From 00ff43625b09ad5d4b2f4222f246edc934bec954 Mon Sep 17 00:00:00 2001 From: foxplaying Date: Wed, 19 Nov 2025 18:21:40 +0800 Subject: [PATCH] =?UTF-8?q?=E9=80=82=E9=85=8DGemini3?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- requirements.txt | 4 +- src/llm_models/model_client/gemini_client.py | 69 ++++++++++++++------ template/model_config_template.toml | 23 ++++++- 3 files changed, 73 insertions(+), 23 deletions(-) diff --git a/requirements.txt b/requirements.txt index 8a0d22c2..37cb855d 100644 --- a/requirements.txt +++ b/requirements.txt @@ -3,7 +3,7 @@ aiohttp-cors>=0.8.1 colorama>=0.4.6 faiss-cpu>=1.11.0 fastapi>=0.116.0 -google-genai>=1.39.1 +google-genai>=1.51.0 jieba>=0.42.1 json-repair>=0.47.6 maim-message @@ -27,4 +27,4 @@ tomlkit>=0.13.3 urllib3>=2.5.0 uvicorn>=0.35.0 msgpack -zstandard \ No newline at end of file +zstandard diff --git a/src/llm_models/model_client/gemini_client.py b/src/llm_models/model_client/gemini_client.py index 444c5671..e630784a 100644 --- a/src/llm_models/model_client/gemini_client.py +++ b/src/llm_models/model_client/gemini_client.py @@ -256,15 +256,16 @@ def _build_stream_api_resp( reason = str(fr) break - if str(reason).endswith("MAX_TOKENS"): + if reason and "MAX_TOKENS" in reason: + model_dbg = getattr(last_resp, "model_version", None) or getattr(last_resp, "modelVersion", None) has_visible_output = bool(resp.content and resp.content.strip()) if has_visible_output: - logger.warning( - "⚠ Gemini 响应因达到 max_tokens 限制被部分截断，\n" - " 可能会对回复内容造成影响，建议修改模型 max_tokens 配置！" + logger.info( + f"模型 {model_dbg} 因达到 max_tokens 限制被部分截断，\n" + "可能会对回复内容造成影响，适当调宽模型 max_tokens 配置！" ) else: - logger.warning("⚠ Gemini 响应因达到 max_tokens 限制被截断，\n 请修改模型 max_tokens 配置！") + logger.warning(f"⚠ 模型 {model_dbg} 因达到 max_tokens 限制被截断，\n 请修改模型 max_tokens 配置！") if not resp.content and not resp.tool_calls: if not getattr(resp, "reasoning_content", None): @@ -396,13 +397,14 @@ def _default_normal_response_parser( if not has_real_output and getattr(resp, "text", None): has_real_output = True + model_dbg = getattr(resp, "model_version", None) or getattr(resp, "modelVersion", None) if has_real_output: - logger.warning( - "⚠ Gemini 响应因达到 max_tokens 限制被部分截断，\n" - " 可能会对回复内容造成影响，建议修改模型 max_tokens 配置！" + logger.info( + f"模型 {model_dbg} 因达到 max_tokens 限制被部分截断，\n" + "可能会对回复内容造成影响，适当调宽模型 max_tokens 配置！" ) else: - logger.warning("⚠ Gemini 响应因达到 max_tokens 限制被截断，\n 请修改模型 max_tokens 配置！") + logger.warning(f"⚠ 模型 {model_dbg} 因达到 max_tokens 限制被截断，\n 请修改模型 max_tokens 配置！") return api_response, _usage_record except Exception as e: @@ -448,6 +450,20 @@ class GeminiClient(BaseClient): """ 按模型限制思考预算范围，仅支持指定的模型（支持带数字后缀的新版本） """ + if model_id.startswith("gemini-3"): + # Gemini 3 使用 thinking_level + thinking_level = "high" # 默认高思考级别 + + if extra_params and "thinking_level" in extra_params: + tl = str(extra_params["thinking_level"]).lower() + if tl in ("low", "high"): + thinking_level = tl + else: + logger.warning(f"无效的 thinking_level: {tl}，已回退为默认值 high") + + # Gemini 3 系列模型不支持禁用思考功能 + return thinking_level + limits = None # 参数传入处理 @@ -543,7 +559,7 @@ class GeminiClient(BaseClient): # 将tool_options转换为Gemini API所需的格式 tools = _convert_tool_options(tool_options) if tool_options else None # 解析并裁剪 thinking_budget - tb = self.clamp_thinking_budget(extra_params, model_info.model_identifier) + think = self.clamp_thinking_budget(extra_params, model_info.model_identifier) # 检测是否为带 -search 的模型 enable_google_search = False model_identifier = model_info.model_identifier @@ -559,12 +575,20 @@ class GeminiClient(BaseClient): "max_output_tokens": max_tokens, "temperature": temperature, "response_modalities": ["TEXT"], - "thinking_config": ThinkingConfig( - include_thoughts=True, - thinking_budget=tb, - ), "safety_settings": gemini_safe_settings, # 防止空回复问题 } + + # 根据模型类型选择 thinking_config 参数 + thinking_kwargs = {"include_thoughts": True} + if model_identifier.startswith("gemini-3"): + thinking_kwargs["thinking_level"] = think + else: + thinking_kwargs["thinking_budget"] = think + try: + generation_config_dict["thinking_config"] = ThinkingConfig(**thinking_kwargs) + except Exception: + logger.warning("当前SDK不支持 thinking_level，请手动更新google-genai库") + generation_config_dict["thinking_config"] = ThinkingConfig(include_thoughts=True) if tools: generation_config_dict["tools"] = Tool(function_declarations=tools) if messages[1]: @@ -707,7 +731,7 @@ class GeminiClient(BaseClient): :return: 转录响应 """ # 解析并裁剪 thinking_budget - tb = self.clamp_thinking_budget(extra_params, model_info.model_identifier) + think = self.clamp_thinking_budget(extra_params, model_info.model_identifier) # 构造 prompt + 音频输入 prompt = "Generate a transcript of the speech. The language of the transcript should **match the language of the speech**." @@ -724,12 +748,19 @@ class GeminiClient(BaseClient): generation_config_dict = { "max_output_tokens": max_tokens, "response_modalities": ["TEXT"], - "thinking_config": ThinkingConfig( - include_thoughts=True, - thinking_budget=tb, - ), "safety_settings": gemini_safe_settings, } + + thinking_kwargs = {"include_thoughts": True} + if model_info.model_identifier.startswith("gemini-3"): + thinking_kwargs["thinking_level"] = think + else: + thinking_kwargs["thinking_budget"] = think + try: + generation_config_dict["thinking_config"] = ThinkingConfig(**thinking_kwargs) + except Exception: + logger.warning("当前SDK不支持 thinking_level，请手动更新google-genai库") + generation_config_dict["thinking_config"] = ThinkingConfig(include_thoughts=True) generate_content_config = GenerateContentConfig(**generation_config_dict) try: diff --git a/template/model_config_template.toml b/template/model_config_template.toml index 6d956ace..8603a33b 100644 --- a/template/model_config_template.toml +++ b/template/model_config_template.toml @@ -1,5 +1,5 @@ [inner] -version = "1.7.8" +version = "1.7.9" # 配置文件版本号迭代规则同bot_config.toml @@ -66,6 +66,26 @@ price_out = 3.0 [models.extra_params] # 可选的额外参数配置 enable_thinking = true # 不启用思考 +[[models]] +model_identifier = "gemini-3-pro-preview" +name = "google-gemini-3-pro-preview" +api_provider = "Google" +price_in = 0.0 +price_out = 0.0 +[models.extra_params] +thinking_level = "high" # 默认高预算 + +[[models]] +model_identifier = "gemini-2.5-pro" +name = "google-gemini-2.5-pro" +api_provider = "Google" +price_in = 0.0 +price_out = 0.0 +[models.extra_params] +thinking_budget = -1 # 由模型自己决定 +# thinking_budget = 0 禁用思考(如果可能) + + [[models]] model_identifier = "deepseek-ai/DeepSeek-R1" name = "siliconflow-deepseek-r1" @@ -73,7 +93,6 @@ api_provider = "SiliconFlow" price_in = 4.0 price_out = 16.0 - [[models]] model_identifier = "Qwen/Qwen3-30B-A3B-Instruct-2507" name = "qwen3-30b"