Fix: Upgrade PEFT and use explicit pixel_values for inference

app.py

@@ -35,10 +35,10 @@ def predict(image):
         image = image.convert("RGB")
         pixel_values = processor(image, return_tensors="pt").pixel_values.to(device)
         
-        # --- ROBUST GENERATE CALL ---
-        # Using inputs= as a keyword argument fixes the PeftModel positional argument error.
+        # --- VISION-SPECIFIC GENERATE CALL ---
+        # Vision models like TrOCR require the 'pixel_values' keyword explicitly
         with torch.no_grad():
-            generated_ids = model.generate(inputs=pixel_values)
+            generated_ids = model.generate(pixel_values=pixel_values)
             generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
         return generated_text
     except Exception as e:

requirements.txt

@@ -1,6 +1,6 @@
 torch
 transformers
-peft
+peft>=0.10.0
 huggingface_hub==0.25.2
 pillow
 safetensors