Frequently Asked Questions

How does cuml.accel work with cudf.pandas?

cuml.accel provides zero-code-change acceleration for scikit-learn and similar packages, while cudf.pandas does the same for pandas. They can be used together from the command line:

python -m cudf.pandas -m cuml.accel script.py

In Jupyter, load both extensions before importing the libraries they accelerate:

%load_ext cudf.pandas
%load_ext cuml.accel

Can I load a serialized model without cuml.accel?

Models serialized while cuml.accel is active may be loaded without the accelerator. They are then restored as their normal scikit-learn, UMAP, or HDBSCAN counterparts. If loaded while cuml.accel is active, they are restored as accelerated proxy models.

Serialization formats are not guaranteed to remain compatible across cuML or dependency versions. For long-term portability, record the training environment and validate loading and inference in the target environment.

Warning

Only unpickle or deserialize models from trusted sources. Python’s pickle module is not secure; malicious data can execute arbitrary code. See the Python pickle documentation for details.

How do I report a bug?

Open a report in the cuML issue tracker. Include a minimal reproducer, dependency versions, input characteristics, and logs from an info or debug run when possible. Please report exceptions, incorrect fallback, measurably worse model quality, and regressions in end-to-end runtime. Import time is expected to be longer with cuml.accel and should be excluded from runtime comparisons.

For help determining what ran on the GPU and why an operation fell back, see Logging and Profiling. The estimator inventory is in Accelerated Estimator Support, along with detailed estimator-specific conditions.