fix: duplicate resize images logic of qwen-vl series models (#12458)

Signed-off-by: yangsijia.614 <yangsijia.614@bytedance.com>

python/sglang/srt/multimodal/processors/dots_vlm.py

@@ -1,16 +1,12 @@
-import asyncio
 import re
 from typing import Dict, List, Union
 
-from PIL import Image
-
 from sglang.srt.models.dots_ocr import DotsOCRForCausalLM
 from sglang.srt.models.dots_vlm import DotsVLMForCausalLM
 from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor,
     MultimodalSpecialTokens,
 )
-from sglang.srt.multimodal.processors.qwen_vl import resize_image_async
 
 
 class DotsVLMImageProcessor(BaseMultimodalProcessor):
@@ -70,18 +66,6 @@ class DotsVLMImageProcessor(BaseMultimodalProcessor):
             multimodal_tokens=self.mm_tokens,
         )
 
-        # Qwen-specific: resize images if they are raw Image objects
-        if base_output.images and isinstance(base_output.images[0], Image.Image):
-            resize_tasks = [
-                resize_image_async(
-                    image,
-                    min_pixels=self.MIN_PIXELS,
-                    max_pixels=self.MAX_PIXELS,
-                    size_factor=self.IMAGE_FACTOR,
-                )
-                for image in base_output.images
-            ]
-            base_output.images = await asyncio.gather(*resize_tasks)
         combined_mm_item, input_ids, _ = self.process_and_combine_mm_data(
             base_output, self.mm_tokens
         )

python/sglang/srt/multimodal/processors/points_v15_chat.py

@@ -1,15 +1,9 @@
 # Copy from qwen_vl.py, adapted for points-v15-chat
 
-import asyncio
 from typing import List, Union
 
-from PIL import Image
-
 from sglang.srt.models.points_v15_chat import POINTSV15ChatModel
-from sglang.srt.multimodal.processors.qwen_vl import (
-    QwenVLImageProcessor,
-    resize_image_async,
-)
+from sglang.srt.multimodal.processors.qwen_vl import QwenVLImageProcessor
 
 
 class POINTSV15ChatProcessor(QwenVLImageProcessor):
@@ -37,10 +31,6 @@ class POINTSV15ChatProcessor(QwenVLImageProcessor):
             multimodal_tokens=self.mm_tokens,
         )
 
-        if base_output.images and isinstance(base_output.images[0], Image.Image):
-            resize_tasks = [resize_image_async(image) for image in base_output.images]
-            base_output.images = await asyncio.gather(*resize_tasks)
-
         mm_items, input_ids, _ = self.process_and_combine_mm_data(
             base_output, self.mm_tokens
         )

python/sglang/srt/multimodal/processors/qwen_vl.py

@@ -1,4 +1,3 @@
-import asyncio
 import math
 import os
 import re
@@ -79,26 +78,6 @@ def smart_resize(
     return h_bar, w_bar
 
 
-def resize_image(
-    image,
-    min_pixels: int = MIN_PIXELS,
-    max_pixels: int = MAX_PIXELS,
-    size_factor: int = IMAGE_FACTOR,
-) -> Image.Image:
-    width, height = image.size
-    min_pixels = min_pixels
-    max_pixels = max_pixels
-    resized_height, resized_width = smart_resize(
-        height,
-        width,
-        factor=size_factor,
-        min_pixels=min_pixels,
-        max_pixels=max_pixels,
-    )
-    image = image.resize((resized_width, resized_height), resample=RESIZE_RESAMPLE)
-    return image
-
-
 def round_by_factor(number: int, factor: int) -> int:
     """Returns the closest integer to 'number' that is divisible by 'factor'."""
     return round(number / factor) * factor
@@ -114,15 +93,6 @@ def floor_by_factor(number: int, factor: int) -> int:
     return math.floor(number / factor) * factor
 
 
-async def resize_image_async(
-    image,
-    min_pixels: int = MIN_PIXELS,
-    max_pixels: int = MAX_PIXELS,
-    size_factor: int = IMAGE_FACTOR,
-):
-    return resize_image(image, min_pixels, max_pixels, size_factor)
-
-
 def smart_nframes(
     ele: dict,
     total_frames: int,
@@ -266,11 +236,6 @@ class QwenVLImageProcessor(SGLangBaseProcessor):
         self.audio_start_token_id = getattr(hf_config, "audio_start_token_id", None)
         self.audio_token_id = getattr(hf_config, "audio_token_id", None)
 
-        self.NUM_TOKEN_PER_FRAME = 770
-        self.IMAGE_FACTOR = 28
-        self.MIN_PIXELS = 4 * 28 * 28
-        self.MAX_PIXELS = 16384 * 28 * 28
-        self.MAX_RATIO = 200
         self.mm_tokens = MultimodalSpecialTokens(
             image_token="<|vision_start|><|image_pad|><|vision_end|>",
             image_token_id=hf_config.image_token_id,
@@ -301,11 +266,6 @@ class QwenVLImageProcessor(SGLangBaseProcessor):
         load_time = time.perf_counter()
         rid = getattr(request_obj, "rid", "anonymous_rid")
 
-        # Qwen-specific: resize images if they are raw Image objects
-        if base_output.images and isinstance(base_output.images[0], Image.Image):
-            resize_tasks = [resize_image_async(image) for image in base_output.images]
-            base_output.images = await asyncio.gather(*resize_tasks)
-
         video_metadata = None
         if base_output.videos:
             videos_processed = [