SYSTEM Cited by 1 source
Wukong (Meta)¶
Definition¶
Wukong is the Meta Ads internal recommendation / ranking model architecture published in 2024 at arXiv:2403.02545. It is the foundational design underpinning Meta's trajectory toward LLM-scale ads ranking, and the architectural basis for Wukong Turbo, the runtime evolution deployed inside Meta Adaptive Ranking Model (Source: sources/2026-03-31-meta-adaptive-ranking-model-bending-the-inference-scaling-curve).
Architectural primitives (summary from Adaptive Ranking Model¶
post)
- Stackable factorisation machines — layered FM blocks for cross-feature interactions, allowing depth to be stacked for richer modelling.
- Sequence learning — over user behaviour histories, the input that the request- oriented sequence scaling optimisation later amortises across ad candidates.
- Cross-layer attention — attention mechanisms spanning architectural depth, not just within individual blocks.
These three primitives are named in the 2026-03-31 Adaptive Ranking Model post; the full architecture is described in the Wukong paper itself (not ingested into this wiki).
Relationship to Wukong Turbo + Adaptive Ranking Model¶
Wukong is the model architecture. Wukong Turbo is the optimised serving-side runtime that layers no-bias numerical stability, FSDP→DDP parameter delegation, and sparsity-based linear simplification on top — refinements that allow Wukong's depth + parameter count to scale without blowing the sub-second latency budget.
Caveats¶
- Stub — this page summarises only what the 2026-03-31 Meta Adaptive Ranking Model post names about Wukong. The full architecture is described in arXiv:2403.02545, which is not ingested into this wiki.
- Candidate for deeper future ingestion if further Wukong or Meta ads-ranking content surfaces.
Seen in¶
- 2026-03-31 Meta — Meta Adaptive Ranking Model — names Wukong as the architectural basis Wukong Turbo builds on (sources/2026-03-31-meta-adaptive-ranking-model-bending-the-inference-scaling-curve).