Update app.py

app.py

@@ -1,22 +1,34 @@
+import os
+# 必须在 import torch 之前设置，防止线程超额分配
+os.environ["OMP_NUM_THREADS"] = "2"
+os.environ["MKL_NUM_THREADS"] = "2"
+os.environ["OPENBLAS_NUM_THREADS"] = "2"
+
 import torch
 import re
 import gradio as gr
 from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
 from threading import Thread
 
-# 1. 设置 CPU 线程数
+# 1. 设置 CPU 线程数（匹配 2 vCPU，避免线程竞争反而变慢）
 torch.set_num_threads(2)
 
-# 2. 加载模型（float32 兼容 CPU，0.5B 体积小速度快）
+# 2. 加载模型
 model_id = "Qwen/Qwen2.5-0.5B-Instruct"
 tokenizer = AutoTokenizer.from_pretrained(model_id)
 model = AutoModelForCausalLM.from_pretrained(
     model_id,
     dtype=torch.float32,
     device_map="cpu",
+    attn_implementation="eager",   # CPU 上 eager 比 sdpa 更快
+    use_cache=True,                # 启用 KV cache，避免重复计算
 )
 
-# 3. 流式生成函数
+# 3. 关闭梯度计算，减少内存和计算开销
+model.eval()
+
+# 4. 流式生成函数
+@torch.inference_mode()
 def stream_response(message, history):
     messages = [{"role": "system", "content": "你是一个乐于助人的 AI 助手。"}]
     for h in history:
@@ -43,10 +55,13 @@ def stream_response(message, history):
     generate_kwargs = {
         **model_inputs,
         "streamer": streamer,
-        "max_new_tokens": 512,
+        "max_new_tokens": 256,       # 从 512 降到 256，减少总计算量
         "do_sample": True,
         "temperature": 0.7,
+        "top_k": 40,                 # 限制候选词范围，减少采样计算
+        "top_p": 0.9,
         "pad_token_id": tokenizer.eos_token_id,
+        "use_cache": True,
     }
 
     thread = Thread(target=model.generate, kwargs=generate_kwargs)
@@ -58,7 +73,7 @@ def stream_response(message, history):
         clean_display_text = re.sub(r'[\x0a\-\*""""#]', '', partial_text)
         yield clean_display_text
 
-# 4. 界面设置
+# 5. 界面设置
 demo = gr.ChatInterface(
     fn=stream_response,
     title="Qwen2.5 流式对话助手",