feature penalty not supported in CUDA

[gnahz04]

Description

It appears that feature_penalty parameter is not activated under CUDA.

Reproducible example

In the following example, I expected CUDA version, similar to the CPU version, to also output feature importance to be "0 600" when feature penalty is 0 on the first feature.

import numpy as np
import lightgbm as lgb

def run_me(cpu_or_cuda='cpu'):
    # 1. Create a synthetic binary classification dataset
    np.random.seed(42)
    N = 1000
    X = np.zeros((N, 2))
    y = np.zeros(N)

    # Feature 0 is a strong predictor (we'll make it correlate  strongly with y)
    # Feature 1 will be a weak predictor (mostly noise)
    X[:, 0] = np.random.randint(0, 2, size=N)        # binary feature
    X[:, 1] = np.random.normal(size=N)               # continuous noise feature
    y = X[:, 0].copy()                               # make label equal to feature 0 (perfect predictor)
    # Add a little randomness to labels to avoid perfect separability
    y[:50] = 1 - y[:50]  # flip first 50 labels to introduce some noise

    # Convert to LightGBM Dataset
    dtrain = lgb.Dataset(X, label=y)

    # 2. Train a baseline model with no feature penalty (using CPU for demo; use device_type='cuda' if available)
    params_no_penalty = {
        "objective": "binary",
        "metric": "auc",
        "device_type": cpu_or_cuda,        # would be "cuda" for GPU with CUDA support
        "verbose": -1,
        "seed": 42
    }
    model_no_penalty = lgb.train(params_no_penalty, dtrain, num_boost_round=20)

    # 3. Train a model with feature penalty to down-weight feature 0
    # We set feature0's penalty to 0 (no gain) and feature1's penalty to 1 (no penalty) for demonstration
    params_with_penalty = {
        "objective": "binary",
        "metric": "auc",
        "device_type": cpu_or_cuda,        # use "cuda" for actual GPU run
        "feature_penalty": [0.0, 1.0],  # penalize feature 0 heavily, feature 1 normal
        "verbose": -1,
        "seed": 42
    }
    model_with_penalty = lgb.train(params_with_penalty, dtrain, num_boost_round=20)

    # 4. Compare feature importances
    print(f"{cpu_or_cuda=}")
    print("Feature importances (no penalty):", model_no_penalty.feature_importance(importance_type="split"))
    print("Feature importances (with penalty):", model_with_penalty.feature_importance(importance_type="split"))

run_me('cpu')
run_me('cuda')

--->

cpu_or_cuda='cpu'
Feature importances (no penalty): [ 20 580]
Feature importances (with penalty): [  0 600]
cpu_or_cuda='cuda'
Feature importances (no penalty): [ 20 580]
Feature importances (with penalty): [ 20 580]

Environment info

LightGBM version or commit hash:
lightgbm 4.6.0.99
Command(s) you used to install LightGBM

pip install -U lightgbm

Additional Comments