rpy2 in API mode cannot be built without R.

[1himan]

Description of the problem

After #13729 is merged, I'm getting the following build error as you can see in ss, after creating a fresh venv:

Inside the pyproject.toml:
dev includes the doc group, and doc currently includes rpy2, which requires a working R installation at build time. See this.

[scott-huberty]

@1himan what version of uv is installed? I cannot replicate the error on install, but I do get that same error when I try to actually run the tutorial added in #13729

In any case, if this is going to be a blocker for contributors, I personally do not feel good about having an implicit R dependency to be able to build the docs. I apologize for not looking closer at #13729 or #13721 and raising my concerns then.

[1himan]

I'm using:
uv 0.11.1 (a6042f67f 2026-03-24 x86_64-unknown-linux-gnu).

Additional info

 mne sys_info
Platform             Linux-6.19.9-arch1-1-x86_64-with-glibc2.43
Python               3.14.3 (main, Feb 13 2026, 15:31:44) [GCC 15.2.1 20260209]
Executable           /usr/bin/python
CPU                  12th Gen Intel(R) Core(TM) i5-12450H (12 cores)
Memory               15.3 GiB

Core
├☑ mne               1.11.0 (latest release)
├☑ numpy             2.4.3 (unknown linalg bindings)
├☑ scipy             1.17.1
└☑ matplotlib        3.10.8 (backend=qtagg)

Numerical (optional)
├☑ sklearn           1.8.0
├☑ pandas            2.3.3
├☑ h5io              0.2.5
├☑ h5py              3.16.0
└☐ unavailable       numba, nibabel, nilearn, dipy, openmeeg, cupy

Visualization (optional)
├☑ qtpy              2.4.3 (PySide6=6.10.2)
├☑ pyqtgraph         0.14.0
├☑ mne-qt-browser    0.7.4
└☐ unavailable       pyvista, pyvistaqt, vtk, ipympl, ipywidgets, trame_client, trame_server, trame_vtk, trame_vuetify

Ecosystem (optional)
├☑ edfio             0.4.13
├☑ pybv              0.7.6
├☑ defusedxml        0.7.1
└☐ unavailable       mne-bids, mne-nirs, mne-features, mne-connectivity, mne-icalabel, mne-bids-pipeline, neo, eeglabio, curryreader, mffpy, antio

[1himan]

okay... even with the latest version
 uv --version uv 0.11.2 (x86_64-unknown-linux-gnu)
I get the same error.

[scott-huberty]

Ah OK I can replicate the error if I install from source (I do not have R installed on my computer either):

uv pip install --no-binary rpy2,rpy2-rinterface,rpy2-robjects rpy2 rpy2-rinterface rpy2-robjects

Looks like there aren't any published Linux wheels for rpy2-rinterface on pypi, which explains why @1himan is hitting this issue when installing the docs group

EDIT: And as the message hints at, the immediate work around is to set export RPY2_CFFI_MODE=ABI

[tsbinns]

I just ran into this also.
I'd say at the minimum there should be a note in the contributing guide about this potentially happening when setting up the dev environment, but better yet if there's a workaround to avoid this in the first place. Not sure what that would look like though.

[drammock]

Discussed this with @tsbinns and @scott-huberty. Plan is for @tsbinns to add a build_script.py to be run by hatchling during install; that script will set the env var so that ABI mode is triggered. I think it should be unnecessary to unset afterward, as the env var will only last the duration of that shell.

If that works, then we probably should also just skip installing r-base in CIs and rely on ABI mode env var.

[drammock]

with that sorted, we should also add some docs (to the example itself, and maybe to the contrib guide) warning that the ABI-mode env var is only a workaround for installation, and that the example will need r-base to be installed in order to be executed (fails with R_HOME is None or similar if it's not). Unfortunately, sphinx-gallery examples don't have a nice skipif mechanism like pytest does.

[tsbinns]

@drammock @scott-huberty Slight hickup: I can't get the rpy2-interface (and other) installs to stop working.

I used export RPY2_CFFI_MODE=ABI yesterday to get the doc group installed, but I can't figure out how to undo this. E.g., if I set export RPY2_CFFI_MODE=API, it doesn't cause the error I got yesterday.

I'll need to track down what's happening before I can test the fix locally.

[drammock]

is it possible you somehow ended up with R installed on your system without realizing it?

[scott-huberty]

@tsbinns This happened to me too - pip was using a cached wheel from my previously successful install (even if I passed the --no-binary flag..)

clearing the cache worked for me: python -m pip cache remove rpy2

EDIT: I guess you can also pass --no-cache-dir flag to the pip install command

[larsoner]

This is also creating a mess in mne-tools/mne-installers#427 and the ABI tweak didn't seem to fix things there when I tried it.

It is so rare for devs to want to build all doc examples nowadays... I think dev shouldn't include rpy2. By extension, doc shouldn't either. I think a separate doc-full could make sense here, and it could inherit the doc group and add any "heavy" dependencies, in this case so far probably just rpy2. In practice maybe only CircleCI will use this group, but having it in pyproject.toml will allow devs to see what they might want to conda install or pip install themselves to get things working in case they want to build all examples locally (I do it maybe once a year nowadays!).

[larsoner]

... rpy2 packages for Windows are no longer being built on conda-forge:

https://github.com/conda-forge/rpy2-feedstock/blob/686bc24189a0b4063487ac6b5e5cd933ffddad01/recipe/meta.yaml#L16

This is perhaps not surprising, since Windows isn't even mentioned in the rpy2 install docs. But this strongly suggests to me that we should not make --group dev (or docs) include it, and just make CircleCI and the rare dev who wants to run all examples (or that particular example) jump through the necessary hoops to do it.