ExFold: Unified Expert Folding for Training-Free MoE Prefill-Decode Acceleration

Researchers propose ExFold, a training-free expert-folding framework that accelerates both prefill and decode phases of Mixture-of-Experts (MoE) inference by projecting excluded expert contributions onto retained experts using calibrated scalar projectors. Implemented as a plug-and-play plugin in vLLM, ExFold delivers up to 1.41x time-to-first-token (TTFT) and 2.45x time-per-output-token (TPOT) speedups while retaining about 99% of original average quality.

arXiv:2608.24938v1 Announce Type: new Abstract: Mixture-of-Experts MoE models scale capacity for strong quality while keeping per-token compute bounded through sparse expert activation. Yet low-latency MoE serving is increasingly challenging, because it spans two inference phases with fundamentally different bottlenecks: prefill is dominated by token-wise expert computation, whereas decode is constrained by memory traffic from the batch-wise activated expert set. However, existing training-free acceleration methods optimize only a single resource proxy, either the experts each token executes or the experts a batch activates, and either discard the excluded experts' contribution or leave it only implicitly approximated. In this paper, we propose ExFold, a unified training-free expert-folding framework for jointly accelerating MoE prefill and decode. ExFold casts both prefill and decode as one budgeted output-approximation problem: execute only a phase-specific constrained expert set while projecting the contribution of budget-excluded experts onto retained experts using calibrated scalar projectors. Motivated by the observation that many expert outputs are directionally aligned but differ in magnitude, ExFold calibrates a pairwise scalar-projector matrix on unlabeled data and uses it at inference time to fold excluded expert contributions into retained experts. Under this view, prefill acceleration becomes token-level Top-K folding, and decode acceleration becomes batch-level expert-pool folding. The two phases differ only in how retained experts are selected, while excluded contributions are recovered by one shared folding mechanism. We implement ExFold as a plug-and-play plugin in vLLM, with a lightweight expert-folding CUDA kernel, delivering up to 1.41x TTFT and 2.45x TPOT speedups while retaining about 99% of the original average quality.