Question 290
A team has corrected this HPE compute solution issue: A research group deploys many high-memory models onto GPUs sized only for the lightest model. Which continuing practice would best prevent the same problem?
Explanation
Correct Answer
Size for simultaneous model residency rather than only one-model tests. This recurring assurance measure is tied to the diagnosed risk: Accelerator memory has not been sized for concurrent resident models and serving overhead. For the active incident, the corrective action remains: Profile combined model memory needs and select suitable supported capacity and placement.
Incorrect Answers
This control protects against a condition in which the hardware choice omits the accelerator resources required by the selected workload. The issue to prevent here is: Accelerator memory has not been sized for concurrent resident models and serving overhead. Size for simultaneous model residency rather than only one-model tests.
Accelerator memory has not been sized for concurrent resident models and serving overhead. The proposed safeguard is intended for another failure mechanism: NIM availability and support depend on the model, container and approved runtime combination. Size for simultaneous model residency rather than only one-model tests.
The prevention measure is relevant to this condition: The physical and management security boundary is not preserved in deployment. Here the recurring weakness is: Accelerator memory has not been sized for concurrent resident models and serving overhead. Size for simultaneous model residency rather than only one-model tests.