gardel/Mofox-Core
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: 流式输出模式增加结束判断与token用量记录

Browse Source
This commit is contained in:
pine
2025-03-11 18:51:28 +08:00
parent 2e3411f103
commit c24bb70291
1 changed files with 20 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/plugins/models/utils_model.py
View File

@@ -216,6 +216,7 @@ class LLM_request:
# 将流式输出转化为非流式输出
if stream_mode:
flag_delta_content_finished = False
accumulated_content = ""
async for line_bytes in response.content:
line = line_bytes.decode("utf-8").strip()
@@ -227,13 +228,25 @@ class LLM_request:
break
try:
chunk = json.loads(data_str)
if flag_delta_content_finished:
usage = chunk.get("usage", None) # 获取tokn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]["finish_reason"]
if finish_reason == "stop":
usage = chunk.get("usage", None)
if 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:
logger.exception("解析流式输出错")
logger.exception("解析流式输出错误")
content = accumulated_content
reasoning_content = ""
think_match = re.search(r'<think>(.*?)</think>', content, re.DOTALL)
@@ -242,7 +255,7 @@ class LLM_request:
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}}]}
"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: