Tensor-Level Inference Scheduling on Consumer Devices
pmttyji · reddit · 2026-07-20
This paper introduces ATSInfer, a hybrid CPU-GPU inference system designed for consumer devices. The authors note that existing offloading solutions mostly operate at the layer or expert level, failing to handle the heterogeneity of different tensors within the same layer or adapt to changing device workloads.
ATSInfer utilizes tensor-granularity offloading, combining static tensor placement, load-aware dynamic transfers, and asynchronous CPU-GPU coordination to allocate storage, data movement, and computation more efficiently on laptops and desktops.
Experiments show:
- prefill throughput increased by up to 1.94×
- decode throughput increased by up to 3.29×
- Improved GPU utilization and better use of PCIe bandwidth
The authors conclude that this design significantly enhances the local LLM deployment experience on personal devices.
Related event: ATSInfer Boosts Local LLM Inference Speed(2 posts)→
More from Infra
- Cloud capacity crisis is now an AI architecture problem, argues new newsletter — DavidLinthicum · 2026-10-26
- SpaceX files $40B GPU financing commitment with Apollo, already at 1.7GW — XFreeze · 2026-10-26
- Meta plans to raise up to $95B to fund massive AI expansion — Polymarket · 2026-10-26
- Should the community archive open model weights in case Hugging Face gets restricted? — Tall-Ad-7742 · 2026-10-26
- What happens to neocloud GPUs when the AI datacenter debt bubble bursts? — bowdoin-yale · 2026-10-26
- Nebius CEO: Debt Markets Closed for Neoclouds Last Week, Leaning on Customer Advances — zephyr_z9 · 2026-10-26