Description
PyProf release for 20.10, is available in the NVIDIA PyTorch NGC containers and directly from the PyProf GitHub page.
New Features
The key features of PyProf v3.5.0 / r20.10 are:
- Latest PyProf version supports PyTorch 1.5.0 (PyTorch 1.6.0 with DLProf) and Nsight Systems 2020.3.2.
- Latest PyProf version is compatible with DLProf v0.16.0 / r20.10.
- Nsight System database lookup improved to speed up the runtime profile analysis time by 50x.
- Node names will now include class info and can be linked back to the original Python source.
Known Issues
- This software only supports PyTorch 1.5.
- Forward-Backward kernel correlation heuristics do not work correctly with PyTorch 1.6.
Recommended work arounds include:
- Use with PyTorch 1.5.
- Use the 20.09-py3 PyTorch NGC container:
docker pull nvcr.io/nvidia/pytorch:20.10-py3
- Use DLProf in the 20.10 NGC PyTorch container or install using Python wheel.
Resolved Issues
- None