Inference Optimization

---
name: inference-optimization
description: Optimize inference using profiling, precision/quantization, batching, cache, kernel/runtime, and model choices without silently violating quality.
---
# Inference Optimization

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 representative prompts/tasks, quality evals, latency/throughput, memory, hardware, model/runtime versions. Do not fill material gaps with assumptions when they can change the result.
3. Establish quality/performance baseline, profile bottlenecks and memory, test one optimization class at a time, re-run task evals, and record hardware-specific tradeoffs.
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
Performance improvement is measured on representative workload and quality regression stays within explicit acceptance bounds.