_list(image_data=image_data)
+        total_frame_count = sum(estimated_frames_list)
+        # a heuristic value, suggesting the maximum fraction of frames to embed from all visual inputs.
+        # e.g., 0.1 suggests that 1 frame out of 10 input frames should be used
+        scaling_factor = min(1.0, MAX_NUM_FRAMES / total_frame_count)
+
+        # Process each input with allocated frames
+        for image_index, (image, estimated_frames) in enumerate(
+            zip(image_data, estimated_frames_list)
+        ):
+            if len(all_frames) >= MAX_NUM_FRAMES:
+                frames_to_process = 0
+            else:
+                frames_to_process = max(1, int(estimated_frames * scaling_factor))
+
+            if frames_to_process == 0:
+                frames = []
+            else:
+                try:
+                    if isinstance(image, str) and image.startswith("video:"):
+                        path = image[len("video:") :]
+                        frames = self.encode_video(
+                            path, frame_count_limit=frames_to_process
+                        )
+                    else:
+                        raw_image, _size = load_image(image)
+                        frames = [raw_image]
+                    if len(frames) == 0:
+                        continue
+                except FileNotFoundError as e:
+                    print(e)
+                    return None
+                image_sizes += frames[0].size * len(frames)
+                image_hashes += [hash(image)] * len(frames)
+                all_frames += frames
+
+            new_text_parts.append(text_parts[image_index])
+            if frames_to_process != 0:
+                new_text_parts.append(image_token * len(frames))
+            assert frames_to_process == len(frames)
+
+        new_text_parts.append(text_parts[-1])
+
+        input_text = "".join(new_text_parts)
+        return BaseImageProcessorOutput(
+            image_hashes, image_sizes, all_frames, input_text
+        )
+
 
 class DummyImageProcessor(BaseImageProcessor):
     def __init__(self):
@@ -248,9 +380,9 @@ class MiniCPMVImageProcessor(BaseImageProcessor):
             text=input_text, images=images, return_tensors="pt"
         )
         return {
-            "input_ids": result["input_ids"],
-            "pixel_values": result["pixel_values"],
-            "tgt_sizes": result["tgt_sizes"],
+            "input_ids": result.input_ids,
+            "pixel_values": result.pixel_values,
+            "tgt_sizes": result.tgt_sizes,
         }
 
     async def _process_images(self, images, input_text):
@@ -278,124 +410,20 @@ class MiniCPMVImageProcessor(BaseImageProcessor):
     ):
         if not image_data:
             return None
-
         if not isinstance(image_data, list):
             image_data = [image_data]
 
-        image_hashes, image_sizes = [], []
-        all_frames = []
-
-        # roughly calculate the max number of frames under the max_req_input_len limit
-        def calculate_max_num_frames() -> int:
-            # Model-specific
-            NUM_TOKEN_PER_FRAME = 330
-
-            ret = (max_req_input_len - len(input_ids)) // NUM_TOKEN_PER_FRAME
-            return min(ret, 100)
-
-        MAX_NUM_FRAMES = calculate_max_num_frames()
-
-        # print(f"MAX_NUM_FRAMES: {MAX_NUM_FRAMES}")
-
-        def get_estimated_frames_list():
-            """
-            estimate the total frame count from all visual input
-            """
-            # Before processing inputs
-            estimated_frames_list = []
-            for image in image_data:
-                if isinstance(image, str) and image.startswith("video:"):
-                    path = image[len("video:") :]
-                    # Estimate frames for the video
-                    vr = VideoReader(path, ctx=cpu(0))
-                    num_frames = len(vr)
-                else:
-                    # For images, each contributes one frame
-                    num_frames = 1
-                estimated_frames_list.append(num_frames)
-
-            return estimated_frames_list
-
-        estimated_frames_list = get_estimated_frames_list()
-        total_frame_count = sum(estimated_frames_list)
-        scaling_factor = min(1.0, MAX_NUM_FRAMES / total_frame_count)
-
-        def encode_video(video_path, frame_count_limit=None):
-            if not os.path.exists(video_path):
-                logger.error(f"Video {video_path} does not exist")
-                return []
-
-            if frame_count_limit == 0:
-                return []
-
-            def uniform_sample(l, n):
-                gap = len(l) / n
-                idxs = [int(i * gap + gap / 2) for i in range(n)]
-                return [l[i] for i in idxs]
-
-            vr = VideoReader(video_path, ctx=cpu(0))
-            sample_fps = round(vr.get_avg_fps() / 1)  # FPS
-            frame_idx = [i for i in range(0, len(vr), sample_fps)]
-            if frame_count_limit is not None and len(frame_idx) > frame_count_limit:
-                frame_idx = uniform_sample(frame_idx, frame_count_limit)
-            frames = vr.get_batch(frame_idx).asnumpy()
-            frames = [Image.fromarray(v.astype("uint8")) for v in frames]
-            return frames
-
-        if isinstance(input_ids, list):
-            assert len(input_ids) and isinstance(input_ids[0], int)
-            input_text = self._processor.tokenizer.decode(input_ids)
-        else:
-            input_text = input_ids
-        # MiniCPMV requires each frame of video as a single image token
-        text_parts = input_text.split(self.IMAGE_TOKEN)
-        new_text_parts = []
-
-        # Process each input with allocated frames
-        for image_index, (image, estimated_frames) in enumerate(
-            zip(image_data, estimated_frames_list)
-        ):
-            if len(all_frames) >= MAX_NUM_FRAMES:
-                frames_to_process = 0
-            else:
-                frames_to_process = max(1, int(estimated_frames * scaling_factor))
-
-            if frames_to_process == 0:
-                frames = []
-            else:
-                try:
-                    if isinstance(image, str) and image.startswith("video:"):
-                        path = image[len("video:") :]
-                        frames = encode_video(path, frame_count_limit=frames_to_process)
-                    else:
-                        raw_image, _size = load_image(image)
-                        frames = [raw_image]
-                    if len(frames) == 0:
-                        continue
-                except FileNotFoundError as e:
-                    print(e)
+        base_output = self.load_images(
+            max_req_input_len, input_ids, image_data, self.IMAGE_TOKEN
+        )
+        if base_output is None:
             return None
-                image_sizes += frames[0].size * len(frames)
-                image_hashes += [hash(image)] * len(frames)
-                all_frames += frames
-
-            assert frames_to_process == len(frames)
-
-            new_text_parts.append(text_parts[image_index])
 
-            if frames_to_process != 0:
-                new_text_parts.append(self.IMAGE_TOKEN * len(frames))
-
-        new_text_parts.append(text_parts[-1])
-
-        input_text = "".join(new_text_parts)
-
-        if len(all_frames) == 0:
+        if len(base_output.all_frames) == 0:
             return None
-        res = await self._process_images(images=all_frames, input_text=input_text)
-        pixel_values = res["pixel_values"]
-        tgt_sizes = res["tgt_sizes"]
-        input_ids = res["input_ids"]
+        res = await self._process_images(
+            images=base_output.all_frames, input_text=base_output.input_text
+        )
 
         # Collect special token ids
         tokenizer = self._processor.tokenizer
@@ -405,10 +433,10 @@ class MiniCPMVImageProcessor(BaseImageProcessor):
             slice_start_id = [tokenizer.slice_start_id]
             slice_end_id = [tokenizer.slice_end_id]
         return {
-            "input_ids": input_ids.flatten().tolist(),
-            "pixel_values": pixel_values,
-            "tgt_sizes": tgt_sizes,
-            "image_hashes": image_hashes,
+            "input_ids": res["input_ids"].flatten().tolist(),
+            "pixel_values": res["pixel_values"],
+            "tgt_sizes": res["tgt_sizes"],
+            "image_hashes": base_output.image_hashes,
             "modalities": request_obj.modalities or ["image"],
             "im_start_id": im_start_id,
             "im_end_id": im_end_id,
@@ -536,13 +564,80 @@ class Qwen2VLImageProcessor(BaseImageProcessor):
         }
 
 
+class Qwen2_5VLImageProcessor(BaseImageProcessor):
+    def __init__(self, hf_config, server_args, _processor):
+        super().__init__(hf_config, server_args, _processor)
+        self.IMAGE_TOKEN = "<|vision_start|><|image_pad|><|vision_end|>"
+        self.IM_START_TOKEN_ID = hf_config.vision_start_token_id
+        self.IM_END_TOKEN_ID = hf_config.vision_end_token_id
+        self.NUM_TOKEN_PER_FRAME = 770
+
+    @staticmethod
+    def _process_images_task(images, input_text):
+        result = global_processor.__call__(
+            text=input_text, images=images, return_tensors="pt"
+        )
+        return {
+            "input_ids": result.input_ids,
+            "pixel_values": result.pixel_values,
+            "image_grid_thws": result.image_grid_thw,
+        }
+
+    async def _process_images(self, images, input_text) -> dict:
+        if self.executor is not None:
+            loop = asyncio.get_event_loop()
+            return await loop.run_in_executor(
+                self.executor,
+                Qwen2_5VLImageProcessor._process_images_task,
+                images,
+                input_text,
+            )
+        else:
+            return self._process_images_task(images, input_text)
+
+    async def process_images_async(
+        self,
+        image_data: List[Union[str, bytes]],
+        input_ids,
+        request_obj,
+        max_req_input_len,
+        *args,
+        **kwargs,
+    ):
+        if not image_data:
+            return None
+        if isinstance(image_data, str):
+            image_data = [image_data]
+
+        image_token = self.IMAGE_TOKEN
+        base_output = self.load_images(
+            max_req_input_len, input_ids, image_data, image_token
+        )
+
+        ret = await self._process_images(base_output.all_frames, base_output.input_text)
+
+        return {
+            "input_ids": ret["input_ids"].flatten().tolist(),
+            "pixel_values": ret["pixel_values"],
+            "image_hashes": base_output.image_hashes,
+            "modalities": request_obj.modalities or ["image"],
+            "image_grid_thws": ret["image_grid_thws"],
+            "im_start_id": self.IM_START_TOKEN_ID,
+            "im_end_id": self.IM_END_TOKEN_ID,
+        }
+
+
 def get_image_processor(
     hf_config, server_args: ServerArgs, processor
 ) -> BaseImageProcessor:
     if "MllamaForConditionalGeneration" in hf_config.architectures:
         return MllamaImageProcessor(hf_config, server_args, processor)
     elif "Qwen2VLForConditionalGeneration" in hf_config.architectures:
-        return Qwen2VLImageProcessor(hf_config, server_args, processor.image_processor)
+
+        return Qwen2VLImageProcessor(hf_config, server_args, processor)
+    elif "Qwen2_5_VLForConditionalGeneration" in hf_config.architectures:
+        return Qwen2_5VLImageProcessor(hf_config, server_args, processor)
+
     elif "MiniCPMV" in hf_config.architectures:
         return MiniCPMVImageProcessor(hf_config, server_args, processor)
     else:

python/sglang/srt/models/qwen2_vl.py

@@ -31,8 +31,9 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from einops import rearrange
+from transformers import Qwen2VLConfig
+from transformers.models.qwen2_vl.configuration_qwen2_vl import Qwen2VLVisionConfig
 
-from sglang.srt.configs import Qwen2VLConfig, Qwen2VLVisionConfig
 from sglang.srt.hf_transformers_utils import get_processor
 from sglang.srt.layers.activation import QuickGELU
 from sglang.srt.layers.attention.vision import VisionAttention

test/srt/test_vision_openai_server.py

@@ -252,6 +252,18 @@ class TestOpenAIVisionServer(unittest.TestCase):
         print("-" * 30)
 
         # Add assertions to validate the video response
+        assert "iPod" in video_response or "device" in video_response, video_response
+        assert (
+            "man" in video_response
+            or "person" in video_response
+            or "individual" in video_response
+        ), video_response
+        assert (
+            "present" in video_response
+            or "examine" in video_response
+            or "display" in video_response
+        )
+        assert "black" in video_response or "dark" in video_response
         self.assertIsNotNone(video_response)
         self.assertGreater(len(video_response), 0)
 
@@ -366,6 +378,30 @@ class TestQWen2VLServer(TestOpenAIVisionServer):
         cls.base_url += "/v1"
 
 
+class TestQWen2_5_VLServer(TestOpenAIVisionServer):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "Qwen/Qwen2.5-VL-7B-Instruct"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            api_key=cls.api_key,
+            other_args=[
+                "--chat-template",
+                "qwen2-vl",
+                # FIXME: workaround to chunked prefill within image embeds
+                "--chunked-prefill-size",
+                "10000",
+                "--mem-fraction-static",
+                "0.4",
+            ],
+        )
+        cls.base_url += "/v1"
+
+
 class TestQWen2VLServerContextLengthIssue(unittest.TestCase):
     @classmethod
     def setUpClass(cls):