MetaAnirudh AgrawalProblem & solution
Design Once, Deploy at Scale: Building a Template-Driven ML Ecosystem
Summary
Meta’s ads ranking ecosystem, processing billions of daily requests, faced a bottleneck where propagating new techniques across a large and fragmented model fleet became harder than the innovation itself. To solve this, they introduced Standard Model Templates (SMT), which standardize model design into a small set of reusable building blocks covering architecture, features, training, and serving. This shifts iteration from per-model to template-level, reducing complexity and enabling multi-model optimization that evaluates techniques across representative models. Over two years, SMT reduced per-model engineering time by 92%, shipped over 5000 technique-model combinations, improved offline model quality by 0.63% at neutral serving capacity, and delivered a 0.86% cumulative online lift covering 70% of models. The approach also provides a structured foundation for agentic ML systems, where agents explore and optimize within bounded template spaces, making automation more effective.
Context
Meta’s ads ranking ecosystem processes billions of daily requests across diverse products and optimization goals, served by a large number of production models. Fragmentation in architectures, features, and settings created high engineering cost, reliability issues, and low predictability for propagating new techniques.
Approach / What changed
Standard Model Templates (SMT) standardize model design into a curated set of reusable building blocks covering architecture, features, training, and serving. Multi-model optimization evaluates candidate techniques across representative models, considering aggregate performance and regression rate, then validates on held-out models before integrating into the next template version.
Takeaways
SMT reduced per-model engineering time by 92% and shipped over 5000 technique-model combinations across Meta’s ads ranking fleet.
Multi-model optimization evaluates techniques across representative models, considering both aggregate performance and regression rate before integration.
SMT provides a structured foundation for agentic ML systems, bounding the search space and enabling more effective automation and exploration.
FAQ
What is a Standard Model Template?
A Standard Model Template is an opinionated, minimal model setup built from shared components such as architecture choices, feature logic, training settings, and serving configs. It groups similar models around a standard design, reducing unnecessary diversity while preserving needed flexibility.
How does multi-model optimization work?
Multi-model optimization evaluates candidate techniques across representative models, searching for configurations that improve aggregate performance while limiting regressions. Techniques that pass validation on held-out models are integrated into the next template version.
What results did Meta achieve with SMT?
Over two years, SMT reduced per-model engineering time by 92%, shipped over 5000 technique-model combinations, improved offline model quality by 0.63% at neutral serving capacity, and delivered a 0.86% cumulative online lift covering 70% of models.
Source and freshness
- Source
- engineering.fb.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.