Decide when to choose vLLM, SGLang, TensorRT-LLM, or a custom stack.

Data parallelism replicates the model, tensor parallelism splits individual operations, pipeline parallelism splits layers, sequence parallelism splits token work, and expert parallelism distributes MoE experts. Real systems combine them to fit memory while minimizing communication bubbles.

Make interfaces and ownership explicit; add versioning, access control, monitoring, canary rollout, rollback, and a plan for delayed labels or human review.