model-serving-deployment

name	model-serving-deployment
description	Deploy and operate model inference with explicit batching, concurrency, memory, routing, health, rollout, fallback, and compatibility behavior.

Model Serving Deployment

Use when this procedure is the primary professional method needed for the assignment.

Procedure

1. Confirm the decision or outcome this work must support, its scope, owner, constraints, and definition of success.
2. Establish the evidence baseline using model size, tokenizer/config, traffic shape, GPU/CPU capacity, latency SLO, and serving runtime. Do not fill material gaps with assumptions when they can change the result.
3. Benchmark representative load, size replicas/accelerators, configure timeouts/batching, expose health/metrics, canary rollout, and validate failure/fallback.
4. Exercise realistic edge, failure, transition, or exception cases that could invalidate the result; record unresolved uncertainty explicitly.
5. Validate the output against the original outcome and any neighboring professional contracts so this skill does not silently absorb another specialist's authority.
6. Record the resulting artifact, measurements, decisions, provenance, and handoff information needed for another owner to reproduce or continue the work.

Quality gate

Serving meets representative quality/latency/throughput goals and has a proven rollback/fallback path.