How to improve the Python import time of your Typer CLI

Feb 17, 2026

tldr: I describe a 6-step method to reduce the import time of Python scripts. Basically, you find the libraries with high import time, you import them inside functions and you also postpone type annotation evaluation. This helped me cut the execution time from 0.584s to 0.075s on a Typer CLI.

When the execution time of your Python CLI is dominated by LLM generations, your Python import time might not matter much.

But you may still want to speed it up a bit. Then asking for the version, help, or getting an error for a wrong argument stays fast. That makes for a better user experience.

This is what I did with my phrasebook CLI, a command line tool that generates translations, audio, and images with the OpenAI API.

I improved its execution time for showing the version from 0.584s to 0.075s. I did it by cutting down its import time with the method below.

6-step method to reduce Python import time¶

Basically, what you want to do is this. Assume your CLI has a --version option, and running it doesn't need heavy libraries and should return fast.

Measure the execution time of your CLI. You can do this with time, like this:
```
$ time uv run mycli.py --version
```
Find the libraries with a high import time using -X importtime a CPython option, like this:
```
$ uv run python -X importtime mycli.py --version 2> importtime.txt
```

Import those libraries inside functions. Instead of this:

from expensive_import_time import bar

def foo():
    x = do_something()
    return bar(x)

Do this:

def foo():
   from expensive_import_time import bar

   x = do_something()
   return bar(x)

You can also think about swapping one library for another one. Pick one with a lower import time.

Postpone type annotation evaluation and store them as strings. Add this line at the top of the file:
```
from __future__ import annotations
```
If you don't do that, you'll get a runtime NameError for type annotations that use classes you now import inside functions (see step 2).
Keep your type checker happy. Import the libraries, now imported inside functions, in the following if block (see step 2), with TYPE_CHECKING import from typing library:
```
from typing import TYPE_CHECKING

if TYPE_CHECKING:
    from expensive_import_time import bar
```
Measure the execution time of your CLI again, as in step 1. Check that it improved.

If you do this, you should see your CLI import time go down.

Example: Reducing import time of a CLI generating translations with OpenAI API¶

Let's look at the following Python CLI. It's an overly simplified version of my phrasebook CLI. It's a typer app. It takes one French word as an argument, translates it to English using the OpenAI API, and saves the translation pair to the translations.csv file. It also has a --version option.

After installing the dependencies and exporting your OpenAI API key, you can translate the word "cusine" ("kitchen") and save it in translations.csv like this:

$ uv init
$ uv install typer openai pandas
$ export OPENAI_API_KEY=<your-api-key>
$ uv run translate_with_slow_import_time.py cuisine

The file translate_with_slow_import_time.py is defined like this:

# translate_with_slow_import_time.py
from pathlib import Path
from typing import Annotated
import logging
import pandas as pd
from pydantic import BaseModel
from openai import OpenAI
import typer

__version__ = "0.1.0"

logger = logging.getLogger(__name__)

TRANSLATIONS_FILE = Path("translation.csv")
COLUMNS = ["french", "english"]

app = typer.Typer(pretty_exceptions_enable=False)


class Translation(BaseModel):
    french: str
    english: str


def generate_translation(word: str, client: OpenAI) -> tuple[str, str] | None:
    response = client.responses.parse(
        model="gpt-5.2",
        instructions="Translate the following French word to English. Return the pair.",
        input=word,
        text_format=Translation,
        max_output_tokens=256,
    )

    if tr := response.output_parsed:
        return (tr.french, tr.english)
    return None


def save_translation(translation: tuple[str, str], translations_file: Path) -> None:
    if translations_file.exists():
        translations_df = pd.read_csv(translations_file, dtype="string")
    else:
        translations_df = pd.DataFrame(columns=pd.Index(COLUMNS), dtype="string")

    new_translation_df = pd.DataFrame(
        [translation], columns=pd.Index(COLUMNS), dtype="string"
    )
    updated = (
        pd.concat([translations_df, new_translation_df], ignore_index=True)
        if not translations_df.empty
        else new_translation_df
    )
    updated.to_csv(translations_file, index=False)


def version_callback(version: bool):
    if version:
        print(f"version {__version__}")
        raise typer.Exit()


@app.command()
def run(
    word: str,
    version: Annotated[
        bool | None,
        typer.Option("--version", callback=version_callback, is_eager=True),
    ] = None,
) -> None:
    client = OpenAI()

    translation = generate_translation(word, client)
    if not translation:
        logger.error(f"Couldn't generate translation for '{word}'.")
    else:
        save_translation(translation, TRANSLATIONS_FILE)


if __name__ == "__main__":
    app()

Step 1 - Measuring the execution time¶

Let's start by measuring the execution time.

$ time uv run translate_with_slow_import_time.py --version
version 0.1.0

real    0m0.578s
user    0m1.687s
sys     0m0.075s

Step 2 - Finding libraries with high import time¶

Now we run the following command. It reports import time for our CLI's libraries, including their dependencies. This gives us a 1400-row table in the importtime.txt file.

$ uv run python -X importtime translate_with_slow_import_time.py --version 2> importtime.txt

Here's a snippet. We keep only the top-level imported libraries:

import time: self [us] | cumulative | imported package
import time:       281 |        848 | _frozen_importlib_external
import time:        83 |        146 | zipimport
import time:       420 |        926 | encodings
import time:       121 |        121 | encodings.utf_8
import time:        66 |         66 | _signal
import time:        99 |        190 | io
import time:       332 |       1454 | site
import time:      1107 |       6819 | pathlib
import time:      1586 |       2117 | typing
import time:      1208 |       4109 | logging
import time:       229 |     183804 | pandas
import time:       180 |      29859 | pydantic
import time:       183 |       1635 | pydantic._internal._config
import time:      2246 |       3140 | pydantic._internal._decorators
import time:       207 |        971 | pydantic._internal._fields
import time:       194 |        714 | pydantic._internal._mock_val_ser
import time:       366 |       7539 | pydantic._internal._model_construction
import time:       374 |     221018 | openai
import time:       144 |       4681 | typer

The libraries with the biggest impact on import time are openai at 221ms, pandas at 183ms, and pydantic at 29ms.

In the next step, we'll import them inside functions.

Note that we use pandas only to read and write CSV files. So we could swap it for the built-in csv library.

Step 3 - Importing libraries inside functions¶

Next we'll tweak the CLI. We'll save the updated version as translate_with_better_import_time.py.

Here we omit the parts of the code that stay the same. We only show the new places where we import openai, pandas, and pydantic. For pydantic, we also move the Translation model definition inside the function, along with the pydantic import.

# translate_with_better_import_time.py
from pathlib import Path
from typing import Annotated
import logging
import typer

# ...

def generate_translation(word: str, client: OpenAI) -> tuple[str, str] | None:
    from pydantic import BaseModel

    class Translation(BaseModel):
        french: str
        english: str
    # ...


def save_translation(translation: tuple[str, str], translations_file: Path) -> None:
    import pandas as pd
    # ...

# ...

@app.command()
def run(...) -> None:
    from openai import OpenAI
    # ...


if __name__ == "__main__":
    app()

If we try to run it, we get an expected NameError:

$ uv run translate_with_better_import_time.py --version
Traceback (most recent call last):
  File "/home/tony/work/python/py0024-import-time-typer-app-openai-pandas-pydantic-importtime/translate_with_better_import_time.py", line 17, in <module>
    def generate_translation(word: str, client: OpenAI) -> tuple[str, str] | None:
                                                ^^^^^^
NameError: name 'OpenAI' is not defined

Step 4 - Postponing type annotation evaluation and storing them as strings¶

To fix that NameError, we add this line at the top of the script:

from __future__ import annotations

Then the CLI works fine:

$ uv run translate_with_better_import_time.py --version
version 0.1.0

Step 5 - Making the type checker happy¶

Now we can see that the ty type checker isn't happy anymore:

$ ty check translate_with_better_import_time.py
error[unresolved-reference]: Name `OpenAI` used when not defined
  --> translate_with_better_import_time.py:18:45
   |
18 | def generate_translation(word: str, client: OpenAI) -> tuple[str, str] | None:
   |                                             ^^^^^^
19 |     from pydantic import BaseModel
   |
info: rule `unresolved-reference` is enabled by default

Found 1 diagnostic

To fix this, we add the following to our script:

from typing import TYPE_CHECKING

if TYPE_CHECKING:
    from openai import OpenAI

So the final file is:

# translate_with_better_import_time.py
from __future__ import annotations
from pathlib import Path
from typing import Annotated, TYPE_CHECKING
import logging
import typer

__version__ = "0.1.0"

if TYPE_CHECKING:
    from openai import OpenAI

logger = logging.getLogger(__name__)

TRANSLATIONS_FILE = Path("translation.csv")
COLUMNS = ["french", "english"]

app = typer.Typer(pretty_exceptions_enable=False)


def generate_translation(word: str, client: OpenAI) -> tuple[str, str] | None:
    from pydantic import BaseModel

    class Translation(BaseModel):
        french: str
        english: str

    response = client.responses.parse(
        model="gpt-5.2",
        instructions="Translate the following French word to English. Return the pair.",
        input=word,
        text_format=Translation,
        max_output_tokens=256,
    )

    if tr := response.output_parsed:
        return (tr.french, tr.english)
    return None


def save_translation(translation: tuple[str, str], translations_file: Path) -> None:
    import pandas as pd


    if translations_file.exists():
        translations_df = pd.read_csv(translations_file, dtype="string")
    else:
        translations_df = pd.DataFrame(columns=pd.Index(COLUMNS), dtype="string")

    new_translation_df = pd.DataFrame(
        [translation], columns=pd.Index(COLUMNS), dtype="string"
    )
    updated = (
        pd.concat([translations_df, new_translation_df], ignore_index=True)
        if not translations_df.empty
        else new_translation_df
    )
    updated.to_csv(translations_file, index=False)


def version_callback(version: bool):
    if version:
        print(f"version {__version__}")
        raise typer.Exit()


@app.command()
def run(
    word: str,
    version: Annotated[
        bool | None,
        typer.Option("--version", callback=version_callback, is_eager=True),
    ] = None,
) -> None:
    from openai import OpenAI

    client = OpenAI()

    translation = generate_translation(word, client)
    if not translation:
        logger.error(f"Couldn't generate translation for '{word}'.")
    else:
        save_translation(translation, TRANSLATIONS_FILE)


if __name__ == "__main__":
    app()

Then type checking is happy:

$ ty check translate_with_better_import_time.py
All checks passed!

Step 6 - Measuring the execution time¶

Finally, we can check that we've improved our execution time. Elapsed time drops from 0.578s to 0.075s. We got there by cutting down the import time of our CLI.

$ time uv run translate_with_better_import_time.py --version
version 0.1.0

real    0m0.075s
user    0m0.053s
sys     0m0.022s

That's all I have for today! Talk soon 👋