Support lazy importing submodules in attach

[sjmonson]

Adds a lazy_submodules: bool argument to attach which substitutes importlib.import_module for the load function. Found this useful for bundling a bunch of functions/objects with their required extras. E.g.

pkg/extras/__init__.py

import lazy_loader as lazy

__getattr__, __dir__, __all__ = lazy.attach(
    __name__,
    submodules=["group_one", "group_two"],
    lazy_submodules=True,
)

pkg/extras/group_one.py

try:
    import numpy as np
except ImportError as e:
    raise ImportError("Install pkg[group_one] to use these features") from e

def add(x np.int64, y: np.int64) -> np.int64:
    ...

[stefanv]

Thanks for the contribution @sjmonson. I don't see a test that exercises the new behavior?

[sjmonson]

@stefanv Oh sorry I forgot about this PR. It actually turned out that while this idea worked in theory, in practice it is pretty fragile and packages like PyTorch can break it by trying to walk sys.modules.

I ended up solving this problem a different way by making a new version of attach which gets layered under a normal attach. You can see it in action here or here. I can make a new PR with those changes if you're interested; for now I'll just close this PR.

[stefanv]

Thanks for the update, Samuel!

Can you explain a bit about what https://github.com/vllm-project/guidellm/blob/d07502b51b4c30d7acdc17ee8990c3899978cb08/src/guidellm/utils/lazy_loader.py#L163 does?

[sjmonson]

Sure, basically the goal is to bring a series of imports in a attributes on the current module without importing the source module in advance. Since we don't know if each attribute we are importing is a attribute of the source module or a submodule of the source module we need to try a couple different things.

src/guidellm/extras/vision.py is a good example. We essentially want a module that looks like:

import imageio.v3 as iio
from PIL import Image as PILImage
from PIL.Image import Image

__all__ = ["Image", "PILImage", "iio"]

imageio.v3 is a module we bind to the iio attribute, PIL.Image is a module we bind to the PILImage attribute, and PIL.Image.Image is an attribute of the PIL.Image module that we bind to Image locally.

At the start of _attach_extras_attrs function we make a map which is <locally bound name>: (<source module>, <source attr name>). In the __getattr__ method after we check if the looked up attribute is in our bound name list, we attempt to import the source module, if that succeeds we try to getattr the source attr name and if that fails we try to import it as a module instead. This is here to ensure that attributes are cached in the module's __dict__ and IIRC ImportErrors are re-thown as AttributeErrors because PyTorch had some weird module walking behavior that would break if an ImportError occurred when accessing an attribute. Its been a bit so I can't remember the specifics.