Databricks ML Professional Practice Question 30: Workload Type GPU
Databricks Certified Machine Learning Professional · Test 3 · question 10 of 10
A transformer-based model that was trained on GPUs is deployed to a Model Serving endpoint created with the default configuration. Latency per request is an order of magnitude worse than the team measured in their GPU notebook, and it does not improve under light load.
What should they change on the endpoint?
Show answer and explanation
Answer:
workload_type selects the *kind* of compute behind a served model, defaulting to CPU. Deploying a model built for GPU inference onto CPU compute is the standard cause of an order-of-magnitude latency regression that light load does not relieve, because the bottleneck is per-request computation rather than queuing. Setting workload_type to a GPU option (GPU_SMALL, GPU_MEDIUM, MULTIGPU_MEDIUM or GPU_MEDIUM_8) places the served model on GPU compute. workload_type and workload_size are independent and are usually set together.
Option A confuses the two parameters. workload_size controls provisioned concurrency, how many requests can be served at once, so a larger size adds CPU replicas without introducing a single GPU.
Option B addresses idle cost, not compute type, and makes latency worse rather than better by adding a cold start after each idle period.
Option C scales throughput for concurrent traffic. The symptom is explicitly present under light load, so the endpoint is not queue-bound and more concurrency changes nothing.
Reference: Create custom model serving endpoints