Skyline: Interactive In-Editor Computational Performance Profiling for Deep Neural Network Training
summary
Skyline is an interactive in-editor profiler implemented as an Atom plugin for PyTorch that builds empirical models predicting throughput and GPU memory versus batch size and updates profiles in the background when code changes. Visualizations link to specific lines of Python code and are manipulatable visualizations that, when dragged, write suggested batch size changes into the code. An exploratory study with seven graduate students shows useful results—three volunteers want to adopt it—and Skyline is open source and installable via pip and Atom package manager.
AI-written summary · about
chapters (8)
- 0:00Introduction to Skyline
- 0:51Why deep neural networks need profiling
- 3:29How deep neural network training works
- 6:46Challenges with existing profiling tools
- 9:18Key properties of DNN training
- 10:29Skyline features and benefits
- 12:16User study and evaluation
- 13:00Key takeaways and conclusion
related talks
21mImplementing your AI strategy with PyTorch Lightning in the cloud
10mLightning Talk: Profiling and Memory Debugging Tools for Distributed ML Workloads on GPUs
21mBuild PyTorch to Understand PyTorch
12mPyTorch Partner Collaborations
05mStill Burning GPUs On Debugging? Scale AI In One Line
01h 05mLIghtning Talks
26mThunder: Distribute and Optimize Your PyTorch Models With Zero...
20mLearn You a PyTorch!