name: IntegrationTest

on: workflow_dispatch

jobs:
  IntegrationTest:
    runs-on: [ self-hosted, A100 ]
    defaults:
      run:
        shell: bash
    strategy:
      matrix:
        cuda: [ cuda11.8, cuda12.1 ]

    container:
      image: "ghcr.io/microsoft/mscclpp/mscclpp:dev-${{ matrix.cuda }}"
      options: --privileged --ipc=host --gpus=all --ulimit memlock=-1:-1

    steps:
      - name: Checkout
        uses: actions/checkout@v4

      - name: Build
        run: |
          mkdir build && cd build
          MPI_HOME=/usr/local/mpi cmake -DCMAKE_BUILD_TYPE=Release ..
          make -j

      - name: Lock GPU clock frequency
        run: |
          sudo nvidia-smi -pm 1
          for i in $(seq 0 $(( $(nvidia-smi -L | wc -l) - 1 ))); do
            sudo nvidia-smi -ac $(nvidia-smi --query-gpu=clocks.max.memory,clocks.max.sm --format=csv,noheader,nounits -i $i | sed 's/\ //') -i $i
          done

      - name: Run mscclpp AllGather test
        run: |
          set -e
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allgather_test_perf -b 1K -e 1G -f 2 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allgather_test_perf -b 1K -e 1G -f 2 -k 1 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allgather_test_perf -b 1K -e 1G -f 2 -k 2 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allgather_test_perf -b 1K -e 1G -f 2 -k 3 -o output.jsonl

      - name: Run mscclpp SendRecv test
        run: |
          set -e
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/sendrecv_test_perf -b 1K -e 1G -f 2 -o output.jsonl

      - name: Run mscclpp AllReduce test
        run: |
          set -e
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allreduce_test_perf -b 1K -e 1G -f 2 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allreduce_test_perf -b 1K -e 1G -f 2 -k 1 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allreduce_test_perf -b 1K -e 1G -f 2 -k 2 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allreduce_test_perf -b 1K -e 1G -f 2 -k 3 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allreduce_test_perf -b 1K -e 1G -f 2 -k 4 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allreduce_test_perf -b 12M -e 48M -i 3145728 2 -k 5 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/allreduce_test_perf -b 24K -e 768K -i 24576 -k 6 -w 100 -n 100 -o output.jsonl

      - name: Run mscclpp AllToAll test
        run: |
          set -e
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/alltoall_test_perf -b 1K -e 1G -f 2 -o output.jsonl
          mpirun --allow-run-as-root -np 8 --bind-to numa -x MSCCLPP_DEBUG=WARN ./build/test/mscclpp-test/alltoall_test_perf -b 1K -e 1G -f 2 -k 1 -o output.jsonl

      - name: Check collective primitives performance
        run: |
          set -e
          python3 test/mscclpp-test/check_perf_result.py --perf-file output.jsonl --baseline-file test/deploy/perf_ndmv4.jsonl