LoRA 专家的混合体

Mixture of LoRA experts combines two established ideas. The first is LoRA, low-rank adaptation, introduced by Edward Hu and colleagues in June 2021. Instead of retraining all the weights of a large model, LoRA freezes them and learns a small correction for selected weight matrices, written as the product of two thin matrices. For GPT-3 175B, the LoRA paper reports 10,000 times fewer trainable parameters and three times less GPU memory than full fine-tuning, with no added inference delay once the correction is merged into the weights. The second idea is the mixture of experts (MoE): a model contains several sub-networks, called experts, and a small gating network, or router, chooses which ones to use for each input. Noam Shazeer and colleagues showed in 2017 that sparsely gated MoE layers with up to thousands of experts could raise model capacity more than 1,000-fold with little extra computation.

A mixture of LoRA experts uses LoRA adapters as the experts. Step by step: a pretrained model is frozen; at chosen layers, usually the attention projections or the feed-forward block, several LoRA adapters are attached side by side; a router, typically one small linear layer followed by a softmax, reads the current token's hidden state and outputs a weight for each adapter; each adapter computes its low-rank correction; and the corrections are added to the frozen layer's output in proportion to the router's weights. With sparse routing, only the top one or two adapters are computed for each token. During training only the adapters and router are updated.

The reason to do this rather than train one bigger adapter is interference. When one adapter is trained on a mixture of very different tasks, such as coding, translation and arithmetic, the tasks compete for the same few parameters. Separate experts let different kinds of input use different parameters. LoRAMoE, posted in December 2023 by Shihan Dou and colleagues, targeted a specific version of this problem: scaling up instruction data can damage the world knowledge a model learned in pretraining. It splits its LoRA experts so some focus on preserving existing knowledge and others on new tasks, and reports gains on downstream tasks while retaining stored knowledge.

The approach developed through several papers. In September 2023 Ted Zadouri and colleagues at Cohere, in "Pushing Mixture of Experts to the Limit", built mixtures of lightweight experts including LoRA-based ones and updated less than 1% of an 11-billion-parameter model, outperforming standard parameter-efficient methods and matching full fine-tuning on instruction tuning. LoraHub, posted in July 2023 by Chengsong Huang and colleagues, took a different route: it combines LoRA modules already trained on separate tasks, choosing fixed mixing weights from a few examples of a new task with a gradient-free search. MixLoRA, from April 2024, added top-k routed LoRA experts inside feed-forward blocks plus separate LoRA adapters on attention, with a load-balancing loss so the router does not overuse a few experts. In April 2024 Xun Wu, Shaohan Huang and Furu Wei described Mixture of LoRA Experts (MoLE), which learns gating over existing trained LoRAs and was tested in both language and vision-and-language settings.

Compared with the alternatives, the trade-offs are clear. Full fine-tuning updates every weight and adapts best, but needs far more memory and produces a full copy of the model per task. A single LoRA adapter is cheapest to train and serve but can struggle with many mixed tasks. Merging several LoRAs by simple weighted addition, a common shortcut, is free at inference but, according to the MoLE paper, can weaken the base model's generation and blur each LoRA's distinct behavior. Keeping separate adapters and switching between them by hand works when the task is known in advance, but requires an outside decision about which adapter to use. A learned router makes that choice automatically, per token, at the cost of extra machinery.

The approach has real limitations. Routers can collapse, sending almost all tokens to one or two experts and leaving the others untrained, which is why methods such as MixLoRA add load-balancing losses. Input-dependent routing removes one of LoRA's main advantages: because the mixture changes token by token, the adapters cannot be merged into the base weights ahead of time, so inference does extra work. Batching requests that use different adapters complicates serving, which is the problem S-LoRA addresses. Reported gains are mostly self-reported on academic benchmarks with specific base models, and results vary with the number of experts, rank, routing method and which layers receive experts.

In practice the costs are modest. Each LoRA expert adds two thin matrices per adapted weight matrix, so even several experts usually total well under the 1% of model parameters reported in the Cohere work. Training fits on the same hardware used for ordinary LoRA fine-tuning, plus a little memory for the extra adapters and router. The main engineering decisions are how many experts to use, their rank, whether routing is dense or top-k, and which layers to adapt. Serving many adapters efficiently benefits from systems built for it; S-LoRA reports serving thousands of concurrent adapters with up to four times the throughput of naive approaches in Hugging Face PEFT and vLLM.