Title: Programming and profiling pytorch applications
Aishwariya Chakraborty aishwariya.chakraborty1@ibm.com
Manas Jha manas.j@ibm.com
Priyanka Naik priyanka.naik@ibm.com
<link to content> (tba)
Topics:
- GenAI workload profiling — tools and techniques
- nvidia-smi
- Nsys
- pytorch profiler
- Usage and exercises
- The what and the why
- torch dynamo and AOT Autograd
- FX graphs, Graph breaks, Dynamic shapes
- Different compilation backends
- Basic usage example and demonstration of various steps
- Performance improvement using torch compile
Pre-requisites/background:
- Pytorch basics
- Pytorch setup on laptop/local machine
Expected outcomes
Participants will
- Observe and measure resource utilization of execution with ML models
- Be able to analyze performance of models using torch profiler
- Understand the eager and compile modes of model execution
- Understand the framework and usage of torch.compile
Reference material
1. https://docs.pytorch.org/docs/2.7/torch.compiler.html
2. https://docs.pytorch.org/docs/stable/profiler.html