Dlprof demo

  • Code in this demo

  • Documentation
    User Guide
    DLProf Viewer

  • MNIST example code
    PyTorch MNIST example

  • DLProf installation (you can skip DLProf and PyTorch installations if you use the .yml file from Github)
    pip install nvidia-pyindex
    pip install nvidia-dlprof[pytorch] #For PyTorch
    pip install nvidia-dlprofviewer #DLProf Viewer
  • PyTorch installation
    conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
  • Profiling PyTorch with nvidia_dlprof_pytorch_nvtx add the following lines to your PyTorch network
    import nvidia_dlprof_pytorch_nvtx

    You should also run the training/inference loop with PyTorch’s NVTX Context Manager with the following:

    with torch.autograd.profiler.emit_nvtx():
  • Training on GPU without DLProf
    python mnist.py
  • Training on GPU with DLProf
    dlprof --mode=pytorch python mnist.py
  • Sample DLProf terminal output
    alt text

  • Files
    total 249M
    drwxr-xr-x 15 jingchao jingchao 4.0K Jan 18 23:17 ../
    -rw-rw-r--  1 jingchao jingchao  24M Jan 18 23:35 nsys_profile.qdrep
    -rw-r--r--  1 jingchao jingchao 193M Jan 18 23:35 nsys_profile.sqlite
    -rw-r--r--  1 jingchao jingchao  33M Jan 18 23:36 dlprof_dldb.sqlite
    -rw-rw-r--  1 jingchao jingchao 5.5K Jan 18 23:40 mnist.py
    drwxrwxr-x  2 jingchao jingchao 4.0K Jan 18 23:40 ./
  • Visualize results
    $ dlprofviewer dlprof_dldb.sqlite 
    [dlprofviewer-04:46:36 AM UTC] dlprofviewer running at http://localhost:8000

    alt text