LightMem/src/lightmem/memory_toolkits/token_monitor.py

from __future__ import annotations
import functools
import json
from collections import deque
from threading import RLock, Lock
from litellm import token_counter as litellm_token_counter
from litellm.types.utils import SelectTokenizerResponse
from litellm import encoding
import tiktoken 
from tokenizers import Tokenizer
from datetime import datetime
import inspect
from typing import (
    Any, 
    Callable, 
    Dict, 
    List, 
    Optional, 
    Tuple,
)

def get_tokenizer_for_model(model: str) -> SelectTokenizerResponse:
    """Get the tokenizer for a model."""
    try: 
        tokenizer = Tokenizer.from_pretrained(model)
        return SelectTokenizerResponse(
            type="huggingface_tokenizer", 
            tokenizer=tokenizer,
        ) 
    except:
        print(
            f"Cannot load native huggingface tokenizer for {model}, "
            "using tiktoken tokenizer instead."
        )
        # See https://github.com/BerriAI/litellm/blob/main/litellm/litellm_core_utils/token_counter.py#L504.
        try:
            tokenizer = tiktoken.encoding_for_model(model)
        except KeyError:
            print(
                f"Cannot load tiktoken tokenizer for {model}, " 
                "using litellm's default tokenizer instead."
            )
            tokenizer = encoding
        return SelectTokenizerResponse(
            type="openai_tokenizer",
            tokenizer=tokenizer,
        )

class CostState: 
    """Cost state for a specific LLM model."""

    def __init__(
        self, 
        input_tokens: int = 0,
        output_tokens: int = 0,
        total_time: float = 0.0, 
        window_size: int = 100_000,
        total_count: int = 0,
        histories: Optional[List[Dict[str, List[Dict[str, str]] | str | float | int]]] = None,
    ) -> CostState:
        self.input_tokens = input_tokens
        self.output_tokens = output_tokens
        self.total_time = total_time
        self.total_count = total_count
        self.histories = (
            deque(maxlen=window_size) 
            if histories is None else 
            deque(histories, maxlen=window_size)
        )
        self._lock = RLock() 

    @property 
    def total_tokens(self) -> int:
        """Compute the total number of tokens."""
        with self._lock:
            return self.input_tokens + self.output_tokens

    @property
    def average_input_tokens(self) -> float:
        """Compute the average number of input tokens per call."""
        with self._lock:
            return self.input_tokens / max(self.total_count, 1)

    @property
    def average_output_tokens(self) -> float:
        """Compute the average number of output tokens per call."""
        with self._lock:
            return self.output_tokens / max(self.total_count, 1)
        
    @property
    def average_tokens_per_call(self) -> float:
        """Compute the average number of tokens per call."""
        with self._lock:
            return self.total_tokens / max(self.total_count, 1)

    @property
    def average_time_per_call(self) -> float:
        """Compute the average time per call."""
        with self._lock:
            return self.total_time / max(self.total_count, 1)
    
    def to_dict(self) -> Dict[str, Any]:
        """Convert the cost state to a dictionary."""
        with self._lock:
            return {
                "total_count": self.total_count,
                "total_tokens": self.total_tokens,
                "average_input_tokens": self.average_input_tokens,
                "average_output_tokens": self.average_output_tokens,
                "average_tokens_per_call": self.average_tokens_per_call,
                "average_time_per_call": self.average_time_per_call,
                "histories": list(self.histories),
                "total_time": self.total_time,
                "input_tokens": self.input_tokens,
                "output_tokens": self.output_tokens,
                "window_size": self.histories.maxlen
            }

    def update(
        self, 
        input_tokens: int, 
        output_tokens: int,
        total_time: float,
        histories: List[Dict[str, List[Dict[str, str]] | str | float | int]],
    ) -> None:
        """Update the cost state."""    
        with self._lock:
            self.input_tokens += input_tokens
            self.output_tokens += output_tokens
            self.total_time += total_time
            self.total_count += len(histories)
            self.histories.extend(histories)

    def to_json(self) -> str:
        """Convert the cost state to a JSON string."""
        return json.dumps(
            self.to_dict(), 
            indent=4, 
            sort_keys=True,
            ensure_ascii=False, 
        )
    
    @classmethod
    def from_dict(cls, data: Dict[str, Any]) -> CostState:
        """Create a cost state from a dictionary."""
        allowed = [
            "input_tokens", 
            "output_tokens", 
            "total_time", 
            "window_size", 
            "total_count", 
            "histories", 
        ]
        kwargs = {k: data[k] for k in allowed if k in data}
        return cls(**kwargs)
    
    @classmethod
    def from_json(cls, json_str: str) -> CostState:
        """Create a cost state from a JSON string."""
        return cls.from_dict(json.loads(json_str))

class CostStateManager:
    """Global manager for per-model CostState and tokenizers.

    This class cannot be instantiated. Use classmethods only.
    """
    _states: Dict[str, CostState | Dict[str, CostState]] = {}
    _tokenizers: Dict[str, SelectTokenizerResponse] = {}
    _lock: Lock = Lock()

    def __init__(self) -> None:
        raise OSError("`CostStateManager` is designed to manage global cost states.")

    @classmethod
    def register(
        cls,
        model: str,
        state: Optional[CostState | Dict[str, CostState]] = None,
        tokenizer: Optional[SelectTokenizerResponse] = None,
        exist_ok: bool = False,
    ) -> None:
        """Register an existing CostState and optional tokenizer for a model."""
        with cls._lock:
            if model in cls._states and not exist_ok:
                raise ValueError(f"Model {model} already registered. Please pick another name.")
            # In the process of initialization, we can register a single model with a single `CostState`. 
            # However, in the process of runtime, the number of `CostState` may be more than one. 
            cls._states[model] = state or CostState()
            if tokenizer is not None:
                cls._tokenizers[model] = tokenizer
            else:
                cls._tokenizers[model] = get_tokenizer_for_model(model)

    @classmethod
    def get(cls, model: str) -> CostState | Dict[str, CostState]:
        """Get the CostState for a model."""
        with cls._lock:
            if model not in cls._states:
                raise KeyError(f"Model {model} is not registered. Please register it first.")
            return cls._states[model]

    @classmethod
    def update(
        cls,
        model: str,
        input_output_pair: Dict[str, Dict[str, List[Dict[str, str]] | str | float | int]],
        **kwargs
    ) -> None:
        """Update model's cost state by computing tokens via LiteLLM and appending history."""
        if "input" not in input_output_pair or "output" not in input_output_pair:
            raise ValueError("`input_output_pair` must contain 'input' and 'output'.")
        if "elapsed" not in input_output_pair or not isinstance(input_output_pair["elapsed"], (int, float)):
            raise ValueError("'elapsed' must be provided as float seconds.")

        input_dict, output_dict = input_output_pair["input"], input_output_pair["output"]
        if "messages" not in input_dict or "messages" not in output_dict:
            raise ValueError("'input' and 'output' must contain 'messages'.")
        
        has_operation_type = "metadata" in input_dict and "op_type" in input_dict["metadata"]
        with cls._lock:
            cost_state = cls._states.get(model, None)
            if cost_state is None:
                raise KeyError(f"Model {model} is not registered. Please register it first.")
            tokenizer = cls._tokenizers.get(model)
            if has_operation_type:
                op_type = input_dict["metadata"]["op_type"]
                if isinstance(cost_state, CostState):
                    if len(cost_state.to_dict()["histories"]) > 0:
                        raise ValueError(
                            "Previous update operations do not contain an operation type. "
                            "However, the current update operation contains an operation type. "
                            "This is not allowed. Please make sure the `input_output_pair` is consistent "
                            "with the previous update operations."
                        )
                    else:
                        cls._states[model] = {}
                        cost_state = cls._states[model]
                if op_type not in cost_state:
                    cost_state[op_type] = CostState()
                cost_state = cost_state[op_type]
            elif isinstance(cost_state, dict):
                raise ValueError(
                    "Previous update operations contain different operation types "
                    "or the type of update operation has been inferred during registration. "
                    "However, the current update operation doesn't contain an operation type. "
                    "This is not allowed. Please make sure the `input_output_pair` is consistent "
                    "with the previous update operations."
                )

        inp = input_dict["messages"]
        out = output_dict["messages"]
        if not (isinstance(inp, (list, str)) and isinstance(out, (list, str))):
            raise TypeError("'messages' must be list[dict] or str for both input and output.")
        
        if isinstance(inp, list):
            input_tokens = litellm_token_counter(
                model=model, 
                custom_tokenizer=tokenizer, 
                messages=inp,
                **kwargs
            )
        else:
            input_tokens = litellm_token_counter(
                model=model, 
                custom_tokenizer=tokenizer, 
                text=inp,
                **kwargs
            )
        input_dict["input_tokens"] = input_tokens
        # NOTE: when we compute the output tokens, we don't consider 
        # `tools`, `tool_choice`, `use_default_image_token_count`, and `default_token_count`
        # as they are taken into account when we compute the input tokens.
        if isinstance(out, list):
            output_tokens = litellm_token_counter(
                model=model, 
                custom_tokenizer=tokenizer, 
                messages=out,
            )
        else:
            output_tokens = litellm_token_counter(
                model=model, 
                custom_tokenizer=tokenizer, 
                text=out,
            )
        output_dict["output_tokens"] = output_tokens

        # Update the corresponding cost state
        cost_state.update(
            input_tokens=input_tokens,
            output_tokens=output_tokens,
            total_time=input_output_pair["elapsed"],
            histories=[input_output_pair],
        )

    @classmethod
    def reset(cls) -> None:
        """Reset all cost states."""
        with cls._lock:
            cls._states.clear()
            cls._tokenizers.clear()
    
    @classmethod
    def save_to_json_file(cls, filename: str) -> None:
        """Save all cost states to a JSON file."""
        with cls._lock:
            output_dict = {} 
            for model, state in cls._states.items():
                if isinstance(state, CostState):
                    output_dict[model] = state.to_dict()
                else:
                    output_dict[model] = {op: cs.to_dict() for op, cs in state.items()}
            with open(f"{filename}.json", 'w') as f:
                json.dump(
                    output_dict, 
                    f, 
                    indent=4, 
                    ensure_ascii=False, 
                    sort_keys=True, 
                )

def token_monitor(
    extract_model_name: Callable[..., Tuple[str, Dict[str, Any]]],
    extract_input_dict: Callable[..., Dict[str, List[Dict[str, str]] | str | float | int]],
    extract_output_dict: Callable[..., Dict[str, List[Dict[str, str]] | str | float | int]], 
) -> Callable:
    """
    Decorator to monitor token usage and latency for LLM API calls.
    
    This decorator wraps sync or async callables, extracts model name and I/O payloads,
    computes input/output tokens via LiteLLM, measures elapsed time, and appends a
    structured record to the per-model `CostState` managed by `CostStateManager`.
    The target function must complete successfully for an update to be recorded.
    
    Parameters
    ----------
    extract_model_name : Callable[..., Tuple[str, Dict[str, Any]]]
        A callable that returns a tuple ``(model_name, metadata)``.
        - ``model_name``: The model identifier passed to LiteLLM's token counter.
        - ``metadata``: Extra keyword-arguments forwarded to LiteLLM (e.g., ``custom_tokenizer``).

    extract_input_dict : Callable[..., Dict[str, List[Dict[str, str]] | str | float | int]]
        A callable that builds the input dictionary. It must include a ``'messages'`` key
        whose value is either ``list[dict]`` (OpenAI-style chat format) or ``str``.
        A ``'timestamp'`` string will be injected by the decorator.

    extract_output_dict : Callable[..., Dict[str, List[Dict[str, str]] | str | float | int]]
        A callable that builds the output dictionary from the function result. It must include
        a ``'messages'`` key. A ``'timestamp'`` string will be injected by the decorator.
        
    Returns
    -------
    Callable
        A wrapper that preserves the original function's signature and supports both
        synchronous and asynchronous callables.
    
    Notes
    -----
    - Before using the decorator, register the model via
      ``CostStateManager.register(model, state=..., tokenizer=...)``. Otherwise, an update
      will raise ``KeyError``.
    - The record pushed to the cost state has the following schema::
        {
            "input": <input_dict>,
            "output": <output_dict>,
            "elapsed": <float seconds>,
            "function_name": <str>,
            "is_success": <bool>,
        }
    - Token counting is performed by LiteLLM's ``token_counter`` and can leverage a
      ``custom_tokenizer`` provided through the returned ``metadata`` from ``extract_model_name``.
    - For async functions, the wrapper awaits the coroutine and then performs accounting.
    
    Examples
    --------
    Synchronous usage::
        
        @token_monitor(
            extract_model_name=lambda *args, **kwargs: ("gpt-4o-mini", {"custom_tokenizer": None}),
            extract_input_dict=lambda *args, **kwargs: {"messages": kwargs["messages"]},
            extract_output_dict=lambda result: {"messages": result["messages"]},
        )
        def call_llm(messages):
            ...
        
    Asynchronous usage::
        
        @token_monitor(
            extract_model_name=lambda *args, **kwargs: ("claude-3-sonnet", {}),
            extract_input_dict=lambda *args, **kwargs: {"messages": kwargs["messages"]},
            extract_output_dict=lambda result: {"messages": result["messages"]},
        )
        async def a_call_llm(messages):
            ...
    """
    def decorator(func: Callable[..., Any]) -> Callable[..., Any]:
        if inspect.iscoroutinefunction(func):
            @functools.wraps(func)
            async def awrapper(*args, **kwargs):
                model_name, metadata = extract_model_name(*args, **kwargs)
                input_dict = extract_input_dict(*args, **kwargs)
                start_time = datetime.now().astimezone()
                input_dict["timestamp"] = start_time.strftime("%Y-%m-%d %H:%M:%S %z")
                try:
                    result = await func(*args, **kwargs)
                except Exception as e:
                    print(f"Error in {func.__name__}: \n\t{e.__class__.__name__}: {e}")
                finally: 
                    end_time = datetime.now().astimezone()
                    output_dict = extract_output_dict(result if "result" in locals() else None)
                    output_dict["timestamp"] = end_time.strftime("%Y-%m-%d %H:%M:%S %z")
                    CostStateManager.update(
                        model_name,
                        {
                            "input": input_dict,
                            "output": output_dict,
                            "elapsed": (end_time - start_time).total_seconds(),
                            "function_name": func.__name__,
                            "is_success": "result" in locals(),
                        },
                        **metadata,
                    )
                return result
            return awrapper
        else:
            @functools.wraps(func)
            def wrapper(*args: Any, **kwargs: Any) -> Any:
                # Extract the model name and metadata used during the token computation 
                # The extraction function should be provided by the user
                model_name, metadata = extract_model_name(*args, **kwargs)

                # Extract the input dictionary 
                input_dict = extract_input_dict(*args, **kwargs)
                start_time = datetime.now().astimezone() 
                input_dict["timestamp"] = start_time.strftime("%Y-%m-%d %H:%M:%S %z")

                try:
                    # Run the original function
                    result = func(*args, **kwargs)
                except Exception as e:
                    print(f"Error in {func.__name__}: \n\t{e.__class__.__name__}: {e}")
                finally:
                    end_time = datetime.now().astimezone() 
                    # Extract the output dictionary
                    output_dict = extract_output_dict(result if "result" in locals() else None)
                    output_dict["timestamp"] = end_time.strftime("%Y-%m-%d %H:%M:%S %z")
                    # Update the cost state
                    CostStateManager.update(
                        model_name,
                        {
                            "input": input_dict,
                            "output": output_dict,
                            "elapsed": (end_time - start_time).total_seconds(),
                            "function_name": func.__name__,
                            "is_success": "result" in locals(),
                        },
                        **metadata,
                    )
                # When the function is not successful, it will throw an error.
                # This behavior is expected as the memory will not be saved due to the error.
                return result
            return wrapper
        
    return decorator