ChatGLM改成多进程运行

crazy_functions/crazy_utils.py

@@ -66,7 +66,7 @@ def request_gpt_model_in_new_thread_with_ui_alive(
     chatbot.append([inputs_show_user, ""])
     yield from update_ui(chatbot=chatbot, history=[]) # 刷新界面
     executor = ThreadPoolExecutor(max_workers=16)
-    mutable = ["", time.time()]
+    mutable = ["", time.time(), ""]
     def _req_gpt(inputs, history, sys_prompt):
         retry_op = retry_times_at_unknown_error
         exceeded_cnt = 0

crazy_functions/询问多个大语言模型.py

@@ -20,7 +20,8 @@ def 同时问询(txt, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt
     gpt_say = yield from request_gpt_model_in_new_thread_with_ui_alive(
         inputs=txt, inputs_show_user=txt, 
         llm_kwargs=llm_kwargs, chatbot=chatbot, history=history, 
-        sys_prompt=system_prompt
+        sys_prompt=system_prompt,
+        retry_times_at_unknown_error=0
     )
 
     history.append(txt)

docs/Dockerfile+ChatGLM

@@ -24,7 +24,7 @@ RUN curl -sS https://bootstrap.pypa.io/get-pip.py | python3.8
 
 # 下载分支
 WORKDIR /gpt
-RUN $useProxyNetwork git clone https://github.com/binary-husky/chatgpt_academic.git -b v3.0
+RUN $useProxyNetwork git clone https://github.com/binary-husky/chatgpt_academic.git -b v3.1
 WORKDIR /gpt/chatgpt_academic
 RUN $useProxyNetwork python3 -m pip install -r requirements.txt
 RUN $useProxyNetwork python3 -m pip install -r request_llm/requirements_chatglm.txt

main.py

@@ -1,56 +1,57 @@
 import os; os.environ['no_proxy'] = '*' # 避免代理网络产生意外污染
-import gradio as gr
-from request_llm.bridge_all import predict
-from toolbox import format_io, find_free_port, on_file_uploaded, on_report_generated, get_conf, ArgsGeneralWrapper, DummyWith
 
-# 建议您复制一个config_private.py放自己的秘密, 如API和代理网址, 避免不小心传github被别人看到
+def main():
-proxies, WEB_PORT, LLM_MODEL, CONCURRENT_COUNT, AUTHENTICATION, CHATBOT_HEIGHT, LAYOUT, API_KEY = \
+    import gradio as gr
+    from request_llm.bridge_all import predict
+    from toolbox import format_io, find_free_port, on_file_uploaded, on_report_generated, get_conf, ArgsGeneralWrapper, DummyWith
+    # 建议您复制一个config_private.py放自己的秘密, 如API和代理网址, 避免不小心传github被别人看到
+    proxies, WEB_PORT, LLM_MODEL, CONCURRENT_COUNT, AUTHENTICATION, CHATBOT_HEIGHT, LAYOUT, API_KEY = \
         get_conf('proxies', 'WEB_PORT', 'LLM_MODEL', 'CONCURRENT_COUNT', 'AUTHENTICATION', 'CHATBOT_HEIGHT', 'LAYOUT', 'API_KEY')
 
-# 如果WEB_PORT是-1, 则随机选取WEB端口
+    # 如果WEB_PORT是-1, 则随机选取WEB端口
-PORT = find_free_port() if WEB_PORT <= 0 else WEB_PORT
+    PORT = find_free_port() if WEB_PORT <= 0 else WEB_PORT
-if not AUTHENTICATION: AUTHENTICATION = None
+    if not AUTHENTICATION: AUTHENTICATION = None
 
-from check_proxy import get_current_version
+    from check_proxy import get_current_version
-initial_prompt = "Serve me as a writing and programming assistant."
+    initial_prompt = "Serve me as a writing and programming assistant."
-title_html = f"<h1 align=\"center\">ChatGPT 学术优化 {get_current_version()}</h1>"
+    title_html = f"<h1 align=\"center\">ChatGPT 学术优化 {get_current_version()}</h1>"
-description =  """代码开源和更新[地址🚀](https://github.com/binary-husky/chatgpt_academic)，感谢热情的[开发者们❤️](https://github.com/binary-husky/chatgpt_academic/graphs/contributors)"""
+    description =  """代码开源和更新[地址🚀](https://github.com/binary-husky/chatgpt_academic)，感谢热情的[开发者们❤️](https://github.com/binary-husky/chatgpt_academic/graphs/contributors)"""
 
-# 问询记录, python 版本建议3.9+（越新越好）
+    # 问询记录, python 版本建议3.9+（越新越好）
-import logging
+    import logging
-os.makedirs("gpt_log", exist_ok=True)
+    os.makedirs("gpt_log", exist_ok=True)
-try:logging.basicConfig(filename="gpt_log/chat_secrets.log", level=logging.INFO, encoding="utf-8")
+    try:logging.basicConfig(filename="gpt_log/chat_secrets.log", level=logging.INFO, encoding="utf-8")
-except:logging.basicConfig(filename="gpt_log/chat_secrets.log", level=logging.INFO)
+    except:logging.basicConfig(filename="gpt_log/chat_secrets.log", level=logging.INFO)
-print("所有问询记录将自动保存在本地目录./gpt_log/chat_secrets.log, 请注意自我隐私保护哦！")
+    print("所有问询记录将自动保存在本地目录./gpt_log/chat_secrets.log, 请注意自我隐私保护哦！")
 
-# 一些普通功能模块
+    # 一些普通功能模块
-from core_functional import get_core_functions
+    from core_functional import get_core_functions
-functional = get_core_functions()
+    functional = get_core_functions()
 
-# 高级函数插件
+    # 高级函数插件
-from crazy_functional import get_crazy_functions
+    from crazy_functional import get_crazy_functions
-crazy_fns = get_crazy_functions()
+    crazy_fns = get_crazy_functions()
 
-# 处理markdown文本格式的转变
+    # 处理markdown文本格式的转变
-gr.Chatbot.postprocess = format_io
+    gr.Chatbot.postprocess = format_io
 
-# 做一些外观色彩上的调整
+    # 做一些外观色彩上的调整
-from theme import adjust_theme, advanced_css
+    from theme import adjust_theme, advanced_css
-set_theme = adjust_theme()
+    set_theme = adjust_theme()
 
-# 代理与自动更新
+    # 代理与自动更新
-from check_proxy import check_proxy, auto_update
+    from check_proxy import check_proxy, auto_update
-proxy_info = check_proxy(proxies)
+    proxy_info = check_proxy(proxies)
 
-gr_L1 = lambda: gr.Row().style()
+    gr_L1 = lambda: gr.Row().style()
-gr_L2 = lambda scale: gr.Column(scale=scale)
+    gr_L2 = lambda scale: gr.Column(scale=scale)
-if LAYOUT == "TOP-DOWN": 
+    if LAYOUT == "TOP-DOWN": 
         gr_L1 = lambda: DummyWith()
         gr_L2 = lambda scale: gr.Row()
         CHATBOT_HEIGHT /= 2
 
-cancel_handles = []
+    cancel_handles = []
-with gr.Blocks(title="ChatGPT 学术优化", theme=set_theme, analytics_enabled=False, css=advanced_css) as demo:
+    with gr.Blocks(title="ChatGPT 学术优化", theme=set_theme, analytics_enabled=False, css=advanced_css) as demo:
         gr.HTML(title_html)
         cookies = gr.State({'api_key': API_KEY, 'llm_model': LLM_MODEL})
         with gr_L1():
@@ -161,8 +162,9 @@ with gr.Blocks(title="ChatGPT 学术优化", theme=set_theme, analytics_enabled=
         # 终止按钮的回调函数注册
         stopBtn.click(fn=None, inputs=None, outputs=None, cancels=cancel_handles)
         stopBtn2.click(fn=None, inputs=None, outputs=None, cancels=cancel_handles)
-# gradio的inbrowser触发不太稳定，回滚代码到原始的浏览器打开函数
+
-def auto_opentab_delay():
+    # gradio的inbrowser触发不太稳定，回滚代码到原始的浏览器打开函数
+    def auto_opentab_delay():
         import threading, webbrowser, time
         print(f"如果浏览器没有自动打开，请复制并转到以下URL：")
         print(f"\t（亮色主题）: http://localhost:{PORT}")
@@ -173,5 +175,8 @@ def auto_opentab_delay():
         threading.Thread(target=open, name="open-browser", daemon=True).start()
         threading.Thread(target=auto_update, name="self-upgrade", daemon=True).start()
 
-auto_opentab_delay()
+    auto_opentab_delay()
-demo.queue(concurrency_count=CONCURRENT_COUNT).launch(server_name="0.0.0.0", server_port=PORT, auth=AUTHENTICATION)
+    demo.queue(concurrency_count=CONCURRENT_COUNT).launch(server_name="0.0.0.0", server_port=PORT, auth=AUTHENTICATION)
+
+if __name__ == "__main__":
+    main()

request_llm/bridge_all.py

@@ -31,6 +31,24 @@ methods = {
     "tgui-ui": tgui_ui,
 }
 
+def LLM_CATCH_EXCEPTION(f):
+    """
+        装饰器函数，将错误显示出来
+    """
+    def decorated(inputs, llm_kwargs, history, sys_prompt, observe_window, console_slience):
+        try:
+            return f(inputs, llm_kwargs, history, sys_prompt, observe_window, console_slience)
+        except Exception as e:
+            from toolbox import get_conf
+            import traceback
+            proxies, = get_conf('proxies')
+            tb_str = '\n```\n' + traceback.format_exc() + '\n```\n'
+            observe_window[0] = tb_str
+            return tb_str
+    return decorated
+
+colors = ['#FF00FF', '#00FFFF', '#FF0000''#990099', '#009999', '#990044']
+
 def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, observe_window, console_slience=False):
     """
         发送至LLM，等待回复，一次性完成，不显示中间过程。但内部用stream的方法避免中途网线被掐。
@@ -62,17 +80,13 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, obser
         return method(inputs, llm_kwargs, history, sys_prompt, observe_window, console_slience)
     else:
         # 如果同时询问多个大语言模型：
-        executor = ThreadPoolExecutor(max_workers=16)
+        executor = ThreadPoolExecutor(max_workers=4)
         models = model.split('&')
         n_model = len(models)
         
         window_len = len(observe_window)
-        if window_len==0:
+        assert window_len==3
-            window_mutex = [[] for _ in range(n_model)] + [True]
+        window_mutex = [["", time.time(), ""] for _ in range(n_model)] + [True]
-        elif window_len==1:
-            window_mutex = [[""] for _ in range(n_model)] + [True]
-        elif window_len==2:
-            window_mutex = [["", time.time()] for _ in range(n_model)] + [True]
 
         futures = []
         for i in range(n_model):
@@ -85,12 +99,12 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, obser
                 method = methods['tgui-no-ui']
             llm_kwargs_feedin = copy.deepcopy(llm_kwargs)
             llm_kwargs_feedin['llm_model'] = model
-            future = executor.submit(method, inputs, llm_kwargs_feedin, history, sys_prompt, window_mutex[i], console_slience)
+            future = executor.submit(LLM_CATCH_EXCEPTION(method), inputs, llm_kwargs_feedin, history, sys_prompt, window_mutex[i], console_slience)
             futures.append(future)
 
         def mutex_manager(window_mutex, observe_window):
             while True:
-                time.sleep(0.2)
+                time.sleep(0.5)
                 if not window_mutex[-1]: break
                 # 看门狗（watchdog）
                 for i in range(n_model): 
@@ -98,8 +112,8 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, obser
                 # 观察窗（window）
                 chat_string = []
                 for i in range(n_model):
-                    chat_string.append( f"[{str(models[i])} 说]: {window_mutex[i][0]}" )
+                    chat_string.append( f"【{str(models[i])} 说】: <font color=\"{colors[i]}\"> {window_mutex[i][0]} </font>" )
-                res = '\n\n---\n\n'.join(chat_string)
+                res = '<br/><br/>\n\n---\n\n'.join(chat_string)
                 # # # # # # # # # # #
                 observe_window[0] = res
 
@@ -107,10 +121,18 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, obser
         t_model.start()
 
         return_string_collect = []
+        while True:
+            worker_done = [h.done() for h in futures]
+            if all(worker_done):
+                executor.shutdown()
+                break
+            time.sleep(1)
+
         for i, future in enumerate(futures):  # wait and get
-            return_string_collect.append( f"[{str(models[i])} 说]: {future.result()}" )
+            return_string_collect.append( f"【{str(models[i])} 说】: <font color=\"{colors[i]}\"> {future.result()} </font>" )
+
         window_mutex[-1] = False # stop mutex thread
-        res = '\n\n---\n\n'.join(return_string_collect)
+        res = '<br/>\n\n---\n\n'.join(return_string_collect)
         return res
 
 

request_llm/bridge_chatglm.py

@@ -3,35 +3,69 @@ from transformers import AutoModel, AutoTokenizer
 import time
 import importlib
 from toolbox import update_ui, get_conf
+from multiprocessing import Process, Pipe
 
+#################################################################################
+class GetGLMHandle(Process):
+    def __init__(self):
+        super().__init__(daemon=True)
+        self.parent, self.child = Pipe()
+        self.chatglm_model = None
+        self.chatglm_tokenizer = None
+        self.start()
+        print('初始化')
         
-global chatglm_model, chatglm_tokenizer
+    def ready(self):
+        return self.chatglm_model is not None
 
-chatglm_model = None
+    def run(self):
-chatglm_tokenizer = None
+        while True:
-
+            try:
-def model_loader():
+                if self.chatglm_model is None:
-    global chatglm_model, chatglm_tokenizer
+                    self.chatglm_tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
-    if chatglm_tokenizer is None:
-        chatglm_tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
-    if chatglm_model is None: # 尚未加载
                     device, = get_conf('LOCAL_MODEL_DEVICE')
                     if device=='cpu':
-            chatglm_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).float()
+                        self.chatglm_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).float()
                     else:
-            chatglm_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
+                        self.chatglm_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
-        chatglm_model = chatglm_model.eval()
+                    self.chatglm_model = self.chatglm_model.eval()
-    chatglm_model = chatglm_model.eval()
+                    break
+                else:
+                    break
+            except:
+                pass
+        while True:
+            kwargs = self.child.recv()
+            try:
+                for response, history in self.chatglm_model.stream_chat(self.chatglm_tokenizer, **kwargs):
+                    self.child.send(response)
+            except:
+                self.child.send('[Local Message] Call ChatGLM fail.')
+            self.child.send('[Finish]')
 
+    def stream_chat(self, **kwargs):
+        self.parent.send(kwargs)
+        while True:
+            res = self.parent.recv()
+            if res != '[Finish]':
+                yield res
+            else:
+                break
+        return
+    
+global glm_handle
+glm_handle = None
+#################################################################################
 def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="", observe_window=None, console_slience=False):
     """
+        多线程方法
         函数的说明请见 request_llm/bridge_all.py
     """
-    global chatglm_model, chatglm_tokenizer
+    global glm_handle
-    if chatglm_model is None:
+    if glm_handle is None:
-        observe_window[0] = "ChatGLM尚未加载，加载需要一段时间 ……"
+        glm_handle = GetGLMHandle()
+        observe_window[0] = "ChatGLM尚未加载，加载需要一段时间。注意，取决于`config.py`的配置，ChatGLM消耗大量的内存（CPU）或显存（GPU），也许会导致低配计算机卡死 ……"
 
-    model_loader()
     # chatglm 没有 sys_prompt 接口，因此把prompt加入 history
     history_feedin = []
     for i in range(len(history)//2):
@@ -40,29 +74,27 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="",
 
     watch_dog_patience = 5 # 看门狗 (watchdog) 的耐心, 设置5秒即可
     response = ""
-    for response, history in chatglm_model.stream_chat(chatglm_tokenizer, inputs, history=history_feedin, max_length=llm_kwargs['max_length'],
+    for response in glm_handle.stream_chat(query=inputs, history=history_feedin, max_length=llm_kwargs['max_length'], top_p=llm_kwargs['top_p'], temperature=llm_kwargs['temperature']):
-                                                       top_p=llm_kwargs['top_p'], temperature=llm_kwargs['temperature']):
-        # 观测窗，把已经获取的数据显示出去
         observe_window[0] = response
-        # 看门狗 (watchdog)，如果超过期限没有喂狗，则终止
         if len(observe_window) >= 2:  
             if (time.time()-observe_window[1]) > watch_dog_patience:
                 raise RuntimeError("程序终止。")
-        # if not console_slience:
-        #     print(response)
     return response
 
 
+
 def predict(inputs, llm_kwargs, plugin_kwargs, chatbot, history=[], system_prompt='', stream = True, additional_fn=None):
     """
+        单线程方法
         函数的说明请见 request_llm/bridge_all.py
     """
-    global chatglm_model, chatglm_tokenizer
     chatbot.append((inputs, ""))
-    if chatglm_model is None:
+
-        chatbot[-1] = (inputs, "ChatGLM尚未加载，加载需要一段时间 ……")
+    global glm_handle
+    if glm_handle is None:
+        glm_handle = GetGLMHandle()
+        chatbot[-1] = (inputs, "ChatGLM尚未加载，加载需要一段时间。注意，取决于`config.py`的配置，ChatGLM消耗大量的内存（CPU）或显存（GPU），也许会导致低配计算机卡死 ……")
         yield from update_ui(chatbot=chatbot, history=[])
-    model_loader()
 
     if additional_fn is not None:
         import core_functional
@@ -71,13 +103,11 @@ def predict(inputs, llm_kwargs, plugin_kwargs, chatbot, history=[], system_promp
         if "PreProcess" in core_functional[additional_fn]: inputs = core_functional[additional_fn]["PreProcess"](inputs)  # 获取预处理函数（如果有的话）
         inputs = core_functional[additional_fn]["Prefix"] + inputs + core_functional[additional_fn]["Suffix"]
 
-
     history_feedin = []
     for i in range(len(history)//2):
         history_feedin.append(["What can I do?", system_prompt] )
         history_feedin.append([history[2*i], history[2*i+1]] )
 
-    for response, history in chatglm_model.stream_chat(chatglm_tokenizer, inputs, history=history_feedin, max_length=llm_kwargs['max_length'],
+    for response in glm_handle.stream_chat(query=inputs, history=history_feedin, max_length=llm_kwargs['max_length'], top_p=llm_kwargs['top_p'], temperature=llm_kwargs['temperature']):
-                                                       top_p=llm_kwargs['top_p'], temperature=llm_kwargs['temperature']):
         chatbot[-1] = (inputs, response)
         yield from update_ui(chatbot=chatbot, history=history)

version

@@ -1,5 +1,5 @@
 {
-  "version": 3.0,
+  "version": 3.1,
   "show_feature": true,
   "new_feature": "支持ChatGLM <-> 支持多LLM模型同时对话"
 }