Mixture-of-Experts (MoE) architectures offer a general solution to the high inference costs of large language models (LLMs) via sparse routing, bringing faster and more accurate models, at the cost of massive parameter counts. For example, the SwitchTransformer-c2048 model has 1.6 trillion parameters, requiring 3.2TB of accelerator memory to run efficiently, which makes practical deployment challenging and expensive. In this paper, we present a solution to this memory problem, in form of a new compression and execution framework called QMoE. Specifically, QMoE consists of a scalable algorithm which accurately compresses trillion-parameter MoEs to less than 1 bit per parameter, in a custom format co-designed with bespoke GPU decoding kernels to facilitate efficient end-to-end compressed inference, with minor runtime overheads relative to uncompressed execution. Concretely, QMoE can compress the 1.6 trillion parameter SwitchTransformer-c2048 model to less than 160GB (20x compression, 0.8 bits per parameter) at only minor accuracy loss, in less than a day on a single GPU. This enables, for the first time, the execution of a trillion-parameter model on affordable commodity hardware, like a single server with 4x NVIDIA A6000 or 8x NVIDIA 3090 GPUs, at less than 5% runtime overhead relative to ideal uncompressed inference. The source code and compressed models are available at github.com/IST-DASLab/qmoe.

Microsoft today released an updated version of its DeepSpeed library that introduces a new approach to training AI models containing trillions of parameters, the variables internal to the model that inform its predictions. The company claims the technique, dubbed 3D parallelism, adapts to the varying needs of workload requirements to power extremely large models while balancing scaling efficiency. Single massive AI models with billions of parameters have achieved great strides in a range of challenging domains. Studies show they perform well because they can absorb the nuances of language, grammar, knowledge, concepts, and context, enabling them to summarize speeches, moderate content in live gaming chats, parse complex legal documents, and even generate code from scouring GitHub. But training the models requires enormous computational resources.