Update rpc.py

rpc.py

@@ -199,8 +199,6 @@ def generate(text, use_rpc=True, max_tokens=128):
             tok = tf.argmax(res, axis=-1).numpy().tolist()
         
         enc_text += [tok]
-        new_text = tokenizer.decode(enc_text)
-        tok_str = new_text[len(text):]
-        text = new_text
-
-        yield tok_str
+        response = tokenizer.decode(enc_text)
+    
+        yield response