Docs

silvic trains and serves LightGBM-compatible models through its Rust crates and Python package. The ray extra adds dependencies for distributed training with silvic.ray:

pip install silvic
pip install "silvic[ray]"

Training accepts LightGBM parameter names:

import silvic
import silvic.serve

params = {"objective": "lambdarank", "num_leaves": 53, "device_type": "auto"}
train_set = silvic.Dataset(
    X, label=y, group=query_sizes, categorical_feature=[131, 132]
)
booster = silvic.train(params, train_set, num_boost_round=1500)
scores = booster.predict(X_test)
fast_scores = silvic.serve.Predictor(booster).predict(X_test)

Requirements

• Python 3.10+
• Source builds Rust 1.94.1+
• Metal macOS · Apple silicon
• CUDA Linux · NVIDIA driver + NVRTC (CUDA 12)
• Neuron Linux · Neuron driver/runtime + compiler

Package and crates

silvic-python

The silvic Python package: LightGBM’s Python API, scikit-learn estimators, and training across Ray actors.

silvic-core

Parameters, binning, objectives, metrics, leaf-wise trees on integer histograms, LightGBM model text, and the CPU backend.

silvic-gpu

The Metal and CUDA backends, from one kernel source compiled at run time.

silvic-neuron

AWS Neuron training and explicit float32 serving, with NKI kernels and the native runtime.

silvic-net

TCP ring collectives for distributed training.

silvic-serve

Exact scoring of LightGBM models on CPU SIMD lanes or explicitly selected NVIDIA CUDA devices.

silvic-triton

A Triton Inference Server backend serving LightGBM models with silvic-serve, where the FIL backend’s configs serve unchanged but for its name.