fix: remove dependency on latest transformers impl (#3635)

python/sglang/srt/configs/qwen2_5_vl_config.py

@@ -48,13 +48,16 @@ from transformers.image_utils import (
     validate_preprocess_arguments,
 )
 from transformers.modeling_rope_utils import rope_config_validation
-from transformers.models.mllama.image_processing_mllama import is_valid_list_of_images
 from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize
 from transformers.processing_utils import ProcessingKwargs, Unpack, VideosKwargs
 from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
 from transformers.utils.constants import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
 
 
+def is_valid_list_of_images(images: List):
+    return images and all(is_valid_image(image) for image in images)
+
+
 class Qwen2_5_VLVisionConfig(PretrainedConfig):
     model_type = "qwen2_5_vl"
     base_config_key = "vision_config"
@@ -999,5 +1002,5 @@ class Qwen2_5_VLImageProcessor(BaseImageProcessor):
         return BatchFeature(data=data, tensor_type=return_tensors)
 
 
-AutoImageProcessor.register(Qwen2_5_VLConfig, Qwen2_5_VLImageProcessor)
+AutoImageProcessor.register(Qwen2_5_VLConfig, None, Qwen2_5_VLImageProcessor, None)
 AutoProcessor.register(Qwen2_5_VLConfig, Qwen2_5_VLProcessor)

python/sglang/srt/models/qwen2_vl.py

@@ -559,7 +559,6 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                     ]
                     image_embeds_offset += num_image_tokens
 
-        input_ids = None
         hidden_states = self.model(
             input_ids=input_ids,
             positions=positions,