Context-aware Adaptive Surgery: A Fast and Effective Framework for Adaptative Model Partition
summary
A retraining-free framework called CAS builds a Partition State Graph and runs a graph-based adaptive DNN surgery algorithm that starts from the last solution and walks neighboring states, exploiting the neighbor effect (adjacent partition points show similar resource use). It achieves rapid re-partitioning (~10 ms) for large models across six DNNs and multiple datasets, keeps inference latency comparable to baselines, and sustains real-time adaptation in a three-hour live deployment.
AI-written summary · about
related talks
15mADaPS: Adaptive Data Partitioning to Parallelize CNN Inference on Resource-Constrained(…)
17mModel-Switching: Dealing with Fluctuating Workloads in Machine-Learning-as-a-Service
15mEffectively Scheduling Computational Graphs of Deep Neural Networks toward Their...
01h 02mMulti-Dimension Mapping of Embedded Applications
17mSPLIDT: Partitioned Decision Trees for Scalable Stateful Inference at Line Rate
23mDecentralized Application-Level Adaptive Scheduling for Multi-Instance DNNs on...
21mArchitecture-Preserving Provable Repair of Deep Neural Networks
05mPlatform-Agnostic Lightweight Deep Learning for Garbage Collection...