Mistify: Automating DNN Model Porting for On-Device Inference at the Edge
summary
Mistify builds an automated server-side pipeline that converts a cloud model into many edge-ready variants by exposing a single configuration file and running an adaptation executor finite-state machine (initialization, search, measure, adjust). It consolidates goals into an ordered batch to avoid redundant searches, generates multi-branch models for instant runtime switching on the client, and launches background re-tailoring. Evaluation reports up to 70× speedup with 100+ goals and over 10× reduction in porting time with orders-of-magnitude less manual code change.
AI-written summary · about
chapters (7)
- 0:00The challenges of model porting
- 3:20Introducing Mystify
- 4:32Adaptation executor design
- 6:16Scaling with collective adaptation
- 7:57Handling runtime dynamics
- 9:17System workflow and integration
- 10:07Performance results
related talks
12mPreview: ML Serving
21mEdge Inference on Unknown Models
20mThe future of Edge Agents
36mApplied AI on the edge
51mSpeeding up deployment of machine learning models
27mModel Serving at the Edge Made Easier
25mLeveraging PyTorch for Generative AI in Distributed Edge Clouds
38measily create, deploy, and monitor mobile AI models