Ml training optimization

Ml Training Optimization

Overview

Use this skill to improve training throughput and cost while preserving model quality and stability.

Scope Boundaries

• Use this skill when the task matches the trigger condition described in description.
• Do not use this skill when the primary task falls outside this skill's domain.

Shared References

• Convergence and budget rules:
  • references/convergence-and-budget-rules.md

Templates And Assets

• Training optimization plan:
  • assets/training-optimization-plan-template.md

Inputs To Gather

• Baseline runtime/cost/convergence behavior.
• Resource constraints and training budget.
• Quality guardrails to prevent regressions.
• Candidate optimization levers (data, algorithm, infra).

Deliverables

• Optimization plan with prioritized interventions.
• Resource and convergence validation results.
• Cost/quality trade-off report.

Workflow

1. Capture baseline and bottlenecks in assets/training-optimization-plan-template.md.
2. Apply references/convergence-and-budget-rules.md to bound risk.
3. Run targeted optimizations with controlled experiments.
4. Validate quality guardrails and budget impact.
5. Publish adopted changes and rollback criteria.

Quality Standard

• Optimization decisions preserve target quality.
• Convergence behavior remains stable.
• Cost and runtime improvements are measurable.

Failure Conditions

• Stop when optimization degrades quality beyond guardrails.
• Stop when instability increases despite speed gains.
• Escalate when budget constraints remain unmet.