NCA-ADS - Introductory MLOps Practices (10% of the exam) - Section 6.3
Manage models, including saving and loading.
Save and load trained models using serialisation formats appropriate to the library, such as pickle, joblib, or framework-native checkpoints. Apply versioning practices so that the model artefact, its training data snapshot, and its hyperparameter configuration remain linked.
An engineer saves a fitted cuML model on a GPU workstation using joblib and ships the file to a colleague who tries to load it. Loading fails. The colleague has the same cuML and CUDA library versions installed but is working on a machine that has no GPU. Which explanation is the most likely root cause of the load failure?
AA cuML model holds GPU-resident state and depends on the CUDA runtime and a device at load time, so deserialising it on a machine with no GPU cannot restore that device state.check_circle Correct
BThe joblib file is corrupted in transit, because joblib archives cannot be transferred between machines without first re-exporting them to a plain text format.
Cjoblib strips the model's learned parameters during saving, so any loaded cuML model must be refitted before use regardless of the hardware on the loading machine.
DMatching cuML versions are insufficient, because the file can only load when the colleague also has the identical CPU model and clock speed as the machine that saved it.
Loading a GPU-native cuML model requires a CUDA device at load time, so saved models are not portable to GPU-less hosts. cuML estimators store device-resident arrays and rely on the CUDA runtime to reconstruct that state when deserialised. A machine without a GPU cannot allocate the device memory the saved object expects, so the load fails even when library versions match.
Why A is correct: Correct: cuML estimators are GPU-native and their persisted state assumes a CUDA device is available; restoring the object on a host with no GPU cannot reconstruct the device-side arrays, so the load fails.
Why B is wrong: Tempting because transfer corruption is a real failure mode, but joblib produces ordinary binary files that transfer like any other file; the relevant difference here is the absence of a GPU, not the transport mechanism.
Why C is wrong: Tempting because some workflows do require refitting, but joblib serialises the fitted parameters faithfully; it does not strip learned state, and a GPU-equipped machine would load the same file without needing a refit.
Why D is wrong: Tempting because environment matching matters for reproducibility, but deserialisation does not depend on identical CPU hardware; the decisive missing component here is a CUDA-capable GPU, not a matching processor.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-ADS bank for this domain.
The raw training dataset, because a saved model cannot generate predictions unless the original training rows travel alongside it
errorWhy it is wrong: Tempting because reproducibility discussions stress keeping training data, but a serialised model already contains learned parameters and does not need the raw training rows to score new data, so this is not the cause.
The registry stored the model in ONNX format, which strips XGBoost-specific sub-module dependencies on export
errorWhy it is wrong: ONNX is a separate export format used to achieve cross-framework portability; the question describes loading from a model registry without specifying ONNX, and ONNX models load through the onnxruntime package rather than XGBoost sub-modules, so this explanation does not fit the error described
A scatter plot of the training data, because the loader inspects the saved plot image to reconstruct the pipeline's fitted parameters before inference can begin.
errorWhy it is wrong: Tempting because exploratory plots are common project artefacts, but a plot carries no fitted state and is never read by a loader; the serialised pipeline already contains the learned parameters.