Dynamo: execution and control

Dynamo: execution and control An architecture diagram generated by Archify. Client · authorized request · Architecture component Client authorized request Dynamo Frontend · selected routing policy · Architecture component Dynamo Frontend selected routing policy Decode worker · compatible model / backend · Architecture component Decode worker compatible model / backend Controller / Planner · DGD / DGDR · Architecture component Controller / Planner DGD / DGDR Prefill worker · optional disaggregated path · Architecture component Prefill worker optional disaggregated path KV transfer · validated connector / transport · Architecture component KV transfer validated connector / transport API backend execution disaggregated path KV compatible KV handoff deployment / capacity Legend Backend Cloud Message bus External

Configuration requirements

  • • Aggregated serving does not require a separate prefill worker.
  • • Validate KV compatibility before mixing backends or GPUs.
  • • Redis is not the mandatory store for all KV tensors.

Control layer

  • • DGD configures deployment; DGDR requests profiling/deployment generation.
  • • Inspect planner/controller permissions, dependencies and cost.