I trained an XGBoost model using GPU acceleration with this code:

model = XGBClassifier(n_jobs=-1, tree_method='gpu_hist', gpu_id=0)

I also used scikit-learn's RandomizedSearchCV for hyperparameter tuning:

grid = RandomizedSearchCV(
    model,
    param_distributions=param_dist,
    cv=list(cross_validation(y_train, ids_train)),
    n_iter=25,
    scoring='balanced_accuracy',
    error_score=0,
    verbose=3,
)

After training, I saved the model with:

grid.best_estimator_.save_model(path)

Can I run inference on this model using only a CPU? Thanks a lot!

Absolutely yes! You can run inference on a CPU for your GPU-trained XGBoost model without any issues—here's a breakdown of why and how to do it:

1. Model Format is Hardware-Agnostic #

The GPU acceleration you used during training (tree_method='gpu_hist') only speeds up the tree-building process. The final saved model stores the same core structure (tree splits, leaf weights, etc.) as a model trained on CPU. XGBoost doesn't tie the model file to the training hardware, so it's fully compatible with CPU inference.

2. Loading and Running Inference on CPU #

You don't need any GPU-specific code to load or run the model. Here are two simple ways to do it:

Using XGBClassifier (scikit-learn API)

from xgboost import XGBClassifier

# Load the model—no GPU parameters needed
model = XGBClassifier()
model.load_model(path)

# Run inference on your test data
predictions = model.predict(X_test)

Using XGBoost's Low-Level Booster API

If you prefer the native XGBoost interface:

import xgboost as xgb

# Load the booster
model = xgb.Booster()
model.load_model(path)

# Convert test data to DMatrix (required for the Booster API)
dmatrix_test = xgb.DMatrix(X_test)

# Run inference
predictions = model.predict(dmatrix_test)

3. Optional: Explicitly Enforce CPU Usage #

While CPU is the default when no GPU is detected, you can explicitly set the tree method to a CPU-compatible option if you want to be certain:

model = XGBClassifier(tree_method='hist')  # Explicit CPU tree method
model.load_model(path)

4. Hyperparameter Tuning Doesn't Impact Compatibility #

Your RandomizedSearchCV workflow doesn't change anything here—grid.best_estimator_ is a standard XGBoost model instance, and saving/loading it works the same regardless of how it was trained.

内容的提问来源于stack exchange，提问作者Petrus