A SageMaker endpoint running the built-in Image Classification model on an ml.m5.xlarge instance is experiencing unacceptable inference latency, and ModelLatency is the dominant contributor. Latency worsens when multiple users call the endpoint concurrently. Which action will reduce inference latency for these researchers?

Attaching an Amazon Elastic Inference (EI) accelerator is the correct solution because EI accelerators are designed to speed up deep learning inference by providing GPU-powered acceleration at a lower cost than full GPU instances. Since ModelLatency is the dominant contributor and worsens with concurrent users, this indicates the model inference itself is the bottleneck, and EI can directly address this. Switching to an ml.t3 burstable instance is incorrect; burstable instances are generally not suitable for sustained, high-performance workloads like inference and would likely worsen latency. Enabling SageMaker Autopilot is for model training and hyperparameter tuning, not for optimizing inference latency on an already deployed model. Changing to a memory-optimized instance type is unlikely to help significantly as the problem is inference computation, not memory capacity.