Upload 3 files

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 +1,83 @@
+# -*- coding:utf-8 -*-
+import gradio as gr
+
+
+title = """<h1 align="left" style="min-width:200px; margin-top:0;"> <img src="https://raw.githubusercontent.com/twitter/twemoji/master/assets/svg/1f432.svg" width="32px" style="display: inline"> Chat with Baize </h1>"""
+description_top = """\
+<div align="left">
+<p> Currently Running: <a href="https://huggingface.co/project-baize/baize-v2-7b">baize-v2-7b</a></p>
+<p>
+Disclaimer: The LLaMA model is a third-party version available on Hugging Face model hub. This demo should be used for research purposes only. Commercial use is strictly prohibited. The model output is not censored and the authors do not endorse the opinions in the generated content. Use at your own risk.
+</p >
+</div>
+"""
+description = """\
+<div align="center" style="margin:16px 0">
+The demo is built on <a href="https://github.com/GaiZhenbiao/ChuanhuChatGPT">ChuanhuChatGPT</a>.
+</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 +1,376 @@
+# -*- 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 GenerationConfig, LlamaForCausalLM, LlamaTokenizer
+
+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"```(.*?)(?:```|$)", re.DOTALL)
+    inline_code_pattern = re.compile(r"`(.*?)`", re.DOTALL)
+    code_blocks = code_block_pattern.findall(md_text)
+    non_code_parts = code_block_pattern.split(md_text)[::2]
+
+    result = []
+    for non_code, code in zip(non_code_parts, code_blocks + [""]):
+        if non_code.strip():
+            non_code = normalize_markdown(non_code)
+            if inline_code_pattern.search(non_code):
+                result.append(markdown(non_code, extensions=["tables"]))
+            else:
+                result.append(mdtex2html.convert(non_code, extensions=["tables"]))
+        if code.strip():
+            code = f"\n```{code}\n\n```"
+            code = markdown_to_html_with_syntax_highlight(code)
+            result.append(code)
+    result = "".join(result)
+    result += ALREADY_CONVERTED_MARK
+    return result
+
+def convert_asis(userinput):
+    return f"<p style=\"white-space:pre-wrap;\">{html.escape(userinput)}</p>"+ALREADY_CONVERTED_MARK
+
+def detect_converted_mark(userinput):
+    if userinput.endswith(ALREADY_CONVERTED_MARK):
+        return True
+    else:
+        return False
+
+
+
+def detect_language(code):
+    if code.startswith("\n"):
+        first_line = ""
+    else:
+        first_line = code.strip().split("\n", 1)[0]
+    language = first_line.lower() if first_line else ""
+    code_without_language = code[len(first_line) :].lstrip() if first_line else code
+    return language, code_without_language
+
+def convert_to_markdown(text):
+    text = text.replace("$","&#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 += f"{line}\n"
+        elif in_code_block is True and line.startswith("```"):
+            in_code_block = False
+            markdown_text += f"{line}\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 named Baize (named after a mythical creature in Chinese folklore). Baize is an open-source AI assistant developed by UCSD and Sun Yat-Sen University. 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=None,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",
+        )
+        if adapter_model is not None:
+            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,
+        )
+        if adapter_model is not None:
+            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
+        )
+        if adapter_model is not None:
+            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
+