fix

Author: hiworldwzj

lightllm/models/qwen2_vl/model.py

@@ -31,10 +31,10 @@ class QWen2VLTokenizer(BaseMultiModalTokenizer):
     def __init__(self, tokenizer=None, image_processor=None, **kwargs):
         super().__init__(tokenizer)
         self.image_processor = image_processor
-        self.min_pixel = self.image_processor.image_processor.min_pixels
-        self.max_pixel = self.image_processor.image_processor.max_pixels
-        self.patch_size = self.image_processor.image_processor.patch_size
-        self.merge_size = self.image_processor.image_processor.merge_size
+        self.min_pixel = self.image_processor.min_pixels
+        self.max_pixel = self.image_processor.max_pixels
+        self.patch_size = self.image_processor.patch_size
+        self.merge_size = self.image_processor.merge_size
         self.image_start_id = kwargs["model_cfg"]["vision_start_token_id"]
         self.image_end_id = kwargs["model_cfg"]["vision_end_token_id"]
         self.image_token_id = kwargs["model_cfg"]["image_token_id"]

lightllm/server/tokenizer.py

@@ -89,8 +89,8 @@ def get_tokenizer(
     elif model_type in ["qwen2_vl", "qwen2_5_vl"] and "vision_config" in model_cfg:
         from transformers import AutoProcessor
 
-        image_processor = AutoProcessor.from_pretrained(tokenizer_name)
-        tokenizer = QWen2VLTokenizer(tokenizer=tokenizer, image_processor=image_processor, model_cfg=model_cfg)
+        processor = AutoProcessor.from_pretrained(tokenizer_name)
+        tokenizer = QWen2VLTokenizer(tokenizer=tokenizer, image_processor=processor.image_processor, model_cfg=model_cfg)
     elif model_type == "internvl_chat":
         tokenizer = InternvlTokenizer(tokenizer, model_cfg, weight_dir=tokenizer_name)
     elif model_type == "gemma3":