feat；模型选择现在可以使用完全随机的策略

Update model_config_template.toml

src/config/api_ada_configs.py

@@ -97,6 +97,9 @@ class TaskConfig(ConfigBase):
     slow_threshold: float = 15.0
     """慢请求阈值（秒），超过此值会输出警告日志"""
 
+    selection_strategy: str = field(default="balance")
+    """模型选择策略：balance（负载均衡）或 random（随机选择）"""
+
 
 @dataclass
 class ModelTaskConfig(ConfigBase):

src/llm_models/utils_model.py

@@ -1,6 +1,7 @@
 import re
 import asyncio
 import time
+import random
 
 from enum import Enum
 from rich.traceback import install
@@ -266,7 +267,7 @@ class LLMRequest:
 
     def _select_model(self, exclude_models: Optional[Set[str]] = None) -> Tuple[ModelInfo, APIProvider, BaseClient]:
         """
-        根据总tokens和惩罚值选择的模型
+        根据配置的策略选择模型：balance（负载均衡）或 random（随机选择）
         """
         available_models = {
             model: scores
@@ -276,15 +277,30 @@ class LLMRequest:
         if not available_models:
             raise RuntimeError("没有可用的模型可供选择。所有模型均已尝试失败。")
 
-        least_used_model_name = min(
+        strategy = self.model_for_task.selection_strategy.lower()
-            available_models,
+        
-            key=lambda k: available_models[k][0] + available_models[k][1] * 300 + available_models[k][2] * 1000,
+        if strategy == "random":
-        )
+            # 随机选择策略
-        model_info = model_config.get_model_info(least_used_model_name)
+            selected_model_name = random.choice(list(available_models.keys()))
+        elif strategy == "balance":
+            # 负载均衡策略：根据总tokens和惩罚值选择
+            selected_model_name = min(
+                available_models,
+                key=lambda k: available_models[k][0] + available_models[k][1] * 300 + available_models[k][2] * 1000,
+            )
+        else:
+            # 默认使用负载均衡策略
+            logger.warning(f"未知的选择策略 '{strategy}'，使用默认的负载均衡策略")
+            selected_model_name = min(
+                available_models,
+                key=lambda k: available_models[k][0] + available_models[k][1] * 300 + available_models[k][2] * 1000,
+            )
+        
+        model_info = model_config.get_model_info(selected_model_name)
         api_provider = model_config.get_provider(model_info.api_provider)
         force_new_client = self.request_type == "embedding"
         client = client_registry.get_client_class_instance(api_provider, force_new=force_new_client)
-        logger.debug(f"选择请求模型: {model_info.name}")
+        logger.debug(f"选择请求模型: {model_info.name} (策略: {strategy})")
         total_tokens, penalty, usage_penalty = self.model_usage[model_info.name]
         self.model_usage[model_info.name] = (total_tokens, penalty, usage_penalty + 1)
         return model_info, api_provider, client

template/model_config_template.toml

@@ -1,5 +1,5 @@
 [inner]
-version = "1.10.1"
+version = "1.11.0"
 
 # 配置文件版本号迭代规则同bot_config.toml
 
@@ -140,38 +140,45 @@ model_list = ["siliconflow-deepseek-v3.2"] # 使用的模型列表，每个子
 temperature = 0.2                        # 模型温度，新V3建议0.1-0.3
 max_tokens = 4096                         # 最大输出token数
 slow_threshold = 15.0                     # 慢请求阈值（秒），模型等待回复时间超过此值会输出警告日志
+selection_strategy = "random"           # 模型选择策略：random（负载均衡）或 random（随机选择）
 
 [model_task_config.tool_use] #功能模型，需要使用支持工具调用的模型，请使用较快的小模型（调用量较大）
 model_list = ["qwen3-30b","qwen3-next-80b"]
 temperature = 0.7
 max_tokens = 1024
 slow_threshold = 10.0
+selection_strategy = "random"           # 模型选择策略：random（负载均衡）或 random（随机选择）
 
 [model_task_config.replyer] # 首要回复模型，还用于表达方式学习
 model_list = ["siliconflow-deepseek-v3.2","siliconflow-deepseek-v3.2-think","siliconflow-glm-4.6","siliconflow-glm-4.6-think"]
 temperature = 0.3                        # 模型温度，新V3建议0.1-0.3
 max_tokens = 2048
 slow_threshold = 25.0
+selection_strategy = "random"           # 模型选择策略：random（负载均衡）或 random（随机选择）
 
 [model_task_config.planner] #决策：负责决定麦麦该什么时候回复的模型
 model_list = ["siliconflow-deepseek-v3.2"]
 temperature = 0.3
 max_tokens = 800
 slow_threshold = 12.0
+selection_strategy = "random"           # 模型选择策略：random（负载均衡）或 random（随机选择）
 
 [model_task_config.vlm] # 图像识别模型
 model_list = ["qwen3-vl-30"]
 max_tokens = 256
 slow_threshold = 15.0
+selection_strategy = "random"           # 模型选择策略：random（负载均衡）或 random（随机选择）
 
 [model_task_config.voice] # 语音识别模型
 model_list = ["sensevoice-small"]
 slow_threshold = 12.0
+selection_strategy = "random"           # 模型选择策略：random（负载均衡）或 random（随机选择）
 
 # 嵌入模型
 [model_task_config.embedding]
 model_list = ["bge-m3"]
 slow_threshold = 5.0
+selection_strategy = "random"           # 模型选择策略：random（负载均衡）或 random（随机选择）
 
 # ------------LPMM知识库模型------------
 
@@ -180,9 +187,11 @@ model_list = ["siliconflow-deepseek-v3.2"]
 temperature = 0.2
 max_tokens = 800
 slow_threshold = 20.0
+selection_strategy = "random"           # 模型选择策略：random（负载均衡）或 random（随机选择）
 
 [model_task_config.lpmm_rdf_build] # RDF构建模型
 model_list = ["siliconflow-deepseek-v3.2"]
 temperature = 0.2
 max_tokens = 800
 slow_threshold = 20.0
+selection_strategy = "random"           # 模型选择策略：random（负载均衡）或 random（随机选择）