Edge0 cuts memory use to serve large AI models on modest hardware
The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction
Artificial Intelligence
Summary
Running very large AI models often needs a lot of memory, which regular computers don't have. The authors created Edge0, which cleverly guesses what parts of the model will be needed next and loads only those from slower storage like an SSD. This approach lets a big 35-billion parameter model run smoothly using just 3GB of active memory on a normal PC. They also improved quality by adjusting the guesses and reducing errors caused by compressing the model data. The tool is open source and works well on other smaller models too.
What this means in practice
- •For machine learning engineers: Deploy large sparse language models efficiently on consumer-grade hardware with limited memory.
- •For embedded ai developers: Run high-parameter AI models in resource-constrained environments by streaming required parts from SSD with trained routing.
Authors
Yu Lin, Yiming Wang, Runyuan Cai, Hanze Liu, Xiaodong Zeng
Abstract
Mixture-of-experts (MoE) inference on consumer hardware is bounded by weight memory: a 35B-class model is 19.5GB at 4-bit, and sparsity shrinks the compute per token, not the bytes that must be held. Naive offloading to SSD does not help on its own, because layer N+1's experts must be chosen before layer N's output exists, so the reads cannot start early enough to hide behind compute. We present Edge0, a streaming MoE inference engine that closes the gap with a prerouter: a per-layer head predicts the next layer's routing one token ahead, and the prediction is consumed as the routing itself, so the staged expert set equals the routed set and nothing is dropped. An unmerged recovery LoRA, trained on the student path, pays back the quality lost to int4 quantization and routing replacement. On a single 24GB machine, Edge0 serves a 35B MoE at 20tok/s inside 3GiB of peak active memory, within a few points of its fp16 teacher on average across five public benchmarks. An 8B tier runs on the same framework, and the framework, checkpoints, and adapters are open source.