Update app.py

app.py

@@ -6,10 +6,10 @@ import time
 
 import huggingface_hub as hf_hub
 # 初始化 OpenVINO 模型
-#model_id = "OpenVINO/Qwen3-4B-int8-ov"
-#model_path = "Qwen3-4B-int8-ov"
-model_id = "OpenVINO/Phi-4-mini-instruct-int4-ov"
-model_path = "Phi-4-mini-instruct-int4-ov"
+model_id = "OpenVINO/Qwen3-0.6B-int4-ov"
+model_path = "Qwen3-0.6B-int4-ov"
+#model_id = "OpenVINO/Phi-4-mini-instruct-int4-ov"
+#model_path = "Phi-4-mini-instruct-int4-ov"
 
 
 
@@ -40,7 +40,8 @@ def generate_stream(prompt):
     def worker():
         # 在背景 thread 中做推論
         nonlocal tps_result
-        gen_result = pipe.generate([prompt], streamer=streamer, max_new_tokens=4096)
+        prompt = prompt + "/no_think"
+        gen_result = pipe.generate([prompt], streamer=streamer, max_new_tokens=32767)
         tps = gen_result.perf_metrics.get_throughput().mean
         tps_result = f"{tps:.2f} tokens/s"        
         q.put(None)  # 結束符號