gamfit

Formula-based generalized additive models for Python, backed by a Rust engine.

gamfit fits Gaussian, binomial (including Bernoulli marginal-slope), Poisson, negative-binomial, Gamma, Beta, Tweedie, and multinomial GLMs with smooth terms, random effects, bounded/constrained coefficients, location-scale extensions, survival likelihoods, and flexible/learnable links. Smoothing parameters are selected by REML or LAML. Posterior sampling uses NUTS where supported, and a Gaussian Laplace approximation otherwise.

Manifold smooths handle predictor spaces that wrap or close: circles, cylinders, tori, and the sphere (intrinsic Wahba and spherical-harmonic kernels), plus periodic tensor products and boundary-conditioned B-splines. The Möbius example in the gallery is a 4π-periodic double-cover parameterization, not a twisted Möbius-strip basis.

rotating recovery of a trefoil knot, latent-free loop, wobbly cylinder, lumpy sphere, bumpy torus, and Möbius double-cover from noisy 3-D point clouds

Docs: https://gamfit.readthedocs.io/.

Install

uv add gamfit

Wheels are published for Linux (x86_64, aarch64), macOS (x86_64, Apple silicon), and Windows. No Rust toolchain is required.

Example

import gamfit

# Smooth fits need enough rows for the basis to be identified; ~20 rows
# is the minimum the default `s(x)` basis (cubic B-spline) is well-posed
# on. Use more rows when the signal is noisier.
train = [
    {"y": 1.05, "x": 0.0}, {"y": 1.32, "x": 0.5}, {"y": 1.78, "x": 1.0},
    {"y": 2.41, "x": 1.5}, {"y": 3.10, "x": 2.0}, {"y": 3.95, "x": 2.5},
    {"y": 4.80, "x": 3.0}, {"y": 5.62, "x": 3.5}, {"y": 6.25, "x": 4.0},
    {"y": 6.71, "x": 4.5}, {"y": 6.94, "x": 5.0}, {"y": 6.88, "x": 5.5},
    {"y": 6.55, "x": 6.0}, {"y": 5.99, "x": 6.5}, {"y": 5.20, "x": 7.0},
    {"y": 4.30, "x": 7.5}, {"y": 3.42, "x": 8.0}, {"y": 2.65, "x": 8.5},
    {"y": 2.10, "x": 9.0}, {"y": 1.82, "x": 9.5},
]

model = gamfit.fit(train, "y ~ s(x)")
print(model.predict([{"x": 1.5}, {"x": 5.0}], interval=0.95))
print(model.summary())
model.save("model.gam")

pandas, polars, pyarrow, numpy, dict-of-columns, and list-of-records inputs are all accepted without conversion.

Features

Polyharmonic / Duchon smooths combine magnitude, gradient, and curvature penalty operators on the same basis. P-spline and thin-plate smooths use their standard derivative penalties. Each penalized block has its own smoothing parameter.
Flexible link functions: flexible(base) adds a spline offset on a base link; blended(...) learns a mixture weight; sas and beta-logistic learn shape parameters.
Surface smooths in arbitrary dimension: thin-plate, Duchon (scale-free by default, hybrid with length_scale=...), and Matérn, with automatic knot placement.
Tensor-product and manifold smooths: te(...) / ti(...) B-spline tensors, periodic 1-D, cylinder / torus tensor products, intrinsic sphere (Wahba kernel or spherical harmonics), and boundary-conditioned B-splines.
Dispersion GAMLSS for Gamma, Beta, negative-binomial, and Tweedie via noise_formula=.
Per-axis anisotropy inside a single joint smooth.
Shape-constrained smooths: s(x, shape=monotone_increasing), convex, concave.
Difference smooths: by= factor smooths plus covariance-aware model.difference_smooth(...) contrasts with optional simultaneous bands.
Marginal-slope models that separate baseline risk from a calibrated score's effect, for Bernoulli and survival outcomes.
Survival in several likelihood modes (transformation, Weibull, location-scale, marginal-slope, latent-Gaussian frailty) plus competing-risks cumulative-incidence functions.
Response geometry for spherical and compositional outcomes via Fréchet-mean tangent-space GAMs.
Posterior sampling via NUTS where supported, Gaussian Laplace otherwise, behind one API; conformal prediction intervals via interval="conformal".

API examples

import gamfit
from gamfit.sklearn import GAMRegressor, GAMClassifier

# Validate before you fit
gamfit.validate_formula(train, "y ~ s(x) + group(site)")

# Posterior sampling and mean bands
posterior = model.sample(train, seed=42)
bands = posterior.predict(test, level=0.95)

# Survival
gamfit.fit(df,
    "Surv(entry, exit, event) ~ s(age) + bmi + timewiggle(internal_knots=6)",
    survival_likelihood="transformation",
    baseline_target="weibull",
)

# scikit-learn
est = GAMRegressor(formula="y ~ s(x)")
est.fit(X, y)

# Diagnose, plot, report
model.diagnose(train).metrics
model.plot(train, x="x", kind="prediction")
model.report("report.html")

Public API

Symbol	Purpose
`gamfit.fit(data, formula, **kwargs)`	Fit a model.
`gamfit.load(path)` / `gamfit.loads(bytes)`	Reload a saved model.
`gamfit.load_posterior(path)`	Reload a `PosteriorSamples` archive.
`gamfit.validate_formula(data, formula, ...)`	Type-check a formula without fitting.
`gamfit.build_info()`	Native extension build metadata.
`gamfit.cuda_diagnostics()` / `gamfit.format_cuda_diagnostics()`	CUDA probe results.
`gamfit.explain_error(exc)`	Human-readable hint for a gamfit exception.
`gamfit.Model`	Fitted model: `predict`, `summary`, `check`, `diagnose`, `plot`, `report`, `sample`, `save`.
`gamfit.SurvivalPrediction`	Per-row hazard / survival surface.
`gamfit.CompetingRisksPrediction`, `competing_risks_cif`	Competing-risks CIF evaluation.
`gamfit.MultinomialModel`	Multinomial-logit / softmax model.
`gamfit.SamplingConfig`, `PosteriorSamples`, `PosteriorPredictive`, `PairedPosteriorSamples`	Posterior interface.
`gamfit.ResponseGeometryModel`, `sphere_frechet_mean`, `simplex_frechet_mean`, `alr`, `clr`, `closure`	Response-geometry utilities.
`gamfit.smooth.Duchon`, `Matern`, `BSpline`, `TensorBSpline`, `MeasureJet`, `Sphere`	Smooth descriptors for `smooths=` and torch.
`gamfit.sklearn.GAMRegressor` / `GAMClassifier`	scikit-learn estimators.

Full reference: https://gamfit.readthedocs.io/en/latest/api-reference/.

Optional extras

uv add "gamfit[pandas]"     # pandas + pyarrow input/output
uv add "gamfit[plot]"       # matplotlib-based plotting
uv add "gamfit[sklearn]"    # scikit-learn integration
uv add "gamfit[cuda]"       # NVIDIA CUDA 12 wheel libraries on Linux x86_64
uv add "gamfit[all]"        # pandas + plot + sklearn extras
uv add torch                # PyTorch bridge dependency

GPU acceleration

CUDA support (cuBLAS / cuSOLVER / cuSPARSE) is built into the same wheel; there is no separate gamfit-gpu package. Install gamfit[cuda] on Linux x86_64 when you want PyPI's NVIDIA CUDA 12 runtime libraries instead of a system CUDA toolkit. Per-op dispatch thresholds are derived at probe time from measured GPU FP64 throughput, CPU FP64 throughput, and PCIe bandwidth, so small kernels stay on the CPU. Inspect the calibrated thresholds with gamfit.build_info()["cuda_diagnostics"] or gamfit.format_cuda_diagnostics().

The wheel uses the CUDA 12 ABI. If PyTorch has already mapped a complete CUDA stack, gamfit continues that same stack rather than preloading a second system toolkit. Without an existing stack it loads one complete system or packaged NVIDIA stack. The GPU probe refuses a partial or mixed mapped stack because CUDA context and library-handle ownership cannot be safely split across implementations.

License

AGPL-3.0-or-later. See LICENSE.

Metadata

Release files for gamfit 0.1.267

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for gamfit 0.1.267
File	Size	Uploaded
gamfit-0.1.267.tar.gz	17.2 MB	Sep 9, 2026	Details

Built distributions (wheels)

Table of built distributions (wheels) for gamfit 0.1.267
File	Interpreter	ABI	Platform	Reset
gamfit-0.1.267-cp310-abi3-win_amd64.whl 36.5 MB Sep 9, 2026	CPython 3.10	abi3	Windows x86-64	Details
gamfit-0.1.267-cp310-abi3-musllinux_1_2_x86_64.whl 34.3 MB Sep 9, 2026	CPython 3.10	abi3	Linux musl 1.2+ x86-64	Details
gamfit-0.1.267-cp310-abi3-musllinux_1_2_aarch64.whl 30.9 MB Sep 9, 2026	CPython 3.10	abi3	Linux musl 1.2+ ARM64	Details
gamfit-0.1.267-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl 34.3 MB Sep 9, 2026	CPython 3.10	abi3	Linux glibc 2.17+ x86-64	Details
gamfit-0.1.267-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl 30.7 MB Sep 9, 2026	CPython 3.10	abi3	Linux glibc 2.17+ ARM64	Details
gamfit-0.1.267-cp310-abi3-macosx_11_0_arm64.whl 29.5 MB Sep 9, 2026	CPython 3.10	abi3	macOS 11.0+ ARM64	Details
gamfit-0.1.267-cp310-abi3-macosx_10_12_x86_64.whl 32.3 MB Sep 9, 2026	CPython 3.10	abi3	macOS 10.12+ x86-64	Details

Total release size: 245.8 MB

Release files / gamfit-0.1.267.tar.gz

Download URL	gamfit-0.1.267.tar.gz
Size	17.2 MB
Tags	Source
SHA-256 checksum How to use checksums	`4c2238ede25070609d50d168e6bd9d69f306043a4c848c0a1b7523e0c338ad36`
BLAKE2b-256 checksum How to use checksums	`00c6645080253f5ddf5d3d3bcd7d7963dd699a3804554e3779cb046998e44b78`
Upload date	Sep 9, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`

Release files / gamfit-0.1.267-cp310-abi3-win_amd64.whl

Download URL	gamfit-0.1.267-cp310-abi3-win_amd64.whl
Size	36.5 MB
Tags	CPython 3.10 Windows x86-64 abi3
SHA-256 checksum How to use checksums	`faaddcff87f022b3be39c76ea67f6afc6caca10e33686641ba99df56b6b008cb`
BLAKE2b-256 checksum How to use checksums	`a0d1f5502d6ed35036611c2ed077bef923f6c4740e8f48978ad7c10882cc03ff`
Upload date	Sep 9, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`

Release files / gamfit-0.1.267-cp310-abi3-musllinux_1_2_x86_64.whl

Download URL	gamfit-0.1.267-cp310-abi3-musllinux_1_2_x86_64.whl
Size	34.3 MB
Tags	CPython 3.10 Linux musl 1.2+ x86-64 abi3
SHA-256 checksum How to use checksums	`36fabb32445ba622251caeca38c240e9bc3fe29f6d51ef1c4ddb1dcb2d30dc8f`
BLAKE2b-256 checksum How to use checksums	`a041cec2a7c054e3659c71f80f4b47f1c6e390cd9d82e8401738ba93085a37bf`
Upload date	Sep 9, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`

Release files / gamfit-0.1.267-cp310-abi3-musllinux_1_2_aarch64.whl

Download URL	gamfit-0.1.267-cp310-abi3-musllinux_1_2_aarch64.whl
Size	30.9 MB
Tags	CPython 3.10 Linux musl 1.2+ ARM64 abi3
SHA-256 checksum How to use checksums	`6005db809c222f8c73f52246528b9915459339f5107ff4d82bec3bd2c9371a0c`
BLAKE2b-256 checksum How to use checksums	`c3326a452b5df9336e4d12a893be1e7ac49f9c49732bb4c19ffa77a73be766d8`
Upload date	Sep 9, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`

Release files / gamfit-0.1.267-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

Download URL	gamfit-0.1.267-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
Size	34.3 MB
Tags	CPython 3.10 Linux glibc 2.17+ x86-64 abi3
SHA-256 checksum How to use checksums	`112b6dc0f1d770f4ad0a05ac5c81ae982db84a5329edcefaf862da1881c7216a`
BLAKE2b-256 checksum How to use checksums	`5c6d70603feeeae0285c2cb86841d26cd567e195c9f0cd6752ecac50668b8cce`
Upload date	Sep 9, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`

Release files / gamfit-0.1.267-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl

Download URL	gamfit-0.1.267-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
Size	30.7 MB
Tags	CPython 3.10 Linux glibc 2.17+ ARM64 abi3
SHA-256 checksum How to use checksums	`add9b2d22ec036da617ed45dcbf9cc6d722d98a494b1d176e85b5a44ee7e07c5`
BLAKE2b-256 checksum How to use checksums	`39737b3115dbe8aba6668dfeb0f11441dda57e46d916a94bba8667372b37c092`
Upload date	Sep 9, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`

Release files / gamfit-0.1.267-cp310-abi3-macosx_11_0_arm64.whl

Download URL	gamfit-0.1.267-cp310-abi3-macosx_11_0_arm64.whl
Size	29.5 MB
Tags	CPython 3.10 abi3 macOS 11.0+ ARM64
SHA-256 checksum How to use checksums	`e3140de0311b3af00260a249fd43a4b883c08a13e7e08277ac0d3128a65d4b36`
BLAKE2b-256 checksum How to use checksums	`1b58968a3e9829b0ab92f3fb9ed49f8b3076b1f51381c5468684ca65325d5c5f`
Upload date	Sep 9, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`

Release files / gamfit-0.1.267-cp310-abi3-macosx_10_12_x86_64.whl

Download URL	gamfit-0.1.267-cp310-abi3-macosx_10_12_x86_64.whl
Size	32.3 MB
Tags	CPython 3.10 abi3 macOS 10.12+ x86-64
SHA-256 checksum How to use checksums	`4e4b7185b087e271bb3af69542a68c1210c265f8135445e3d26890d978cefc28`
BLAKE2b-256 checksum How to use checksums	`ed11ffa657b238767f3b85580e52929ec7fe04b3cc2a700e68245968210cc5ad`
Upload date	Sep 9, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`