Update app.py

app.py

@@ -51,7 +51,9 @@ config.top_p = 0.9
 config.top_k = 30
 
 pipe = ov_genai.LLMPipeline(model_path, "CPU")
-pipe.get_tokenizer().set_chat_template(pipe.get_tokenizer().chat_template)
+tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+#pipe.get_tokenizer().set_chat_template(pipe.get_tokenizer().chat_template)
 
 # 讀取剛剛存的 txt 檔案
 documents = SimpleDirectoryReader("./data").load_data()
@@ -98,7 +100,12 @@ def generate_stream(prompt):
 
     def worker():
         nonlocal tps_result
-        gen_result = pipe.generate([final_prompt], streamer=streamer, config=config)
+        text = tokenizer.apply_chat_template(
+            final_prompt,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        gen_result = pipe.generate([text], streamer=streamer, config=config)
         tps = gen_result.perf_metrics.get_throughput().mean
         tps_result = f"{tps:.2f} tokens/s"
         q.put(None)  # 結束符號