Upload app.py with huggingface_hub

app.py

@@ -4,17 +4,25 @@ import gradio as gr
 import re
 from PIL import Image
 
-from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
+from transformers import AutoTokenizer, AutoImageProcessor, Qwen2_5_VLForConditionalGeneration, AutoModelForVision2Seq
 
 MODEL_ID = "openbmb/EVisRAG-7B"
 
-processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True, padding_side="left")
+# Use the model class directly (avoids AutoProcessor video-processor auto-detection bug in transformers 5.x)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True, padding_side="left")
+image_processor = AutoImageProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
+
+# Construct the processor from the model class which knows how to combine them
 model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     MODEL_ID,
     torch_dtype=torch.bfloat16,
     attn_implementation="sdpa",
 ).to("cuda").eval()
 
+# Build a processor from the tokenizer + image_processor
+from transformers import Qwen2_5_VLProcessor
+processor = Qwen2_5_VLProcessor(image_processor=image_processor, tokenizer=tokenizer)
+
 
 def evidence_prompt(query: str) -> str:
     """Build the evidence-guided multi-image reasoning prompt used by EVisRAG."""