add interactive mode

app.py

@@ -30,31 +30,17 @@ def chat(user_input, max_new_tokens=512):
     #create streamer
     streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)
 
-    def get_start_idx(response, input):
-        match = re.search(re.escape(response), input)
-        if not match:
-            return -1
-        return match.end()
-
     def generate():
         model.generate(**inputs, max_new_tokens=max_new_tokens, streamer=streamer)
 
     thread = Thread(target=generate)
     thread.start()
 
-    start_idx = -1
+    start_idx = len(inputs)
     generated_text = ""
     for new_text in streamer:
         generated_text += new_text
-
-        if (start_idx == -1):
-            start_idx = get_start_idx(generated_text, user_input)
-            if (start_idx != -1):
-                start_idx += len("ASSISTANT")
-        #print(generated_text)
-        #yield generated_text
-        if (start_idx > 0):
-            yield generated_text[start_idx:]
+        yield generated_text[start_idx:]
 
     thread.join()