Upload 19 files

app.py

@@ -0,0 +1,244 @@
+# -*- coding:utf-8 -*-
+import os
+import logging
+import sys
+import gradio as gr
+import torch
+import gc
+from app_modules.utils import *
+from app_modules.presets import *
+from app_modules.overwrites import *
+
+import os
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+
+
+logging.basicConfig(
+    level=logging.DEBUG,
+    format="%(asctime)s [%(levelname)s] [%(filename)s:%(lineno)d] %(message)s",
+)
+
+# base_model = "decapoda-research/llama-7b-hf"
+# adapter_model = "project-baize/baize-lora-7B"
+
+base_model = "facebook/opt-1.3b"
+# adapter_model = "msuhail97/opt-1.3b-lora"
+# tokenizer, model, device = load_tokenizer_and_model(base_model, adapter_model)           
+
+# finetune_model_path = "/home2/xqchang/Chat-web/ft_models/ft-opt-1.3b"
+tokenizer, model, device = load_finetune_tokenizer_and_model(base_model)
+
+
+total_count = 0
+def predict(text,
+            chatbot,
+            history,
+            top_p,
+            temperature,
+            max_length_tokens,
+            max_context_length_tokens,):
+    if text=="":
+        yield chatbot,history,"Empty context."
+        return 
+    try:
+        model
+    except:
+        yield [[text,"No Model Found"]],[],"No Model Found"
+        return
+
+    inputs = generate_prompt_with_history(text,history,tokenizer,max_length=max_context_length_tokens)
+    if inputs is None:
+        yield chatbot,history,"Input too long."
+        return 
+    else:
+        prompt,inputs=inputs
+        begin_length = len(prompt)
+    input_ids = inputs["input_ids"][:,-max_context_length_tokens:].to(device)
+    torch.cuda.empty_cache()
+    global total_count
+    total_count += 1
+    print(total_count)
+    if total_count % 50 == 0 :
+        os.system("nvidia-smi")
+    with torch.no_grad():
+        for x in greedy_search(input_ids,model,tokenizer,stop_words=["[|Human|]", "[|AI|]"],max_length=max_length_tokens,temperature=temperature,top_p=top_p):
+            if is_stop_word_or_prefix(x,["[|Human|]", "[|AI|]"]) is False:
+                if "[|Human|]" in x:
+                    x = x[:x.index("[|Human|]")].strip()
+                if "[|AI|]" in x:
+                    x = x[:x.index("[|AI|]")].strip() 
+                x = x.strip()   
+                a, b=   [[y[0],convert_to_markdown(y[1])] for y in history]+[[text, convert_to_markdown(x)]],history + [[text,x]]
+                yield a, b, "Generating..."
+            if shared_state.interrupted:
+                shared_state.recover()
+                try:
+                    yield a, b, "Stop: Success"
+                    return
+                except:
+                    pass
+    del input_ids
+    gc.collect()
+    torch.cuda.empty_cache()
+    #print(text)
+    #print(x)
+    #print("="*80)
+    try:
+        yield a,b,"Generate: Success"
+    except:
+        pass
+        
+def retry(
+        text,
+        chatbot,
+        history,
+        top_p,
+        temperature,
+        max_length_tokens,
+        max_context_length_tokens,
+        ):
+    logging.info("Retry...")
+    if len(history) == 0:
+        yield chatbot, history, f"Empty context"
+        return
+    chatbot.pop()
+    inputs = history.pop()[0]
+    for x in predict(inputs,chatbot,history,top_p,temperature,max_length_tokens,max_context_length_tokens):
+        yield x
+
+
+gr.Chatbot.postprocess = postprocess
+
+with open("assets/custom.css", "r", encoding="utf-8") as f:
+    customCSS = f.read()
+
+with gr.Blocks(css=customCSS, theme=small_and_beautiful_theme) as demo:
+    history = gr.State([])
+    user_question = gr.State("")
+    with gr.Row():
+        gr.HTML(title)
+        status_display = gr.Markdown("Success", elem_id="status_display")
+    gr.Markdown(description_top)
+    with gr.Row(scale=1).style(equal_height=True):
+        with gr.Column(scale=5):
+            with gr.Row(scale=1):
+                chatbot = gr.Chatbot(elem_id="chuanhu_chatbot").style(height="100%")
+            with gr.Row(scale=1):
+                with gr.Column(scale=12):
+                    user_input = gr.Textbox(
+                        show_label=False, placeholder="Enter text"
+                    ).style(container=False)
+                with gr.Column(min_width=70, scale=1):
+                    submitBtn = gr.Button("Send")
+                with gr.Column(min_width=70, scale=1):
+                    cancelBtn = gr.Button("Stop")
+            with gr.Row(scale=1):
+                emptyBtn = gr.Button(
+                    "🧹 New Conversation",
+                )
+                retryBtn = gr.Button("🔄 Regenerate")
+                delLastBtn = gr.Button("🗑️ Remove Last Turn") 
+        with gr.Column():
+            with gr.Column(min_width=50, scale=1):
+                with gr.Tab(label="Parameter Setting"):
+                    gr.Markdown("# Parameters")
+                    top_p = gr.Slider(
+                        minimum=-0,
+                        maximum=1.0,
+                        value=0.95,
+                        step=0.05,
+                        interactive=True,
+                        label="Top-p",
+                    )
+                    temperature = gr.Slider(
+                        minimum=0.1,
+                        maximum=2.0,
+                        value=1,
+                        step=0.1,
+                        interactive=True,
+                        label="Temperature",
+                    )
+                    max_length_tokens = gr.Slider(
+                        minimum=0,
+                        maximum=512,
+                        value=256,
+                        step=8,
+                        interactive=True,
+                        label="Max Generation Tokens",
+                    )
+                    max_context_length_tokens = gr.Slider(
+                        minimum=0,
+                        maximum=4096,
+                        value=2048,
+                        step=128,
+                        interactive=True,
+                        label="Max History Tokens",
+                    )
+    # gr.Markdown(description)
+
+    predict_args = dict(
+        fn=predict,
+        inputs=[
+            user_question,
+            chatbot,
+            history,
+            top_p,
+            temperature,
+            max_length_tokens,
+            max_context_length_tokens,
+        ],
+        outputs=[chatbot, history, status_display],
+        show_progress=True,
+    )
+    retry_args = dict(
+        fn=retry,
+        inputs=[
+            user_input,
+            chatbot,
+            history,
+            top_p,
+            temperature,
+            max_length_tokens,
+            max_context_length_tokens,
+        ],
+        outputs=[chatbot, history, status_display],
+        show_progress=True,
+    )
+
+    reset_args = dict(
+        fn=reset_textbox, inputs=[], outputs=[user_input, status_display]
+    )
+ 
+    # Chatbot
+    transfer_input_args = dict(
+        fn=transfer_input, inputs=[user_input], outputs=[user_question, user_input, submitBtn], show_progress=True
+    )
+
+    predict_event1 = user_input.submit(**transfer_input_args).then(**predict_args)
+
+    predict_event2 = submitBtn.click(**transfer_input_args).then(**predict_args)
+
+    emptyBtn.click(
+        reset_state,
+        outputs=[chatbot, history, status_display],
+        show_progress=True,
+    )
+    emptyBtn.click(**reset_args)
+
+    predict_event3 = retryBtn.click(**retry_args)
+
+    delLastBtn.click(
+        delete_last_conversation,
+        [chatbot, history],
+        [chatbot, history, status_display],
+        show_progress=True,
+    )
+    cancelBtn.click(
+        cancel_outputing, [], [status_display], 
+        cancels=[
+            predict_event1,predict_event2,predict_event3
+        ]
+    )    
+# demo.title = "Baize"
+
+demo.queue(concurrency_count=1).launch(server_name="127.0.0.1", server_port=7860, share=True)

app_modules/overwrites.py

@@ -0,0 +1,57 @@
+from __future__ import annotations
+import logging
+
+from llama_index import Prompt
+from typing import List, Tuple
+import mdtex2html
+
+from app_modules.presets import *
+from app_modules.utils import *
+
+def compact_text_chunks(self, prompt: Prompt, text_chunks: List[str]) -> List[str]:
+    logging.debug("Compacting text chunks...🚀🚀🚀")
+    combined_str = [c.strip() for c in text_chunks if c.strip()]
+    combined_str = [f"[{index+1}] {c}" for index, c in enumerate(combined_str)]
+    combined_str = "\n\n".join(combined_str)
+    # resplit based on self.max_chunk_overlap
+    text_splitter = self.get_text_splitter_given_prompt(prompt, 1, padding=1)
+    return text_splitter.split_text(combined_str)
+
+
+def postprocess(
+    self, y: List[Tuple[str | None, str | None]]
+) -> List[Tuple[str | None, str | None]]:
+    """
+    Parameters:
+        y: List of tuples representing the message and response pairs. Each message and response should be a string, which may be in Markdown format.
+    Returns:
+        List of tuples representing the message and response. Each message and response will be a string of HTML.
+    """
+    if y is None or y == []:
+        return []
+    temp = []
+    for x in y:
+        user, bot = x
+        if not detect_converted_mark(user):
+            user = convert_asis(user)
+        if not detect_converted_mark(bot):
+            bot = convert_mdtext(bot)
+        temp.append((user, bot))
+    return temp
+
+with open("./assets/custom.js", "r", encoding="utf-8") as f, open("./assets/Kelpy-Codos.js", "r", encoding="utf-8") as f2:
+    customJS = f.read()
+    kelpyCodos = f2.read()
+
+def reload_javascript():
+    print("Reloading javascript...")
+    js = f'<script>{customJS}</script><script>{kelpyCodos}</script>'
+    def template_response(*args, **kwargs):
+        res = GradioTemplateResponseOriginal(*args, **kwargs)
+        res.body = res.body.replace(b'</html>', f'{js}</html>'.encode("utf8"))
+        res.init_headers()
+        return res
+
+    gr.routes.templates.TemplateResponse = template_response
+
+GradioTemplateResponseOriginal = gr.routes.templates.TemplateResponse

app_modules/presets.py

@@ -0,0 +1,78 @@
+# -*- coding:utf-8 -*-
+import gradio as gr
+
+
+title = """<h1 align="left" style="min-width:200px; margin-top:0;"> OPT </h1>"""
+description_top = """\
+<div align="left">
+<p>
+Disclaimer: The OPT model is a third-party version available on Hugging Face model hub.
+</p >
+</div>
+"""
+
+CONCURRENT_COUNT = 100
+
+
+ALREADY_CONVERTED_MARK = "<!-- ALREADY CONVERTED BY PARSER. -->"
+
+small_and_beautiful_theme = gr.themes.Soft(
+        primary_hue=gr.themes.Color(
+            c50="#02C160",
+            c100="rgba(2, 193, 96, 0.2)",
+            c200="#02C160",
+            c300="rgba(2, 193, 96, 0.32)",
+            c400="rgba(2, 193, 96, 0.32)",
+            c500="rgba(2, 193, 96, 1.0)",
+            c600="rgba(2, 193, 96, 1.0)",
+            c700="rgba(2, 193, 96, 0.32)",
+            c800="rgba(2, 193, 96, 0.32)",
+            c900="#02C160",
+            c950="#02C160",
+        ),
+        secondary_hue=gr.themes.Color(
+            c50="#576b95",
+            c100="#576b95",
+            c200="#576b95",
+            c300="#576b95",
+            c400="#576b95",
+            c500="#576b95",
+            c600="#576b95",
+            c700="#576b95",
+            c800="#576b95",
+            c900="#576b95",
+            c950="#576b95",
+        ),
+        neutral_hue=gr.themes.Color(
+            name="gray",
+            c50="#f9fafb",
+            c100="#f3f4f6",
+            c200="#e5e7eb",
+            c300="#d1d5db",
+            c400="#B2B2B2",
+            c500="#808080",
+            c600="#636363",
+            c700="#515151",
+            c800="#393939",
+            c900="#272727",
+            c950="#171717",
+        ),
+        radius_size=gr.themes.sizes.radius_sm,
+    ).set(
+        button_primary_background_fill="#06AE56",
+        button_primary_background_fill_dark="#06AE56",
+        button_primary_background_fill_hover="#07C863",
+        button_primary_border_color="#06AE56",
+        button_primary_border_color_dark="#06AE56",
+        button_primary_text_color="#FFFFFF",
+        button_primary_text_color_dark="#FFFFFF",
+        button_secondary_background_fill="#F2F2F2",
+        button_secondary_background_fill_dark="#2B2B2B",
+        button_secondary_text_color="#393939",
+        button_secondary_text_color_dark="#FFFFFF",
+        # background_fill_primary="#F7F7F7",
+        # background_fill_primary_dark="#1F1F1F",
+        block_title_text_color="*primary_500",
+        block_title_background_fill="*primary_100",
+        input_background_fill="#F6F6F6",
+    )

app_modules/utils.py

@@ -0,0 +1,454 @@
+# -*- coding:utf-8 -*-
+from __future__ import annotations
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Tuple, Type
+import logging
+import json
+import os
+import datetime
+import hashlib
+import csv
+import requests
+import re
+import html
+import markdown2
+import torch 
+import sys
+import gc
+from pygments.lexers import guess_lexer, ClassNotFound
+
+import gradio as gr
+from pypinyin import lazy_pinyin
+import tiktoken
+import mdtex2html
+from markdown import markdown
+from pygments import highlight
+from pygments.lexers import guess_lexer,get_lexer_by_name
+from pygments.formatters import HtmlFormatter
+import transformers
+from peft import PeftModel
+from transformers import LlamaForCausalLM, LlamaTokenizer
+
+from transformers import (
+    AutoTokenizer, 
+    AutoModelForCausalLM,
+)
+
+from app_modules.presets import *
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s [%(levelname)s] [%(filename)s:%(lineno)d] %(message)s",
+)
+
+
+def markdown_to_html_with_syntax_highlight(md_str):
+    def replacer(match):
+        lang = match.group(1) or "text"
+        code = match.group(2)
+        lang = lang.strip()
+        #print(1,lang)
+        if lang=="text":
+            lexer = guess_lexer(code)
+            lang = lexer.name
+            #print(2,lang)
+        try:
+            lexer = get_lexer_by_name(lang, stripall=True)
+        except ValueError:
+            lexer = get_lexer_by_name("python", stripall=True)
+        formatter = HtmlFormatter()
+        #print(3,lexer.name)
+        highlighted_code = highlight(code, lexer, formatter)
+
+        return f'<pre><code class="{lang}">{highlighted_code}</code></pre>'
+
+    code_block_pattern = r"```(\w+)?\n([\s\S]+?)\n```"
+    md_str = re.sub(code_block_pattern, replacer, md_str, flags=re.MULTILINE)
+
+    html_str = markdown(md_str)
+    return html_str
+
+
+def normalize_markdown(md_text: str) -> str:
+    lines = md_text.split("\n")
+    normalized_lines = []
+    inside_list = False
+
+    for i, line in enumerate(lines):
+        if re.match(r"^(\d+\.|-|\*|\+)\s", line.strip()):
+            if not inside_list and i > 0 and lines[i - 1].strip() != "":
+                normalized_lines.append("")
+            inside_list = True
+            normalized_lines.append(line)
+        elif inside_list and line.strip() == "":
+            if i < len(lines) - 1 and not re.match(
+                r"^(\d+\.|-|\*|\+)\s", lines[i + 1].strip()
+            ):
+                normalized_lines.append(line)
+            continue
+        else:
+            inside_list = False
+            normalized_lines.append(line)
+
+    return "\n".join(normalized_lines)
+
+
+def convert_mdtext(md_text):
+    code_block_pattern = re.compile(r"```(.*?)(?:```|$)", re.DOTALL)
+    inline_code_pattern = re.compile(r"`(.*?)`", re.DOTALL)
+    code_blocks = code_block_pattern.findall(md_text)
+    non_code_parts = code_block_pattern.split(md_text)[::2]
+
+    result = []
+    for non_code, code in zip(non_code_parts, code_blocks + [""]):
+        if non_code.strip():
+            non_code = normalize_markdown(non_code)
+            if inline_code_pattern.search(non_code):
+                result.append(markdown(non_code, extensions=["tables"]))
+            else:
+                result.append(mdtex2html.convert(non_code, extensions=["tables"]))
+        if code.strip():
+            # _, code = detect_language(code)  # 暂时去除代码高亮功能，因为在大段代码的情况下会出现问题
+            # code = code.replace("\n\n", "\n") # 暂时去除代码中的空行，因为在大段代码的情况下会出现问题
+            code = f"\n```{code}\n\n```"
+            code = markdown_to_html_with_syntax_highlight(code)
+            result.append(code)
+    result = "".join(result)
+    result += ALREADY_CONVERTED_MARK
+    return result
+
+def convert_asis(userinput):
+    return f"<p style=\"white-space:pre-wrap;\">{html.escape(userinput)}</p>"+ALREADY_CONVERTED_MARK
+
+def detect_converted_mark(userinput):
+    if userinput.endswith(ALREADY_CONVERTED_MARK):
+        return True
+    else:
+        return False
+
+
+
+def detect_language(code):
+    if code.startswith("\n"):
+        first_line = ""
+    else:
+        first_line = code.strip().split("\n", 1)[0]
+    language = first_line.lower() if first_line else ""
+    code_without_language = code[len(first_line) :].lstrip() if first_line else code
+    return language, code_without_language
+
+def convert_to_markdown(text):
+    text = text.replace("$","&#36;")
+    def replace_leading_tabs_and_spaces(line):
+        new_line = []
+        
+        for char in line:
+            if char == "\t":
+                new_line.append("&#9;")
+            elif char == " ":
+                new_line.append("&nbsp;")
+            else:
+                break
+        return "".join(new_line) + line[len(new_line):]
+
+    markdown_text = ""
+    lines = text.split("\n")
+    in_code_block = False
+
+    for line in lines:
+        if in_code_block is False and line.startswith("```"):
+            in_code_block = True
+            markdown_text += "```\n"
+        elif in_code_block is True and line.startswith("```"):
+            in_code_block = False
+            markdown_text += "```\n"
+        elif in_code_block:
+            markdown_text += f"{line}\n"
+        else:
+            line = replace_leading_tabs_and_spaces(line)
+            line = re.sub(r"^(#)", r"\\\1", line)
+            markdown_text += f"{line}  \n"
+
+    return markdown_text
+
+def add_language_tag(text):
+    def detect_language(code_block):
+        try:
+            lexer = guess_lexer(code_block)
+            return lexer.name.lower()
+        except ClassNotFound:
+            return ""
+
+    code_block_pattern = re.compile(r"(```)(\w*\n[^`]+```)", re.MULTILINE)
+
+    def replacement(match):
+        code_block = match.group(2)
+        if match.group(2).startswith("\n"):
+            language = detect_language(code_block)
+            if language:
+                return f"```{language}{code_block}```"
+            else:
+                return f"```\n{code_block}```"
+        else:
+            return match.group(1) + code_block + "```"
+
+    text2 = code_block_pattern.sub(replacement, text)
+    return text2
+
+def delete_last_conversation(chatbot, history):
+    if len(chatbot) > 0:
+        chatbot.pop()
+
+    if len(history) > 0:
+        history.pop()
+        
+    return (
+        chatbot,
+        history,
+        "Delete Done",
+    )
+
+def reset_state():
+    return [], [], "Reset Done"
+
+def reset_textbox():
+    return gr.update(value=""),""
+
+def cancel_outputing():
+    return "Stop Done"
+
+def transfer_input(inputs):
+    # 一次性返回，降低延迟
+    textbox = reset_textbox()
+    return (
+        inputs,
+        gr.update(value=""),
+        gr.Button.update(visible=True),
+    )
+
+
+class State:
+    interrupted = False
+
+    def interrupt(self):
+        self.interrupted = True
+
+    def recover(self):
+        self.interrupted = False
+shared_state = State()
+
+
+
+
+
+# Greedy Search
+def greedy_search(input_ids: torch.Tensor,
+                  model: torch.nn.Module,
+                  tokenizer: transformers.PreTrainedTokenizer,
+                  stop_words: list,
+                  max_length: int,
+                  temperature: float = 1.0,
+                  top_p: float = 1.0,
+                  top_k: int = 25) -> Iterator[str]:
+    generated_tokens = []
+    past_key_values = None
+    current_length = 1
+    for i in range(max_length):
+        with torch.no_grad():
+            if past_key_values is None:
+                outputs = model(input_ids)
+            else:
+                outputs = model(input_ids[:, -1:], past_key_values=past_key_values)
+            logits = outputs.logits[:, -1, :]                       
+            past_key_values = outputs.past_key_values
+
+            # apply temperature
+            logits /= temperature
+    
+            probs = torch.softmax(logits, dim=-1)
+            # apply top_p
+            probs_sort, probs_idx = torch.sort(probs, dim=-1, descending=True)
+            probs_sum = torch.cumsum(probs_sort, dim=-1)
+            mask = probs_sum - probs_sort > top_p
+            probs_sort[mask] = 0.0
+    
+            # apply top_k
+            # if top_k is not None:
+            #    probs_sort1, _ = torch.topk(probs_sort, top_k)
+            #    min_top_probs_sort = torch.min(probs_sort1, dim=-1, keepdim=True).values
+            #    probs_sort = torch.where(probs_sort < min_top_probs_sort, torch.full_like(probs_sort, float(0.0)), probs_sort)
+    
+            probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
+            next_token = torch.multinomial(probs_sort, num_samples=1)
+            next_token = torch.gather(probs_idx, -1, next_token)
+    
+            input_ids = torch.cat((input_ids, next_token), dim=-1)
+    
+            generated_tokens.append(next_token[0].item())
+            text = tokenizer.decode(generated_tokens)
+    
+            yield text
+            if any([x in text for x in stop_words]):
+                del past_key_values
+                del logits
+                del probs
+                del probs_sort
+                del probs_idx
+                del probs_sum
+                gc.collect()
+                return 
+
+def generate_prompt_with_history(text,history,tokenizer,max_length=2048):
+    prompt = "The following is a conversation between a human and an AI assistant. The human and the AI assistant take turns chatting. Human statements start with [|Human|] and AI assistant statements start with [|AI|]. The AI assistant always provides responses in as much detail as possible, and in Markdown format. The AI assistant always declines to engage with topics, questions and instructions related to unethical, controversial, or sensitive issues. Complete the transcript in exactly that format.\n[|Human|]Hello!\n[|AI|]Hi!"   
+    history = ["\n[|Human|]{}\n[|AI|]{}".format(x[0],x[1]) for x in history]
+    history.append("\n[|Human|]{}\n[|AI|]".format(text))
+    history_text = ""
+    flag = False
+    for x in history[::-1]:
+        if tokenizer(prompt+history_text+x, return_tensors="pt")['input_ids'].size(-1) <= max_length:
+            history_text = x + history_text
+            flag = True
+        else:
+            break
+    if flag:
+        return  prompt+history_text,tokenizer(prompt+history_text, return_tensors="pt")
+    else:
+        return None
+
+
+def is_stop_word_or_prefix(s: str, stop_words: list) -> bool:
+    for stop_word in stop_words:
+        if s.endswith(stop_word):
+            return True
+        for i in range(1, len(stop_word)):
+            if s.endswith(stop_word[:i]):
+                return True
+    return False
+
+
+
+def load_tokenizer_and_model(base_model, adapter_model, load_8bit=False):         #base_model, adapter_model, load_8bit=False
+    if torch.cuda.is_available():
+        device = "cuda"
+    else:
+        device = "cpu"
+
+    try:
+        if torch.backends.mps.is_available():
+            device = "mps"
+    except:  # noqa: E722
+        pass
+    # tokenizer = LlamaTokenizer.from_pretrained(base_model)
+    # if device == "cuda":
+    #     model = LlamaForCausalLM.from_pretrained(
+    #         base_model,
+    #         load_in_8bit=load_8bit,
+    #         torch_dtype=torch.float16,
+    #         device_map="auto",
+    #     )
+    #     model = PeftModel.from_pretrained(
+    #         model,
+    #         adapter_model,
+    #         torch_dtype=torch.float16,
+    #     )
+    # elif device == "mps":
+    #     model = LlamaForCausalLM.from_pretrained(
+    #         base_model,
+    #         device_map={"": device},
+    #         torch_dtype=torch.float16,
+    #     )
+    #     model = PeftModel.from_pretrained(
+    #         model,
+    #         adapter_model,
+    #         device_map={"": device},
+    #         torch_dtype=torch.float16,
+    #     )
+    # else:
+    #     model = LlamaForCausalLM.from_pretrained(
+    #         base_model, device_map={"": device}, low_cpu_mem_usage=True
+    #     )
+    #     model = PeftModel.from_pretrained(
+    #         model,
+    #         adapter_model,
+    #         device_map={"": device},
+    #     )
+
+    tokenizer = AutoTokenizer.from_pretrained(base_model)
+    if device == "cuda":
+        model = AutoModelForCausalLM.from_pretrained(
+            base_model,
+            load_in_8bit=load_8bit,
+            torch_dtype=torch.float16,
+            device_map="auto",
+        )
+        model = PeftModel.from_pretrained(
+            model,
+            adapter_model,
+            torch_dtype=torch.float16,
+        )
+    elif device == "mps":
+        model = AutoModelForCausalLM.from_pretrained(
+            base_model,
+            device_map={"": device},
+            torch_dtype=torch.float16,
+        )
+        model = PeftModel.from_pretrained(
+            model,
+            adapter_model,
+            device_map={"": device},
+            torch_dtype=torch.float16,
+        )
+    else:
+        model = AutoModelForCausalLM.from_pretrained(
+            base_model, device_map={"": device}, low_cpu_mem_usage=True
+        )
+        model = PeftModel.from_pretrained(
+            model,
+            adapter_model,
+            device_map={"": device},
+        )
+
+    if not load_8bit:
+        model.half()  # seems to fix bugs for some users.
+
+    model.eval()
+    return tokenizer, model, device
+
+
+
+def load_finetune_tokenizer_and_model(base_model_name_or_path, load_8bit=False):       
+    if torch.cuda.is_available():
+        device = "cuda"
+    else:
+        device = "cpu"
+
+    try:
+        if torch.backends.mps.is_available():
+            device = "mps"
+    except:  # noqa: E722
+        pass
+
+    tokenizer = AutoTokenizer.from_pretrained(base_model_name_or_path)
+    if device == "cuda":
+        model = AutoModelForCausalLM.from_pretrained(
+            base_model_name_or_path,
+            load_in_8bit=load_8bit,
+            torch_dtype=torch.float16,
+            device_map="auto",
+        )
+    elif device == "mps":
+        model = AutoModelForCausalLM.from_pretrained(
+            base_model_name_or_path,
+            device_map={"": device},
+            torch_dtype=torch.float16,
+        )
+    else:
+        model = AutoModelForCausalLM.from_pretrained(
+            base_model_name_or_path, device_map={"": device}, low_cpu_mem_usage=True
+        )
+
+    if not load_8bit:
+        model.half()  # seems to fix bugs for some users.
+
+    model.eval()
+    return tokenizer, model, device
+

assets/Kelpy-Codos.js

@@ -0,0 +1,76 @@
+// ==UserScript==
+// @name         Kelpy Codos
+// @namespace    https://github.com/Keldos-Li/Kelpy-Codos
+// @version      1.0.5
+// @author       Keldos; https://keldos.me/
+// @description  Add copy button to PRE tags before CODE tag, for Chuanhu ChatGPT especially. 
+//               Based on Chuanhu ChatGPT version: ac04408 (2023-3-22)
+// @license      GPL-3.0
+// @grant        none
+// ==/UserScript==
+
+(function () {
+    'use strict';
+
+    function addCopyButton(pre) {
+        var code = pre.querySelector('code');
+        if (!code) {
+            return; // 如果没有找到 <code> 元素，则不添加按钮
+        }
+        var firstChild = code.firstChild;
+        if (!firstChild) {
+            return; // 如果 <code> 元素没有子节点，则不添加按钮
+        }
+        var button = document.createElement('button');
+        button.textContent = '\uD83D\uDCCE'; // 使用 📎 符号作为“复制”按钮的文本
+        button.style.position = 'relative';
+        button.style.float = 'right';
+        button.style.fontSize = '1em'; // 可选：调整按钮大小
+        button.style.background = 'none'; // 可选：去掉背景颜色
+        button.style.border = 'none'; // 可选：去掉边框
+        button.style.cursor = 'pointer'; // 可选：显示指针样式
+        button.addEventListener('click', function () {
+            var range = document.createRange();
+            range.selectNodeContents(code);
+            range.setStartBefore(firstChild); // 将范围设置为第一个子节点之前
+            var selection = window.getSelection();
+            selection.removeAllRanges();
+            selection.addRange(range);
+
+            try {
+                var success = document.execCommand('copy');
+                if (success) {
+                    button.textContent = '\u2714';
+                    setTimeout(function () {
+                        button.textContent = '\uD83D\uDCCE'; // 恢复按钮为“复制”
+                    }, 2000);
+                } else {
+                    button.textContent = '\u2716';
+                }
+            } catch (e) {
+                console.error(e);
+                button.textContent = '\u2716';
+            }
+
+            selection.removeAllRanges();
+        });
+        code.insertBefore(button, firstChild); // 将按钮插入到第一个子元素之前
+    }
+
+    function handleNewElements(mutationsList, observer) {
+        for (var mutation of mutationsList) {
+            if (mutation.type === 'childList') {
+                for (var node of mutation.addedNodes) {
+                    if (node.nodeName === 'PRE') {
+                        addCopyButton(node);
+                    }
+                }
+            }
+        }
+    }
+
+    var observer = new MutationObserver(handleNewElements);
+    observer.observe(document.documentElement, { childList: true, subtree: true });
+
+    document.querySelectorAll('pre').forEach(addCopyButton);
+})();

assets/custom.css

@@ -0,0 +1,190 @@
+:root {
+    --chatbot-color-light: #F3F3F3;
+    --chatbot-color-dark: #121111;
+}
+
+/* status_display */
+#status_display {
+    display: flex;
+    min-height: 2.5em;
+    align-items: flex-end;
+    justify-content: flex-end;
+}
+#status_display p {
+    font-size: .85em;
+    font-family: monospace;
+    color: var(--body-text-color-subdued);
+}
+
+
+
+/* usage_display */
+#usage_display {
+    height: 1em;
+}
+#usage_display p{
+    padding: 0 1em;
+    font-size: .85em;
+    font-family: monospace;
+    color: var(--body-text-color-subdued);
+}
+/* list */
+ol:not(.options), ul:not(.options) {
+    padding-inline-start: 2em !important;
+}
+
+/* Thank @Keldos-Li for fixing it */
+/* Light mode (default) */
+#chuanhu_chatbot {
+    background-color: var(--chatbot-color-light) !important;
+    color: #000000 !important;
+}
+[data-testid = "bot"] {
+    background-color: #FFFFFF !important;
+}
+[data-testid = "user"] {
+    background-color: #95EC69 !important;
+}
+
+/* Dark mode */
+.dark #chuanhu_chatbot {
+    background-color: var(--chatbot-color-dark) !important;
+    color: #FFFFFF !important;
+}
+.dark [data-testid = "bot"] {
+    background-color: #2C2C2C !important;
+}
+.dark [data-testid = "user"] {
+    background-color: #26B561 !important;
+}
+
+#chuanhu_chatbot {
+    height: 100%;
+    min-height: 400px;
+}
+
+[class *= "message"] {
+    border-radius: var(--radius-xl) !important;
+    border: none;
+    padding: var(--spacing-xl) !important;
+    font-size: var(--text-md) !important;
+    line-height: var(--line-md) !important;
+    min-height: calc(var(--text-md)*var(--line-md) + 2*var(--spacing-xl));
+    min-width: calc(var(--text-md)*var(--line-md) + 2*var(--spacing-xl));
+}
+[data-testid = "bot"] {
+    max-width: 85%;
+    border-bottom-left-radius: 0 !important;
+}
+[data-testid = "user"] {
+    max-width: 85%;
+    width: auto !important;
+    border-bottom-right-radius: 0 !important;
+}
+/* Table */
+table {
+    margin: 1em 0;
+    border-collapse: collapse;
+    empty-cells: show;
+}
+td,th {
+    border: 1.2px solid var(--border-color-primary) !important;
+    padding: 0.2em;
+}
+thead {
+    background-color: rgba(175,184,193,0.2);
+}
+thead th {
+    padding: .5em .2em;
+}
+/* Inline code */
+code {
+    display: inline;
+    white-space: break-spaces;
+    border-radius: 6px;
+    margin: 0 2px 0 2px;
+    padding: .2em .4em .1em .4em;
+    background-color: rgba(175,184,193,0.2);
+}
+/* Code block */
+pre code {
+    display: block;
+    overflow: auto;
+    white-space: pre;
+    background-color: hsla(0, 0%, 0%, 80%)!important;
+    border-radius: 10px;
+    padding: 1.4em 1.2em 0em 1.4em;
+    margin: 1.2em 2em 1.2em 0.5em;
+    color: #FFF;
+    box-shadow: 6px 6px 16px hsla(0, 0%, 0%, 0.2);
+}
+/* Hightlight */
+.highlight .hll { background-color: #49483e }
+.highlight .c { color: #75715e } /* Comment */
+.highlight .err { color: #960050; background-color: #1e0010 } /* Error */
+.highlight .k { color: #66d9ef } /* Keyword */
+.highlight .l { color: #ae81ff } /* Literal */
+.highlight .n { color: #f8f8f2 } /* Name */
+.highlight .o { color: #f92672 } /* Operator */
+.highlight .p { color: #f8f8f2 } /* Punctuation */
+.highlight .ch { color: #75715e } /* Comment.Hashbang */
+.highlight .cm { color: #75715e } /* Comment.Multiline */
+.highlight .cp { color: #75715e } /* Comment.Preproc */
+.highlight .cpf { color: #75715e } /* Comment.PreprocFile */
+.highlight .c1 { color: #75715e } /* Comment.Single */
+.highlight .cs { color: #75715e } /* Comment.Special */
+.highlight .gd { color: #f92672 } /* Generic.Deleted */
+.highlight .ge { font-style: italic } /* Generic.Emph */
+.highlight .gi { color: #a6e22e } /* Generic.Inserted */
+.highlight .gs { font-weight: bold } /* Generic.Strong */
+.highlight .gu { color: #75715e } /* Generic.Subheading */
+.highlight .kc { color: #66d9ef } /* Keyword.Constant */
+.highlight .kd { color: #66d9ef } /* Keyword.Declaration */
+.highlight .kn { color: #f92672 } /* Keyword.Namespace */
+.highlight .kp { color: #66d9ef } /* Keyword.Pseudo */
+.highlight .kr { color: #66d9ef } /* Keyword.Reserved */
+.highlight .kt { color: #66d9ef } /* Keyword.Type */
+.highlight .ld { color: #e6db74 } /* Literal.Date */
+.highlight .m { color: #ae81ff } /* Literal.Number */
+.highlight .s { color: #e6db74 } /* Literal.String */
+.highlight .na { color: #a6e22e } /* Name.Attribute */
+.highlight .nb { color: #f8f8f2 } /* Name.Builtin */
+.highlight .nc { color: #a6e22e } /* Name.Class */
+.highlight .no { color: #66d9ef } /* Name.Constant */
+.highlight .nd { color: #a6e22e } /* Name.Decorator */
+.highlight .ni { color: #f8f8f2 } /* Name.Entity */
+.highlight .ne { color: #a6e22e } /* Name.Exception */
+.highlight .nf { color: #a6e22e } /* Name.Function */
+.highlight .nl { color: #f8f8f2 } /* Name.Label */
+.highlight .nn { color: #f8f8f2 } /* Name.Namespace */
+.highlight .nx { color: #a6e22e } /* Name.Other */
+.highlight .py { color: #f8f8f2 } /* Name.Property */
+.highlight .nt { color: #f92672 } /* Name.Tag */
+.highlight .nv { color: #f8f8f2 } /* Name.Variable */
+.highlight .ow { color: #f92672 } /* Operator.Word */
+.highlight .w { color: #f8f8f2 } /* Text.Whitespace */
+.highlight .mb { color: #ae81ff } /* Literal.Number.Bin */
+.highlight .mf { color: #ae81ff } /* Literal.Number.Float */
+.highlight .mh { color: #ae81ff } /* Literal.Number.Hex */
+.highlight .mi { color: #ae81ff } /* Literal.Number.Integer */
+.highlight .mo { color: #ae81ff } /* Literal.Number.Oct */
+.highlight .sa { color: #e6db74 } /* Literal.String.Affix */
+.highlight .sb { color: #e6db74 } /* Literal.String.Backtick */
+.highlight .sc { color: #e6db74 } /* Literal.String.Char */
+.highlight .dl { color: #e6db74 } /* Literal.String.Delimiter */
+.highlight .sd { color: #e6db74 } /* Literal.String.Doc */
+.highlight .s2 { color: #e6db74 } /* Literal.String.Double */
+.highlight .se { color: #ae81ff } /* Literal.String.Escape */
+.highlight .sh { color: #e6db74 } /* Literal.String.Heredoc */
+.highlight .si { color: #e6db74 } /* Literal.String.Interpol */
+.highlight .sx { color: #e6db74 } /* Literal.String.Other */
+.highlight .sr { color: #e6db74 } /* Literal.String.Regex */
+.highlight .s1 { color: #e6db74 } /* Literal.String.Single */
+.highlight .ss { color: #e6db74 } /* Literal.String.Symbol */
+.highlight .bp { color: #f8f8f2 } /* Name.Builtin.Pseudo */
+.highlight .fm { color: #a6e22e } /* Name.Function.Magic */
+.highlight .vc { color: #f8f8f2 } /* Name.Variable.Class */
+.highlight .vg { color: #f8f8f2 } /* Name.Variable.Global */
+.highlight .vi { color: #f8f8f2 } /* Name.Variable.Instance */
+.highlight .vm { color: #f8f8f2 } /* Name.Variable.Magic */
+.highlight .il { color: #ae81ff } /* Literal.Number.Integer.Long */

assets/custom.js

@@ -0,0 +1 @@
+// custom javascript here

requirements.txt

@@ -0,0 +1,17 @@
+gradio @ https://gradio-builds.s3.amazonaws.com/reset-iterator/attempt-01/gradio-3.24.1-py3-none-any.whl
+mdtex2html
+pypinyin
+tiktoken
+socksio
+tqdm
+colorama
+duckduckgo_search
+Pygments
+llama_index
+langchain
+markdown
+markdown2
+torch
+git+https://github.com/huggingface/peft.git
+git+https://github.com/huggingface/transformers.git
+SentencePiece