View a markdown version of this page

ProductionVariantRoutingConfig - Amazon SageMaker

ProductionVariantRoutingConfig

Settings that control how the endpoint routes incoming traffic to the instances that the endpoint hosts.

Contents

RoutingStrategy

Sets how the endpoint routes incoming traffic:

  • LEAST_OUTSTANDING_REQUESTS: The endpoint routes requests to the specific instances that have more capacity to process them.

  • RANDOM: The endpoint routes each request to a randomly chosen instance.

  • PREFIX_AWARE: The endpoint routes requests that share the same prompt prefix to the same instance and, when that instance hosts more than one copy of the invoked inference component, to the same copy. When the number of in-flight requests on the selected instance or copy reaches the ConcurrencyThreshold value, the endpoint routes the request to another copy on that instance or to another instance. The endpoint applies prefix-aware routing only when two or more instances can serve the request.

  • KV_AWARE: The endpoint routes each request to an instance that already holds the start of the request's prompt in its key-value (KV) cache, and weighs the length of that match against how busy each instance is. On an endpoint with inference components, the endpoint chooses among the copies of the invoked inference component in the same way. When the endpoint can't find an instance or copy that holds the start of the prompt and isn't too busy, it chooses an instance as it does with LEAST_OUTSTANDING_REQUESTS. KV_AWARE works only with supported containers, such as the SageMaker AI vLLM Deep Learning Container.

Type: String

Valid Values: LEAST_OUTSTANDING_REQUESTS | RANDOM | PREFIX_AWARE | KV_AWARE

Required: Yes

PrefixAwareRoutingConfig

The configuration for prefix-aware routing. Specify this parameter only when you set RoutingStrategy to PREFIX_AWARE.

Type: PrefixAwareRoutingConfig object

Required: No

See Also

For more information about using this API in one of the language-specific AWS SDKs, see the following: