Fix dots.ocr model error by switching to Qwen2-VL-2B-Instruct

app.py

@@ -11,7 +11,7 @@ from transformers import AutoModelForCausalLM, AutoProcessor
 from qwen_vl_utils import process_vision_info
 
 
-MODEL_ID = os.getenv("OCR_MODEL_ID", "rednote-hilab/dots.ocr")
+MODEL_ID = os.getenv("OCR_MODEL_ID", "Qwen/Qwen2-VL-2B-Instruct")
 DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
 # Use /tmp for model cache (always writable in containers)
@@ -39,7 +39,7 @@ def load_model():
 
 processor, model = load_model()
 
-app = FastAPI(title="TextSense OCR (DotsOCR)")
+app = FastAPI(title="TextSense OCR (Qwen2-VL)")
 
 
 def read_image_from_upload(upload: UploadFile) -> Image.Image:
@@ -83,10 +83,10 @@ async def extract(
         else:
             return JSONResponse({"error": "No image provided. Provide 'image' file or 'image_url'."}, status_code=400)
 
-        # Prepare chat-style inputs for dots.ocr
+        # Prepare chat-style inputs for Qwen2-VL
         prompt = (
-            "Extract all text from the image in natural reading order. "
-            "Return plain text only, no additional commentary or formatting."
+            "Please extract all text from this image in natural reading order. "
+            "Provide only the extracted text without any additional commentary or formatting."
         )
         messages = [
             {

requirements.txt

@@ -16,4 +16,3 @@ sentencepiece==0.2.0
 safetensors==0.4.3
 protobuf==3.20.3
 qwen-vl-utils==0.0.8
-flash-attn==2.6.3