Is there an easy way to indicate a python_source requires a torch CUDAExtension to be built?

You can always add explicit dependencies on a target.

Is there an nvcc build rule in pants I can use?

If you are using

python_sources

, you can use the

overrides

field to add the dependency to the

dependencies

field of one of the

python_source

targets generated from it.

Gotcha. Is it alternatively possible to have a python_source dependency that must be called? For background, I'm trying to port a setup.py to pants and it calls CUDAExtension from torch (similar to making a Cython extension). I want to be able to build this CUDAEXtension and package it with pex.

What do you mean by "a python_source dependency that must be called"? In what context would you want it called? Is it more like a build script?

Similar to manually calling setup.py

Also

adhoc_tool

accepts any runnable target and

python_source

targets are runnable.

It sounds like I'll have to invest some time in

adhoc_tool

. Are there any cython + pants examples I could look at? That should be a similar enough use case to base a solution on.

Alternate solution: Can you just publish the wheel for setup.py distribution to a private PyPI repository? And then have Pants generate its lockfile from that.

Unfortunately, no.

Re Cython, this (still open) issue is probably of interest: https://github.com/pantsbuild/pants/issues/14752

https://pantsbuild.slack.com/archives/C046T6T9U/p1706019547628299?thread_ts=1706019330.046929&cid=C046T6T9U seems relevant

python_distribution

might work, although it produces a wheel as its output. I don't know how that interacts with using a lockfile.