diff --git a/apps/multi_roles_chat_room/app.py b/apps/multi_roles_chat_room/app.py index dc4f497b5..2cdb96f24 100644 --- a/apps/multi_roles_chat_room/app.py +++ b/apps/multi_roles_chat_room/app.py @@ -1,13 +1,17 @@ +import os import re +from typing import Optional import gradio as gr import json import modelscope_studio as mgr +from modelscope_agent.multi_agents_utils.executors.ray import RayTaskExecutor from role_core import (chat_progress, init_all_remote_actors, start_chat_with_topic) from story_holder import get_avatar_by_name, get_story_by_id, stories chat_history = [] +RayTaskExecutor.init_ray() # 发送消息的函数 @@ -54,6 +58,17 @@ def end_topic(): return '', 'topic ended。' +def check_uuid(uid: Optional[str]) -> str: + """Checks whether a UUID is provided or generates a default one.""" + if not uid or uid == '': + if os.getenv('MODELSCOPE_ENVIRONMENT') == 'studio': + import gradio as gr + + raise gr.Error('Please login first') + uid = 'local_user' + return uid + + def format_entry_html(): base_html = '''
@@ -89,6 +104,8 @@ def format_entry_html(): with demo: state = gr.State({}) story_state = gr.State('1') # default value + uuid = gr.Textbox(label='modelscope_uuid', visible=False) + with gr.Column(visible=True) as entry: gr.Markdown('## 选择一个场景进入聊天吧～') entry_btn = gr.Button( @@ -197,10 +214,13 @@ def back(): back_btn.click(fn=back, inputs=[], outputs=[entry, content]) def start_chat(username, from_user, topic, _state, _chatbot, _input, - _story_state, select_model): + _story_state, select_model, _uuid): + # get uuid + _uuid = check_uuid(_uuid) + user_task_id = f'{_uuid[:6]}_{_story_state}' roles = get_story_by_id(_story_state)['roles'] _state = init_all_remote_actors(roles, username, _state, _story_state, - select_model) + select_model, user_task_id) _state = start_chat_with_topic(from_user, topic, _state) init_chat = [{ @@ -328,7 +348,7 @@ def send_message(_chatbot, _input, _state): fn=start_chat, inputs=[ user_select, start_topic_from, start_topic_input, state, - user_chatbot, preview_chat_input, story_state, select_model + user_chatbot, preview_chat_input, story_state, select_model, uuid ], outputs=[user_chatbot, preview_chat_input, state]) diff --git a/apps/multi_roles_chat_room/requirements.txt b/apps/multi_roles_chat_room/requirements.txt index 143544549..51a17116e 100644 --- a/apps/multi_roles_chat_room/requirements.txt +++ b/apps/multi_roles_chat_room/requirements.txt @@ -1 +1,3 @@ gradio==4.8 +modelscope-agent==0.4.1 +modelscope_studio diff --git a/apps/multi_roles_chat_room/role_core.py b/apps/multi_roles_chat_room/role_core.py index 8c2b314a9..821d038d3 100644 --- a/apps/multi_roles_chat_room/role_core.py +++ b/apps/multi_roles_chat_room/role_core.py @@ -8,7 +8,6 @@ from modelscope_agent import create_component from modelscope_agent.agent_env_util import AgentEnvMixin from modelscope_agent.agents import MultiRolePlay -from modelscope_agent.multi_agents_utils.executors.ray import RayTaskExecutor from modelscope_agent.task_center import TaskCenter from story_holder import get_story_by_id @@ -16,22 +15,21 @@ # instruction prompt ROLE_INSTRUCTION_PROMPT = """<|im_start|>system -你是{role}，请你根据对话情节设定，和你的对话角色设定，继续当前的对话，说话要符合你的角色设定 - -【你的角色设定】 -{role_description} - -【对话情节设定】 -{story} +你是{role}，请你根据对话情节设定，当前的对话记录和你的对话角色设定，继续当前的对话，说话要符合你的角色设定，不要重复chat history中的内容 【注意事项】 1. 根据user的历史对话回复时，不要重复说历史对话里的句子，要保持生成内容的有趣，丰富度，多样性 2. 长话短说，不要说太多话，不要超过50字 +3. 符合当前对话情节设定，围绕对话情节，你是{role}，不要模仿其他角色，也不要重复chat history中的内容 + +【对话场景】 +{story} + +【chat history】 +chat_records 【你的角色设定】 -{role_description} -<|im_end|> -""" +{role_description}<|im_end|>""" CHATROOM_INSTRUCTION_PROMPT = """<|im_start|>system 你是一个小说作家，请你根据对话场景、人物介绍及最近的对话记录，选择继续对话的下一个角色。注意，对话历史是以群聊的形式展现，因此角色可能会@某个人表示对这个人说话。 @@ -42,17 +40,17 @@ 【人物介绍】 {all_roles_info} -【对话记录】 +【chat history】 chat_records 【最新对话】 recent_records 【注意事项】 -根据对话记录和最新对话 +根据chat history和最新对话 1. 当主角{user_role}说话中提到某个角色，你需要只让提到的角色接话。 2. 不要选【最新对话】里的角色发言 -3. 要让每个角色有平等的对话机会，多选一些【对话记录】没有出现的角色，要求情节充满戏剧性。 +3. 要让每个角色有平等的对话机会，多选一些chat history没有出现的角色，要求情节充满戏剧性。 4. 只写角色名字即可，每次最多选两个角色，尽量多的选择主角，当前对话的主角是{user_role} 【回复格式】 @@ -91,7 +89,8 @@ def generate_role_instruction(role, story_info): return instruction -def upsert_role(new_user, user_char, human_input_mode, story_info, llm_config): +def upsert_role(new_user, user_char, human_input_mode, story_info, llm_config, + _uid): role = create_component( MultiRolePlay, name=new_user, @@ -101,7 +100,8 @@ def upsert_role(new_user, user_char, human_input_mode, story_info, llm_config): llm=llm_config, function_list=function_list, instruction=generate_role_instruction(new_user, story_info), - human_input_mode=human_input_mode) + human_input_mode=human_input_mode, + prefix_name=_uid) return role @@ -127,9 +127,7 @@ def change_user_role(user_role, state): def init_all_remote_actors(_roles, user_role, _state, _story_state, - select_model): - RayTaskExecutor.init_ray() - + select_model, _uid): story_info = get_story_by_id(_story_state) llm_config['model'] = select_model @@ -139,7 +137,7 @@ def init_all_remote_actors(_roles, user_role, _state, _story_state, return state task_center = create_component( - TaskCenter, name='Task_Center', remote=REMOTE_MODE) + TaskCenter, name='Task_Center', remote=REMOTE_MODE, prefix_name=_uid) # init all agents and task center role_agents = [] @@ -148,7 +146,7 @@ def init_all_remote_actors(_roles, user_role, _state, _story_state, if role == user_role: human_input_mode = 'ON' role_agent = upsert_role(role, _roles[role], human_input_mode, - story_info, llm_config) + story_info, llm_config, _uid) role_agents.append(role_agent) @@ -156,7 +154,6 @@ def init_all_remote_actors(_roles, user_role, _state, _story_state, MultiRolePlay, name='chat_room', remote=True, - role='chat_room', llm=llm_config, function_list=function_list, instruction=CHATROOM_INSTRUCTION_PROMPT.format( @@ -164,9 +161,11 @@ def init_all_remote_actors(_roles, user_role, _state, _story_state, story=story_info['story'], user_role=user_role), is_watcher=True, - use_history=False) + use_history=False, + prefix_name=_uid) - logging.warning(msg=f'time:{time.time()} done create task center') + logging.warning( + msg=f'time:{time.time()} done create task center with uid: {_uid}') ray.get(task_center.add_agents.remote(role_agents)) ray.get(task_center.add_agents.remote([chat_room])) diff --git a/apps/multi_roles_chat_room/story_holder.py b/apps/multi_roles_chat_room/story_holder.py index be90e4623..2129a21e2 100644 --- a/apps/multi_roles_chat_room/story_holder.py +++ b/apps/multi_roles_chat_room/story_holder.py @@ -31,8 +31,7 @@ stories = [ { 'id': '1', - 'cover': - '//img.alicdn.com/imgextra/i1/O1CN01UHwXNQ2780lrVHY6n_!!6000000007751-0-tps-1024-512.jpg', + 'cover': '//s21.ax1x.com/2024/04/16/pFxG1zj.jpg', 'title': '我被美女包围了', 'description': '用户是男主角顾易，与多位长相、性格都大相径庭的美女相识', 'roles': ROLES_1, @@ -41,8 +40,7 @@ }, { 'id': '2', - 'cover': - '//img.alicdn.com/imgextra/i1/O1CN01UHwXNQ2780lrVHY6n_!!6000000007751-0-tps-1024-512.jpg', + 'cover': '//s21.ax1x.com/2024/04/16/pFxGgw6.png', 'title': '我是雷军，雷中有“电”，军下有“车”', 'description': '用户是男主角雷军，小米创始人，最近发布了小米SU7', 'roles': ROLES_2, diff --git a/modelscope_agent/__init__.py b/modelscope_agent/__init__.py index bc51dd488..ae04173cb 100644 --- a/modelscope_agent/__init__.py +++ b/modelscope_agent/__init__.py @@ -1,13 +1,19 @@ from .agent import Agent -def _create_remote(cls, name, max_concurrency=1, *args, **kwargs): +def _create_remote(cls, + name, + max_concurrency=1, + force_new=False, + *args, + **kwargs): ''' Create a remote actor by ray Args: cls: the class to be created name: the name of ray actor, also the role name max_concurrency: max concurrency of the actor + focus_new: force to create a new actor *args: **kwargs: @@ -15,7 +21,16 @@ def _create_remote(cls, name, max_concurrency=1, *args, **kwargs): ''' import ray - + try: + # try to get an existing actor + ray_actor = ray.get_actor(name) + if force_new: + ray.kill(ray_actor) + else: + return ray_actor + except ValueError: + pass + # if failed, create a new actor return ray.remote( name=name, max_concurrency=max_concurrency)(cls).remote(*args, **kwargs) @@ -39,11 +54,15 @@ def create_component(cls, name, remote=False, max_concurrency=1, + prefix_name=None, *args, **kwargs): kwargs['remote'] = remote kwargs['role'] = name + kwargs['prefix_name'] = prefix_name if remote: + if prefix_name is not None: + name = f'{prefix_name}_{name}' return _create_remote(cls, name, max_concurrency, *args, **kwargs) else: return _create_local(cls, *args, **kwargs) diff --git a/modelscope_agent/agent.py b/modelscope_agent/agent.py index f07a12983..88ff1f58b 100644 --- a/modelscope_agent/agent.py +++ b/modelscope_agent/agent.py @@ -3,7 +3,7 @@ from modelscope_agent.llm import get_chat_model from modelscope_agent.llm.base import BaseChatModel -from modelscope_agent.tools import TOOL_REGISTRY +from modelscope_agent.tools.base import TOOL_REGISTRY from modelscope_agent.utils.utils import has_chinese_chars diff --git a/modelscope_agent/agent_env_util.py b/modelscope_agent/agent_env_util.py index e478b7452..ebce48517 100644 --- a/modelscope_agent/agent_env_util.py +++ b/modelscope_agent/agent_env_util.py @@ -156,9 +156,6 @@ def step(self, history = [] if self.use_history: history = self.memory.get_history() - logger.info( - f'reach here 2 {result}, self.human_input_mode {self.human_input_mode}' - ) # run generation for frame in self.run( prompt, @@ -266,10 +263,10 @@ def pull(self): else: return '' - def convert_to_string(self, messages: List[Message], max_turn=5): + def convert_to_string(self, messages: List[Message], max_turn=15): prompt_template = """{conversation_history}""" conversation_history = '' - for item in messages[:max_turn]: + for item in messages[-1 * max_turn:]: conversation_history += f'{item.sent_from}: {item.content}\n' return prompt_template.format( conversation_history=conversation_history.strip()) diff --git a/modelscope_agent/constants.py b/modelscope_agent/constants.py index 77921d2bd..ca54a0235 100644 --- a/modelscope_agent/constants.py +++ b/modelscope_agent/constants.py @@ -4,3 +4,6 @@ DEFAULT_LOG_STORAGE_PATH = DEFAULT_AGENT_ROOT / 'log' DEFAULT_SEND_TO = 'all' USER_REQUIREMENT = 'user_requirement' +ENVIRONMENT_NAME = 'env' +AGENT_REGISTRY_NAME = 'agent_center' +TASK_CENTER_NAME = 'task_center' diff --git a/modelscope_agent/llm/dashscope.py b/modelscope_agent/llm/dashscope.py index 6bf4ec535..2d22a0708 100644 --- a/modelscope_agent/llm/dashscope.py +++ b/modelscope_agent/llm/dashscope.py @@ -209,14 +209,52 @@ def build_raw_prompt(self, messages: list): prompt = prompt[:-len(f'{im_end}')] return prompt + def build_multi_role_raw_prompt(self, messages: list): + prompt = '' + im_start = '<|im_start|>' + im_end = '<|im_end|>' + print('build_raw_prompt', messages) + if messages[0]['role'] == 'system': + system_prompt = messages[0]['content'] + else: + system_prompt = f'{im_start}system\nYou are a helpful assistant.{im_end}' + + # select user + if 'recent_records' in system_prompt and 'chat_records' in system_prompt: + chat_records = messages[-1]['content'].strip() + recent_records = chat_records.split('\n')[-1] + prompt = f'{system_prompt.replace("chat_records", chat_records).replace("recent_records", recent_records)}<|im_start|>assistant\n' # noqa E501 + else: + try: + re_pattern_config = re.compile(pattern=r'你是([\s\S]+)，请你根据对话') + res = re_pattern_config.search(system_prompt) + cur_role_name = res.group(1).strip() + except Exception: + cur_role_name = 'assistant' + print('cur_role_name: ', cur_role_name) + prompt = system_prompt + content = messages[-1]['content'].lstrip('\n').rstrip() + if 'chat_records' in prompt: + prompt = f'{prompt.replace("chat_records", content)}\n<|im_start|>{cur_role_name}\n' + else: + prompt = f'{prompt}user\n{content}<|im_end|>\n<|im_start|>assistant\n{cur_role_name}: ' -@register_llm('dashscope_qwen_spark') -class QwenSparkAtDS(DashScopeLLM): + print('prompt: ', [prompt]) + return prompt def _chat_stream(self, messages: List[Dict], stop: Optional[List[str]] = None, **kwargs) -> Iterator[str]: + if self.model == 'qwen-spark-plus': + return self._chat_stream_with_raw_prompt(messages, stop, **kwargs) + else: + return super()._chat_stream(messages, stop, **kwargs) + + def _chat_stream_with_raw_prompt(self, + messages: List[Dict], + stop: Optional[List[str]] = None, + **kwargs) -> Iterator[str]: stop = stop or [] stop.append('<|im_end|>') generation_input = { @@ -226,10 +264,12 @@ def _chat_stream(self, 'stop_str': word, 'mode': 'exclude' } for word in stop], - 'top_p': kwargs.get('top_p', 0.8), + 'top_p': kwargs.get('top_p', 0.95), + 'temperature': kwargs.get('temperature', 0.92), 'result_format': 'message', 'stream': True, 'use_raw_prompt': True, + 'max_length': 100 } logger.query_info( @@ -240,39 +280,5 @@ def _chat_stream(self, generation_input['temperature'] = kwargs.get('temperature') if kwargs.get('seed', None): generation_input['seed'] = kwargs.get('seed') - logger.info(f'######## input{generation_input}') - response = dashscope.Generation.call(**generation_input) - logger.info(f'######## response{response}') - return stream_output(response, **kwargs) - - def build_multi_role_raw_prompt(self, messages: list): - prompt = '' - im_start = '<|im_start|>' - im_end = '<|im_end|>' - print('build_raw_prompt', messages) - if messages[0]['role'] == 'system': - system_prompt = messages[0]['content'] - else: - system_prompt = f'{im_start}system\nYou are a helpful assistant.{im_end}' - - # select user - if 'chat_records' in system_prompt: - chat_records = messages[-1]['content'].strip() - recent_records = chat_records.split('\n')[-1] - prompt = f'{system_prompt.replace("chat_records", chat_records).replace("recent_records", recent_records)}<|im_start|>assistant\n' # noqa E501 - else: - try: - re_pattern_config = re.compile(pattern=r'你是([\s\S]+)，请你根据对话') - res = re_pattern_config.search(system_prompt) - cur_role_name = res.group(1).strip() - except Exception: - cur_role_name = 'assistant' - print('cur_role_name: ', cur_role_name) - prompt = system_prompt - content = messages[-1]['content'].lstrip('\n').rstrip() - prompt = f'{prompt}user\n{content}<|im_end|>\n<|im_start|>assistant\n{cur_role_name}: ' - - print('prompt: ', [prompt]) - return prompt diff --git a/modelscope_agent/multi_agents_utils/README.md b/modelscope_agent/multi_agents_utils/README.md index 415dbbcb5..08340c5cc 100644 --- a/modelscope_agent/multi_agents_utils/README.md +++ b/modelscope_agent/multi_agents_utils/README.md @@ -423,6 +423,17 @@ With the above inputs, the step could be used in different scenarios of multi-ag For example, in a three-man debate scenario, the `step` method could be used like this: ```python +# add new role +role_template_hillary = 'you are the former secretary of state Hillary Clinton, and you are debating with former president Donald Trump and current president Joe Biden with couple of topics' + +hillary_clinton = create_component( + RolePlay, + name='hillary_clinton', + remote=REMOTE_MODE, + llm=llm_config, + function_list=function_list, + instruction=role_template_hillary) + # initialize the agents task_center.add_agents([joe_biden, donald_trump, hillary_clinton]) @@ -434,9 +445,12 @@ for frame in task_center.step(send_to='hillary_clinton'): print(frame) # in 2nd step, allow only donald_trump to response the topic -for frame in task_center.step(allower_roles='donald_trump'): +for frame in task_center.step(allowed_roles='donald_trump'): print(frame) ``` +Notice that the `frame` will only show the message from different agents with format `<[role_name]>: [message stream]` +user should take care of the message format in the `step` method. + The above case show how to use `send_to` and `allowed_roles` in the `step` method to control the communication between agents in a multi-agent task. There is another case, in a chatbot mode, the `user_response` could be used to let the user response in this step to replace the llm output, if user_agent is in this step. @@ -456,17 +470,20 @@ user = create_component( task_center.add_agents([joe_biden, donald_trump, hillary_clinton, user]) # let joe_biden start the topic -task_center.send_task_request('what is the best solution to land on moon?', send_to='joe_biden') +task_center.send_task_request('what is the best solution to land on moon, in one sentence?', send_to='joe_biden') # in 1st step, let joe_biden send his opinion to all agents. for frame in task_center.step(): print(frame) # in 2nd step, allow only user to response the topic, with user_response -for frame in task_center.step(allower_roles='user', user_response='I dont agree with you about the landing project'): - print(frame) -assert frame == 'I dont agree with you about the landing project' +result = '' +for frame in task_center.step(allowed_roles='user', user_response='I dont agree with you about the landing project'): + result += frame + print(frame) +# the input from outside will not print out here as the user_response is set +assert result == '' ``` The user response will be used in this step to replace the llm output, because `user` is a human agent. diff --git a/modelscope_agent/multi_agents_utils/README_CN.md b/modelscope_agent/multi_agents_utils/README_CN.md index 40b45c6c8..fa4baf826 100644 --- a/modelscope_agent/multi_agents_utils/README_CN.md +++ b/modelscope_agent/multi_agents_utils/README_CN.md @@ -382,6 +382,17 @@ step方法将使每个agent在这一步骤中作出响应，响应将是一个 有了以上的参数，step方法可以在不同场景的multi-agent中使用。 例如，在一个三人辩论场景中，step方法可以像这样使用： ```python +# add new role +role_template_hillary = 'you are the former secretary of state Hillary Clinton, and you are debating with former president Donald Trump and current president Joe Biden with couple of topics' + +hillary_clinton = create_component( + RolePlay, + name='hillary_clinton', + remote=REMOTE_MODE, + llm=llm_config, + function_list=function_list, + instruction=role_template_hillary) + # initialize the agents task_center.add_agents([joe_biden, donald_trump, hillary_clinton]) @@ -393,9 +404,13 @@ for frame in task_center.step(send_to='hillary_clinton'): print(frame) # in 2nd step, allow only donald_trump to response the topic -for frame in task_center.step(allower_roles='donald_trump'): +for frame in task_center.step(allowed_roles='donald_trump'): print(frame) ``` +*请注意*，在`frame`中只会显示来自不同agent的消息，并且格式为`<[role_name]>: [message stream]`。 +用户需要根据自己的业务需求在step方法中处理future输出的格式。 + + 上述案例展示了如何在multi-agent任务中使用step方法中的send_to和allowed_roles来控制agent之间的通信。 在另一个情况下，在聊天机器人模式中，如果本步骤中包含user-agent，可以使用user_response让用户在这一步骤中进行输入，以取代LLM（大型语言模型）的输出。 @@ -423,9 +438,13 @@ for frame in task_center.step(): print(frame) # in 2nd step, allow only user to response the topic, with user_response -for frame in task_center.step(allower_roles='user', user_response='I dont agree with you about the landing project'): - print(frame) -assert frame == 'I dont agree with you about the landing project' +result = '' +for frame in task_center.step(allowed_roles='user', user_response='I dont agree with you about the landing project'): + result += frame + print(frame) + +# the input from outside will not print out here as the user_response is set +assert result == '' ``` 可以看到，用户的响应将在这个步骤中被使用，以取代大型语言模型（LLM）的输出，因为名为`user`的agent 是一个user-agent。 diff --git a/modelscope_agent/schemas.py b/modelscope_agent/schemas.py index 10a89543d..4d40559bc 100644 --- a/modelscope_agent/schemas.py +++ b/modelscope_agent/schemas.py @@ -1,5 +1,6 @@ -from typing import List, Set, Union +from typing import List, Union +from modelscope_agent.constants import DEFAULT_SEND_TO from pydantic import BaseModel @@ -10,7 +11,7 @@ class Message(BaseModel): role: str = 'user' # user, assistant, system, tool content: str = '' sent_from: str = '' - send_to: Union[str, Set[str]] = {'all'} + send_to: Union[str, List[str]] = DEFAULT_SEND_TO class Document(BaseModel): diff --git a/modelscope_agent/task_center.py b/modelscope_agent/task_center.py index 2196b0f37..614b37efe 100644 --- a/modelscope_agent/task_center.py +++ b/modelscope_agent/task_center.py @@ -3,7 +3,8 @@ from modelscope_agent import create_component from modelscope_agent.agent import Agent from modelscope_agent.agents_registry import AgentRegistry -from modelscope_agent.constants import DEFAULT_SEND_TO, USER_REQUIREMENT +from modelscope_agent.constants import (AGENT_REGISTRY_NAME, DEFAULT_SEND_TO, + ENVIRONMENT_NAME, USER_REQUIREMENT) from modelscope_agent.environment import Environment from modelscope_agent.schemas import Message from modelscope_agent.utils.logger import agent_logger as logger @@ -11,16 +12,24 @@ class TaskCenter: - def __init__(self, remote=False, **kwargs): + def __init__(self, remote=False, prefix_name=None, **kwargs): if remote: from modelscope_agent.multi_agents_utils.executors.ray import RayTaskExecutor self.task_executor = RayTaskExecutor else: from modelscope_agent.multi_agents_utils.executors.local import LocalTaskExecutor self.task_executor = LocalTaskExecutor - self.env = create_component(Environment, 'env', remote) - self.agent_registry = create_component(AgentRegistry, 'agent_center', - remote) + # used to create the environment and agent registry with specific prefix + self.env = create_component( + cls=Environment, + name=ENVIRONMENT_NAME, + remote=remote, + prefix_name=prefix_name) + self.agent_registry = create_component( + cls=AgentRegistry, + name=AGENT_REGISTRY_NAME, + remote=remote, + prefix_name=prefix_name) self.remote = remote def add_agents(self, agents: List[Agent]): @@ -133,6 +142,8 @@ def step(self, if len(allowed_roles) == 0: roles = self.task_executor.get_notified_roles(self.env) else: + if isinstance(allowed_roles, str): + allowed_roles = [allowed_roles] roles = allowed_roles if len(roles) == 0: