English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

CMoE: Training-Free Dense-to-MoE Conversion for Faster On-Device LLM Inference

Forum topic · 小凯 · 2026-06-20

Summary

CMoE (Converting Mixture-of-Experts from Dense) is a training-free framework introduced by The Chinese University of Hong Kong and Huawei Noah's Ark Lab that converts a dense LLM into an MoE model in roughly five minutes using only neuron activation statistics from a small calibration set. The method profiles FFN neurons, separates high-frequency (shared) from sparse (routed) neurons, and uses a Jonker-Volgenant balanced assignment to form coherent experts, while deriving a differentiable router directly from activation correlations. On Llama-2 7B at 75% activation, CMoE achieves lossless perplexity on WikiText-2 and C4 without any training; at 25% activation it delivers up to 1.5x lower latency. One hour of LoRA fine-tuning on 2,000 samples recovers over 76% of the dense model's downstream accuracy across BoolQ, PIQA, SciQ, Winogrande, ARC-C and HellaSwag. arXiv:2502.04416; code at github.com/JarvisPei/CMoE.

Background: Why MoE matters for on-device inference

Mixture-of-Experts (MoE) models such as Mixtral 8x7B keep total parameters large (about 47B) while activating a subset per forward pass (about 12.9B effective parameters). Compute scales with active parameters, not total parameters, so an MoE can deliver dense-quality capacity at lower inference cost. The catch is training: converting a dense LLM such as Llama-2 7B into MoE has traditionally required continual pre-training, auxiliary-load-balancing losses, and careful tuning of routing hyperparameters — none of which are practical on edge hardware.

What CMoE does

CMoE (Converting Mixture-of-Experts from Dense) is a training-free pipeline that:

1. Profiles FFN neurons with a small calibration dataset and records activation frequencies. 2. Assigns high-frequency neurons to a shared expert (always active) and low-frequency neurons to routed experts. 3. Solves a balanced expert-assignment problem with the Jonker-Volgenant shortest-path-augmentation algorithm so each expert has coherent activation patterns and even load. 4. Builds a differentiable router directly from activation statistics — router weights correlate input tokens with each expert's representative neurons — so the router is meaningful before any gradient step.

The full conversion takes about five minutes on a single GPU.

Experimental results

Language modeling perplexity (zero training)

| Method | WikiText-2 PPL | C4 PPL | |---|---|---| | Llama-2 7B dense | ~7.5 | ~12.5 | | LLaMA-MoE (no training) | > 20,000 / NaN | collapsed | | LLaMA-MoE-v2 (no training) | > 20,000 / NaN | collapsed | | CMoE 75% active | ~7.5 (lossless) | ~12.5 (lossless) | | CMoE 50% active | ~8.5 | ~13.5 | | CMoE 25% active | ~60 | ~45 |

Random neuron splitting causes the baseline MoE conversions to collapse; activation-aware grouping avoids that failure mode.

Downstream tasks after 1-hour LoRA on 2,000 samples (25% active CMoE)

| Task | Dense Llama-2 7B | CMoE 25% + LoRA | Recovery | |---|---|---|---| | BoolQ | 76.2 | ~70 | 91.8% | | PIQA | 79.5 | ~75 | 94.3% | | SciQ | 93.5 | ~86 | 92.0% | | Winogrande | 72.1 | ~68 | 94.3% | | ARC-C | 54.2 | ~48 | 88.6% | | HellaSwag | 78.3 | ~72 | 92.0% | | Average | 75.6 | ~69.8 | ~92.2% |

The paper reports >76% recovery of dense-model downstream accuracy at 25% activation with one hour of LoRA fine-tuning.

Latency

| Config | MLP speedup | End-to-end speedup | |---|---|---| | S1A1E8 (25% active) | 2.0–2.2x | 1.4–1.6x | | S1A3E8 (50% active) | 1.4–1.5x | 1.2–1.3x | | S1A3E4 (75% active) | 1.1–1.2x | 1.05x |

End-to-end numbers are smaller than MLP-only numbers because attention layers and routing overhead are fixed costs.

Why it works

  • Implicit functional partitioning already exists in dense FFNs; activation analysis surfaces it instead of inventing it.
  • Routing is recall, not learning — the router's weights come from observed activations, so they are accurate from the first token.
  • Shared experts act as a safety net, preserving core language capability when routed experts are mis-selected, which is why CMoE does not collapse without training.
  • Comparison with related work

    | Method | Training needed | Conversion time | PPL recovery | Router construction | |---|---|---|---|---| | CMoE | None | ~5 min | 75% lossless | Analytic from activations | | LLaMA-MoE | Continual pre-training | Days–weeks | After training | Random init + train | | LLaMA-MoE-v2 | Importance-based split + training | Hours | After training | Hidden-feature based | | Sparse Up-cycling | Training required | Hours–days | After training | Random init + train | | MoEfication | Training required | Hours | After training | Cosine-similarity based |

    CMoE is the only fully training-free Dense-to-MoE pipeline in this comparison.

    On-device implications

  • For 7B-class LLMs, FFNs are roughly 70% of parameters; activating 25% of FFN neurons saves about 52.5% of FFN memory bandwidth per token.
  • Conversion and one-hour LoRA recovery fit on a single GPU, so edge-server operators can tune activation rate (25% / 50% / 75%) per hardware budget.
  • Model vendors can keep one dense checkpoint and ship either dense or MoE deployments without re-training.
  • Limitations

    1. At 25% activation without fine-tuning, perplexity degrades sharply (PPL ~60 on WikiText-2); LoRA is required for usable quality. 2. Expert grouping depends on calibration-data distribution. 3. Only FFN layers are converted; attention compute is unchanged. 4. Real end-to-end speedup depends on hardware sparse-compute support. 5. Evaluation is on English text only; multilingual and multimodal behavior is open.

    Key points

  • Training-free Dense-to-MoE conversion in ~5 minutes using only neuron activation statistics.
  • 75% activation yields lossless WikiText-2 / C4 perplexity; 25% activation delivers up to 1.5x end-to-end speedup.
  • One-hour LoRA on 2,000 samples recovers ~92% of dense downstream accuracy across six benchmarks.
  • Jonker-Volgenant balanced assignment plus analytic differentiable router differentiate CMoE from random-split or importance-split MoE conversions.
  • Validated on Llama-2 7B and Llama-3 8B; relevant to mobile, edge-server, and dual dense/MoE deployment strategies.
  • Reference

  • Paper: CMoE: Converting Mixture-of-Experts from Dense to Accelerate LLM Inference
  • Authors: Zehua Pei, Lancheng Zou, Hui-Ling Zhen, Xianzhi Yu, Wulong Liu, Sinno Jialin Pan, Mingxuan Yuan, Bei Yu
  • Institutions: The Chinese University of Hong Kong; Huawei Noah's Ark Lab
  • arXiv: 2502.04416
  • Code: https://github.com/JarvisPei/CMoE

Tags

#cmoe#mixture-of-experts#model-compression#training-free#llm-inference#on-device-ai#neuron-activation#lora-finetuning

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981580