Upload 4 files

.flake8

@@ -0,0 +1,21 @@
+[flake8]
+ignore =
+  # E203 whitespace before ':'
+  E203
+  D203,
+  # line too long
+  E501
+per-file-ignores =
+  # imported but unused
+  # __init__.py: F401
+  test_*.py: F401
+exclude =
+  .git,
+  __pycache__,
+  docs/source/conf.py,
+  old,
+  build,
+  dist,
+  .venv
+  pad*.py
+max-complexity = 25

app.py

@@ -1,14 +1,41 @@
-from transformers import AutoModel, AutoTokenizer
+"""Credit to https://github.com/THUDM/ChatGLM2-6B/blob/main/web_demo.py while mistakes are mine."""
+# pylint: disable=broad-exception-caught, redefined-outer-name, missing-function-docstring, missing-module-docstring, too-many-arguments, line-too-long, invalid-name, redefined-builtin, redefined-argument-from-local
+# import gradio as gr
+
+# model_name = "models/THUDM/chatglm2-6b-int4"
+# gr.load(model_name).lauch()
+
+# %%writefile demo-4bit.py
+
+import os
+import time
+from textwrap import dedent
+
 import gradio as gr
 import mdtex2html
+import torch
+from loguru import logger
+from transformers import AutoModel, AutoTokenizer
+
+# fix timezone in Linux
+os.environ["TZ"] = "Asia/Shanghai"
+try:
+    time.tzset()  # type: ignore # pylint: disable=no-member
+except Exception:
+    # Windows
+    logger.warning("Windows, cant run time.tzset()")
 
-checkpoint = "fb700/chatglm-fitness-RLHF"
 
-tokenizer = AutoTokenizer.from_pretrained(checkpoint, trust_remote_code=True)
-model = AutoModel.from_pretrained(checkpoint, trust_remote_code=True).quantize(4).half().cuda()
+ 
+model_name = "fb700/chatglm-fitness-RLHF" 
+
+RETRY_FLAG = False
+
+tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+model = AutoModel.from_pretrained(model_name, trust_remote_code=True).quantize(4).half().cuda()
 model = model.eval()
 
-"""Override Chatbot.postprocess"""
+_ = """Override Chatbot.postprocess"""
 
 
 def postprocess(self, y):
@@ -26,22 +53,21 @@ gr.Chatbot.postprocess = postprocess
 
 
 def parse_text(text):
-    """copy from https://github.com/GaiZhenbiao/ChuanhuChatGPT/"""
     lines = text.split("\n")
     lines = [line for line in lines if line != ""]
     count = 0
     for i, line in enumerate(lines):
         if "```" in line:
             count += 1
-            items = line.split('`')
+            items = line.split("`")
             if count % 2 == 1:
                 lines[i] = f'<pre><code class="language-{items[-1]}">'
             else:
-                lines[i] = f'<br></code></pre>'
+                lines[i] = "<br></code></pre>"
         else:
             if i > 0:
                 if count % 2 == 1:
-                    line = line.replace("`", "\`")
+                    line = line.replace("`", r"\`")
                     line = line.replace("<", "&lt;")
                     line = line.replace(">", "&gt;")
                     line = line.replace(" ", "&nbsp;")
@@ -53,51 +79,310 @@ def parse_text(text):
                     line = line.replace("(", "&#40;")
                     line = line.replace(")", "&#41;")
                     line = line.replace("$", "&#36;")
-                lines[i] = "<br>"+line
+                lines[i] = "<br>" + line
     text = "".join(lines)
     return text
 
 
-def predict(input, chatbot, max_length, top_p, temperature, history):
-    chatbot.append((parse_text(input), ""))
-    for response, history in model.stream_chat(tokenizer, input, history, max_length=max_length, top_p=top_p,
-                                               temperature=temperature):
-        chatbot[-1] = (parse_text(input), parse_text(response))       
+def predict(
+    RETRY_FLAG, input, chatbot, max_length, top_p, temperature, history, past_key_values
+):
+    try:
+        chatbot.append((parse_text(input), ""))
+    except Exception as exc:
+        logger.error(exc)
+        logger.debug(f"{chatbot=}")
+        _ = """
+        if chatbot:
+            chatbot[-1] = (parse_text(input), str(exc))
+            yield chatbot, history, past_key_values
+        # """
+        yield chatbot, history, past_key_values
+
+    for response, history, past_key_values in model.stream_chat(
+        tokenizer,
+        input,
+        history,
+        past_key_values=past_key_values,
+        return_past_key_values=True,
+        max_length=max_length,
+        top_p=top_p,
+        temperature=temperature,
+    ):
+        chatbot[-1] = (parse_text(input), parse_text(response))
+
+        yield chatbot, history, past_key_values
+
+
+def trans_api(input, max_length=40960, top_p=0.7, temperature=0.95):
+    if max_length < 10:
+        max_length = 40960
+    if top_p < 0.1 or top_p > 1:
+        top_p = 0.7
+    if temperature <= 0 or temperature > 1:
+        temperature = 0.01
+    try:
+        res, _ = model.chat(
+            tokenizer,
+            input,
+            history=[],
+            past_key_values=None,
+            max_length=max_length,
+            top_p=top_p,
+            temperature=temperature,
+        )
+        # logger.debug(f"{res=} \n{_=}")
+    except Exception as exc:
+        logger.error(f"{exc=}")
+        res = str(exc)
 
-        yield chatbot, history
+    return res
 
 
 def reset_user_input():
-    return gr.update(value='')
+    return gr.update(value="")
 
 
 def reset_state():
-    return [], []
+    return [], [], None
 
 
-with gr.Blocks() as demo:
-    gr.HTML("""<h1 align="center">帛凡 ChatGLM-6b-fitness-RLHF  演示 </h1>""")
+# Delete last turn
+def delete_last_turn(chat, history):
+    if chat and history:
+        chat.pop(-1)
+        history.pop(-1)
+    return chat, history
 
+
+# Regenerate response
+def retry_last_answer(
+    user_input, chatbot, max_length, top_p, temperature, history, past_key_values
+):
+    if chatbot and history:
+        # Removing the previous conversation from chat
+        chatbot.pop(-1)
+        # Setting up a flag to capture a retry
+        RETRY_FLAG = True
+        # Getting last message from user
+        user_input = history[-1][0]
+        # Removing bot response from the history
+        history.pop(-1)
+
+    yield from predict(
+        RETRY_FLAG,  # type: ignore
+        user_input,
+        chatbot,
+        max_length,
+        top_p,
+        temperature,
+        history,
+        past_key_values,
+    )
+
+
+with gr.Blocks(title="ChatGLM2-6B-int4", theme=gr.themes.Soft(text_size="sm")) as demo:
+    # gr.HTML("""<h1 align="center">ChatGLM2-6B-int4</h1>""")
+    gr.HTML(
+        """<center><a href="https://huggingface.co/spaces/mikeee/chatglm2-6b-4bit?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>It's beyond Fitness,模型由[帛凡]基于ChatGLM-6b进行微调后，在健康（全科）、心理等领域达至少60分的专业水准，而且中文总结能力超越了GPT3.5各版本。</center>"""
+        """<center><免责声明：本应用仅为模型能力演示，无任何商业行为，部署资源为huggingface官方免费提供，任何通过此项目产生的知识仅用于学术参考，作者和网站均不承担任何责任   。</center>"""
+        """<h1 align="center">帛凡 Fitness AI 演示</h1>"""
+    )
+
+    with gr.Accordion("🎈 Info", open=False):
+        _ = f"""
+            ## {model_name}
+
+            ChatGLM-6B 是开源中英双语对话模型，本次训练基于ChatGLM-6B 的第一代版本，在保留了初代模型对话流畅、部署门槛较低等众多优秀特性的基础之上开展训练。
+                       
+            本项目经过多位网友实测，中文总结能力超越了GPT3.5各版本，健康咨询水平优于其它同量级模型，且经优化目前可以支持无限context，远大于4k、8K、16K......，可能是任何个人和中小企业首选模型。
+            
+            *首先，用40万条高质量数据进行强化训练，以提高模型的基础能力；
+            
+            *第二，使用30万条人类反馈数据，构建一个表达方式规范优雅的语言模式（RM模型）；
+            
+            *第三，在保留SFT阶段三分之一训练数据的同时，增加了30万条fitness数据，叠加RM模型，对ChatGLM-6B进行强化训练。
+            
+            通过训练我们对模型有了更深刻的认知，LLM在一直在进化，好的方法和数据可以挖掘出模型的更大潜能。
+            训练中特别强化了中英文学术论文的翻译和总结，可以成为普通用户和科研人员的得力助手。 
+            
+            免责声明：本应用仅为模型能力演示，无任何商业行为，部署资源为huggingface官方免费提供，任何通过此项目产生的知识仅用于学术参考，作者和网站均不承担任何责任   。      
+            
+            The T4 GPU is sponsored by a community GPU grant from Huggingface. Thanks a lot!
+            
+            [模型下载地址](https://huggingface.co/fb700/chatglm-fitness-RLHF)             
+
+
+            """
+        gr.Markdown(dedent(_))
     chatbot = gr.Chatbot()
     with gr.Row():
         with gr.Column(scale=4):
             with gr.Column(scale=12):
-                user_input = gr.Textbox(show_label=False, placeholder="Input...", lines=10).style(
-                    container=False)
+                user_input = gr.Textbox(
+                    show_label=False,
+                    placeholder="Input...",
+                ).style(container=False)
+                RETRY_FLAG = gr.Checkbox(value=False, visible=False)
             with gr.Column(min_width=32, scale=1):
-                submitBtn = gr.Button("Submit", variant="primary")
+                with gr.Row():
+                    submitBtn = gr.Button("Submit", variant="primary")
+                    deleteBtn = gr.Button("删除最后一条对话", variant="secondary")
+                    retryBtn = gr.Button("重新生成Regenerate", variant="secondary")
         with gr.Column(scale=1):
             emptyBtn = gr.Button("Clear History")
-            max_length = gr.Slider(0, 40960, value=20480, step=1.0, label="Maximum length", interactive=True)
-            top_p = gr.Slider(0, 1, value=0.7, step=0.01, label="Top P", interactive=True)
-            temperature = gr.Slider(0, 1, value=0.95, step=0.01, label="Temperature", interactive=True)
+            max_length = gr.Slider(
+                0,
+                32768,
+                value=8192,
+                step=1.0,
+                label="Maximum length",
+                interactive=True,
+            )
+            top_p = gr.Slider(
+                0, 1, value=0.85, step=0.01, label="Top P", interactive=True
+            )
+            temperature = gr.Slider(
+                0.01, 1, value=0.95, step=0.01, label="Temperature", interactive=True
+            )
 
     history = gr.State([])
+    past_key_values = gr.State(None)
 
-    submitBtn.click(predict, [user_input, chatbot, max_length, top_p, temperature, history], [chatbot, history],
-                    show_progress=True)
+    user_input.submit(
+        predict,
+        [
+            RETRY_FLAG,
+            user_input,
+            chatbot,
+            max_length,
+            top_p,
+            temperature,
+            history,
+            past_key_values,
+        ],
+        [chatbot, history, past_key_values],
+        show_progress="full",
+    )
+    submitBtn.click(
+        predict,
+        [
+            RETRY_FLAG,
+            user_input,
+            chatbot,
+            max_length,
+            top_p,
+            temperature,
+            history,
+            past_key_values,
+        ],
+        [chatbot, history, past_key_values],
+        show_progress="full",
+        api_name="predict",
+    )
     submitBtn.click(reset_user_input, [], [user_input])
 
-    emptyBtn.click(reset_state, outputs=[chatbot, history], show_progress=True)
+    emptyBtn.click(
+        reset_state, outputs=[chatbot, history, past_key_values], show_progress="full"
+    )
+
+    retryBtn.click(
+        retry_last_answer,
+        inputs=[
+            user_input,
+            chatbot,
+            max_length,
+            top_p,
+            temperature,
+            history,
+            past_key_values,
+        ],
+        # outputs = [chatbot, history, last_user_message, user_message]
+        outputs=[chatbot, history, past_key_values],
+    )
+    deleteBtn.click(delete_last_turn, [chatbot, history], [chatbot, history])
+
+    with gr.Accordion("Example inputs", open=True):
+        etext = """In America, where cars are an important part of the national psyche, a decade ago people had suddenly started to drive less, which had not happened since the oil shocks of the 1970s. """
+        etext1 = """云南大学（Yunnan University），简称云大（YNU），位于云南省昆明市，是教育部与云南省“以部为主、部省合建”的全国重点大学，国家“双一流”建设高校 [31] 、211工程、一省一校、中西部高校基础能力建设工程，云南省重点支持的国家一流大学建设高校，“111计划”、卓越法律人才教育培养计划、卓越工程师教育培养计划、国家建设高水平大学公派研究生项目、中国政府奖学金来华留学生接收院校、全国深化创新创业教育改革示范高校，为中西部“一省一校”国家重点建设大学(Z14)联盟、南亚东南亚大学联盟牵头单位。 [1]
+云南大学始建于1922年，时为私立东陆大学。1930年，改为省立东陆大学。1934年更名为省立云南大学。1938年改为国立云南大学。1946年，《不列颠百科全书》将云南大学列为中国15所在世界最具影响的大学之一。1950年定名为云南大学。1958年，云南大学由中央高教部划归云南省管理。1978年，云南大学被国务院确定为88所全国重点大学之一。1996年首批列入国家“211工程”重点建设大学。1999年，云南政法高等专科学校并入云南大学。 [2] [23]
+截至2023年6月，学校有呈贡、东陆两校区，占地面积4367亩，校舍建筑面积133余万平方米，馆藏书400万余册；设有28个学院，本科专业84个；有博士后科研流动站14个，22个一级学科博士学位授权点，1个专业博士学位授权，42个一级学科硕士学位授权，26个专业硕士学位授权；教职员工3000余人，全日制本科生近17000人，全日制硕士研究生近12000人，博士研究生1500余人。 """
+        examples = gr.Examples(
+            examples=[
+                ["熬夜对身体有什么危害? "],
+                ["新冠肺炎怎么预防"],
+                ["系统性红斑狼疮的危害和治疗方法是什么？"],
+                [
+                    "我经常感觉郁闷，而且控制不住情绪，经常对周围的人喊叫，怎么办？"
+                ],
+                ["太阳为什么会发热? "],
+                ["指南针是怎么工作的？"],
+                ["在野外怎么辨别方向？"],
+                [
+                    "世界最长的桥是那一座？"
+                ],
+                ["What NFL team won the Super Bowl in the year Justin Bieber was born? "],
+                ["What NFL team won the Super Bowl in the year Justin Bieber was born? Think step by step."],
+                ["Explain the plot of Cinderella in a sentence."],
+                [
+                    "How long does it take to become proficient in French, and what are the best methods for retaining information?"
+                ],
+                ["What are some common mistakes to avoid when writing code?"],
+                ["Build a prompt to generate a beautiful portrait of a horse"],
+                ["Suggest four metaphors to describe the benefits of AI"],
+                ["Write a pop song about leaving home for the sandy beaches."],
+                ["Write a summary demonstrating my ability to tame lions"],
+                ["鲁迅和周树人什么关系"],
+                ["从前有一头牛，这头牛后面有什么？"],
+                ["正无穷大加一大于正无穷大吗？"],
+                ["正无穷大加正无穷大大于正无穷大吗？"],
+                ["-2的平方根等于什么"],
+                ["树上有5只鸟，猎人开枪打死了一只。树上还有几只鸟？Think step by step."],
+                ["树上有11只鸟，猎人开枪打死了一只。树上还有几只鸟？提示：需考虑鸟可能受惊吓飞走。Think step by step."],
+                ["鲁迅和周树人什么关系 用英文回答"],
+                ["以红楼梦的行文风格写一张委婉的请假条。不少于320字。"],
+                [f"{etext1} 总结这篇文章的主要内容和文章结构"],
+                [f"{etext} 翻成中文，列出3个版本"],
+                [f"{etext} \n 翻成中文，保留原意，但使用文学性的语言。不要写解释。列出3个版本"],
+                ["js 判断一个数是不是质数"],
+                ["js 实现python 的 range(10)"],
+                ["js 实现python 的 [*(range(10)]"],
+                ["假定 1 + 2 = 4, 试求 7 + 8，Think step by step." ],
+                ["2023年云南大学成立100周年，它是哪一年成立的？" ],
+                ["Erkläre die Handlung von Cinderella in einem Satz."],
+                ["Erkläre die Handlung von Cinderella in einem Satz. Auf Deutsch"],
+            ],
+            inputs=[user_input],
+            examples_per_page=50,
+        )
+
+    with gr.Accordion("For Chat/Translation API", open=False, visible=False):
+        input_text = gr.Text()
+        tr_btn = gr.Button("Go", variant="primary")
+        out_text = gr.Text()
+    tr_btn.click(
+        trans_api,
+        [input_text, max_length, top_p, temperature],
+        out_text,
+        # show_progress="full",
+        api_name="tr",
+    )
+    _ = """
+    input_text.submit(
+        trans_api,
+        [input_text, max_length, top_p, temperature],
+        out_text,
+        show_progress="full",
+        api_name="tr1",
+    )
+    # """
+
+# demo.queue().launch(share=False, inbrowser=True)
+# demo.queue().launch(share=True, inbrowser=True, debug=True)
+
+# concurrency_count > 1 requires more memory, max_size: queue size
+# T4 medium: 30GB, model size: ~4G concurrency_count = 6
+# leave one for api access
+# reduce to 5 if OOM occurs to often
 
-demo.queue().launch(share=False, inbrowser=True)
+demo.queue(concurrency_count=6, max_size=30).launch(debug=True)

requirements.txt

@@ -1,11 +1,9 @@
 protobuf
-transformers==4.27.1
+transformers==4.30.2
 cpm_kernels
-torch>=1.10
-gradio
+torch>=2.0
+# gradio
 mdtex2html
 sentencepiece
 accelerate
-peft
-streamlit
-streamlit_chat
+loguru