Support polars, etc. through narwhals (#957)

* Start working on narwhalizing glum again

* Some fixes

* Make category expansion work

* Delete temporary file

* Some additional tests and test updates

* Add narwhals as a dependency

* `feature_dtypes_` backwards compatibility

* I don't think this was doing anything

* Add back the check for non-contiguous arrays

* Use the same default

* Disclaimer (should we just deprecate/ignore this argument?)

* Bump minimum tabmat (and, by necessity, sklearn and python) version

* Bump minimum versions

* Avoid large diff

* Make sure manifests are consistent

* Adjust tests

* Polars golden master tests

* Additional polars tests

* Add support for unpickling old models

* Polar's global string cache bites again

* Enums are fine, though

* Version bumps

* Different runner

* Different runners again

* Review comment

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Remove CV pickle compatibility test

The GeneralizedLinearRegressorCV fitting behavior changed intentionally
on main (alpha path computation + test weight normalization), so the
v3.0 pickle predictions no longer match a fresh fit. Regenerate the
non-CV pickle with actual glum 3.0.0 and drop the CV test case.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Regenerate v3.0 pickle using glum 3.0.0 + pandas 1.4.4

The previous pickle was generated with pandas 3.0.1, causing a
StringDtype constructor mismatch when loaded in the oldies environment
(pandas 1.4.4). Regenerate using matching versions.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* pre-commit

* oops

* Add changelog entry for Polars support

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: 3 people

.github/workflows/package.yml

@@ -22,7 +22,7 @@ jobs:
         include:
           - { variant-file: linux_64_numpy2.0python3.10.____cpython,  target-platform: linux-64,  os: ubuntu-latest,  rattler-build-args: '' }
           - { variant-file: linux_64_numpy2python3.13.____cp313,      target-platform: linux-64,  os: ubuntu-latest,  rattler-build-args: '' }
-          - { variant-file: osx_64_numpy2.0python3.10.____cpython,    target-platform: osx-64,    os: macos-latest,   rattler-build-args: '' }
+          - { variant-file: osx_64_numpy2.0python3.10.____cpython,    target-platform: osx-64,    os: macos-15-intel, rattler-build-args: '' }
           - { variant-file: osx_arm64_numpy2.0python3.10.____cpython, target-platform: osx-arm64, os: macos-latest,   rattler-build-args: '' }
           - { variant-file: osx_arm64_numpy2python3.13.____cp313,     target-platform: osx-arm64, os: macos-latest,   rattler-build-args: '' }
           - { variant-file: win_64_numpy2.0python3.10.____cpython,    target-platform: win-64,    os: windows-latest, rattler-build-args: '' }

CHANGELOG.rst

@@ -13,6 +13,7 @@ Changelog
 **New features:**
 
 - Add ``solver="closed-form"`` for Gaussian identity-link models, using an analytical normal-equations solution for ridge/OLS, auto-selecting it under ``solver="auto"`` for unconstrained no-L1 cases, and falling back to least-squares for singular or ill-conditioned systems.
+- :class:`~glum.GeneralizedLinearRegressor` and :class:`~glum.GeneralizedLinearRegressorCV` now accept `Polars <https://pola.rs>`_ DataFrames as input, in addition to pandas DataFrames and numpy arrays.
 
 3.1.3 - 2025-02-18
 ------------------

conda.recipe/recipe.yaml

@@ -36,21 +36,25 @@ requirements:
     - if: osx
       then:
         - llvm-openmp
-    - numpy
+    - numpy >=1.24
     - pip
-    - scikit-learn >=0.23
+    - scikit-learn >=1.1.0
     - setuptools
     - setuptools-scm
   run:
     - python
-    - formulaic >=0.6
+    - formulaic >=1.2.0
     - joblib
+    - narwhals >=2.0.0
     - numexpr
+    - numpy >=1.24
     - packaging
-    - pandas
-    - scikit-learn >=0.23
-    - scipy
-    - tabmat >=4.0.0
+    - pandas >=1.4
+    - pyarrow
+    - scikit-learn >=1.1.0
+    - scipy >=1.8.0
+    - tabmat >=4.2.0
+    - tqdm
 
 tests:
   - python:

pixi.lock

@@ -48,18 +48,20 @@ cmd = """
 c-compiler = "*"
 cxx-compiler = "*"
 cython = "*"
-formulaic = "*"
+formulaic = ">=1.2.0"
+narwhals = ">=2.0.0"
 numexpr = "*"
+numpy = ">=1.24"
 packaging = "*"
 pandas = ">=1.4"
 pip = "*"
 pyarrow = "*"
-python = ">=3.9"
-scikit-learn = ">=0.23"
-scipy = "*"
+python = ">=3.10"
+scikit-learn = ">=1.1.0"
+scipy = ">=1.8.0"
 setuptools = ">=61"
 setuptools-scm = ">=8.1"
-tabmat = ">=4.0.0"
+tabmat = ">=4.2.0"
 tqdm = "*"
 wheel = "*"
 
@@ -77,6 +79,7 @@ attrs = "*"
 click = "*"
 git_root = "*"
 mypy = "*"
+polars = "*"
 psutil = "*"
 pytest = "*"
 pytest-xdist = "*"
@@ -146,15 +149,13 @@ cxx-compiler = "*"
 cython = "!=3.0.4"
 make = "*"
 mako = "*"
-narwhals = ">=1.4.1"
+narwhals = ">=2.0.0"
 pip = "*"
 setuptools-scm = "*"
 xsimd = "<11|>12.1"
 [feature.build-tabmat.target.unix.dependencies]
 jemalloc-local = "*"
 
-[feature.py39.dependencies]
-python = "3.9.*"
 [feature.py310.dependencies]
 python = "3.10.*"
 [feature.py311.dependencies]
@@ -165,12 +166,14 @@ python = "3.12.*"
 python = "3.13.*"
 
 [feature.oldies.dependencies]
-formulaic = "0.6.*"
+formulaic = "1.2.*"
+narwhals = "2.0.*"
+numpy = "1.24.*"
 pandas = "1.4.*"
-python = "3.9.*"
-scikit-learn = "0.24.*"
-scipy = "1.7.*"
-tabmat = "4.0.*"
+python = "3.10.*"
+scikit-learn = "==1.1.0"
+scipy = "1.8.*"
+tabmat = "==4.2.0"
 
 [environments]
 benchmark = ["benchmark"]
@@ -189,4 +192,3 @@ py310 = ["py310", "test"]
 py311 = ["py311", "test"]
 py312 = ["py312", "test"]
 py313 = ["py313", "test"]
-py39 = ["py39", "test"]

pixi.toml

@@ -31,7 +31,7 @@ select = [
 known-first-party = ["glum", "glum_benchmarks"]
 
 [tool.mypy]
-python_version = "3.9"
+python_version = '3.10'
 exclude = [
   "^tests/",
   "^\\.pixi/",

pyproject.toml

@@ -63,24 +63,27 @@
     license="BSD",
     classifiers=[  # Optional
         "Programming Language :: Python :: 3",
-        "Programming Language :: Python :: 3.9",
         "Programming Language :: Python :: 3.10",
         "Programming Language :: Python :: 3.11",
         "Programming Language :: Python :: 3.12",
         "Programming Language :: Python :: 3.13",
     ],
     package_dir={"": "src"},
     packages=find_packages(where="src", include=["glum"]),
-    python_requires=">=3.9",
+    python_requires=">=3.10",
     install_requires=[
+        "formulaic>=1.2.0",
         "joblib",
+        "narwhals>=2.0.0",
         "numexpr",
-        "numpy",
-        "pandas",
-        "scikit-learn>=0.23",
-        "scipy",
-        "formulaic>=0.6",
-        "tabmat>=4.0.0",
+        "numpy>=1.24",
+        "packaging",
+        "pandas>=1.4",
+        "pyarrow",
+        "scikit-learn>=1.1.0",
+        "scipy>=1.8.0",
+        "tabmat>=4.1.5",
+        "tqdm",
     ],
     entry_points=None,
     ext_modules=cythonize(

setup.py

@@ -1574,7 +1574,7 @@ def guess_intercept(
     if (not isinstance(link, IdentityLink)) and (len(np.unique(y)) == 1):
         raise ValueError("No variation in `y`. Coefficients can't be estimated.")
 
-    avg_y: float = np.average(y, weights=sample_weight)
+    avg_y = np.average(y, weights=sample_weight)
 
     if isinstance(link, IdentityLink):
         # This is only correct for the normal. For other distributions, the
@@ -1585,7 +1585,7 @@ def guess_intercept(
 
         avg_eta = eta if np.isscalar(eta) else np.average(eta, weights=sample_weight)
 
-        return avg_y - avg_eta
+        return avg_y - avg_eta  # type: ignore[operator]
 
     elif isinstance(link, LogLink):
         # This is only correct for Tweedie