Why it matters

Ray is popular ML orchestration. Understanding shapes design.

Advertisement

The architecture

Head node + workers.

Actors + tasks.

Placement groups for GPU.

Ray architectureRay headcoordinatorWorkersrun tasks + actorsPlacement groupGPU allocationRay Train orchestrates PyTorch DDP; Ray Serve for LLM serving
Ray.
Advertisement

How it works end to end

Ray Train: distributed training.

Ray Serve: model serving.

Ray Tune: hyperparameter.

KubeRay: run Ray on K8s.