alpaca_turbo.py

"""
Alpaca Turbo
"""
import json
import logging
import os
import platform
import signal
from time import time

import psutil
from interact import Process as process
from rich import print as eprint
from rich.logging import RichHandler
from rich.progress import track

# from pwn import  process

# pylint: disable=C0103
# pylint: disable=C0114
# pylint: disable=C0116


class AssistantSettings:
    """Settings handler for assistant"""

    def __init__(self, assistant) -> None:
        self.assistant = assistant

    def load_settings(self):
        if os.path.exists("settings.dat"):
            with open("settings.dat", "r") as file:
                settings = json.load(file)
                _  =eprint(settings) if self.assistant.DEBUG else None
                self.assistant.seed = settings["seed"]
                self.assistant.top_k = settings["top_k"]
                self.assistant.top_p = settings["top_p"]
                self.assistant.temp = settings["temp"]
                self.assistant.threads = settings["threads"]
                self.assistant.repeat_penalty = settings["repeat_penalty"]
                self.assistant.repeat_last_n = settings["repeat_last_n"]
                self.assistant.model_path = settings["model_path"]

    def save_settings(self):
        settings = {
            "seed": self.assistant.seed,
            "top_k": self.assistant.top_k,
            "top_p": self.assistant.top_p,
            "temp": self.assistant.temp,
            "threads": self.assistant.threads,
            "repeat_penalty": self.assistant.repeat_penalty,
            "repeat_last_n": self.assistant.repeat_last_n,
            "model_path": self.assistant.model_path,
        }
        with open("settings.dat", "w") as file:
            json.dump(settings, file)


class Assistant:
    """Alpaca Assistant"""

    model_path = "~/dalai/alpaca/models/7B/ggml-model-q4_0.bin"

    def __init__(self, auto_load=True, DEBUG=False) -> None:
        self.DEBUG = DEBUG
        self.seed = 888777
        self.threads = 4
        self.n_predict = 200
        self.top_k = 40
        self.top_p = 0.9
        self.temp = 0.5
        self.repeat_last_n = 64
        self.repeat_penalty = 1.3

        if platform.system() == "Windows":
            Assistant.model_path = os.path.expanduser(Assistant.model_path).replace("/", "\\")

        self.model_path = os.path.expanduser(Assistant.model_path)

        self.persona = "chat transcript between human and a bot named devil and the bot remembers everything from previous response"

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

        self.format = (
            """\n### Instruction:\n{instruction}\n\n### Response:\n{response}"""
        )
        self.enable_history = True
        self.is_ready = False

        self.settings = AssistantSettings(self)

        self.end_marker = b"[end of text]"

        self.chat_history = []

    def reload(self):
        try:
            self.program.kill(signal.SIGTERM)
        except:
            pass
        self.is_ready = False
        self.prep_model()

    @staticmethod
    def get_bin_path():
        if os.path.exists("bin/local"):
            return "bin/local"
        system_name = platform.system()
        if system_name == "Linux":
            name = "linux"
        elif system_name == "Windows":
            name = "win.exe"
        elif system_name == "Darwin":
            name = "mac"
        # elif system_name == "Android":
        #     return "Android"
        else:
            exit()

        return os.path.join("bin", name)

    @property
    def command(self):
        command = [
            Assistant.get_bin_path(),
            # "--color",
            # "-i",
            "--seed",
            f"{self.seed}",
            "-t",
            f"{self.threads}",
            "--top_k",
            f"{self.top_k}",
            "--top_p",
            f"{self.top_p}",
            "--repeat_last_n",
            f"{self.repeat_last_n}",
            "--repeat_penalty",
            f"{self.repeat_penalty}",
            "-m",
            f"{self.model_path}",
            "--interactive-start",
        ]
        return command

    def prep_bot_input(self):
        """
        prep_bot_input
        """
        prompt = self.persona + "\n" + self.prompt
        history = self.chat_history if self.enable_history else [self.chat_history[-1]]
        for instr, resp in history:
            prompt += self.format.format(instruction=instr, response=resp)
        prompt = prompt.strip("\n")
        prompt = prompt.replace("\n", "\\\n")
        # print("======")
        # print(prompt)
        # print("======")
        return prompt

    def prep_model(self):
        if self.is_ready:
            return None
        _ = (
            ""
            if os.path.exists(self.model_path)
            else print("Set the model path in settings")
        )
        if not os.path.exists(self.model_path):
            return

        if os.path.exists("./pid"):
            try:
                with open("./pid") as file:
                    pid = int(file.readline())
                    if psutil.pid_exists(pid):
                        os.kill(pid, signal.SIGTERM)
                    # os.remove("./pid")
            except (ProcessLookupError, FileNotFoundError):
                pass
        tstart = time()
        cmd = self.command
        _ = eprint(cmd) if self.DEBUG else None
        self.program = process(cmd, timeout=600)
        self.program.readline()
        self.program.recvuntil(b".")

        model_done = False
        for _ in track(range(40), "Loading Model"):
            data = self.program.recv(1).decode("utf-8") if not model_done else None
            model_done = True if data == "d" else model_done
            if model_done:
                continue
        self.program.recvuntil("\n")
        self.is_ready = True
        tend = time()
        eprint(f"Model Loaded in {(tend-tstart)} s")

    def streamer(
        self,
        stuff_to_type,
        pre_recv_hook=None,
        post_recv_hook=None,
    ):
        _ = self.prep_model() if not self.is_ready else None

        self.program.recvuntil(">")

        opts = stuff_to_type.split("\n")
        for opt in opts:
            self.program.sendline(opt)

        while True:
            # _ = pre_recv_hook(self.program) if pre_recv_hook is not None else None
            yield self.program.recv(1)
            # _ = post_recv_hook(self.program) if pre_recv_hook is not None else None

    def ask_bot(self, question, answer=""):
        self.chat_history.append((question, answer))
        inp_to_send = self.prep_bot_input()

        opt_stream = self.streamer(inp_to_send)
        tstart = time()

        buffer = b""

        try:
            isfirst = True
            marker_detected = b""
            char_old = b""
            for char in opt_stream:
                buffer += char  # update the buffer

                if isfirst:
                    t_firstchar = time()
                    wcount = len(question.replace("\n", " ").split(" "))
                    if self.DEBUG:
                        eprint(f"Size of Input: {len(question)} chars || {wcount} words")
                        eprint(
                            f"Time taken to analyze the user input {t_firstchar-tstart} s"
                        )
                    isfirst = False
                else:
                    # Detect end of text if detected try to confirm else reset
                    if char == b"[" or marker_detected:
                        marker_detected += char
                        if marker_detected in self.end_marker[: len(marker_detected)]:
                            continue
                        marker_detected = b""

                    if self.end_marker in buffer:
                        buffer = buffer.replace(b"[end of text]", b"")
                        tend = time()
                        wcount = len(buffer.replace(b"\n", b" ").split(b" "))
                        if self.DEBUG == True:
                            eprint(f"Size of output: {len(buffer)} chars || {wcount} words") 
                            eprint(f"Time taken to for generation {(tend-tstart)} s")
                        break
                try:
                    # Load the full buffer
                    char = char_old + char

                    # print single printable chars
                    if len(char) == 1 and char[0] <= 0x7E and char[0] >= 0x21:
                        char = char.decode("utf-8")
                        char_old = b""
                    elif len(char) in [4, 6]:  # If 4 byte code or 6 byte code
                        char = char.decode("utf-8")
                        char_old = b""
                    else:
                        char_old = char
                        continue
                except UnicodeDecodeError:
                    char_old = char
                    continue
                # print(char, end="")
                yield char

        except (KeyboardInterrupt, EOFError):
            print("Stooping")

        self.chat_history[-1] = (question, buffer.decode("utf-8").strip("\n"))
        return buffer

    def ask_bot_old(self, question, answer=""):
        """
        run
        """
        tend = 0
        _ = self.prep_model() if not self.is_ready else None
        tstart = time()

        self.program.recvuntil(">")

        self.chat_history.append((question, answer))

        opts = self.prep_bot_input.split("\n")
        for opt in opts:
            self.program.sendline(opt)

        data = None

        try:
            marker_detected = b""
            char = self.program.recv(1)
            tfirstchar = time()
            wcount = len(question.replace("\n", " ").split(" "))
            if self.DEBUG:
                eprint(f"Size of Input: {len(question)} chars || {wcount} words")
                eprint(f"Time taken to analyze the user input {(tfirstchar-tstart)} s")
            data = char
            yield char.decode("utf-8")
            while True:
                char = self.program.recv(1)

                data += char

                if char == b"[" or marker_detected:
                    marker_detected += char
                    if marker_detected in self.end_marker:
                        continue
                    marker_detected = b""

                if self.end_marker in data:
                    data = data.replace(b"[end of text]", b"")
                    tend = time()
                    wcount = len(data.replace(b"\n", b" ").split(b" "))
                    if self.DEBUG:
                        eprint(f"Size of output: {len(data)} chars || {wcount} words")
                        eprint(f"Time taken to for generation {(tend-tstart)} s")
                    break

                yield char.decode("utf-8")
        except (KeyboardInterrupt, EOFError):
            print("Stooping")

        self.chat_history[-1] = (question, data.decode("utf-8").strip("\n"))

        return data

    @staticmethod
    def repl(debug=False):
        assistant = Assistant(DEBUG=debug)
        assistant.prep_model()
        while True:
            _  = eprint(assistant.chat_history) if debug else None
            resp = assistant.ask_bot(input(">>> "), "")

            for char in resp:
                print(char, end="")
            print()


def health_checks():
    FORMAT = "%(message)s"
    logging.basicConfig(
        level="NOTSET", format=FORMAT, datefmt="[%X]", handlers=[RichHandler()]
    )
    log = logging.getLogger("rich")

    log.info("Running health checks ")
    test_assistant = Assistant(auto_load=False)

    # check if binary is available for system
    log.info("Checking for dependencies")
    if os.path.exists(Assistant.get_bin_path()):
        log.info("Found binary")
    else:
        log.fatal("Binary Not Found")
        log.info("Check https://github.com/ViperX7/alpaca.cpp")
        log.info("put the compiled file in (./bin/main or ./bin/main.exe )")
        exit()

    log.info("checking if system is supported")
    try:
        prog = process(Assistant.get_bin_path())
        log.info("Supported system")
    except OSError:
        log.fatal("Binary Not supported on this system")
        log.info("Check https://github.com/ViperX7/alpaca.cpp")
        log.info("put the compiled file in (./bin/main or ./bin/main.exe )")
        exit()

    log.info("Looking for the models to load")
    if os.path.exists(os.path.expanduser(Assistant.model_path)):
        log.info(f"Found Model {Assistant.model_path}")
        sz = os.path.getsize(test_assistant.model_path) // (1024 * 1024)
        log.info(f"size of your model is {sz} MB (approx)")
    else:
        log.fatal(
            "model not found you need to download the models and set the model path in settings"
        )

    if os.path.exists("./pid"):
        log.info("Other checks")
        log.fatal("Already running another instance or dirty exit last time")
        with open("./pid") as file:
            pid = int(file.readline())
        log.info("Attempting to kill the process")
        os.kill(pid, signal.SIGTERM)
        os.remove("./pid")
        log.info("Fixed the Issue Now Retry running")
        exit()

    memstat = psutil.virtual_memory()
    log.info("checking memory")
    log.info(f"Total memory {memstat.total//(1024*1024)} MB")
    log.info(f"Used memory {memstat.used//(1024*1024)} MB")
    log.info(f"Free memory {memstat.free//(1024*1024)} MB")
    exit()
    log.level = 5


# health_checks()

def main():
    import sys
    debug = "-d" in sys.argv
    Assistant.repl(debug)




assistant =  main() if __name__ == "__main__" else None