Update app.py

app.py

@@ -154,19 +154,16 @@ class DAggerEncoder:
             print(f"⚠️ 上传失败: {e}")
 
 # ==========================================
-# [单局 DAgger 引擎]
+# [单局 DAgger 引擎] (已剥离模型加载)
 # ==========================================
-def play_dagger_game(encoder: DAggerEncoder):
+def play_dagger_game(encoder: DAggerEncoder, student_bots: dict, teacher_bots: dict):
     env = RiichiEnv(game_mode="3p-red-half", rule=GameRule.default_tenhou())
     
-    # 每次开局，必须重新实例化 Bot，防止底层 Rust 状态机污染
-    # Student 负责打牌，Teacher 负责指点
-    student_bots = {i: load_model_student(i, STUDENT_MODEL) for i in range(3)}
-    teacher_bots = {i: load_model_teacher(i, TEACHER_MODEL) for i in range(3)}
-    
-    # 纯 Python 状态机，负责截获 Student 视角的 1012维 Obs
+    # ⚠️ Python 层的特征状态机极其轻量，为了绝对安全，每局仍然重新实例化
     python_states = {i: sanma_state.PlayerState(i) for i in range(3)}
     
+    # env.reset() 会生成包含 "start_game" 的初始事件流
+    # 底层的 Rust Bot 收到 "start_game" 后会自动清空上一局的残余记忆
     obs_dict = env.reset()
     
     while not env.done():
@@ -181,10 +178,10 @@ def play_dagger_game(encoder: DAggerEncoder):
                 # 1. 更新 Python 状态机以生成 1012 维特征
                 cans = python_states[pid].update(event_patched)
                 
-                # 2. 幽灵教师思考 (但不参与实战控制)
+                # 2. 幽灵教师思考 (常驻内存的 Bot 接收事件)
                 teacher_resp_str = teacher_bots[pid].react(event_patched)
                 
-                # 3. 如果到了需要决策的节点，进行 DAgger 抓取
+                # 3. 抓取 DAgger 数据
                 if cans.can_act:
                     obs_s, mask_s = python_states[pid].encode_obs(4, False)
                     
@@ -193,16 +190,15 @@ def play_dagger_game(encoder: DAggerEncoder):
                         action_dict = json.loads(teacher_resp) if teacher_resp else None
                         label_idx = action_to_label(pid, action_dict)
                         
-                        # 只有当存在有效决策空间时，才记录
                         if int(np.count_nonzero(mask_s)) > 1:
                             encoder.append(obs_s, mask_s, label_idx)
                     except Exception:
                         pass
                 
-                # 4. 学生思考 (真正掌控局面)
+                # 4. 学生思考 (常驻内存的 Bot 接收事件)
                 student_resp_str = student_bots[pid].react(event_patched)
                 
-                # 提取学生最后一次有效响应，喂给环境
+                # 提取学生动作
                 if student_resp_str and '"type":"none"' not in student_resp_str.replace(' ', ''):
                     student_final_action = env_obs.select_action_from_mjai(patch_resp_fast(student_resp_str))
             
@@ -211,27 +207,36 @@ def play_dagger_game(encoder: DAggerEncoder):
                 
             actions_to_env[pid] = student_final_action
 
-        # 环境按照学生(劣势)的动作继续
+        # 环境步进
         obs_dict = env.step(actions_to_env)
         
     worker_status["games_played"] += 1
 
 # ==========================================
-# [后台循环]
+# [后台循环] (常驻内存优化版)
 # ==========================================
 def background_dagger_loop():
     sync_models_from_hub()
     encoder = DAggerEncoder(chunk_size=2048)
-    worker_status["status"] = "Generating DAgger Data..."
     
-    # 限制单线程跑牌，防止内存爆炸，反正是在 HF Space 白嫖
+    worker_status["status"] = "Loading Models into Memory (One-time)..."
+    print("🧠 正在将 Student 和 Teacher 模型加载进内存 (常驻)...")
+    
+    # 🚀 提速核心：在这里只加载一次，然后反复复用！
+    # 只要保证每局开头环境喂给它们 "start_game"，Rust 层就会自动初始化
+    student_bots = {i: load_model_student(i, STUDENT_MODEL) for i in range(3)}
+    teacher_bots = {i: load_model_teacher(i, TEACHER_MODEL) for i in range(3)}
+    
+    worker_status["status"] = "Generating DAgger Data at High Speed..."
+    print("⚡ 模型加载完毕，数据工厂开始全速运转！")
+    
     while EVAL_RUNNING:
         try:
-            play_dagger_game(encoder)
+            # 将内存中的 Bot 引用传给打牌函数
+            play_dagger_game(encoder, student_bots, teacher_bots)
         except Exception as e:
             print(f"Game crashed: {e}")
             traceback.print_exc()
-            time.sleep(1)
 
 # ==========================================
 # [前端 UI]