refactor(LLM config): Restructure LLM configuration and management

config.yaml

@@ -1,11 +1,25 @@
 # * Settings marked with * are advanced settings that won't appear in the Streamlit page and can only be modified manually in config.py
 version: "2.0.0"
 ## ======================== Basic Settings ======================== ##
-# API settings
-api:
-  key: 'YOUR_API_KEY'
-  base_url: 'https://api.302.ai'
-  model: 'claude-3-5-sonnet-20240620'
+# LLM 模型配置部分
+llm_models:
+  default_model:  # 默认模型配置
+    key: 'YOUR_API_KEY'
+    base_url: 'https://api.302.ai'
+    model: 'claude-3-5-sonnet-20240620'
+  fast_model:  # 快速模型配置
+    key: 'YOUR_API_KEY'
+    base_url: 'https://api.302.ai'
+    model: 'Yi-Lightning'
+
+# 各阶段使用的模型配置
+llm_stages:
+  align: 'fast_model'  # 字幕对齐
+  split: 'fast_model'  # 字幕分割
+  summarize: 'default_model'  # 字幕总结
+  translate_faithfulness: 'default_model'  # 字幕翻译（精确）
+  translate_expressiveness: 'default_model'  # 字幕翻译（信达雅）
+  reduce: 'fast_model'  # 字幕缩减
 
 # Language settings, written into the prompt, can be described in natural language
 target_language: '简体中文'
@@ -34,7 +48,7 @@ subtitle:
   target_multiplier: 1.2
 
 # *Number of LLM multi-threaded accesses
-max_workers: 5
+max_workers: 8
 # *Maximum number of words for the first rough cut, below 18 will cut too finely affecting translation, above 22 is too long and will make subsequent subtitle splitting difficult to align
 max_split_length: 20
 
@@ -129,6 +143,7 @@ llm_support_json:
 - 'gemini-1.5-flash-latest'
 - 'gemini-1.5-pro-latest'
 - 'gemini-1.5-pro-002'
+- 'Yi-Lightning'
 
 # have problems
 # - 'Qwen/Qwen2.5-72B-Instruct'

core/ask_gpt.py

@@ -30,7 +30,7 @@ def save_log(model, prompt, response, log_title = 'default', message = None):
     with open(log_file, 'w', encoding='utf-8') as f:
         json.dump(logs, f, ensure_ascii=False, indent=4)
 
-def check_ask_gpt_history(prompt, model, log_title):
+def check_ask_gpt_history(prompt, log_title):
     # check if the prompt has been asked before
     if not os.path.exists(LOG_FOLDER):
         return False
@@ -43,11 +43,26 @@ def check_ask_gpt_history(prompt, model, log_title):
                     return item["response"]
     return False
 
-def ask_gpt(prompt, response_json=True, valid_def=None, log_title='default'):
-    api_set = load_key("api")
+def ask_gpt(prompt, response_json=True, valid_def=None, log_title='default', check_api=False, api_set=None):
     llm_support_json = load_key("llm_support_json")
+    if check_api:   
+        api_set = api_set
+    else:
+        if log_title == 'sentence_splitbymeaning':
+           api_set = load_key(f"llm_models.{load_key('llm_stages')['split']}")
+        elif log_title == 'align_subs':
+           api_set = load_key(f"llm_models.{load_key('llm_stages')['align']}")
+        elif log_title == 'summary':
+           api_set = load_key(f"llm_models.{load_key('llm_stages')['summarize']}")
+        elif log_title == 'translate_faithfulness':
+           api_set = load_key(f"llm_models.{load_key('llm_stages')['translate_faithfulness']}")
+        elif log_title == 'translate_expressiveness':
+           api_set = load_key(f"llm_models.{load_key('llm_stages')['translate_expressiveness']}")
+        elif log_title == 'subtitle_trim':
+           api_set = load_key(f"llm_models.{load_key('llm_stages')['reduce']}")
+
     with LOCK:
-        history_response = check_ask_gpt_history(prompt, api_set["model"], log_title)
+        history_response = check_ask_gpt_history(prompt, log_title)
         if history_response:
             return history_response
 
@@ -109,4 +124,4 @@ def ask_gpt(prompt, response_json=True, valid_def=None, log_title='default'):
 
 
 if __name__ == '__main__':
-    print(ask_gpt('hi there hey response in json format, just return 200.' , response_json=True, log_title=None))
+    print(ask_gpt('hi there hey response in json format, just return 200.' , response_json=True, log_title='None', api_set=load_key("llm_models.default_model"), check_api=True))

core/config_utils.py

@@ -26,25 +26,37 @@ def load_key(key: str) -> Any:
     return value
 
 def update_key(key: str, new_value: Any) -> bool:
+    """Update a key in the config file. If the key doesn't exist, it will be created.
+
+    Args:
+        key: Dot-separated key path (e.g. "llm_models.default_model.key")
+        new_value: Value to set
+
+    Returns:
+        bool: True if successful
+    """
     with config_lock:
         with open(CONFIG_PATH, 'r', encoding='utf-8') as file:
             data = yaml.load(file)
 
         keys = key.split('.')
         current = data
+
+        # 遍历除最后一个key外的所有key
         for k in keys[:-1]:
-            if isinstance(current, dict) and k in current:
-                current = current[k]
-            else:
-                return False
-
-        if isinstance(current, dict) and keys[-1] in current:
-            current[keys[-1]] = new_value
-            with open(CONFIG_PATH, 'w', encoding='utf-8') as file:
-                yaml.dump(data, file)
-            return True
-        else:
-            raise KeyError(f"Key '{keys[-1]}' not found in configuration")
+            # 如果key不存在，创建一个新的字典
+            if k not in current:
+                current[k] = {}
+            current = current[k]
+
+        # 设置最终的值
+        current[keys[-1]] = new_value
+
+        # 保存更新后的配置
+        with open(CONFIG_PATH, 'w', encoding='utf-8') as file:
+            yaml.dump(data, file)
+        return True
+
 
 # basic utils
 def get_joiner(language):

i18n/中文/config.yaml

@@ -1,11 +1,27 @@
 # * 标有 * 的设置是高级设置，不会出现在 Streamlit 页面中，只能在 config.py 中手动修改
 version: "2.0.0"
 ## ======================== 基本设置 ======================== ##
-# API 设置
-api:
-  key: 'YOUR_API_KEY'
-  base_url: 'https://api.siliconflow.cn'
-  model: 'Qwen/Qwen2.5-72B-Instruct'
+# LLM 模型配置部分
+llm_models:
+  default_model:  # 默认模型配置
+    key: 'YOUR_API_KEY'
+    base_url: 'https://api.302.ai'
+    model: 'claude-3-5-sonnet-20240620'
+  fast_model:  # 快速模型配置
+    key: 'YOUR_API_KEY'
+    base_url: 'https://api.302.ai'
+    model: 'Yi-Lightning'
+
+
+# 各阶段使用的模型配置
+llm_stages:
+  align: 'fast_model'  # 字幕对齐
+  split: 'fast_model'  # 字幕分割
+  summarize: 'default_model'  # 字幕总结
+  translate_faithfulness: 'default_model'  # 字幕翻译（精确）
+  translate_expressiveness: 'default_model'  # 字幕翻译（信达雅）
+  reduce: 'fast_model'  # 字幕缩减
+
 
 # 语言设置，写入提示词，可以用自然语言描述
 target_language: '简体中文'
@@ -129,6 +145,7 @@ llm_support_json:
 - 'gemini-1.5-flash-latest'
 - 'gemini-1.5-pro-latest'
 - 'gemini-1.5-pro-002'
+- 'Yi-Lightning'
 
 # 存在问题
 # - 'Qwen/Qwen2.5-72B-Instruct'
@@ -158,4 +175,4 @@ language_split_with_space:
 # 不使用空格作为分隔符的语言
 language_split_without_space:
 - 'zh'
-- 'ja'
+- 'ja'

i18n/中文/st_components/sidebar_setting.py

@@ -6,24 +6,91 @@
 
 def config_input(label, key, help=None):
     """Generic config input handler"""
-    val = st.text_input(label, value=load_key(key), help=help)
+    val = st.text_input(
+        label, 
+        value=load_key(key), 
+        help=help,
+        key=f"config_input_{key}"
+    )
     if val != load_key(key):
         update_key(key, val)
     return val
 
 def page_setting():
-    with st.expander("LLM 配置", expanded=True):
-        config_input("API_KEY", "api.key")
-        config_input("BASE_URL", "api.base_url", help="API请求的基础URL")
+    # LLM配置部分
+    st.subheader("LLM配置")
+
+    # 模型配置管理
+    st.subheader("模型配置")
+
+    # 获取现有模型列表
+    models = load_key("llm_models")
+
+    # 添加新模型按钮
+    if st.button("➕ 添加新模型"):
+        new_model_name = f"model_{len(models)}"
+        update_key(f"llm_models.{new_model_name}", {
+            "key": "",
+            "base_url": "",
+            "model": ""
+        })
+        st.rerun()
+
+    # 显示现有模型配置
+    for model_name in models:
+        st.markdown(f"### 📑 {model_name}")
+        # 模型名称编辑（可选）
+        new_name = st.text_input("模型名称", value=model_name, key=f"name_{model_name}")
 
-        c1, c2 = st.columns([4, 1])
-        with c1:
-            config_input("模型", "api.model")
+        # 基础配置
+        config_input(f"API密钥", f"llm_models.{model_name}.key")
+        config_input(f"基础URL", f"llm_models.{model_name}.base_url")
+        config_input(f"模型", f"llm_models.{model_name}.model")
+
+        # 测试和删除按钮
+        c1, c2, c3 = st.columns([3, 1, 1])
         with c2:
-            if st.button("📡", key="api"):
-                st.toast("API密钥有效" if check_api() else "API密钥无效", 
-                        icon="✅" if check_api() else "❌")
+            if st.button("🔍 测试", key=f"test_{model_name}"):
+                api_set = load_key(f"llm_models.{model_name}")
+                st.toast(
+                    "API密钥有效" if check_api(api_set) else "API密钥无效",
+                    icon="✅" if check_api(api_set) else "❌"
+                )
+        with c3:
+            if len(models) > 1 and st.button("🗑️ 删除", key=f"delete_{model_name}"):
+                models_dict = load_key("llm_models")
+                del models_dict[model_name]
+                update_key("llm_models", models_dict)
+                st.rerun()
+        st.divider()
+
+    # 阶段配置
+    st.subheader("阶段配置")
+    stages = {
+        "align": "字幕对齐",
+        "split": "字幕分割",
+        "summarize": "总结",
+        "translate_faithfulness": "翻译（精确）",
+        "translate_expressiveness": "翻译（优雅）",
+        "reduce": "字幕缩减"
+    }
 
+    for stage_key, stage_name in stages.items():
+        col1, col2 = st.columns([3, 2])
+        with col1:
+            st.write(stage_name)
+        with col2:
+            current_model = load_key(f"llm_stages.{stage_key}")
+            selected_model = st.selectbox(
+                "模型",
+                options=list(models.keys()),
+                index=list(models.keys()).index(current_model),
+                key=f"stage_{stage_key}",
+                label_visibility="collapsed"
+            )
+            if selected_model != current_model:
+                update_key(f"llm_stages.{stage_key}", selected_model)
+
     with st.expander("转写和字幕设置", expanded=True):
         c1, c2 = st.columns(2)
         with c1:
@@ -131,10 +198,16 @@ def page_setting():
             if selected_refer_mode != load_key("gpt_sovits.refer_mode"):
                 update_key("gpt_sovits.refer_mode", selected_refer_mode)
 
-def check_api():
+def check_api(api_set):
+    """检查API配置是否有效"""
     try:
-        resp = ask_gpt("This is a test, response 'message':'success' in json format.", 
-                      response_json=True, log_title='None')
+        resp = ask_gpt(
+            "This is a test, response 'message':'success' in json format.",
+            response_json=True,
+            log_title='None',
+            check_api=True,
+            api_set=api_set
+        )
         return resp.get('message') == 'success'
     except Exception:
         return False