FLAMLOpenAIChatAtomicFlow.py

import pprint
from copy import deepcopy

import hydra
import logging

import colorama
import time

from typing import List, Dict, Optional, Any, Callable, Tuple

from flaml import tune, BlendSearch

from langchain import PromptTemplate
import langchain
from langchain.schema import HumanMessage, AIMessage, SystemMessage

from flows.history import FlowHistory
from flows.message_annotators.abstract import MessageAnnotator
from flows.base_flows.abstract import AtomicFlow
from flows.datasets import GenericDemonstrationsDataset

from flows import utils
from flows.messages.chat_message import ChatMessage
from flows.utils.caching_utils import flow_run_cache

log = utils.get_pylogger(__name__)
logger = log


class FLAMLOpenAIChatAtomicFlow(AtomicFlow):
    model_name: str
    generation_parameters: Dict

    system_message_prompt_template: PromptTemplate
    human_message_prompt_template: PromptTemplate

    system_name: str = "system"
    user_name: str = "user"
    assistant_name: str = "assistant"

    n_api_retries: int = 6
    wait_time_between_retries: int = 20

    query_message_prompt_template: Optional[PromptTemplate] = None
    demonstrations: GenericDemonstrationsDataset = None
    demonstrations_response_template: PromptTemplate = None
    response_annotators: Optional[Dict[str, MessageAnnotator]] = {}

    default_search_space = {
        # "model": tune.choice(
        #     [
        #         # "text-ada-001",
        #         # "text-babbage-001",
        #         # "text-davinci-003",
        #         "gpt-3.5-turbo",
        #         # "gpt-4",
        #     ]
        # ),
        "temperature_or_top_p": tune.choice(
            [
                {"temperature": tune.uniform(0, 2)},
                {"top_p": tune.uniform(0, 1)},
            ]
        ),
        "max_tokens": tune.lograndint(1000, 4000),
        # we use langchain api, https://github.com/hwchase17/langchain/blob/master/langchain/chat_models/base.py#L201
        # it only take the first generation as the output, thus n is not relevant
        # "n": tune.randint(1, 100), 
    }

    def __init__(self, **kwargs):
        self._validate_parameters(kwargs)
        super().__init__(**kwargs)

        assert self.flow_config["name"] not in [
            "system",
            "user",
            "assistant",
        ], f"Flow name '{self.flow_config['name']}' cannot be 'system', 'user' or 'assistant'"

    def set_up_flow_state(self):
        super().set_up_flow_state()
        self.flow_state["conversation_initialized"] = False

    @classmethod
    def _validate_parameters(cls, kwargs):
        # ToDo: Deal with this in a cleaner way (with less repetition)
        super()._validate_parameters(kwargs)

        # ~~~ Model generation ~~~
        if "model_name" not in kwargs["flow_config"]:
            raise KeyError("model_name not specified in the flow_config.")

        if "generation_parameters" not in kwargs["flow_config"]:
            raise KeyError("generation_parameters not specified in the flow_config.")

        # ~~~ Prompting ~~~
        if "system_message_prompt_template" not in kwargs:
            raise KeyError("system_message_prompt_template not passed to the constructor.")

        if "query_message_prompt_template" not in kwargs:
            raise KeyError("query_message_prompt_template not passed to the constructor.")

        if "human_message_prompt_template" not in kwargs:
            raise KeyError("human_message_prompt_template not passed to the constructor.")

    @classmethod
    def _set_up_prompts(cls, config):
        kwargs = {}

        kwargs["system_message_prompt_template"] = \
            hydra.utils.instantiate(config['system_message_prompt_template'], _convert_="partial")
        kwargs["query_message_prompt_template"] = \
            hydra.utils.instantiate(config['query_message_prompt_template'], _convert_="partial")
        kwargs["human_message_prompt_template"] = \
            hydra.utils.instantiate(config['human_message_prompt_template'], _convert_="partial")

        return kwargs

    @classmethod
    def _set_up_demonstration_templates(cls, config):
        kwargs = {}

        if "demonstrations_response_template" in config:
            kwargs["demonstrations_response_template"] = \
                hydra.utils.instantiate(config['demonstrations_response_template'], _convert_="partial")

        return kwargs

    @classmethod
    def _set_up_response_annotators(cls, config):
        response_annotators = config.get("response_annotators", {})
        if len(response_annotators) > 0:
            for key, config in response_annotators.items():
                if isinstance(config, MessageAnnotator):
                    response_annotators[key] = config
                else:
                    response_annotators[key] = hydra.utils.instantiate(config, _convert_="partial")
        return {"response_annotators": response_annotators}

    @classmethod
    def instantiate_from_config(cls, config):
        flow_config = deepcopy(config)

        kwargs = {"flow_config": flow_config}

        # ~~~ Set up prompts ~~~
        kwargs.update(cls._set_up_prompts(flow_config))

        # ~~~ Set up demonstration templates ~~~
        kwargs.update(cls._set_up_demonstration_templates(flow_config))

        # ~~~ Set up response annotators ~~~
        kwargs.update(cls._set_up_response_annotators(flow_config))

        # ~~~ Instantiate flow ~~~
        return cls(**kwargs)

    def _is_conversation_initialized(self):
        return self.flow_state["conversation_initialized"]

    def expected_inputs_given_state(self):
        if self._is_conversation_initialized():
            return ["query"]
        else:
            return self.flow_config["expected_inputs"]

    @staticmethod
    def _get_message(prompt_template, input_data: Dict[str, Any]):
        template_kwargs = {}
        for input_variable in prompt_template.input_variables:
            template_kwargs[input_variable] = input_data[input_variable]

        msg_content = prompt_template.format(**template_kwargs)
        return msg_content

    def _get_demonstration_query_message_content(self, sample_data: Dict):
        input_variables = self.query_message_prompt_template.input_variables
        return self.query_message_prompt_template.format(**{k: sample_data[k] for k in input_variables}), []

    def _get_demonstration_response_message_content(self, sample_data: Dict):
        input_variables = self.demonstrations_response_template.input_variables
        return self.demonstrations_response_template.format(**{k: sample_data[k] for k in input_variables}), []

    def _get_annotator_with_key(self, key: str):
        for _, ra in self.response_annotators.items():
            if ra.key == key:
                return ra

    def _response_parsing(self, response: str, expected_outputs: List[str]):
        target_annotators = [ra for _, ra in self.response_annotators.items() if ra.key in expected_outputs]

        if len(target_annotators) == 0:
            return {expected_outputs[0]: response}

        parsed_outputs = {}
        for ra in target_annotators:
            parsed_out = ra(response)
            parsed_outputs.update(parsed_out)
        return parsed_outputs

    def _add_demonstrations(self):
        if self.demonstrations is not None:
            for example in self.demonstrations:
                query, parents = self._get_demonstration_query_message_content(example)
                response, parents = self._get_demonstration_response_message_content(example)

                self._log_chat_message(content=query,
                                       message_creator=self.user_name,
                                       parent_message_ids=parents)

                self._log_chat_message(content=response,
                                       message_creator=self.assistant_name,
                                       parent_message_ids=parents)

    def _log_chat_message(self, message_creator: str, content: str, parent_message_ids: List[str] = None):
        chat_message = ChatMessage(
            message_creator=message_creator,
            parent_message_ids=parent_message_ids,
            flow_runner=self.flow_config["name"],
            flow_run_id=self.flow_run_id,
            content=content
        )
        return self._log_message(chat_message)

    def _initialize_conversation(self, input_data: Dict[str, Any]):
        # ~~~ Add the system message ~~~
        system_message_content = self._get_message(self.system_message_prompt_template, input_data)

        self._log_chat_message(content=system_message_content,
                               message_creator=self.system_name)

        # ~~~ Add the demonstration query-response tuples (if any) ~~~
        self._add_demonstrations()
        self._update_state(update_data={"conversation_initialized": True})

    def get_conversation_messages(self, message_format: Optional[str] = None):
        messages = self.flow_state["history"].get_chat_messages()

        if message_format is None:
            return messages

        elif message_format == "open_ai":
            processed_messages = []

            for message in messages:
                if message.message_creator == self.system_name:
                    processed_messages.append(SystemMessage(content=message.content))
                elif message.message_creator == self.assistant_name:
                    processed_messages.append(AIMessage(content=message.content))
                elif message.message_creator == self.user_name:
                    processed_messages.append(HumanMessage(content=message.content))
                else:
                    raise ValueError(f"Unknown name: {message.message_creator}")
            return processed_messages
        else:
            raise ValueError(
                f"Currently supported conversation message formats: 'open_ai'. '{message_format}' is not supported")

    def _call(self):
        api_key = self.flow_state["api_key"]

        backend = langchain.chat_models.ChatOpenAI(
            model_name=self.flow_config["model_name"],
            openai_api_key=api_key,
            **self.flow_config["generation_parameters"],
        )

        messages = self.get_conversation_messages(
            message_format="open_ai"
        )

        _success = False
        attempts = 1
        error = None
        response = None
        while attempts <= self.n_api_retries:
            try:
                response = backend(messages).content
                _success = True
                break
            except Exception as e:
                log.error(
                    f"Error {attempts} in calling backend: {e}. Key used: `{api_key}`. "
                    f"Retrying in {self.wait_time_between_retries} seconds..."
                )
                log.error(
                    f"API call raised Exception with the following arguments arguments: "
                    f"\n{self.flow_state['history'].to_string()}"
                )
                attempts += 1
                time.sleep(self.wait_time_between_retries)
                error = e

        if not _success:
            raise error

        if self.flow_config["verbose"]:
            messages_str = self.flow_state["history"].to_string()
            log.info(
                f"\n{colorama.Fore.MAGENTA}~~~ History [{self.flow_config['name']}] ~~~\n"
                f"{colorama.Style.RESET_ALL}{messages_str}"
            )

        return response

    def _prepare_conversation(self, input_data: Dict[str, Any]):
        if self._is_conversation_initialized():
            # ~~~ Check that the message has a `query` field ~~~
            user_message_content = self.human_message_prompt_template.format(query=input_data["query"])

        else:
            self._initialize_conversation(input_data)
            user_message_content = self._get_message(self.query_message_prompt_template, input_data)

        self._log_chat_message(message_creator=self.user_name,
                               content=user_message_content)

    @flow_run_cache()
    def run(self, input_data: Dict[str, Any], expected_outputs: List[str]) -> Dict[str, Any]:
        # ~~~ Chat-specific preparation ~~~
        self._prepare_conversation(input_data)

        # ~~~ Call ~~~
        response = self._call()
        answer_message = self._log_chat_message(
            message_creator=self.flow_config["assistant_name"],
            content=response
        )

        # ~~~ Response parsing ~~~
        parsed_outputs = self._response_parsing(
            response=response,
            expected_outputs=expected_outputs
        )
        self._update_state(update_data=parsed_outputs)

        if self.flow_config["verbose"]:
            parsed_output_messages_str = pprint.pformat({k: m for k, m in parsed_outputs.items()},
                                                        indent=4)
            log.info(
                f"\n{colorama.Fore.MAGENTA}~~~ "
                f"Response [{answer_message.message_creator} -- "
                f"{answer_message.message_id} -- "
                f"{answer_message.flow_run_id}] ~~~"
                f"\n{colorama.Fore.YELLOW}Content: {answer_message}{colorama.Style.RESET_ALL}"
                f"\n{colorama.Fore.YELLOW}Parsed Outputs: {parsed_output_messages_str}{colorama.Style.RESET_ALL}"
            )

        # ~~~ The final answer should be in self.flow_state, thus allow_class_namespace=False ~~~
        return self._get_keys_from_state(keys=expected_outputs, allow_class_namespace=False)

    @classmethod
    def tune(
        cls,
        tune_dps: List[Dict],
        metric: str,
        mode: str,
        eval_func: Callable,
        api_key: str,
        log_file_name: Optional[str] = None, # TODO(yeeef)
        inference_budget: Optional[float] = None,
        optimization_budget: Optional[float] = None,
        num_samples: Optional[int] = 1,
        logging_level: Optional[int] = logging.WARN, # TODO(yeeef)
        initial_flow_config: Optional[Dict] = None, # if not supplied will use default flow config of the class (xxx.yaml)
        **config,
    ) -> Tuple[Dict, Any]: # tune.ExperimentAnalysis
        """
        Args:
            - tune_dps (list): The list of data points to tune the hyperparameters.
            - metric (str): The metric to optimize.
            - mode (str): The optimization mode, "min" or "max.
            - eval_func (Callable): The evaluation function for responses.
                The function should take a response and a data point as input,
                and return a dict of metrics.
            - log_file_name (str, optional): The log file.
            - inference_budget (float, optional): The inference budget, dollar per instance.
            - optimization_budget (float, optional): The optimization budget, dollar in total.
            - num_samples (int, optional): The number of samples to evaluate.
                -1 means no hard restriction in the number of trials
                and the actual number is decided by optimization_budget. Defaults to 1.
            - logging_level (optional): logging level. Defaults to logging.WARNING.
            - **config (dict): The search space to update over the default search.
                For prompt, please provide a string/Callable or a list of strings/Callables.
                    - If prompt is provided for chat models, it will be converted to messages under role "user".
                    - Do not provide both prompt and messages for chat models, but provide either of them.
                    - A string template will be used to generate a prompt for each data instance
                      using `prompt.format(**data)`.
                    - A callable template will be used to generate a prompt for each data instance
                      using `prompt(data)`.
                For stop, please provide a string, a list of strings, or a list of lists of strings.
                For messages (chat models only), please provide a list of messages (for a single chat prefix)
                or a list of lists of messages (for multiple choices of chat prefix to choose from).
                Each message should be a dict with keys "role" and "content". The value of "content" can be a string/Callable template.

        Returns:
            - dict: The optimized hyperparameter setting.
            - tune.ExperimentAnalysis: The tuning results.
        """

        initial_flow_config = initial_flow_config or cls.get_config()
        space = cls.default_search_space.copy()

        if config is not None:
            space.update(config)
            if "messages" in space:
                space.pop("prompt", None)
            temperature = space.pop("temperature", None)
            top_p = space.pop("top_p", None)
            if temperature is not None and top_p is None:
                space["temperature_or_top_p"] = {"temperature": temperature}
            elif temperature is None and top_p is not None:
                space["temperature_or_top_p"] = {"top_p": top_p}
            elif temperature is not None and top_p is not None:
                space.pop("temperature_or_top_p")
                space["temperature"] = temperature
                space["top_p"] = top_p
                log.warning("temperature and top_p are not recommended to vary together.")
        
        # Note: currently we fix the model rather than make it tunable
        search_alg = BlendSearch(
            cost_attr="cost",
            cost_budget=optimization_budget,
            metric=metric,
            mode=mode,
            space=space,
        )
    
        # Args:
        # evaluation_function: A user-defined evaluation function.
        #     It takes a configuration as input, outputs a evaluation
        #     result (can be a numerical value or a dictionary of string
        #     and numerical value pairs) for the input configuration.
        #     For machine learning tasks, it usually involves training and
        #     scoring a machine learning model, e.g., through validation loss.


        def updated_flow_config_with_search_config(flow_config: Dict[str, Any], search_config: Dict[str, Any]):
            """
            inputs are immutable
            """
            flow_config = deepcopy(flow_config)
            search_config = deepcopy(search_config)

            temperature_or_top_p = search_config.pop("temperature_or_top_p", None)
            if temperature_or_top_p is not None:
                search_config.update(temperature_or_top_p)

            flow_config["model_name"] = search_config.get("model", flow_config["model_name"])
            generation_parameters = flow_config["generation_parameters"]
            for generation_parameter in generation_parameters:
                if generation_parameter == "model_kwargs":
                    continue
                if generation_parameter in search_config:
                    generation_parameters[generation_parameter] = search_config[generation_parameter]

            model_kwargs = generation_parameters["model_kwargs"]
            for model_kwarg in model_kwargs:
                if model_kwarg in search_config:
                    model_kwargs[model_kwarg] = search_config[model_kwarg]

            return flow_config

        def tune_run_eval(search_config: Dict[str, Any]) -> Dict[str, float]:
            """
            evaluation_function: A user-defined evaluation function.
                It takes a configuration as input, outputs a evaluation
                result (can be a numerical value or a dictionary of string
                and numerical value pairs) for the input configuration.
                For machine learning tasks, it usually involves training and
                scoring a machine learning model, e.g., through validation loss.
            """
            # extract the flow_construct_kwargs from search_config
            """
            {'expected_inputs': [], 'expected_outputs': [], 'flow_type': 'Flow', 'verbose': True, 'dry_run': False, 'namespace_clearing_after_run': True, 'n_api_retries': 6, 'wait_time_between_retries': 20, 'system_name': 'system', 'user_name': 'user', 'assistant_name': 'assistant', 'response_annotators': {'code_extractor': <flows.message_annotators.regex_extractor_first.RegexFirstOccurrenceExtractor object at 0x7f532121bc70>}, 'query_message_prompt_template': {'_target_': 'langchain.PromptTemplate', 'template': '# Problem statement\n{{problem_description}}\n\n# Input description\n{{input_description}}\n\n# Output description\n{{output_description}}\n\n{{io_examples_and_explanation}}\n\n\nThe input should be read from the standard input and the output should be passed to the standard output.\nReturn Python code that solves the problem. Reply in the following format:\n```python\n{{code_placeholder}}\n```', 'input_variables': ['problem_description', 'input_description', 'output_description', 'io_examples_and_explanation'], 'partial_variables': {'code_placeholder': '{{python_code}}'}, 'template_format': 'jinja2'}, 'demonstrations': None, 'demonstrations_response_template': None, 'name': 'CodeAgent', 'description': 'ToDO: add description', 'model_name': 'gpt-3.5-turbo', 'generation_parameters': {'n': 1, 'max_tokens': 3000, 'temperature': 0.3, 'model_kwargs': {'top_p': 0.2, 'frequency_penalty': 0, 'presence_penalty': 0}}, 'system_message_prompt_template': {'_target_': 'langchain.PromptTemplate', 'template': 'Your goal is to provide executable Python code that solves a competitive programming problem. The code should correctly handle all corner cases in order to pass the hidden test cases, which are used to evaluate the correctness of the solution.\n\nThe user will specify the problem by providing you with:\n  - the problem statement\n  - input description\n  - output description\n  - example test cases\n  - (optional) explanation of the test cases\n\nThe user will provide you with a task and an output format that you will strictly follow.', 'input_variables': [], 'template_format': 'jinja2'}, 'human_message_prompt_template': {'_target_': 'langchain.PromptTemplate', 'template': '{{query}}', 'input_variables': ['query'], 'template_format': 'jinja2'}}
            """
            log.info(f"Tunning with config: {search_config}")
            # TODO: the code currently only works when there is no subspace, i.e. there is only one model to tune with
            # align search_config with flow_config
            updated_flow_config = updated_flow_config_with_search_config(flow_config=initial_flow_config, search_config=search_config)
            log.info(f"Updated flow_config: {updated_flow_config}")
            # flow_launcher = FlowAPILauncher(flow, 1, False, 3, 0, ["code"]) TODO: maybe refactor with flow_launcher

            # TODO: limitations: langchain api call does not give us the cost of the api call, and only give us
            #       one result no matter the n
            final_metrics = {}
            for sample in tune_dps:
                sample["api_key"] = api_key
                # log.info(f"sample: {sample}")
                flow = cls.instantiate_from_config(updated_flow_config)
                task_message = flow.package_task_message(recipient_flow=flow,
                                                            task_name="run_task",
                                                            task_data=sample,
                                                            expected_outputs=["code"])
                output_message = flow(task_message)
                # log.info(f"output_message: {output_message}")

                metrics = eval_func(output_message.data['code'], sample)
                log.info(f"metrics for dp: {metrics}")
                if not final_metrics:
                    final_metrics = metrics
                else:
                    for k, v in metrics.items():
                        final_metrics[k] += v
            log.info(f"final metric {final_metrics} for this config {search_config}")
            return final_metrics

        analysis = tune.run(
            tune_run_eval,
            search_alg=search_alg,
            num_samples=num_samples,
            log_file_name=log_file_name,
            verbose=3,
        )
        best_search_config = analysis.best_config
        flow_config = updated_flow_config_with_search_config(initial_flow_config, best_search_config)
        log.info(f"best search config found: {best_search_config}, analysis: {analysis.best_result}")
        return flow_config, analysis