44h ago · arXiv (AI/CL/LG, keyword-filtered) →
On-device inference requires kernels that make effective use of local accelerators.