BoF: Angel 3.0: A Full Stack Machine Learning Platform
summary
Angel 3.0 adds auto feature engineering, Bayesian hyperparameter tuning and a cross-platform serving system that handles Angel, Spark, XGBoost and PyTorch models; a new PyTorch engine uses PyTorch for forward/backward passes while Angel’s parameter server stores and updates weights. The in-house math library claims ~10× speed vs Spark MLlib, model training drops from >10 hours to ~1 hour on a 63,000‑feature recommender (≈10×), a GNN job falls from >20h to ~5h (≈4×), and serving hits ~1,900 QPS vs 1,008 for TensorFlow Serving.
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:11What is Angel? Angel is a large scale distributed machine learning platform
- 3:40Angel Architecture
- 4:38Angel outside Tencent
- 6:51Open source & Papers
- 7:54Angel 3.0 features
- 8:54Auto Feature Engineering (1/2)
- 15:48Spark On Angel (SONA, 2/3)
- 22:37Hyper Parameter Tuning(1/2)
- 27:11Angel Serving
- 31:30Short Video Recommendation in Tencent
- 33:28Financial Anti-Fraud in Tencent
related talks
19mBeam Up Your GenAI Usage: Usability, Efficiency, Reliability
38mFantastic ML apps and how to build them
53mData Engineering and AI
31mUse-Case Driven ML Platforms
20mAccelerating innovation using tiered AI Architecture
16mAmeet Talwalkar, Assistant Professor of Computer Science, UCLA @ MLconf SF Part 2
44mPanel: ML Platform on Ray
33mScaling PyTorch Training with Ray Data on Uber's Michelangelo AI Platform