Latest Profiler APIs and Best Practices
summary
PyTorch Profiler measures runtimes and tensor memory per operator and now supports record_shapes true to group by input shapes, profile_memory true to report CPU/CUDA memory (the example shows most memory coming from the empty operator and from resize), and with_stack true to attach stack traces so you can group by top and stack entries. It also saves traces in Chrome JSON format and reveals two distinct call sites for convolution inside the ResNet script.
AI-written summary · about
chapters (4)
related talks
12mPyTorch Partner Collaborations
10mLightning Talk: Profiling and Memory Debugging Tools for Distributed ML Workloads on GPUs
30mProfiling CPU and Memory on Linux, with Opensource Graphical Tools
12mContinuous Profiling for GPUs
41mDeep dive into Android Studio Profilers
43mExploring Deep Learning Framework PyTorch
19mProfiling Your Code: 5 Tips to Significantly Boost Performance
11mPer-iteration PyTorch Visibility...