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

arXiv: https://arxiv.org/abs/2608.24938 · HTML: https://mlsanity.com/paper/2608.24938/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Efficient MoE training and inference systems, one of 12 directions in Mixture-of-experts.
- Builds on 17 of its 27 references that are in the map.
- First posted on arXiv 24 Aug 2026.

## Placed in

- Mixture-of-experts / [Efficient MoE training and inference systems](https://mlsanity.com/moe/efficient-moe-training-and-inference-systems/)

## Builds on (most cited first, 17 of 17)

- [Efficient Memory Management for Large Language Model Serving with PagedAttention](https://mlsanity.com/paper/2309.06180/) (2023)
- [Measuring Mathematical Problem Solving With the MATH Dataset](https://mlsanity.com/paper/2103.03874/) (2021)
- [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://mlsanity.com/paper/2311.12022/) (2023)
- [LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code](https://mlsanity.com/paper/2403.07974/) (2024)
- [MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark](https://mlsanity.com/paper/2406.01574/) (2024)
- [Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation](https://mlsanity.com/paper/2305.01210/) (2023)
- [DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model](https://mlsanity.com/paper/2405.04434/) (2024)
- [DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence](https://mlsanity.com/paper/2606.19348/) (2026)
- [Merge, Then Compress: Demystify Efficient SMoE with Hints from Its Routing Policy](https://mlsanity.com/paper/2310.01334/) (2023)
- [Harder Tasks Need More Experts: Dynamic Routing in MoE Models](https://mlsanity.com/paper/2403.07652/) (2024)
- [REAP the Experts: Why Pruning Prevails for One-Shot MoE compression](https://mlsanity.com/paper/2510.13999/) (2025)
- [Sub-MoE: Efficient Mixture-of-Expert LLMs Compression via Subspace Expert Merging](https://mlsanity.com/paper/2506.23266/) (2025)
- [Lynx: Enabling Efficient MoE Inference through Dynamic Batch-Aware Expert Selection](https://mlsanity.com/paper/2411.08982/) (2024)
- [Pre3: Enabling Deterministic Pushdown Automata for Faster Structured LLM Generation](https://mlsanity.com/paper/2506.03887/) (2025)
- [MoDES: Accelerating Mixture-of-Experts Multimodal Large Language Models via Dynamic Expert Skipping](https://mlsanity.com/paper/2511.15690/) (2025)
- [TokenFlow: Responsive LLM Text Streaming Serving under Request Burst via Preemptive Scheduling](https://mlsanity.com/paper/2510.02758/) (2025)
- [REAM: Merging Improves Pruning of Experts in LLMs](https://mlsanity.com/paper/2604.04356/) (2026)

## Built on by (most cited first, 2 of 2)

- [SlimWise: Decoupling Expert Pruning Across Prefill and Decode for Efficient MoE Serving](https://mlsanity.com/paper/2609.34117/) (2026)
- [BASE: Batch-Aware Selection of Experts Using Predicted Removal Error for Efficient MoE Decoding](https://mlsanity.com/paper/2609.36222/) (2026)

## Abstract (arXiv)

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.