The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction
Edge0 streams 35B-parameter MoE inference from SSD on a 24GB machine at 20 tokens/s using a one-token-ahead prerouter, with framework and checkpoints open-sourced.
The paper (arXiv 2609.18063) presents Edge0, a streaming MoE inference engine whose per-layer prerouter predicts the next layer's expert routing one token ahead so SSD reads can be pre-staged with nothing dropped. An unmerged recovery LoRA trained on the student path compensates for quality lost to int4 quantization and routing replacement. On a single 24GB machine it serves a 35B MoE at 20 tok/s within 3GiB of peak active memory, within a few points of its fp16 teacher across five public benchmarks; framework, checkpoints, and adapters are open source.
- Per-layer prerouter predicts the next layer's expert routing one token ahead, enabling pre-staged SSD reads with zero dropped experts.
- Unmerged recovery LoRA trained on the student path recovers quality lost to int4 quantization and routing replacement.
- Serves a 35B MoE at 20 tok/s in 3GiB peak active memory on 24GB hardware, near fp16 teacher accuracy.
- An 8B tier is supported; framework, checkpoints, and adapters are released open source.
Full article182 words · extracted from huggingface.co · click to collapse
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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.18063