AFD-Ledger cuts MoE deployment search costs by 83.5%

Serving large Mixture-of-Experts (MoE) models efficiently is a growing challenge, and Attention-Feed-Forward Network Disaggregation (AFD) has emerged as a promising architecture. However, teams face a critical question: under the same model, workload, latency target, and hardware budget, does AFD actually outperform the best collocated deployment? Answering this requires jointly optimizing hardware assignment and deployment organization, making exhaustive provisioning prohibitively expensive. Now, a team of researchers has built AFD-Ledger, an offline analytical provisioning system designed to settle this decision quickly and accurately.

AFD-Ledger works by independently provisioning both AFD and collocated deployments using an analytical execution model and an evaluation-bounded hardware search. The system avoids brute-force exploration: across deployment spaces where exhaustive provisioning is feasible, AFD-Ledger cuts the number of complete deployment evaluations by 68.8%–83.5% while still recovering the globally optimal deployment. This makes provisioning tractable at scale, a key hurdle for real-world MoE serving. The researchers validated the system on three physical LongCat 2.0 deployments, where it preserved the correct architecture decision while predicting AFD-to-collocated throughput within 6.6%–9.6% of actual measurement.

The paper also delivers several practical insights. Homogeneous AFD improves fixed-budget throughput in only a minority of studied settings, meaning disaggregation isn't automatically the winner. Heterogeneous AFD, on the other hand, requires deployment-level hardware complementarity rather than simple rule-of-thumb device selection. Additionally, role-specific hardware upgrades matter mainly when they enable better deployment organizations by crossing capability–price boundaries. For engineers building MoE serving infrastructure, AFD-Ledger offers a validated framework to answer the AFD-vs-collocated question with measurable confidence, not guesswork.