COPE picks the right mixture-of-experts offloading policy before runtime, cutting implementation effort by up to 74%
AspireAI community · arXiv
Mixture-of-experts models are sparse enough to keep most expert weights in CPU memory and transfer or execute routed experts on demand, which makes memory-efficient deployment practical. The problem is that existing MoE offloading systems target specific model-platform configurations, and differences in model sparsity, CPU and GPU throughput, VRAM capacity and host-device bandwidth can reorder the policy ranking. COPE, the Composable Offloading Policy Emulator, selects the best-performing policy for a target model-platform configuration before runtime integration. It has two components: a Policy Matrix that factors the design space into expert residency and MoE execution strategy, which enables mechanism reuse across candidates and exposes four new candidates; and a Virtual Heterogeneous Machine that profiles combined-operator schedules on the target platform and combines those measurements with offline routing traces to estimate candidate performance. Against the evaluated baselines, COPE improves fastest-policy selection accuracy by 24% to 100% and reduces performance-prediction error by 1.7 to 6.8 times, and by identifying the preferred policy ahead of integration it further reduces policy-specific implementation effort by up to 74%. The limitation is that the reported accuracy and error improvements are relative to the evaluated baselines on the configurations profiled, not universal across models and hardware.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.