Update app.py

app.py

@@ -84,6 +84,7 @@ def generate(
 
     
     retriever_inputs = retriever_tokenizer([message], max_length=1024, truncation=True, return_tensors="pt")
+    retriever_inputs = {k:v.to(retriever.device) for k,v in retriever_inputs.items()}
     with torch.no_grad():
         embeddings = retriever(**retriever_inputs).pooler_output
         embeddings = embeddings.cpu().numpy()