From 915d1ef2b0cdde1f553eeb73d0feac401711202a Mon Sep 17 00:00:00 2001 From: Liskarm Date: Mon, 26 Jan 2026 04:44:09 +0800 Subject: [PATCH] feat: add sequential model selection strategy - Add 'sequential' strategy to select models in order - Only try next model when current one is unavailable - Update frontend UI to include sequential option - Update config template with new strategy documentation --- .../config/model/components/TaskConfigCard.tsx | 3 ++- src/config/api_ada_configs.py | 2 +- src/llm_models/utils_model.py | 11 ++++++++++- template/model_config_template.toml | 18 +++++++++--------- 4 files changed, 22 insertions(+), 12 deletions(-) diff --git a/dashboard/src/routes/config/model/components/TaskConfigCard.tsx b/dashboard/src/routes/config/model/components/TaskConfigCard.tsx index 51c63b1c..0311e3d0 100644 --- a/dashboard/src/routes/config/model/components/TaskConfigCard.tsx +++ b/dashboard/src/routes/config/model/components/TaskConfigCard.tsx @@ -143,10 +143,11 @@ export const TaskConfigCard = React.memo(function TaskConfigCard({ 负载均衡（balance） 随机选择（random） + 顺序优先（sequential）

- 负载均衡：优先选择使用次数少的模型。随机选择：完全随机从模型列表中选择 + 负载均衡：优先选择使用次数少的模型。随机选择：完全随机选择。顺序优先：按列表顺序，前者不可用才用后者

diff --git a/src/config/api_ada_configs.py b/src/config/api_ada_configs.py index f450e91f..8f45635c 100644 --- a/src/config/api_ada_configs.py +++ b/src/config/api_ada_configs.py @@ -98,7 +98,7 @@ class TaskConfig(ConfigBase): """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）""" + """模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先）""" @dataclass diff --git a/src/llm_models/utils_model.py b/src/llm_models/utils_model.py index 7fe46fd5..861611ab 100644 --- a/src/llm_models/utils_model.py +++ b/src/llm_models/utils_model.py @@ -267,7 +267,7 @@ class LLMRequest: def _select_model(self, exclude_models: Optional[Set[str]] = None) -> Tuple[ModelInfo, APIProvider, BaseClient]: """ - 根据配置的策略选择模型：balance（负载均衡）或 random（随机选择） + 根据配置的策略选择模型：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） """ available_models = { model: scores @@ -282,6 +282,15 @@ class LLMRequest: if strategy == "random": # 随机选择策略 selected_model_name = random.choice(list(available_models.keys())) + elif strategy == "sequential": + # 顺序优先策略：按照model_list中的顺序，选择第一个可用的模型 + selected_model_name = None + for model_name in self.model_for_task.model_list: + if model_name in available_models: + selected_model_name = model_name + break + if selected_model_name is None: + raise RuntimeError("没有可用的模型可供选择。所有模型均已尝试失败。") elif strategy == "balance": # 负载均衡策略：根据总tokens和惩罚值选择 selected_model_name = min( diff --git a/template/model_config_template.toml b/template/model_config_template.toml index 4e97cdb3..d5bc4ec3 100644 --- a/template/model_config_template.toml +++ b/template/model_config_template.toml @@ -140,45 +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：balance（负载均衡）或 random（随机选择） +selection_strategy = "random" # 模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） [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：balance（负载均衡）或 random（随机选择） +selection_strategy = "random" # 模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） [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：balance（负载均衡）或 random（随机选择） +selection_strategy = "random" # 模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） [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：balance（负载均衡）或 random（随机选择） +selection_strategy = "random" # 模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） [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：balance（负载均衡）或 random（随机选择） +selection_strategy = "random" # 模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） [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：balance（负载均衡）或 random（随机选择） +selection_strategy = "random" # 模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） # 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：balance（负载均衡）或 random（随机选择） +selection_strategy = "random" # 模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） # ------------LPMM知识库模型------------ @@ -187,11 +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：balance（负载均衡）或 random（随机选择） +selection_strategy = "random" # 模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） [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：balance（负载均衡）或 random（随机选择） \ No newline at end of file +selection_strategy = "random" # 模型选择策略：balance（负载均衡）、random（随机选择）或 sequential（顺序优先） \ No newline at end of file