train on CUDA and load on CPU but got an EXC_BAD_ACCESS crash

[warscape]

trained my XGBoost model on CUDA, saved in json format and then tried to load model on macOS, got an EXC_BAD_ACCESS crash.
The python versions on both platforms are 3.10.x, and XGBoost versions are both 3.0.3.

Full crash log attached.

Python-2026-01-21-163640.log

[warscape]

Tried to modify model params and save another copy in deprecated format, it can be successfully loaded on macOS but will fail at the predict call.

new_params = {'device': 'cpu', 'tree_method': 'hist', 'predictor': 'cpu_predictor'} model = xgb.XGBClassifier(params =new_params) model.load_model(_modelPath) model.set_params(**new_params) model.save_model(_cpuModelPath)

Here are crash logs:

Python-2026-01-21-110924.log
Python-2026-01-21-095943.log

[trivialfis]

I couldn't reproduce the issue by loading the GPU-trained model with a CPU-only XGBoost build. By default, the output from save_model doesn't contain any device-related information (you can try to search "cuda" or "gpu" in the JSON file), I suspect the issue is caused by something else.

'predictor': 'cpu_predictor'

XGBoost 3.x doesn't have a predictor parameter.

@hcho3 could you please take a look with a Mac when you are available?

[warscape]

@trivialfis thanks for the quick response!
The predictor parameter has no effect, - I got a warning in console log saying this is an unknown parameter and is ignored.

The xgboost I used on macOS is just xgboost, not xgboost-cpu, and it works well when doing training with CPU.
Do you think I should try xgboost-cpu?

[trivialfis]

For macos, xgboost is the same as xgboost-cpu, since there's no CUDA support there.

I'm just trying to reproduce the issue (but haven't been able to).

[warscape]

@trivialfis @hcho3 any updates on this issue? 🙏

[trivialfis]

I will try set up a temporary CI job and see if we can reproduce there.