Collecting Profiling Data
# Install instrumentation packages. pip install pytorch-2.1.0/dist/torch-*.whl --force-reinstall --no-deps pip install torch_npu/dist/torch_npu-*.whl --force-reinstall --no-deps # Set key environment variables. export OMP_PROC_BIND=false export LLVM_PROFILE_FILE=/tmp/profile/default_%m.profraw # Ensure that the /tmp/profile directory is empty. # Execute an actual training task to collect performance data. Example: bash run_model.sh ...
Parent topic: Compilation Optimization of torch and torch_npu