models.py

from __future__ import annotations

import json
import random
import re
from abc import ABC, abstractmethod
from typing import List, Dict, Union, Optional

from huggingface_hub import InferenceClient
from tenacity import retry, stop_after_attempt, wait_random_exponential
from transformers import AutoTokenizer

from config import *

ROLE_SYSTEM = 'system'
ROLE_USER = 'user'
ROLE_ASSISTANT = 'assistant'

SUPPORTED_MISTRAL_MODELS = ['mistralai/Mixtral-8x7B-Instruct-v0.1', 'mistralai/Mistral-7B-Instruct-v0.2']
SUPPORTED_NOUS_MODELS = ['NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO']
SUPPORTED_LLAMA_MODELS = ['meta-llama/Llama-2-70b-chat-hf',
                          'meta-llama/Llama-2-13b-chat-hf',
                          'meta-llama/Llama-2-7b-chat-hf']
ALL_SUPPORTED_MODELS = SUPPORTED_MISTRAL_MODELS + SUPPORTED_NOUS_MODELS + SUPPORTED_LLAMA_MODELS    


def select_model(model_name: str, system_prompt: str, **kwargs) -> Model:
    if model_name in SUPPORTED_MISTRAL_MODELS:
        return MistralModel(system_prompt, model_name)
    elif model_name in SUPPORTED_NOUS_MODELS:
        return NousHermesModel(system_prompt, model_name)
    elif model_name in SUPPORTED_LLAMA_MODELS:
        return LlamaModel(system_prompt, model_name)
    else:
        raise ValueError(f'Model {model_name} not supported')


class Model(ABC):
    name: str
    messages: List[Dict[str, str]]
    system_prompt: str

    def __init__(self, model_name: str, system_prompt: str):
        self.name = model_name
        self.system_prompt = system_prompt
        self.messages = [
            {'role': ROLE_SYSTEM, 'content': system_prompt}
        ]

    @abstractmethod
    def __call__(self, *args, **kwargs) -> Union[str, Dict]:
        raise NotImplementedError

    def add_message(self, role: str, content: str):
        assert role in [ROLE_SYSTEM, ROLE_USER, ROLE_ASSISTANT]
        self.messages.append({'role': role, 'content': content})

    def clear_conversations(self):
        self.messages.clear()
        self.add_message(ROLE_SYSTEM, self.system_prompt)

    def __str__(self) -> str:
        return self.name

    def __repr__(self) -> str:
        return self.name


class HFAPIModel(Model):

    def __call__(self, user_prompt: str, *args,
                 use_json: bool = False,
                 temperature: float = 0,
                 timeout: float = None,
                 cache: bool = False,
                 json_retry_count: int = 5,
                 **kwargs) -> Union[str, Dict]:
        """
        Returns the model's response.
        If use_json = True, will try its best to return a json dict, but not guaranteed.
        If we cannot parse the JSON, we will return the response string directly.
        """
        self.add_message(ROLE_USER, user_prompt)
        response = self.get_response(temperature, use_json, timeout, cache)
        if use_json:
            for i in range(json_retry_count):
                # cache only if both instruct to do and first try
                response = self.get_response(temperature, use_json, timeout, cache and i == 0)
                json_obj = self.find_first_valid_json(response)
                if json_obj is not None:
                    response = json_obj
                    break
        self.add_message(ROLE_ASSISTANT, response)
        return response

    @retry(stop=stop_after_attempt(5), wait=wait_random_exponential(max=10), reraise=True)  # retry if exception
    def get_response(self, temperature: float, use_json: bool, timeout: float, cache: bool) -> str:
        client = InferenceClient(model=self.name, timeout=timeout)
        # client = InferenceClient(model=self.name, token=random.choice(HF_API_TOKENS), timeout=timeout)
        if not cache:
            client.headers["x-use-cache"] = "0"
        # print(self.formatter(self.messages))  # debug
        r = client.text_generation(self.format_messages(),
                                   do_sample=temperature > 0,
                                   temperature=temperature if temperature > 0 else None,
                                   max_new_tokens=4096)
        return r

    @abstractmethod
    def format_messages(self) -> str:
        raise NotImplementedError

    def get_short_name(self) -> str:
        """
        Returns the last part of the model name.
        For example, "mistralai/Mixtral-8x7B-Instruct-v0.1" -> "Mixtral-8x7B-Instruct-v0.1"
        """
        return self.name.split('/')[-1]

    @staticmethod
    def find_first_valid_json(s) -> Optional[Dict]:
        s = re.sub(r'\\(?!["\\/bfnrt]|u[0-9a-fA-F]{4})', lambda m: m.group(0)[1:], s)  # remove all invalid escapes chars
        for i in range(len(s)):
            if s[i] != '{':
                continue
            for j in range(i + 1, len(s) + 1):
                if s[j - 1] != '}':
                    continue
                try:
                    potential_json = s[i:j]
                    json_obj = json.loads(potential_json, strict=False)
                    return json_obj  # Return the first valid JSON object found
                except json.JSONDecodeError:
                    pass  # Continue searching if JSON decoding fails
        return None  # Return None if no valid JSON object is found


class MistralModel(HFAPIModel):

    def __init__(self, system_prompt: str, model_name: str = 'mistralai/Mixtral-8x7B-Instruct-v0.1') -> None:
        assert model_name in ['mistralai/Mixtral-8x7B-Instruct-v0.1',
                              'mistralai/Mistral-7B-Instruct-v0.2'], 'Model not supported'
        super().__init__(model_name, system_prompt)

    def format_messages(self) -> str:
        messages = self.messages
        # mistral doesn't support system prompt, so we need to convert it to user prompt
        if messages[0]['role'] == ROLE_SYSTEM:
            assert len(self.messages) >= 2
            messages = [{'role'   : ROLE_USER,
                         'content': messages[0]['content'] + '\n' + messages[1]['content']}] + messages[2:]
        tokenizer = AutoTokenizer.from_pretrained(self.name)
        r = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, max_length=4096)
        # print(r)
        return r


class NousHermesModel(HFAPIModel):

    def __init__(self, system_prompt: str, model_name: str = 'NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO') -> None:
        assert model_name in ['NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO'], 'Model not supported'
        super().__init__(model_name, system_prompt)

    def format_messages(self) -> str:
        messages = self.messages
        assert len(messages) >= 2  # must be at least a system and a user
        assert messages[0]['role'] == ROLE_SYSTEM and messages[1]['role'] == ROLE_USER
        tokenizer = AutoTokenizer.from_pretrained(self.name)
        r = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, max_length=4096)
        # print(r)
        return r


class LlamaModel(HFAPIModel):

    def __init__(self, system_prompt: str, model_name: str = 'meta-llama/Llama-2-70b-chat-hf') -> None:
        assert model_name in ['meta-llama/Llama-2-70b-chat-hf',
                              'meta-llama/Llama-2-13b-chat-hf',
                              'meta-llama/Llama-2-7b-chat-hf'], 'Model not supported'
        super().__init__(model_name, system_prompt)

    def format_messages(self) -> str:
        """
        <s>[INST] <<SYS>>
        {system_prompt}
        <</SYS>>

        {user_message} [/INST]
        """
        messages = self.messages
        assert len(messages) >= 2  # must be at least a system and a user
        r = f'<s>[INST] <<SYS>>\n{messages[0]["content"]}\n<</SYS>>\n\n{messages[1]["content"]} [/INST]'
        for msg in messages[2:]:
            role, content = msg['role'], msg['content']
            if role == ROLE_SYSTEM:
                assert ValueError
            elif role == ROLE_USER:
                if r.endswith('</s>'):
                    r += '<s>'
                r += f'[INST] {content} [/INST]'
            elif role == ROLE_ASSISTANT:
                r += f'{content}</s>'
            else:
                raise ValueError
        return r