Multi-tenant model serving

Design an online model-serving platform that hosts thousands of heterogeneous models (small CPU classifiers up to large GPU transformers) for many internal teams, with per-model latency SLOs from 10ms to 500ms and bursty traffic. The platform must support safe rollouts, multi-version traffic splitting, autoscaling to zero for cold models, and not let one noisy tenant starve another. Cover the serving architecture, how you schedule across CPU/GPU, and the rollout/observability story.