fix:尝试修复炸飞问题
This commit is contained in:
SengokuCola
@@ -198,156 +198,195 @@ class LLM_request:
|
||||
headers["Accept"] = "text/event-stream"
|
||||
|
||||
async with aiohttp.ClientSession() as session:
|
||||
async with session.post(api_url, headers=headers, json=payload) as response:
|
||||
# 处理需要重试的状态码
|
||||
if response.status in policy["retry_codes"]:
|
||||
wait_time = policy["base_wait"] * (2**retry)
|
||||
logger.warning(f"错误码: {response.status}, 等待 {wait_time}秒后重试")
|
||||
if response.status == 413:
|
||||
logger.warning("请求体过大，尝试压缩...")
|
||||
image_base64 = compress_base64_image_by_scale(image_base64)
|
||||
payload = await self._build_payload(prompt, image_base64, image_format)
|
||||
elif response.status in [500, 503]:
|
||||
logger.error(f"错误码: {response.status} - {error_code_mapping.get(response.status)}")
|
||||
raise RuntimeError("服务器负载过高，模型恢复失败QAQ")
|
||||
else:
|
||||
logger.warning(f"请求限制(429)，等待{wait_time}秒后重试...")
|
||||
|
||||
await asyncio.sleep(wait_time)
|
||||
continue
|
||||
elif response.status in policy["abort_codes"]:
|
||||
logger.error(f"错误码: {response.status} - {error_code_mapping.get(response.status)}")
|
||||
# 尝试获取并记录服务器返回的详细错误信息
|
||||
try:
|
||||
error_json = await response.json()
|
||||
if error_json and isinstance(error_json, list) and len(error_json) > 0:
|
||||
for error_item in error_json:
|
||||
if "error" in error_item and isinstance(error_item["error"], dict):
|
||||
error_obj = error_item["error"]
|
||||
error_code = error_obj.get("code")
|
||||
error_message = error_obj.get("message")
|
||||
error_status = error_obj.get("status")
|
||||
logger.error(
|
||||
f"服务器错误详情: 代码={error_code}, 状态={error_status}, "
|
||||
f"消息={error_message}"
|
||||
)
|
||||
elif isinstance(error_json, dict) and "error" in error_json:
|
||||
# 处理单个错误对象的情况
|
||||
error_obj = error_json.get("error", {})
|
||||
error_code = error_obj.get("code")
|
||||
error_message = error_obj.get("message")
|
||||
error_status = error_obj.get("status")
|
||||
logger.error(
|
||||
f"服务器错误详情: 代码={error_code}, 状态={error_status}, 消息={error_message}"
|
||||
)
|
||||
try:
|
||||
async with session.post(api_url, headers=headers, json=payload) as response:
|
||||
# 处理需要重试的状态码
|
||||
if response.status in policy["retry_codes"]:
|
||||
wait_time = policy["base_wait"] * (2**retry)
|
||||
logger.warning(f"错误码: {response.status}, 等待 {wait_time}秒后重试")
|
||||
if response.status == 413:
|
||||
logger.warning("请求体过大，尝试压缩...")
|
||||
image_base64 = compress_base64_image_by_scale(image_base64)
|
||||
payload = await self._build_payload(prompt, image_base64, image_format)
|
||||
elif response.status in [500, 503]:
|
||||
logger.error(f"错误码: {response.status} - {error_code_mapping.get(response.status)}")
|
||||
raise RuntimeError("服务器负载过高，模型恢复失败QAQ")
|
||||
else:
|
||||
# 记录原始错误响应内容
|
||||
logger.error(f"服务器错误响应: {error_json}")
|
||||
except Exception as e:
|
||||
logger.warning(f"无法解析服务器错误响应: {str(e)}")
|
||||
logger.warning(f"请求限制(429)，等待{wait_time}秒后重试...")
|
||||
|
||||
if response.status == 403:
|
||||
# 只针对硅基流动的V3和R1进行降级处理
|
||||
if (
|
||||
self.model_name.startswith("Pro/deepseek-ai")
|
||||
and self.base_url == "https://api.siliconflow.cn/v1/"
|
||||
):
|
||||
old_model_name = self.model_name
|
||||
self.model_name = self.model_name[4:] # 移除"Pro/"前缀
|
||||
logger.warning(f"检测到403错误，模型从 {old_model_name} 降级为 {self.model_name}")
|
||||
|
||||
# 对全局配置进行更新
|
||||
if global_config.llm_normal.get("name") == old_model_name:
|
||||
global_config.llm_normal["name"] = self.model_name
|
||||
logger.warning(f"将全局配置中的 llm_normal 模型临时降级至{self.model_name}")
|
||||
|
||||
if global_config.llm_reasoning.get("name") == old_model_name:
|
||||
global_config.llm_reasoning["name"] = self.model_name
|
||||
logger.warning(f"将全局配置中的 llm_reasoning 模型临时降级至{self.model_name}")
|
||||
|
||||
# 更新payload中的模型名
|
||||
if payload and "model" in payload:
|
||||
payload["model"] = self.model_name
|
||||
|
||||
# 重新尝试请求
|
||||
retry -= 1 # 不计入重试次数
|
||||
continue
|
||||
|
||||
raise RuntimeError(f"请求被拒绝: {error_code_mapping.get(response.status)}")
|
||||
|
||||
response.raise_for_status()
|
||||
reasoning_content = ""
|
||||
|
||||
# 将流式输出转化为非流式输出
|
||||
if stream_mode:
|
||||
flag_delta_content_finished = False
|
||||
accumulated_content = ""
|
||||
usage = None # 初始化usage变量，避免未定义错误
|
||||
|
||||
async for line_bytes in response.content:
|
||||
await asyncio.sleep(wait_time)
|
||||
continue
|
||||
elif response.status in policy["abort_codes"]:
|
||||
logger.error(f"错误码: {response.status} - {error_code_mapping.get(response.status)}")
|
||||
# 尝试获取并记录服务器返回的详细错误信息
|
||||
try:
|
||||
line = line_bytes.decode("utf-8").strip()
|
||||
if not line:
|
||||
error_json = await response.json()
|
||||
if error_json and isinstance(error_json, list) and len(error_json) > 0:
|
||||
for error_item in error_json:
|
||||
if "error" in error_item and isinstance(error_item["error"], dict):
|
||||
error_obj = error_item["error"]
|
||||
error_code = error_obj.get("code")
|
||||
error_message = error_obj.get("message")
|
||||
error_status = error_obj.get("status")
|
||||
logger.error(
|
||||
f"服务器错误详情: 代码={error_code}, 状态={error_status}, "
|
||||
f"消息={error_message}"
|
||||
)
|
||||
elif isinstance(error_json, dict) and "error" in error_json:
|
||||
# 处理单个错误对象的情况
|
||||
error_obj = error_json.get("error", {})
|
||||
error_code = error_obj.get("code")
|
||||
error_message = error_obj.get("message")
|
||||
error_status = error_obj.get("status")
|
||||
logger.error(
|
||||
f"服务器错误详情: 代码={error_code}, 状态={error_status}, 消息={error_message}"
|
||||
)
|
||||
else:
|
||||
# 记录原始错误响应内容
|
||||
logger.error(f"服务器错误响应: {error_json}")
|
||||
except Exception as e:
|
||||
logger.warning(f"无法解析服务器错误响应: {str(e)}")
|
||||
|
||||
if response.status == 403:
|
||||
# 只针对硅基流动的V3和R1进行降级处理
|
||||
if (
|
||||
self.model_name.startswith("Pro/deepseek-ai")
|
||||
and self.base_url == "https://api.siliconflow.cn/v1/"
|
||||
):
|
||||
old_model_name = self.model_name
|
||||
self.model_name = self.model_name[4:] # 移除"Pro/"前缀
|
||||
logger.warning(f"检测到403错误，模型从 {old_model_name} 降级为 {self.model_name}")
|
||||
|
||||
# 对全局配置进行更新
|
||||
if global_config.llm_normal.get("name") == old_model_name:
|
||||
global_config.llm_normal["name"] = self.model_name
|
||||
logger.warning(f"将全局配置中的 llm_normal 模型临时降级至{self.model_name}")
|
||||
|
||||
if global_config.llm_reasoning.get("name") == old_model_name:
|
||||
global_config.llm_reasoning["name"] = self.model_name
|
||||
logger.warning(f"将全局配置中的 llm_reasoning 模型临时降级至{self.model_name}")
|
||||
|
||||
# 更新payload中的模型名
|
||||
if payload and "model" in payload:
|
||||
payload["model"] = self.model_name
|
||||
|
||||
# 重新尝试请求
|
||||
retry -= 1 # 不计入重试次数
|
||||
continue
|
||||
if line.startswith("data:"):
|
||||
data_str = line[5:].strip()
|
||||
if data_str == "[DONE]":
|
||||
break
|
||||
try:
|
||||
chunk = json.loads(data_str)
|
||||
if flag_delta_content_finished:
|
||||
chunk_usage = chunk.get("usage", None)
|
||||
if chunk_usage:
|
||||
usage = chunk_usage # 获取token用量
|
||||
else:
|
||||
delta = chunk["choices"][0]["delta"]
|
||||
delta_content = delta.get("content")
|
||||
if delta_content is None:
|
||||
delta_content = ""
|
||||
accumulated_content += delta_content
|
||||
# 检测流式输出文本是否结束
|
||||
finish_reason = chunk["choices"][0].get("finish_reason")
|
||||
if delta.get("reasoning_content", None):
|
||||
reasoning_content += delta["reasoning_content"]
|
||||
if finish_reason == "stop":
|
||||
|
||||
raise RuntimeError(f"请求被拒绝: {error_code_mapping.get(response.status)}")
|
||||
|
||||
response.raise_for_status()
|
||||
reasoning_content = ""
|
||||
|
||||
# 将流式输出转化为非流式输出
|
||||
if stream_mode:
|
||||
flag_delta_content_finished = False
|
||||
accumulated_content = ""
|
||||
usage = None # 初始化usage变量，避免未定义错误
|
||||
|
||||
async for line_bytes in response.content:
|
||||
try:
|
||||
line = line_bytes.decode("utf-8").strip()
|
||||
if not line:
|
||||
continue
|
||||
if line.startswith("data:"):
|
||||
data_str = line[5:].strip()
|
||||
if data_str == "[DONE]":
|
||||
break
|
||||
try:
|
||||
chunk = json.loads(data_str)
|
||||
if flag_delta_content_finished:
|
||||
chunk_usage = chunk.get("usage", None)
|
||||
if chunk_usage:
|
||||
usage = chunk_usage
|
||||
break
|
||||
# 部分平台在文本输出结束前不会返回token用量，此时需要再获取一次chunk
|
||||
flag_delta_content_finished = True
|
||||
usage = chunk_usage # 获取token用量
|
||||
else:
|
||||
delta = chunk["choices"][0]["delta"]
|
||||
delta_content = delta.get("content")
|
||||
if delta_content is None:
|
||||
delta_content = ""
|
||||
accumulated_content += delta_content
|
||||
# 检测流式输出文本是否结束
|
||||
finish_reason = chunk["choices"][0].get("finish_reason")
|
||||
if delta.get("reasoning_content", None):
|
||||
reasoning_content += delta["reasoning_content"]
|
||||
if finish_reason == "stop":
|
||||
chunk_usage = chunk.get("usage", None)
|
||||
if chunk_usage:
|
||||
usage = chunk_usage
|
||||
break
|
||||
# 部分平台在文本输出结束前不会返回token用量，此时需要再获取一次chunk
|
||||
flag_delta_content_finished = True
|
||||
|
||||
except Exception as e:
|
||||
logger.exception(f"解析流式输出错误: {str(e)}")
|
||||
except GeneratorExit:
|
||||
logger.warning("流式输出被中断")
|
||||
break
|
||||
except Exception as e:
|
||||
logger.error(f"处理流式输出时发生错误: {str(e)}")
|
||||
break
|
||||
content = accumulated_content
|
||||
think_match = re.search(r"<think>(.*?)</think>", content, re.DOTALL)
|
||||
if think_match:
|
||||
reasoning_content = think_match.group(1).strip()
|
||||
content = re.sub(r"<think>.*?</think>", "", content, flags=re.DOTALL).strip()
|
||||
# 构造一个伪result以便调用自定义响应处理器或默认处理器
|
||||
result = {
|
||||
"choices": [{"message": {"content": content, "reasoning_content": reasoning_content}}],
|
||||
"usage": usage,
|
||||
}
|
||||
return (
|
||||
response_handler(result)
|
||||
if response_handler
|
||||
else self._default_response_handler(result, user_id, request_type, endpoint)
|
||||
)
|
||||
except Exception as e:
|
||||
logger.exception(f"解析流式输出错误: {str(e)}")
|
||||
except GeneratorExit:
|
||||
logger.warning("流式输出被中断，正在清理资源...")
|
||||
# 确保资源被正确清理
|
||||
await response.release()
|
||||
# 返回已经累积的内容
|
||||
result = {
|
||||
"choices": [{"message": {"content": accumulated_content, "reasoning_content": reasoning_content}}],
|
||||
"usage": usage,
|
||||
}
|
||||
return (
|
||||
response_handler(result)
|
||||
if response_handler
|
||||
else self._default_response_handler(result, user_id, request_type, endpoint)
|
||||
)
|
||||
except Exception as e:
|
||||
logger.error(f"处理流式输出时发生错误: {str(e)}")
|
||||
# 确保在发生错误时也能正确清理资源
|
||||
try:
|
||||
await response.release()
|
||||
except Exception as cleanup_error:
|
||||
logger.error(f"清理资源时发生错误: {cleanup_error}")
|
||||
# 返回已经累积的内容
|
||||
result = {
|
||||
"choices": [{"message": {"content": accumulated_content, "reasoning_content": reasoning_content}}],
|
||||
"usage": usage,
|
||||
}
|
||||
return (
|
||||
response_handler(result)
|
||||
if response_handler
|
||||
else self._default_response_handler(result, user_id, request_type, endpoint)
|
||||
)
|
||||
content = accumulated_content
|
||||
think_match = re.search(r"<think>(.*?)</think>", content, re.DOTALL)
|
||||
if think_match:
|
||||
reasoning_content = think_match.group(1).strip()
|
||||
content = re.sub(r"<think>.*?</think>", "", content, flags=re.DOTALL).strip()
|
||||
# 构造一个伪result以便调用自定义响应处理器或默认处理器
|
||||
result = {
|
||||
"choices": [{"message": {"content": content, "reasoning_content": reasoning_content}}],
|
||||
"usage": usage,
|
||||
}
|
||||
return (
|
||||
response_handler(result)
|
||||
if response_handler
|
||||
else self._default_response_handler(result, user_id, request_type, endpoint)
|
||||
)
|
||||
else:
|
||||
result = await response.json()
|
||||
# 使用自定义处理器或默认处理
|
||||
return (
|
||||
response_handler(result)
|
||||
if response_handler
|
||||
else self._default_response_handler(result, user_id, request_type, endpoint)
|
||||
)
|
||||
|
||||
except (aiohttp.ClientError, asyncio.TimeoutError) as e:
|
||||
if retry < policy["max_retries"] - 1:
|
||||
wait_time = policy["base_wait"] * (2**retry)
|
||||
logger.error(f"网络错误，等待{wait_time}秒后重试... 错误: {str(e)}")
|
||||
await asyncio.sleep(wait_time)
|
||||
continue
|
||||
else:
|
||||
result = await response.json()
|
||||
# 使用自定义处理器或默认处理
|
||||
return (
|
||||
response_handler(result)
|
||||
if response_handler
|
||||
else self._default_response_handler(result, user_id, request_type, endpoint)
|
||||
)
|
||||
logger.critical(f"网络错误达到最大重试次数: {str(e)}")
|
||||
raise RuntimeError(f"网络请求失败: {str(e)}") from e
|
||||
except Exception as e:
|
||||
logger.critical(f"未预期的错误: {str(e)}")
|
||||
raise RuntimeError(f"请求过程中发生错误: {str(e)}") from e
|
||||
|
||||
except aiohttp.ClientResponseError as e:
|
||||
# 处理aiohttp抛出的响应错误
|
||||
|
||||
Reference in New Issue
Block a user