eval/modules/model_manager.py

import json
import time
import logging
from typing import Optional, Dict, Any, Callable, Tuple
from transformers import AutoTokenizer
from openai import OpenAI
import requests
from functools import wraps
import os

# configure logging
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)

def model_query_decorator(func: Callable) -> Callable:
    """
    model query decorator, handle common pre-processing and retry logic
    get max_tries and time_sleep from self instance dynamically
    """
    @wraps(func)
    def wrapper(self, prompt: str) -> str:
        # pre-processing: check empty prompt
        if not prompt or not prompt.strip():
            logger.warning("Empty prompt, return empty string")
            return "", None
        
        # pre-processing: truncate prompt
        processed_prompt = self._truncate_prompt(prompt.strip())
        
        # get retry parameters from self instance
        max_tries = getattr(self, 'max_tries', 5)
        time_sleep = getattr(self, 'time_sleep', 1.0)

        time.sleep(time_sleep)
        
        # retry mechanism
        last_exception = None
        
        for attempt in range(1, max_tries + 1):
            try:
                logger.info(f"Try {attempt} times...")
                
                # call the specific query method
                answer, thinking = func(self, processed_prompt)
                
                logger.info(f"Query successful, try {attempt} times")
                
                return answer, thinking
                
            except KeyboardInterrupt:
                logger.info("User interrupt")
                raise
            except Exception as e:
                last_exception = e
                logger.warning(f"API error (try {attempt}/{max_tries}): {e}")
            
            # exponential backoff strategy
            if attempt < max_tries:
                sleep_time = time_sleep * (2 ** (attempt - 1))
                logger.info(f"Wait {sleep_time:.1f} seconds and retry...")
                time.sleep(sleep_time)
        
        # all tries failed
        logger.error(f"All {max_tries} tries failed, last error: {last_exception}")
        return "", None
    
    return wrapper


class ModelManagerBase:
    """Base model manager"""

    def __init__(
        self,
        tokenizer_path: str, 
        context_max_length: int,
        url: str, 
        api_key: str,
        temperature: float,
        max_new_tokens: int,
        timeout: int,
        max_tries: int,
        time_sleep: float,
    ):
        # parameter validation
        if not os.path.exists(tokenizer_path):
            raise ValueError("tokenizer_path is not found")
        if context_max_length <= 0:
            raise ValueError("context_max_length must be greater than 0")
        if max_tries <= 0:
            raise ValueError("max_tries must be greater than 0")
        
        self.tokenizer_path = tokenizer_path
        self.context_max_length = context_max_length
        self.url = url
        self.api_key = api_key
        self.temperature = temperature
        self.max_new_tokens = max_new_tokens
        self.timeout = timeout
        self.max_tries = max_tries
        self.time_sleep = time_sleep
        self.tokenizer = self._get_tokenizer()
    
    def _get_tokenizer(self) -> AutoTokenizer:
        """Get tokenizer"""
        try:
            return AutoTokenizer.from_pretrained(
                self.tokenizer_path, 
                trust_remote_code=True
            )
        except Exception as e:
            logger.error(f"Failed to load tokenizer: {e}")
            raise
    
    def _truncate_prompt(self, prompt: str) -> str:
        """Truncate prompt, keep important parts"""
        input_ids = self.tokenizer.encode(prompt)
        
        if len(input_ids) <= self.context_max_length:
            return prompt
        
        truncated_input_ids = input_ids[:self.context_max_length//2] + input_ids[-self.context_max_length//2:]

        truncated_prompt = self.tokenizer.decode(
            truncated_input_ids, 
            skip_special_tokens=True
        )

        return truncated_prompt
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> str:
        """Query LLM model"""
        raise NotImplementedError("Subclass must implement this method")


class ModelManagerMagistral(ModelManagerBase):
    """Magistral model manager"""

    def __init__(
        self,
        model_name: str,
        tokenizer_path: str = "",
        context_max_length: int = 120000, # 128k - 8k
        url: str = "http://127.0.0.1:8000/v1", 
        api_key: str = "EMPTY",
        temperature: float = 0.7,
        max_new_tokens: int = 8192,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
        extra_body: Optional[Dict[str, Any]] = None,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.model_name = model_name
        self.extra_body = extra_body or {}
        self.client = self._create_client()
        self.system_prompt = """First draft your thinking process (inner monologue) until you arrive at a response. Format your response using Markdown, and use LaTeX for any mathematical equations. Write both your thoughts and the response in the same language as the input.\n\nYour thinking process must follow the template below:[THINK]Your thoughts or/and draft, like working through an exercise on scratch paper. Be as casual and as long as you want until you are confident to generate the response. Use the same language as the input.[/THINK]Here, provide a self-contained response."""

    def _create_client(self) -> OpenAI:
        """Create Magistral client"""
        try:
            return OpenAI(
                base_url=self.url,
                api_key=self.api_key
            )
        except Exception as e:
            logger.error(f"Failed to create OpenAI client: {e}")
            raise
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle OpenAI specific logic"""
        completion = self.client.chat.completions.create(
            model=self.model_name,
            messages=[
                {"role": "system", "content": self.system_prompt},
                {"role": "user", "content": processed_prompt}
            ],
            temperature=self.temperature,
            extra_body=self.extra_body,
            max_tokens=self.max_new_tokens,
            timeout=self.timeout,
        )
        answer = completion.choices[0].message.content
        try:
            thinking = completion.choices[0].message.reasoning_content
        except:
            thinking = None
        return answer, thinking

class ModelManagerOpenAI(ModelManagerBase):
    """OpenAI model manager"""

    def __init__(
        self,
        model_name: str,
        tokenizer_path: str = "model/Tokenizers/qwen",
        context_max_length: int = 120000, # 128k - 8k
        url: str = "http://127.0.0.1:8000/v1", 
        api_key: str = "EMPTY",
        temperature: float = 1.0,
        max_new_tokens: int = 8192,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
        extra_body: Optional[Dict[str, Any]] = None,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.model_name = model_name
        self.extra_body = extra_body or {}
        self.client = self._create_client()
    
    def _create_client(self) -> OpenAI:
        """Create OpenAI client"""
        try:
            return OpenAI(
                base_url=self.url,
                api_key=self.api_key
            )
        except Exception as e:
            logger.error(f"Failed to create OpenAI client: {e}")
            raise
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle OpenAI specific logic"""
        completion = self.client.chat.completions.create(
            model=self.model_name,
            messages=[{"role": "user", "content": processed_prompt}],
            temperature=self.temperature,
            extra_body=self.extra_body,
            max_tokens=self.max_new_tokens,
            timeout=self.timeout,
        )
        answer = completion.choices[0].message.content
        try:
            thinking = completion.choices[0].message.reasoning_content
        except:
            thinking = None
        return answer, thinking

class ModelManagerGemini3(ModelManagerBase):
    """Gemini 3.0 model manager"""

    def __init__(
        self,
        tokenizer_path: str = "model/Tokenizers/gemini",
        context_max_length: int = 1000000, # 1M
        url: str = "https://runway.devops.rednote.life/openai/google/v1:generateContent", 
        api_key: str = "162420e2621c480d9f8ab1bb7b8c4c91",
        temperature: float = 1.0, # default 1.0
        max_new_tokens: int = 32768,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'api-key': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle Gemini specific logic"""
        # prepare payload
        payload = json.dumps({
            "contents": [
                {
                    "role": "user",
                    "parts": [
                        {
                            "text": processed_prompt
                        }
                    ]
                }
            ],
            "generationConfig": {
                "maxOutputTokens": self.max_new_tokens,
                "temperature": self.temperature,
                "thinkingConfig": {
                    "thinking_level": "high" # default "high"
                }
            }
        })
        
        # send request
        response = requests.request("POST", self.url, headers=self.headers, data=payload).json()
        
        # extract result
        return response["candidates"][0]["content"]["parts"][0]["text"], None


class ModelManagerGemini25(ModelManagerBase):
    """Gemini 2.5 thinking model manager"""

    def __init__(
        self,
        tokenizer_path: str = "model/Tokenizers/gemini",
        context_max_length: int = 1000000, # 1M
        url: str = "https://runway.devops.rednote.life/openai/google/v1:generateContent", 
        api_key: str = "66c251052f44452a834ce83d0c7fd3ba",
        temperature: float = 1.0, # default 1.0
        max_new_tokens: int = 32768,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'api-key': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle Gemini specific logic"""
        # prepare payload
        payload = json.dumps({
            "contents": [
                {
                    "role": "user",
                    "parts": [
                        {
                            "text": processed_prompt
                        }
                    ]
                }
            ],
            "generationConfig": {
                "maxOutputTokens": self.max_new_tokens,
                "temperature": self.temperature,
                "thinkingConfig": {
                    "thinkingBudget": -1 # enable auto thinking
                }
            }
        })
        
        # send request
        response = requests.request("POST", self.url, headers=self.headers, data=payload).json()
        
        # extract result
        return response["candidates"][0]["content"]["parts"][0]["text"], None


class ModelManagerGemini25FlashNonthinking(ModelManagerBase):
    """Gemini 2.5 Flash nonthinking model manager"""

    def __init__(
        self,
        tokenizer_path: str = "model/Tokenizers/gemini",
        context_max_length: int = 1000000, # 1M
        url: str = "https://runway.devops.rednote.life/openai/google/v1:generateContent", 
        api_key: str = "66c251052f44452a834ce83d0c7fd3ba",
        temperature: float = 1.0, # default 1.0
        max_new_tokens: int = 1024,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'api-key': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle Gemini specific logic"""
        # prepare payload
        payload = json.dumps({
            "contents": [
                {
                    "role": "user",
                    "parts": [
                        {
                            "text": processed_prompt
                        }
                    ]
                }
            ],
            "generationConfig": {
                "maxOutputTokens": self.max_new_tokens,
                "temperature": self.temperature,
                "thinkingConfig": {
                    "thinkingBudget": 0 # disable thinking
                }
            }
        })
        
        # send request
        response = requests.request("POST", self.url, headers=self.headers, data=payload).json()
        
        # extract result
        return response["candidates"][0]["content"]["parts"][0]["text"], None


class ModelManagerGPT5(ModelManagerBase):
    """GPT5 model manager"""

    def __init__(
        self,
        tokenizer_path: str = "model/Tokenizers/gpt",
        context_max_length: int = 262144, # 256k
        url: str = "https://runway.devops.rednote.life/openai/chat/completions?api-version=2025-01-01-preview", 
        api_key: str = "9a7403aa383e4a44a8c0f852710630e0",
        temperature: float = 1.0, # only support 1.0
        max_new_tokens: int = 32768,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'api-key': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle GPT5 specific logic"""
        payload = json.dumps({
            "messages":[
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "text",
                            "text": processed_prompt
                        }
                    ]
                }
            ],
            "max_completion_tokens": self.max_new_tokens,
            "temperature": self.temperature,
        })

        # send request
        response = requests.request("POST", self.url, headers=self.headers, data=payload).json()

        # extract answer
        return response["choices"][0]["message"]["content"], None

class ModelManagerGPT4o(ModelManagerBase):
    """GPT4o model manager"""

    def __init__(
        self,
        tokenizer_path: str = "model/Tokenizers/gpt",
        context_max_length: int = 120000, # 128k - 8k
        url: str = "https://runway.devops.rednote.life/openai/chat/completions?api-version=2025-01-01-preview", 
        api_key: str = "9d876c24a1d74e218e69339258db13a3",
        temperature: float = 1.0,
        max_new_tokens: int = 7168,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'api-key': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle GPT4o specific logic"""
        payload = json.dumps({
            "messages":[
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "text",
                            "text": processed_prompt
                        }
                    ]
                }
            ],
            "max_tokens": self.max_new_tokens,
            "temperature": self.temperature,
        })

        # send request
        response = requests.request("POST", self.url, headers=self.headers, data=payload).json()

        # extract answer
        return response["choices"][0]["message"]["content"], None


class ModelManagerClaude4(ModelManagerBase):
    """Claude 4 thinking model manager"""

    def __init__(
        self,
        tokenizer_path: str = "model/Tokenizers/claude",
        context_max_length: int = 1000000, # 1M
        url: str = "https://runway.devops.rednote.life/openai/bedrock_runtime/model/invoke", 
        api_key: str = "899efa27c7c74654bb561242e1a0e423",
        temperature: float = 1.0, # only support 1.0
        max_new_tokens: int = 32768,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'token': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle Claude4 specific logic"""

        payload = json.dumps({
            "anthropic_version": "bedrock-2023-05-31",
            "max_tokens": self.max_new_tokens,
            "temperature": self.temperature,
            "anthropic_beta": ["context-1m-2025-08-07"], # support 1M context
            "thinking": {
                "type": "enabled",
                "budget_tokens": self.max_new_tokens - 1024
            },
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "text",
                            "text": processed_prompt
                        }
                    ]
                }
            ]
        })

        # send request
        response = requests.post(self.url, headers=self.headers, data=payload).json()

        # extract answer
        return response["content"][1]["text"], response["content"][0]["thinking"]

class ModelManagerClaude4Nonthinking(ModelManagerBase):
    """Claude 4 nonthinking model manager"""

    def __init__(
        self,
        tokenizer_path: str = "model/Tokenizers/claude",
        context_max_length: int = 1000000, # 1M
        url: str = "https://runway.devops.rednote.life/openai/bedrock_runtime/model/invoke", 
        api_key: str = "899efa27c7c74654bb561242e1a0e423",
        temperature: float = 1.0, # only support 1.0
        max_new_tokens: int = 1024,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'token': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle Claude4 specific logic"""

        payload = json.dumps({
            "anthropic_version": "bedrock-2023-05-31",
            "max_tokens": self.max_new_tokens,
            "temperature": self.temperature,
            "anthropic_beta": ["context-1m-2025-08-07"], # support 1M context
            "thinking": {
                "type": "disabled",
            },
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "text",
                            "text": processed_prompt
                        }
                    ]
                }
            ]
        })

        # send request
        response = requests.post(self.url, headers=self.headers, data=payload).json()
        
        # extract answer
        return response["content"][0]["text"], None


class ModelManagerClaude37(ModelManagerBase):
    """Claude 3.7 thinking model manager"""

    def __init__(
        self,
        tokenizer_path: str = "model/Tokenizers/claude",
        context_max_length: int = 200000,
        url: str = "https://runway.devops.rednote.life/openai/bedrock_runtime/model/invoke", 
        api_key: str = "ff15724dce4d4c1e95939efd2f40628f",
        temperature: float = 1.0, # only support 1.0
        max_new_tokens: int = 32768,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'token': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle Claude4 specific logic"""

        payload = json.dumps({
            "anthropic_version": "bedrock-2023-05-31",
            "max_tokens": self.max_new_tokens,
            "temperature": self.temperature,
            "thinking": {
                "type": "enabled",
                "budget_tokens": self.max_new_tokens - 200
            },
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "text",
                            "text": processed_prompt
                        }
                    ]
                }
            ]
        })

        # send request
        response = requests.post(self.url, headers=self.headers, data=payload).json()

        # extract answer
        return response["content"][1]["text"], response["content"][0]["thinking"]

class ModelManagerClaude37Nonthinking(ModelManagerBase):
    """Claude 3.7 nonthinking model manager"""

    def __init__(
        self,
        tokenizer_path: str = "model/Tokenizers/claude",
        context_max_length: int = 200000,
        url: str = "https://runway.devops.rednote.life/openai/bedrock_runtime/model/invoke", 
        api_key: str = "ff15724dce4d4c1e95939efd2f40628f",
        temperature: float = 1.0, # only support 1.0
        max_new_tokens: int = 1024,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'token': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle Claude4 specific logic"""

        payload = json.dumps({
            "anthropic_version": "bedrock-2023-05-31",
            "max_tokens": self.max_new_tokens,
            "temperature": self.temperature,
            "thinking": {
                "type": "disabled",
            },
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "text",
                            "text": processed_prompt
                        }
                    ]
                }
            ]
        })

        # send request
        response = requests.post(self.url, headers=self.headers, data=payload).json()

        # extract answer
        return response["content"][0]["text"], None

class ModelManagerKimi(ModelManagerBase):
    """Kimi model manager"""

    def __init__(
        self,
        tokenizer_path: str = "/cpfs/user/chengfeng/huggingface/models/moonshotai/Kimi-K2-Instruct",
        context_max_length: int = 224000, # 256k - 32k
        url: str = "https://runway.devops.xiaohongshu.com/openai/moonshot/v1/chat/completions", 
        api_key: str = "ea70f961e2e94024b0e8a2037ae9b477",
        temperature: float = 0.6,
        max_new_tokens: int = 32768,
        timeout: int = 1200,
        max_tries: int = 5,
        time_sleep: float = 1.0,
    ):
        super().__init__(tokenizer_path, context_max_length, url, api_key, 
                temperature, max_new_tokens, timeout, max_tries, time_sleep)
        self.headers = {
            'api-key': self.api_key,
            'Content-Type': 'application/json'
        }
    
    @model_query_decorator
    def query(self, processed_prompt: str) -> Tuple[str, Optional[str]]:
        """Query LLM model - only handle Kimi specific logic"""
        payload = json.dumps({
            "model": "kimi-k2-0905-preview",
            "messages":[
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "text",
                            "text": processed_prompt
                        }
                    ]
                }
            ],
            "max_tokens": self.max_new_tokens,
            "temperature": self.temperature,
            "timeout": self.timeout
        })

        # send request
        response = requests.post(self.url, headers=self.headers, data=payload).json()

        # extract answer
        return response["choices"][0]["message"]["content"], None

if __name__ == "__main__":
    model_manager = ModelManagerGemini3()
    answer, thinking = model_manager.query("Hello, how are you?")
    print("Answer:", answer)
    print("Thinking:", thinking)