From the source
Amazon published a walkthrough for deploying LiteLLM with
NVIDIA's open source NeMo Switchyard routing plugin on Amazon Bedrock, so teams can send each coding agent request to a different model.
- The post argues that price per token does not reveal what a whole coding task costs, because the total depends on how many turns a model needs, how much history each turn replays, and how often a task is retried or never finishes.
- It recommends judging routing by cost per solved task, the total cost of runs divided by the number of runs whose tests pass.
- The sample deploys LiteLLM and the Switchyard plugin in one container on Amazon ECS on AWS Fargate in private subnets, with AWS WAF filtering requests and an Application Load Balancer forwarding only three paths.
- The integration supports Stage routing, which reads recent tool activity and results in the request history and picks between two candidate models, and Random routing, which picks from configured models with optional weights for traffic splits and A/B comparisons.
- Neither strategy calls a judge model, and classifier-backed strategies such as Task and Composite routing need intermediate model calls the integration does not support.
- The sample installs switchyard-server 0.3.0 and runs the published LiteLLM v1.102.0 image, and NVIDIA documents switchyard-server 0.3.0 as a demo and evaluation component whose configuration can change between releases.






