MLA-C01 ML Model Development Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Data parallelism using SageMaker distributed data parallelism

  Why it's wrong here

  Data parallelism replicates the entire model on every GPU and splits only the training batch, so a model exceeding one GPU's memory still cannot load. It suits models that fit per device with large datasets; model parallelism partitions layers across GPUs to satisfy the memory constraint.

• ✗

  Switch to a smaller instance type and use horizontal scaling

  Why it's wrong here

  A smaller instance reduces per-device GPU memory further, worsening the constraint, and horizontal scaling adds replicas rather than splitting one model. Model parallelism shards layers across GPUs so a model too large for one device fits; smaller instances suit models already fitting comfortably.

• ✗

  Use multiple training jobs with hyperparameter tuning

  Why it's wrong here

  Hyperparameter tuning launches independent jobs that each train a full model copy, so it cannot split one model across GPUs and leaves the memory constraint unresolved. It suits searching configurations for smaller models; model parallelism shards layers across devices when a single model exceeds one GPU's memory.

• ✓

  Model parallelism using SageMaker distributed model parallelism

  Why this is correct

  Model parallelism shards the model's layers and parameters across multiple GPUs, so weights that cannot fit in one GPU's memory are held collectively. This directly satisfies the stem's constraint that the model does not fit in a single GPU, whereas data parallelism would replicate the full model on every device.

Go deeper

Related to this question