costly 2.2.3

Estimate costs and running times of complex LLM workflows/experiments/pipelines in advance before spending money, via simulations.

costly

Estimate costs and running times of complex LLM workflows/experiments/pipelines in advance before spending money, via simulations. Just put @costly() on the load-bearing function; make sure all functions that call it pass **kwargs to it and call your complex function with simulate=True and some cost_log: Costlog object. See examples.ipynb for more details.

(Actually you don't have to pass cost_log and simulate throughout; you can just define a global cost_log and simulate at the top of the file that contains your @costly functions and pass them as defaults to your @costly functions.)

(Pass @costly(disable_costly=True) to disable costly for a function. E.g. you may set disable_costly to be a global variable and pass it to your @costly decorators.)

https://github.com/abhimanyupallavisudhir/costly

Installation

pip install costly

Usage

See examples.ipynb for a full walkthrough; some examples below.

from costly import Costlog, costly, CostlyResponse
from costly.estimators.llm_api_estimation import LLM_API_Estimation as estimator


@costly()
def chatgpt(input_string: str, model: str) -> str:
    from openai import OpenAI

    client = OpenAI()
    response = client.chat.completions.create(
        model=model, messages=[{"role": "user", "content": input_string}]
    )
    output_string = response.choices[0].message.content
    return output_string


@costly(
    input_tokens=lambda kwargs: LLM_API_Estimation.messages_to_input_tokens(
        kwargs["messages"], kwargs["model"]
    ),
)
def chatgpt_messages(messages: list[dict[str, str]], model: str) -> str:
    from openai import OpenAI

    client = OpenAI()
    response = client.chat.completions.create(model=model, messages=messages)
    output_string = response.choices[0].message.content
    return output_string


@costly()
def chatgpt(input_string: str, model: str) -> str:
    from openai import OpenAI

    client = OpenAI()
    response = client.chat.completions.create(
        model=model,
        messages=[
            {"role": "user", "content": input_string},
        ],
    )

    return CostlyResponse(
        output=response.choices[0].message.content,
        cost_info={
            "input_tokens": response.usage.prompt_tokens,
            "output_tokens": response.usage.completion_tokens,
        },
    ) # in usage, this will still just return the output, not the whole CostlyResponse object

Testing

poetry run pytest -s -m "not slow"
poetry run pytest -s -m "slow"

Tests for instructor currently fail.

TODO

• Make it work with async
• Decide and document what the best way to "propagate" description (for breakdown purposes) through function calls is. Have the user manually write def f(...): ... g(description = kwargs.get("description") + ["f"]? Add a @description("blabla") decorator? Add a @description decorator that automatically appends the function name and arguments into description?
• Better solution for token counting for Chat messages (search HACK in the repo)
• make instructor tests pass https://community.openai.com/t/how-to-calculate-the-tokens-when-using-function-call/266573/11
• Support for locally run LLMs -- ideally need a cost & time estimator that takes into account your machine details, GPU pricing etc.
• support more models

Instructor tests don't really pass but I can kinda live with this. Lmk if anyone has a good way to count tokens from messages that includes tool calling (I'm using this).

FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[PERSONINFO_gpt-4o] - AssertionError: ['Input tokens estimate 65 not within 20pc of truth 83']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[PERSONINFO_gpt-4o-mini] - AssertionError: ['Input tokens estimate 65 not within 20pc of truth 83']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[PERSONINFO_gpt-4-turbo] - AssertionError: ['Input tokens estimate 65 not within 20pc of truth 85']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[PERSONINFO_gpt-3.5-turbo] - AssertionError: ['Input tokens estimate 65 not within 20pc of truth 85']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[FOOMODEL_gpt-4o] - AssertionError: ['Input tokens estimate 77 not within 20pc of truth 108']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[FOOMODEL_gpt-4o-mini] - AssertionError: ['Input tokens estimate 77 not within 20pc of truth 108']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[FOOMODEL_gpt-4-turbo] - AssertionError: ['Input tokens estimate 77 not within 20pc of truth 113']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[FOOMODEL_gpt-3.5-turbo] - AssertionError: ['Input tokens estimate 77 not within 20pc of truth 113']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[BARMODEL_gpt-4o] - AssertionError: ['Input tokens estimate 70 not within 20pc of truth 168']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[BARMODEL_gpt-4o-mini] - AssertionError: ['Input tokens estimate 70 not within 20pc of truth 168']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[BARMODEL_gpt-4-turbo] - AssertionError: ['Input tokens estimate 70 not within 20pc of truth 178']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[BARMODEL_gpt-4] - AssertionError: ['Input tokens estimate 70 not within 20pc of truth 126']
FAILED tests/test_estimators/test_llm_api_estimation.py::test_estimate_contains_exact_instructor[BARMODEL_gpt-3.5-turbo] - AssertionError: ['Input tokens estimate 70 not within 20pc of truth 178']