vibe-profiling

using GCC -finstrument-functions option to do performance profile various of C/C++ programs

This project enables function-level profiling using GCC's -finstrument-functions. It provides detailed insights into execution flow, helping developers analyze performance, debug, and optimize large-scale models like LLaMA for NLP tasks.

Currenly our profiling covers llama.cpp and pytorch

Start

git clone https://github.com/ucas-linux/vibe-profiling.git

cd vibe-profiling

git submodule update --init --recursive

profile llama.cpp

build llama.cpp

sudo apt-get install elfutils libdw-dev

./build-binary.sh

profiling llama.cpp

extern/llama.cpp/buildinstrument/bin/llama-cli -t 1 -m ggml-org_gemma-3-1b-it-GGUF_gemma-3-1b-it-Q4_K_M.gguf

In other terminal

kill -sSIGUSR1 pid of llama-cli

profiling pytorch

install dependent software

sudo apt-get update
sudo apt-get install -y \
  git cmake ninja-build build-essential \
  python3 python3-dev python3-venv python3-pip \
  libopenblas-dev libomp-dev

cd extern/pytorch

git submodule update --init --recursive

python3 -m venv ~/venvs/torch-src

source ~/venvs/torch-src/bin/activate

pip install -r requirements.txt

build pytorch

export CMAKE_BUILD_TYPE=Debug
export CMAKE_GENERATOR=Ninja
export USE_CUDA=0
export USE_ROCM=0
export USE_CUFILE=0
export USE_NCCL=0
export USE_DISTRIBUTED=0
export CMAKE_CXX_FLAGS=-finstrument-functions
export BUILT_TEST=0

python -m pip install --no-build-isolation -v -e .

profiling pytorch

g++ -shared -fPIC src/profilepytorch.cpp -g3 -o src/profilepytorch.so -ldl -ldw -lelf

LD_PRELOAD=./src/profilepytorch.so python src/grad.py