Mixture-of-Experts (MoE) architectures scale language models by activating only a subset of specialized expert networks for each input token, thereby reducing the number of floating-point operations. However, the growing size of modern MoE models causes their full parameter sets to exceed GPU memory capacity; for example, Mixtral-8x7B has 45 billion parameters and requires 87 GB of memory even though only 14 billion parameters are used per token. Existing systems alleviate this limitation by offloading inactive experts to CPU memory, but transferring experts across the PCIe interconnect incurs significant latency (about 10 ms). Prefetching heuristics aim to hide this latency by predicting which experts are needed, but prefetch failures introduce significant stalls and amplify inference latency. In the event of a prefetch failure, prior work offers two primary solutions: either fetch the expert on demand, which incurs a long stall due to the PCIe bottleneck, or drop the expert from the computation, which significantly degrades model accuracy. The critical challenge, therefore, is to maintain both high inference speed and model accuracy when prefetching fails.
View on arXiv