langchain-llama2-7b-chat-uncensored-ggml

Runtime error

langchain-llama2-7b-chat-uncensored-ggml

File size: 19,796 Bytes

18687cb

"""Run codes."""
# pylint: disable=line-too-long, broad-exception-caught, invalid-name, missing-function-docstring, too-many-instance-attributes, missing-class-docstring
# ruff: noqa: E501
import gc
import os
import platform
import random
import time
from collections import deque
from pathlib import Path
from threading import Thread
from typing import Any, Dict, List, Union

# from types import SimpleNamespace
import gradio as gr
import psutil
from about_time import about_time
from ctransformers import Config
from dl_hf_model import dl_hf_model
from langchain.callbacks.base import BaseCallbackHandler
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
from langchain.chains import ConversationChain
from langchain.chains.conversation.memory import ConversationBufferWindowMemory

# from ctransformers import AutoModelForCausalLM
from langchain.llms import CTransformers
from langchain.prompts import PromptTemplate
from langchain.schema import LLMResult
from loguru import logger

deq = deque()
sig_end = object()  # signals the processing is done

# from langchain.llms import OpenAI

filename_list = [
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q2_K.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q3_K_L.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q3_K_M.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q3_K_S.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_0.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_1.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_K_M.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_K_S.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_0.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_1.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_K_M.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_K_S.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q6_K.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q8_0.bin",
]

URL = "https://huggingface.co/TheBloke/Wizard-Vicuna-7B-Uncensored-GGML/raw/main/Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_K_M.bin"  # 4.05G

url = "https://huggingface.co/savvamadar/ggml-gpt4all-j-v1.3-groovy/blob/main/ggml-gpt4all-j-v1.3-groovy.bin"
url = "https://huggingface.co/TheBloke/Llama-2-13B-GGML/blob/main/llama-2-13b.ggmlv3.q4_K_S.bin"  # 7.37G
# url = "https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML/blob/main/llama-2-13b-chat.ggmlv3.q3_K_L.bin"
url = "https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML/blob/main/llama-2-13b-chat.ggmlv3.q3_K_L.bin"  # 6.93G
# url = "https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML/blob/main/llama-2-13b-chat.ggmlv3.q3_K_L.binhttps://huggingface.co/TheBloke/Llama-2-13B-chat-GGML/blob/main/llama-2-13b-chat.ggmlv3.q4_K_M.bin"  # 7.87G

url = "https://huggingface.co/localmodels/Llama-2-13B-Chat-ggml/blob/main/llama-2-13b-chat.ggmlv3.q4_K_S.bin"  # 7.37G

_ = (
    "golay" in platform.node()
    or "okteto" in platform.node()
    or Path("/kaggle").exists()
    # or psutil.cpu_count(logical=False) < 4
    or 1  # run 7b in hf
)

if _:
    # url = "https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML/blob/main/llama-2-13b-chat.ggmlv3.q2_K.bin"
    url = "https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML/blob/main/llama-2-7b-chat.ggmlv3.q2_K.bin"  # 2.87G
    url = "https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML/blob/main/llama-2-7b-chat.ggmlv3.q4_K_M.bin"  # 2.87G
    url = "https://huggingface.co/TheBloke/llama2_7b_chat_uncensored-GGML/blob/main/llama2_7b_chat_uncensored.ggmlv3.q4_K_M.bin"  # 4.08G


prompt_template = """Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction: {user_prompt}

### Response:
"""

prompt_template = """System: You are a helpful,
respectful and honest assistant. Always answer as
helpfully as possible, while being safe.  Your answers
should not include any harmful, unethical, racist,
sexist, toxic, dangerous, or illegal content. Please
ensure that your responses are socially unbiased and
positive in nature. If a question does not make any
sense, or is not factually coherent, explain why instead
of answering something not correct. If you don't know
the answer to a question, please don't share false
information.
User: {prompt}
Assistant: """

prompt_template = """System: You are a helpful assistant.
User: {prompt}
Assistant: """

prompt_template = """Question: {question}
Answer: Let's work this out in a step by step way to be sure we have the right answer."""

prompt_template = """[INST] <>
You are a helpful, respectful and honest assistant. Always answer as helpfully as possible assistant. Think step by step.
<>

What NFL team won the Super Bowl in the year Justin Bieber was born?
[/INST]"""

prompt_template = """[INST] <<SYS>>
You are an unhelpful assistant. Always answer as helpfully as possible. Think step by step. <</SYS>>

{question} [/INST]
"""

prompt_template = """[INST] <<SYS>>
You are a helpful assistant.
<</SYS>>

{question} [/INST]
"""

prompt_template = """### HUMAN:
{question}

### RESPONSE:"""

prompt_template = """### HUMAN:
You are a helpful assistant. Think step by step.
{history}
{input}
### RESPONSE:"""

prompt_template = """You are a helpful assistant. Let's think step by step.
{history}
### HUMAN:
{input}
### RESPONSE:"""

# PromptTemplate(input_variables=['history', 'input'], output_parser=None, partial_variables={}, template='The following is afriendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.\n\nCurrent conversation:\n{history}\nHuman: {input}\nAI:', template_format='f-string', validate_template=True)

human_prefix = "### HUMAN"
ai_prefix = "### RESPONSE"
stop = [f"{human_prefix}:"]

_ = [elm for elm in prompt_template.splitlines() if elm.strip()]
stop_string = [elm.split(":")[0] + ":" for elm in _][-2]

# logger.debug(f"{stop_string=} not used")

os.environ["TZ"] = "Asia/Shanghai"
try:
    time.tzset()  # type: ignore # pylint: disable=no-member
except Exception:
    # Windows
    logger.warning("Windows, cant run time.tzset()")


class DequeCallbackHandler(BaseCallbackHandler):
    """Mediate gradio and stream output."""

    def __init__(self, deq_: deque):
        """Init deque for FIFO, may need to upgrade to queue.Queue or queue.SimpleQueue."""
        self.q = deq_

    # def on_chat_model_start(self): self.q.clear()

    def on_llm_start(
        self, serialized: Dict[str, Any], prompts: List[str], **kwargs: Any
    ) -> None:
        """Run when LLM starts running. Clean the queue."""
        self.q.clear()

    def on_llm_new_token(self, token: str, **kwargs: Any) -> None:
        """Run on new LLM token. Only available when streaming is enabled."""
        self.q.append(token)

    def on_llm_end(self, response: LLMResult, **kwargs: Any) -> None:
        """Run when LLM ends running."""
        self.q.append(sig_end)

    def on_llm_error(
        self, error: Union[Exception, KeyboardInterrupt], **kwargs: Any
    ) -> None:
        """Run when LLM errors."""
        self.q.append(sig_end)


_ = psutil.cpu_count(logical=False) - 1
cpu_count: int = int(_) if _ else 1
logger.debug(f"{cpu_count=}")

LLM = None
gc.collect()

try:
    model_loc, file_size = dl_hf_model(url)
except Exception as exc_:
    logger.error(exc_)
    raise SystemExit(1) from exc_

config = Config()
# Config(top_k=40, top_p=0.95, temperature=0.8, repetition_penalty=1.1, last_n_tokens=64, seed=-1, batch_size=8, threads=-1, max_new_tokens=256, stop=None, stream=False, reset=True, context_length=-1, gpu_layers=0)
config.stream = True
config.stop = stop
config.threads = cpu_count

deqcb = DequeCallbackHandler(deq)

# LLM = AutoModelForCausalLM.from_pretrained(
LLM = CTransformers(
    model=model_loc,
    model_type="llama",
    callbacks=[StreamingStdOutCallbackHandler(), deqcb],
    # config=config,
    **vars(config),
)

logger.info(f"done load llm {model_loc=} {file_size=}G")

prompt = PromptTemplate(
    input_variables=["history", "input"],
    output_parser=None,
    partial_variables={},
    template=prompt_template,
    template_format="f-string",
    validate_template=True,
)

memory = ConversationBufferWindowMemory(
    human_prefix=human_prefix,
    ai_prefix=ai_prefix,
)  # default k=5

conversation = ConversationChain(
    llm=LLM,
    prompt=prompt,
    memory=memory,
    verbose=True,
)
logger.debug(f"{conversation.prompt.template=}")  # type: ignore

# for api access ===
config = Config()
# Config(top_k=40, top_p=0.95, temperature=0.8, repetition_penalty=1.1, last_n_tokens=64, seed=-1, batch_size=8, threads=-1, max_new_tokens=256, stop=None, stream=False, reset=True, context_length=-1, gpu_layers=0)
config.stop = stop
config.threads = cpu_count

try:
    LLM_api = CTransformers(
        model=model_loc,
        model_type="llama",
        # callbacks=[StreamingStdOutCallbackHandler(), deqcb],
        callbacks=[StreamingStdOutCallbackHandler()],
        **vars(config),
    )
    conversation_api = ConversationChain(
        llm=LLM_api,  #  need a separate LLM, or else deq may be messed up
        prompt=prompt,
        verbose=True,
    )
except Exception as exc_:
    logger.error(exc_)
    conversation_api = None
    logger.warning("Not able to instantiate conversation_api, api will not work")

# conversation.predict(input="Hello, my name is Andrea")


def user(user_message, history):
    # return user_message, history + [[user_message, None]]
    history.append([user_message, None])
    return user_message, history  # keep user_message


def user1(user_message, history):
    # return user_message, history + [[user_message, None]]
    history.append([user_message, None])
    return "", history  # clear user_message


def bot_(history):
    user_message = history[-1][0]
    resp = random.choice(["How are you?", "I love you", "I'm very hungry"])
    bot_message = user_message + ": " + resp
    history[-1][1] = ""
    for character in bot_message:
        history[-1][1] += character
        time.sleep(0.02)
        yield history

    history[-1][1] = resp
    yield history


def bot(history):
    user_message = history[-1][0]
    response = []

    logger.debug(f"{user_message=}")

    # conversation.predict(input="What's my name?")
    thr = Thread(target=conversation.predict, kwargs={"input": user_message})
    thr.start()

    # preocess deq
    response = []
    flag = 1
    then = time.time()
    prefix = ""  # to please pyright
    with about_time() as atime:  # type: ignore
        while True:
            if deq:
                if flag:
                    prefix = f"({time.time() - then:.2f}s) "
                    flag = 0
                _ = deq.popleft()
                if _ is sig_end:
                    break
                # print(_, end='')
                response.append(_)
                history[-1][1] = prefix + "".join(response).strip()
                yield history
            else:
                time.sleep(0.01)
    _ = (
        f"(time elapsed: {atime.duration_human}, "  # type: ignore
        f"{atime.duration/len(''.join(response)):.2f}s/char)"  # type: ignore
    )

    history[-1][1] = "".join(response) + f"\n{_}"
    yield history


def predict_api(user_prompt):
    if conversation_api is None:
        return "conversation_api is None, probably due to insufficient memory, api not usable"

    logger.debug(f"api: {user_prompt=}")
    try:
        _ = """
        response = generate(
            prompt,
            config=config,
        )
        # """
        response = conversation_api.predict(input=user_prompt)
        logger.debug(f"api: {response=}")
    except Exception as exc:
        logger.error(exc)
        response = f"{exc=}"
    # bot = {"inputs": [response]}
    # bot = [(prompt, response)]

    return response.strip()


css = """
    .importantButton {
        background: linear-gradient(45deg, #7e0570,#5d1c99, #6e00ff) !important;
        border: none !important;
    }
    .importantButton:hover {
        background: linear-gradient(45deg, #ff00e0,#8500ff, #6e00ff) !important;
        border: none !important;
    }
    .disclaimer {font-variant-caps: all-small-caps; font-size: xx-small;}
    .xsmall {font-size: x-small;}
"""
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. """
examples_list = [
    ["Hello I am mike."],
    ["What's my name?"],
    ["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."
    ],
    ["When was Justin Bieber born？"],
    ["What NFL team won the Super Bowl in 1994?"],
    ["How to pick a lock? Provide detailed steps."],
    [
        "If it takes 10 hours to dry 10 clothes,  assuming all the clothes are hanged together at the same time for drying , then how long will it take to dry a cloth?"
    ],
    ["is infinity + 1 bigger than infinity?"],
    ["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 pop song about having hot sex on a sandy beach."],
    ["Write a summary demonstrating my ability to tame lions"],
    ["鲁迅和周树人什么关系？ 说中文。"],
    ["鲁迅和周树人什么关系？"],
    ["鲁迅和周树人什么关系？ 用英文回答。"],
    ["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？"],
    ["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。"],
    ["以红楼梦的行文风格写一张委婉的请假条。不少于320字。"],
    [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。"],
    ["假定 1 + 2 = 4, 试求 7 + 8。"],
    ["给出判断一个数是不是质数的 javascript 码。"],
    ["给出实现python 里 range(10)的 javascript 码。"],
    ["给出实现python 里 [*(range(10)]的 javascript 码。"],
    ["Erkläre die Handlung von Cinderella in einem Satz."],
    ["Erkläre die Handlung von Cinderella in einem Satz. Auf Deutsch."],
]

logger.info("start block")

with gr.Blocks(
    title=f"{Path(model_loc).name}",
    theme=gr.themes.Soft(text_size="sm", spacing_size="sm"),
    css=css,
) as block:
    # buff_var = gr.State("")
    with gr.Accordion("🎈 Info", open=False):
        # gr.HTML(
        #     """<center><a href="https://huggingface.co/spaces/mikeee/mpt-30b-chat?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate"></a> and spin a CPU UPGRADE to avoid the queue</center>"""
        # )
        gr.Markdown(
            f"""<h5><center>{Path(model_loc).name}</center></h4>
            The bot can conduct multi-turn conversations, i.e. it remembers past dialogs. The process time is longer.
            It typically takes about 120 seconds for the first response to appear.

            Most examples are meant for another model.
            You probably should try to test
            some related prompts.""",
            elem_classes="xsmall",
        )

    chatbot = gr.Chatbot(height=500)

    with gr.Row():
        with gr.Column(scale=5):
            msg = gr.Textbox(
                label="Chat Message Box",
                placeholder="Ask me anything (press Shift+Enter or click Submit to send)",
                show_label=False,
                # container=False,
                lines=6,
                max_lines=30,
                show_copy_button=True,
                # ).style(container=False)
            )
        with gr.Column(scale=1, min_width=50):
            with gr.Row():
                submit = gr.Button("Submit", elem_classes="xsmall")
                stop = gr.Button("Stop", visible=True)
                clear = gr.Button("Clear History", visible=True)
    with gr.Row(visible=False):
        with gr.Accordion("Advanced Options:", open=False):
            with gr.Row():
                with gr.Column(scale=2):
                    system = gr.Textbox(
                        label="System Prompt",
                        value=prompt_template,
                        show_label=False,
                        container=False,
                        # ).style(container=False)
                    )
                with gr.Column():
                    with gr.Row():
                        change = gr.Button("Change System Prompt")
                        reset = gr.Button("Reset System Prompt")

    with gr.Accordion("Example Inputs", open=True):
        examples = gr.Examples(
            examples=examples_list,
            inputs=[msg],
            examples_per_page=40,
        )

    with gr.Accordion("Disclaimer", open=False):
        _ = Path(model_loc).name
        gr.Markdown(
            f"Disclaimer: {_} can produce factually incorrect output, and should not be relied on to produce "
            "factually accurate information. {_} was trained on various public datasets; while great efforts "
            "have been taken to clean the pretraining data, it is possible that this model could generate lewd, "
            "biased, or otherwise offensive outputs.",
            elem_classes=["disclaimer"],
        )

    msg_submit_event = msg.submit(
        # fn=conversation.user_turn,
        fn=user,
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        queue=True,
        show_progress="full",
        # api_name=None,
    ).then(bot, chatbot, chatbot, queue=True)
    submit_click_event = submit.click(
        # fn=lambda x, y: ("",) + user(x, y)[1:],  # clear msg
        fn=user1,  # clear msg
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        queue=True,
        # queue=False,
        show_progress="full",
        # api_name=None,
    ).then(bot, chatbot, chatbot, queue=True)
    stop.click(
        fn=None,
        inputs=None,
        outputs=None,
        cancels=[msg_submit_event, submit_click_event],
        queue=False,
    )

    # TODO: clear conversation memory as well
    clear.click(lambda: None, None, chatbot, queue=False)

    with gr.Accordion("For Chat/Translation API", open=False, visible=False):
        input_text = gr.Text()
        api_btn = gr.Button("Go", variant="primary")
        out_text = gr.Text()

    if conversation_api is not None:
        api_btn.click(
            predict_api,
            input_text,
            out_text,
            api_name="api",
        )

# concurrency_count=5, max_size=20
# max_size=36, concurrency_count=14
# CPU cpu_count=2 16G, model 7G
# CPU UPGRADE cpu_count=8 32G, model 7G

# does not work
_ = """
# _ = int(psutil.virtual_memory().total / 10**9 // file_size - 1)
# concurrency_count = max(_, 1)
if psutil.cpu_count(logical=False) >= 8:
    # concurrency_count = max(int(32 / file_size) - 1, 1)
else:
    # concurrency_count = max(int(16 / file_size) - 1, 1)
# """

concurrency_count = 1
logger.info(f"{concurrency_count=}")

block.queue(concurrency_count=concurrency_count, max_size=5).launch(debug=True)