Profile

ChatLearn provides two ways to profile performance:

1. Torch profiler

2. nsys Note: For large models, the profile result can be very large. It is recommended to reduce the model size when profiling.

Torch Profiler

Users can enable the Torch profiler by configuring the rlhf setting profiler_dir: path_to_profile_dir in the main configuration file of the system.

profiler_dir: path_to_profile_dir

nsys

Users can enable the nsys profiler by configuring the rlhf setting nsys: True in the main configuration file of the system.

runtime:
  nsys: True

When launching the program, nsys startup parameters need to be added before the execution command, as shown in the following example:

nsys profile -w true -t cuda,nvtx,osrt,cudnn,cublas -s none --capture-range=cudaProfilerApi --capture-range-end=stop-shutdown --cudabacktrace=true -x true --force-overwrite true -o my_profile \
python train_rlhf.py XXX