gafime 0.4.7

GPU Accelerated Feature Interaction Mining Engine

GAFIME

GPU-Accelerated Feature Interaction Mining Engine.

GAFIME is a native feature interaction mining engine for tabular and structured machine-learning workflows. Python owns the public API; C++ Core, Rust scheduling, CUDA, Metal, and ROCm/HIP own the hot execution paths.

The engine is built for workloads where interaction candidates, threshold-style regions, and temporal transforms become too expensive to search with ordinary Python loops or model-by-model trial code.

Install

Core package:

pip install gafime

Optional Python integrations:

pip install "gafime[sklearn]"
pip install "gafime[bench]"

Vendor GPU payloads are explicit in the v0.4.7 distribution design. Pip can select wheels by Python, ABI, OS, and CPU architecture, but not by local GPU vendor. CUDA and ROCm payloads therefore need explicit package selection once the split payload packages are published:

pip install "gafime[cuda]"
pip install "gafime[rocm]"  # Linux x86_64 only in v0.4.7

The extras install the separate PyPI payload projects gafime-cuda and gafime-rocm for the same GAFIME release.

Apple Silicon Metal follows the macOS arm64 wheel/platform path.

Detailed install and backend policy:

• docs/backend-selection.md
• BUILD.md

Basic Usage

from gafime import ComputeBudget, EngineConfig, GafimeEngine

config = EngineConfig(
    backend="auto",
    metric_names=("pearson", "r2"),
    budget=ComputeBudget(max_comb_size=2),
)

report = GafimeEngine(config).analyze(X, y, feature_names=feature_names)
print(report.backend)
print(report.interactions[:5])

Generate the reference notebook:

gafime --init

Advanced API tour notebooks are also kept in the repository:

• gafime_full_api_reference_notebook.ipynb

Candidate Families

GAFIME supports:

• continuous interaction candidates,
• soft discrete thresholds, intervals, value-gated thresholds, rectangles, and value-in-rectangle candidates,
• explicit time-series candidates such as lag, delta, velocity, acceleration, rolling mean, rolling standard deviation, and rolling sum,
• scikit-learn transformer integration through gafime.sklearn.GafimeSelector,
• large-file streaming helpers through GafimeStreamer.

GPU discrete feature engineering is soft/vectorized only. Hard discrete mode is a CPU/Core behavior and raises a clear error on GPU backends.

Backend Policy

backend="auto" resolves native backends with platform-aware priority:

• macOS arm64: metal -> core
• Linux/Windows x86_64 with CUDA payloads: cuda -> core
• Linux x86_64 with ROCm payloads: rocm -> core
• Linux/Windows ARM64: core

GAFIME does not initialize every GPU runtime during auto resolution. It uses the installed platform payload selected by the backend resolver, or core when no supported GPU payload is selected.

backend="gpu" is deprecated because it is ambiguous across CUDA, ROCm, and Metal. Use auto, cuda, rocm, metal, or core.

Native Reports

Reports are structured Python objects. Read properties such as:

• report.interactions
• report.decision
• report.backend
• report.warnings

DiagnosticReport.to_dict() remains only as a deprecated export convenience. It should not be used as a runtime data-flow path.

Developer Docker Images

Docker files in this repository are development environments, not distribution images. Normal users should install GAFIME from PyPI wheels.

Available source-build containers:

docker compose run --build gafime-cuda-dev
docker compose run --build gafime-core-smoke

The CUDA development image includes the CUDA toolkit, compiler toolchain, Rust, CMake, GAFIME development/benchmark/scikit-learn dependencies, and the locally staged gafime-cuda payload by default. Extra workstation packages can be added with the EXTRA_PIP_PACKAGES Docker build argument. The Core smoke image is a smaller CPU-native source-build check.

Docker details:

• CONTRIBUTING.md

Project References

• docs/releases/v0.4.7.md
• docs/v0.4.7-rocm-native-backend.md
• docs/backend-selection.md
• gafime_full_api_reference_notebook.ipynb
• CONTRIBUTING.md

Contact

Maintainer: Hamza Usta

Email: [e-mail]