hey 😄 I'm having some issues with transitive dependencies. my torch dependencies from

transformers

is not resolved by

pants

(since

transformers

resolves it at runtime). I can solve that with my

pex_binary

for production by using

inherit_path="fallback"

. How can I achieve the same solution for my dev script that is executed with

python_source

? here is my

BUILD

file:

python_source(
    name="dev",
    dependencies=[":api"],
    source="src/serve.py",
    restartable=True,
)

pex_binary(
    name="prod",
    dependencies=[":api"],
    entry_point="src/serve.py",
    inherit_path="fallback",
)

python_sources(
    name="api",
    sources=["**/*.py"],
)

python_tests(
    name="tests",
    sources=["tests/**/test_*.py"],
)

python_test_utils(
    name="test_utils",
    sources=["tests/conftest.py"],
)

even having torch installed on my virtualenv, Pants for some reason still can't use it

Pants creates its own virtualenvs, so I'm not sure which virtualenv you're referring to?

@happy-kitchen-89482 I would like to avoid having to have torch within my Pants virtualenv mainly because that increases the deployment time of the app drastically

I see, I think. So you want to be able to

pants run

some source file on some existing venv, instead of one that Pants creates?

this can be an option, for sure 😄

Hmm, I think @flat-zoo-31952 has some perspective on this, although I'm not sure this can be done today, but it's a good use case to know about for https://github.com/pantsbuild/pants/discussions/20897

I have to admit I haven’t really followed the progress on provided dependencies

@happy-kitchen-89482 cool 😄 I'll follow that discussion if we can have something like that it would be amazing. For some specific workflows (usually when using ML tools) this would be desirable since torch (and most ML frameworks) are extremely slow to install. We don't need exactly the feature I asked for (being capable of falling back to locally installed modules), what we would need to ensure is avoiding reinstalling it