- line_profiler to profile python code line by line
- typical usage involves adding a
@profiledecorator to the function you want to profile - launch your usual script with
CUDA_LAUNCH_BLOCKING=1 kernprof -lv main.py --flag1 VALUE1 ... - Ctrl+C to interupt after some time
- investigate results !
- (CUDA_LAUNCH_BLOCKING=1 prevents laze execution of CUDA functions, non-blocking mode improves execution times but is less interpretable (e.g. code for a line is actually later, and therefore contribute to the execution line of some other line)
- typical usage involves adding a