Skip to content

Commit

Permalink
feat: 通义千问支持流式输出，另外追加了 阿里云百炼支持的几个其他公司的大模型（百川、月之暗面、Yi）
Browse files Browse the repository at this point in the history
  • Loading branch information
@Ikaros-521
Ikaros-521 committed Aug 8, 2024
1 parent 9a7fc2a commit 7b8a9f2
Show file tree
Hide file tree
Showing 5 changed files with 96 additions and 17 deletions.
3 changes: 2 additions & 1 deletion config.json
View file
Original file line number Diff line number Diff line change
Expand Up @@ -432,7 +432,8 @@
"enable_search": true,
"max_tokens": 1024,
"history_enable": true,
"history_max_len": 300
"history_max_len": 300,
"stream": true
},
"tongyixingchen": {
"access_token": "此处填写你的密钥",
Expand Down
3 changes: 2 additions & 1 deletion config.json.bak
View file
Original file line number Diff line number Diff line change
Expand Up @@ -432,7 +432,8 @@
"enable_search": true,
"max_tokens": 1024,
"history_enable": true,
"history_max_len": 300
"history_max_len": 300,
"stream": true
},
"tongyixingchen": {
"access_token": "此处填写你的密钥",
Expand Down
42 changes: 40 additions & 2 deletions utils/gpt_model/tongyi.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -42,11 +42,12 @@ def __init__(self, data):
except Exception as e:
logger.error(traceback.format_exc())

def get_resp(self, prompt):
def get_resp(self, prompt, stream=False):
"""请求对应接口，获取返回值
Args:
prompt (str): 你的提问
stream (bool, optional): 是否流式返回. Defaults to False.
Returns:
str: 返回的文本回答
Expand All @@ -71,7 +72,7 @@ def get_resp(self, prompt):
from dashscope import Generation
from dashscope.api_entities.dashscope_response import Role

if self.config_data['history_enable'] == False:
if self.config_data['history_enable'] is False:
# 预设不能为空
if self.config_data["preset"] == "":
self.config_data["preset"] = "请做为一个人工智能，回答我的问题"
Expand All @@ -93,7 +94,16 @@ def get_resp(self, prompt):
top_k=self.config_data['top_k'],
enable_search=self.config_data['enable_search'],
max_tokens=self.config_data['max_tokens'],
stream=stream,
)

if response is None:
return None

if stream:
# 返回响应
return response

if response.status_code == HTTPStatus.OK:
logger.debug(response)

Expand Down Expand Up @@ -121,6 +131,34 @@ def get_resp(self, prompt):
logger.error(traceback.format_exc())
return None

# 添加AI返回消息到会话，用于提供上下文记忆
def add_assistant_msg_to_session(self, prompt, message):
try:
if self.config_data["type"] == "api":
from dashscope.api_entities.dashscope_response import Role

# 启用历史就给我记住！
if self.config_data['history_enable']:
self.history.append({'role': Role.USER, 'content': prompt})
self.history.append({'role': Role.ASSISTANT, 'content': message})
while True:
# 获取嵌套列表中所有字符串的字符数
total_chars = sum(len(item['content']) for item in self.history if 'content' in item)
# 如果大于限定最大历史数，就剔除第一个元素
if total_chars > int(self.config_data["history_max_len"]):
self.history.pop(0)
self.history.pop(0)
else:
break

logger.debug(f"history={self.history}")

return {"ret": True}

return {"ret": False}
except Exception as e:
logger.error(traceback.format_exc())
return {"ret": False}

if __name__ == '__main__':
# 配置日志输出格式
Expand Down
36 changes: 28 additions & 8 deletions utils/my_handle.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1606,6 +1606,7 @@ def llm_stream_handle_and_audio_synthesis(self, chat_type, data, type="chat", we
chat_model_methods = {
"chatgpt": lambda: self.chatgpt.get_gpt_resp(data["username"], data["content"], stream=True),
"zhipu": lambda: self.zhipu.get_resp(data["content"], stream=True),
"tongyi": lambda: self.tongyi.get_resp(data["content"], stream=True),
}
elif type == "vision":
pass
Expand All @@ -1628,21 +1629,35 @@ def split_by_chinese_punctuation(s):

if resp is not None:
tmp = ""
# 已经切掉的字符长度，针对一些特殊llm的流式输出，需要去掉前面的字符
cut_len = 0
for chunk in resp:
# logger.warning(chunk)
if chat_type in ["chatgpt", "zhipu"]:
# 流式的内容是追加形式的
tmp += chunk.choices[0].delta.content
resp_content += chunk.choices[0].delta.content
elif chat_type in ["tongyi"]:
# 这个是一直输出全部的内容，所以要切分掉已经处理的文本长度
tmp = chunk.output.choices[0].message.content[cut_len:]
resp_content = chunk.output.choices[0].message.content


# 用于切分，根据中文标点符号切分语句
resp_json = split_by_chinese_punctuation(tmp)
if resp_json["ret"]:
# 切出来的句子
tmp_content = resp_json["content1"]
# 标点符号后的内容包留，用于之后继续追加内容
tmp = resp_json["content2"]

logger.warning(f"句子生成：{tmp_content}")


if chat_type in ["chatgpt", "zhipu"]:
# 标点符号后的内容包留，用于之后继续追加内容
tmp = resp_json["content2"]
elif chat_type in ["tongyi"]:
# 记录 并追加切出的文本长度
cut_len += len(tmp_content)

"""
双重过滤，为您保驾护航
"""
Expand Down Expand Up @@ -1698,11 +1713,15 @@ def split_by_chinese_punctuation(s):

self.audio_synthesis_handle(message)

# logger.info(chunk)
if chunk.choices[0].finish_reason == "stop":
logger.info("流式接收完毕")
break

if chat_type in ["chatgpt", "zhipu"]:
# logger.info(chunk)
if chunk.choices[0].finish_reason == "stop":
logger.info("流式接收完毕")
break
elif chat_type in ["tongyi"]:
if chunk.output.choices[0].finish_reason == "stop":
logger.info("流式接收完毕")
break

# 返回为空，触发异常报警
else:
Expand All @@ -1720,6 +1739,7 @@ def split_by_chinese_punctuation(s):
chat_model_methods = {
"chatgpt": lambda: self.chatgpt.add_assistant_msg_to_session(data["username"], resp_content),
"zhipu": lambda: self.zhipu.add_assistant_msg_to_session(content_bak, resp_content),
"tongyi": lambda: self.tongyi.add_assistant_msg_to_session(content_bak, resp_content),
}
elif type == "vision":
pass
Expand Down
29 changes: 24 additions & 5 deletions webui.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2076,6 +2076,7 @@ def common_textarea_handle(content):
config_data["tongyi"]["enable_search"] = switch_tongyi_enable_search.value
config_data["tongyi"]["history_enable"] = switch_tongyi_history_enable.value
config_data["tongyi"]["history_max_len"] = int(input_tongyi_history_max_len.value)
config_data["tongyi"]["stream"] = switch_tongyi_stream.value

if config.get("webui", "show_card", "llm", "tongyixingchen"):
config_data["tongyixingchen"]["access_token"] = input_tongyixingchen_access_token.value
Expand Down Expand Up @@ -2942,7 +2943,7 @@ def save_config():
'qanything': 'QAnything',
'koboldcpp': 'koboldcpp',
'anythingllm': 'AnythingLLM',
'tongyi': '通义千问',
'tongyi': '通义千问/阿里云百炼',
'gpt4free': 'GPT4Free',
'dify': 'Dify',
'llm_tpu': 'LLM_TPU',
Expand Down Expand Up @@ -4487,7 +4488,7 @@ def anythingllm_get_workspaces_list():

if config.get("webui", "show_card", "llm", "tongyi"):
with ui.card().style(card_css):
ui.label("通义千问")
ui.label("通义千问/阿里云百炼")
with ui.row():
lines = ['web', 'api']
data_json = {}
Expand All @@ -4501,14 +4502,31 @@ def anythingllm_get_workspaces_list():
input_tongyi_cookie_path = ui.input(label='cookie路径', placeholder='web类型下，通义千问登录后，通过浏览器插件Cookie Editor获取Cookie JSON串，然后将数据保存在这个路径的文件中', value=config.get("tongyi", "cookie_path"))
input_tongyi_cookie_path.style("width:400px")
with ui.row():
lines = ['qwen-turbo', 'qwen-plus', 'qwen-max']
lines = [
'qwen-turbo',
'qwen-plus',
'qwen-long',
'qwen-max-longcontext',
'qwen-max',
'qwen-max-0428',
'baichuan2-turbo',
'moonshot-v1-8k',
'moonshot-v1-32k',
'moonshot-v1-128k',
'yi-large',
'yi-large-turbo',
'yi-medium',
]
data_json = {}
for line in lines:
data_json[line] = line
select_tongyi_model = ui.select(
label='类型',
options=data_json,
value=config.get("tongyi", "model")
value=config.get("tongyi", "model"),
with_input=True,
new_value_mode='add-unique',
clearable=True
).style("width:150px")
input_tongyi_api_key = ui.input(label='密钥', value=config.get("tongyi", "api_key"), placeholder='API类型下，DashScope平台申请的API密钥')
input_tongyi_preset = ui.input(label='预设', placeholder='API类型下，用于指定一组预定义的设置，以便模型更好地适应特定的对话场景。', value=config.get("tongyi", "preset")).style("width:500px")
Expand All @@ -4520,7 +4538,8 @@ def anythingllm_get_workspaces_list():
with ui.row():
switch_tongyi_history_enable = ui.switch('上下文记忆', value=config.get("tongyi", "history_enable")).style(switch_internal_css)
input_tongyi_history_max_len = ui.input(label='最大记忆长度', value=config.get("tongyi", "history_max_len"), placeholder='最长能记忆的问答字符串长度，超长会丢弃最早记忆的内容，请慎用！配置过大可能会有丢大米')

switch_tongyi_stream = ui.switch('流式输出', value=config.get("tongyi", "stream")).tooltip("是否开启流式输出，开启后，回答会逐句输出，关闭后，回答会一次性输出。")

if config.get("webui", "show_card", "llm", "gpt4free"):
with ui.card().style(card_css):
ui.label("GPT4Free")
Expand Down

0 comments on commit 7b8a9f2

Please sign in to comment.