Auto Deploy with DGDR

Why it matters

Choosing a serving layout by hand means guessing parallelism and replica counts. A DGDR states the service targets instead and lets a profiler choose the layout, so the cost of each target becomes visible before deployment.

What to practise

Write a DynamoGraphDeploymentRequest with model, a workload shape (isl, osl, requestRate) and sla targets (ttft, itl). Compare searchStrategy: rapid, which simulates, with thorough, which benchmarks real GPUs. Set autoApply: false so the request stops at Ready with the generated DynamoGraphDeployment for review, and follow the phases from Pending through Profiling to Deployed. Watch how unreachable targets push the profiler toward its most expensive layout.

How it fits

This puts 4.5.1’s latency-target sizing into a deployment request. The optional features.planner is the runtime autoscaling of 4.4.3; weigh it against the fixed headroom 4.5.3 funds for failures.