config.yml

# Python CircleCI 2.0 configuration file
#
# Check https://circleci.com/docs/2.0/language-python/ for more details
#
# Adopted from
# https://github.com/facebookresearch/detectron2/blob/main/.circleci/config.yml
#
# Pro tip: download circle ci cli to validate the config locally during development.
#
# To reset/clean the cache update the CACHE_VERSION variable in project settings
# in the fairscale project in CircleCI's web UI. The CACHE_VERSION follows the convention
# v$(FAIRSCALE_VERSION)-${CACHE_NUMBER}. E.g. v0.4.2-1. CACHE_NUMBER must start
# at 1 and increase in whole numbers. When changing the CACHE_VERSION manually
# always set the FAIRSCALE_VERSION value to the fairscale version being tested.
# To reset the cache when not updating the fairscale version, only update the
# CACHE_NUMBER value.

version: 2.1

orbs:
  codecov: codecov/codecov@1.0.2

# -------------------------------------------------------------------------------------
# Environments to run the jobs in
# -------------------------------------------------------------------------------------
cpu_py38: &cpu_py38
  docker:
    - image: circleci/python:3.8
  resource_class: large

cpu_py39: &cpu_py39
  docker:
    - image: circleci/python:3.9
  resource_class: large

cpu_py310: &cpu_py310
  docker:
    - image: circleci/python:3.10
  resource_class: large

# Here is the list of GPU images:
# https://circleci.com/docs/2.0/configuration-reference/#available-linux-gpu-images
# We need to use multiple gpus for several jobs. The resource_class
# values are available here T101565170
# gpu.nvidia.small.multi = 2 gpus with 16 GB ram each
# gpu.nvidia.medium.multi = 4 gpus with 16 GB ram each

gpu_cu_11_2_small_multi: &gpu_cu_11_2_small_multi
  environment:
    CUDA_VERSION: "11.2"
    CUDA_HOME: /usr/local/cuda-11.2
  machine:
    image: ubuntu-2004-cuda-11.2:202103-01
  resource_class: gpu.nvidia.small.multi

gpu_cu_11_2_medium_multi: &gpu_cu_11_2_medium_multi
  environment:
    CUDA_VERSION: "11.2"
    CUDA_HOME: /usr/local/cuda-11.2
  machine:
    image: ubuntu-2004-cuda-11.2:202103-01
  resource_class: gpu.nvidia.medium.multi

# -------------------------------------------------------------------------------------
# Re-usable commands
# -------------------------------------------------------------------------------------
setup_venv: &setup_venv
  - run:
      name: Setup Virtual Env
      working_directory: ~/
      # use bash -x for debug early commands executed in .bashrc.
      shell: /bin/bash
      command: |
        set -e
        set -o pipefail
        python -m venv ~/venv
        echo ". ~/venv/bin/activate" >> $BASH_ENV
        . ~/venv/bin/activate
        python --version
        which python
        which pip
        pip install --upgrade pip

# most recent LTS version
install_dep_pytorch_lts: &install_dep_pytorch_lts
  - run:
      name: Install Dependencies with torch 1.8.2 (LTS)
      command: |
        # check if we have restored venv cache (/home/circleci/venv) correctly, if so, just skip
        if [ -f /home/circleci/venv/check_version.py ]; then python /home/circleci/venv/check_version.py torch eq 1.8 && exit 0; fi
        # start installing
        pip install --progress-bar off torch==1.8.2+cu111 torchvision==0.9.2+cu111 --extra-index-url https://download.pytorch.org/whl/lts/1.8/cu111
        pip install --progress-bar off -r requirements-dev.txt
        pip install --progress-bar off -r requirements-benchmarks.txt
        python -c 'import torch; print("Torch version:", torch.__version__)'
        python -c 'import torch; assert torch.__version__.split(".")[:2] == ["1", "8"], f"wrong torch version {torch.__version__}"'
        python -m torch.utils.collect_env
        wget -O /home/circleci/venv/check_version.py https://raw.githubusercontent.com/min-xu-ai/check_verion/main/check_version.py

# most recent stable version
install_dep_pytorch_stable: &install_dep_pytorch_stable
  - run:
      name: Install Dependencies with torch 1.12.0
      command: |
        # check if we have restored venv cache (/home/circleci/venv) correctly, if so, just skip
        if [ -f /home/circleci/venv/check_version.py ]; then python /home/circleci/venv/check_version.py torch eq 1.11 && exit 0; fi
        # start installing
        pip install --progress-bar off torch==1.12.0+cu113 torchvision==0.13.0+cu113 --extra-index-url https://download.pytorch.org/whl/cu113
        pip install --progress-bar off -r requirements-dev.txt
        pip install --progress-bar off -r requirements-benchmarks.txt
        python -c 'import torch; print("Torch version:", torch.__version__)'
        python -c 'import torch; assert torch.__version__.split(".")[:2] == ["1", "12"], f"wrong torch version {torch.__version__}"'
        python -m torch.utils.collect_env
        wget -O /home/circleci/venv/check_version.py https://raw.githubusercontent.com/min-xu-ai/check_verion/main/check_version.py

install_dep_pytorch_nightly: &install_dep_pytorch_nightly
  - run:
      name: Install Dependencies with a torch nightly preview build
      command: |
        # check if we have restored venv cache (/home/circleci/venv) correctly, if so, just skip
        if [ -f /home/circleci/venv/check_version.py ]; then python /home/circleci/venv/check_version.py torch eq 1.12 && exit 0; fi
        # start installing
        pip install --pre torch==1.13.0.dev20220825+cu113 torchvision==0.14.0.dev20220825+cu113 --extra-index-url https://download.pytorch.org/whl/nightly/cu113
        pip install --progress-bar off -r requirements-dev.txt
        pip install --progress-bar off -r requirements-benchmarks.txt
        python -c 'import torch; print("Torch version:", torch.__version__)'
        python -c 'import torch; assert torch.__version__.split(".")[:2] == ["1", "13"], f"wrong torch version {torch.__version__}"'
        python -m torch.utils.collect_env
        wget -O /home/circleci/venv/check_version.py https://raw.githubusercontent.com/min-xu-ai/check_verion/main/check_version.py

install_repo: &install_repo
  - run:
      name: Install Repository
      command: |
        pip install .
        # Test import.
        python -c 'import sys; sys.path = sys.path[1:]; import fairscale'

check_test_list: &check_test_list
  - run:
      name: Verify that unit test list files are correct
      command: |
        bash ./tests/ci_test_list_check.sh

upload_coverage: &upload_coverage
  - codecov/upload:
      file: 'coverage.xml'
      token: $CODECOV_TOKEN

run_offload_benchmark: &run_offload_benchmark
  - run:
      name: Run Offload Benchmark
      command: |
        PYTHONPATH=. python benchmarks/experimental/offload.py --checkpoint_activation

run_fsdp_benchmark: &run_fsdp_benchmark
  - run:
      name: Run FSDP Benchmark
      command: |
        PYTHONPATH=. python benchmarks/fsdp.py --use_synthetic_data

run_pipe_benchmark: &run_pipe_benchmark
  - run:
      name: Run Pipe Benchmark
      command: |
        PYTHONPATH=. python benchmarks/pipe.py

run_oss_benchmark: &run_oss_benchmark
  - run:
      name: Run OSS Benchmark
      command: |
        PYTHONPATH=. python benchmarks/oss.py --world_size 4 --epochs 2
        PYTHONPATH=. python benchmarks/oss.py --check_regression --world_size 4 --optim_type oss_sharded_ddp --epochs 12

run_oss_gloo: &run_oss_gloo
  - run:
      name: Run OSS with Gloo
      command: |
        PYTHONPATH=. python benchmarks/oss.py --gloo --optim_type oss_ddp --epochs 2
        PYTHONPATH=. python benchmarks/oss.py --gloo --optim_type oss_sharded_ddp --epochs 2

run_oss_amp: &run_oss_amp
   - run:
       name: Run OSS with Torch AMP
       command: |
         PYTHONPATH=. python benchmarks/oss.py --amp --epochs 3 --optim_type oss_sharded_ddp

run_oss_for_each: &run_oss_for_each
   - run:
       name: Run OSS with Torch AMP and ForEach optmizer
       command: |
         PYTHONPATH=. python benchmarks/oss.py --amp --epochs 3 --optim_type oss_sharded_ddp --multi_tensor_optim

run_doc_build: &run_doc_build
   - run:
       name: Testing doc build
       command: |
         cd docs
         pip install --progress-bar off -r requirements.txt
         make help
         make singlehtml | tee make.out
         ! tail make.out | grep -q warning

# This is an alias to run all unit tests possible on a platform.
run_unittests: &run_unittests
   - run:
       name: Run all unit tests.
       # We run all and not stopping on failure on CPU since docker time is cheaper.
       command: |
         ulimit -n 10000
         pytest --junitxml=test-results/junit.xml --verbose --timeout 60 --cov-report=xml --cov=./

commands:

   # This is a command (like a function) that run tests from a given test_list_file.
   # If test_list_file is not given, this results in an error.
   run_unittests_from_list:
     parameters:
       test_list_file:
         type: string
         default: "/dev/non_exist"  # Default to error out
     steps:
       - run:
           name: Run Unit Tests
           command: |
             ulimit -n 10000
             if [ ! -f <<parameters.test_list_file>> ]; then exit 1; fi
             pytest --junitxml=test-results/junit.xml --verbose --timeout 70 --cov-report=xml --cov=./ `cat <<parameters.test_list_file>>`

   setup_pyenv:
     parameters:
       version:
         type: string
     steps:
       # Cache the pyenv download directory to avoid re-downloading over and over.
       - restore_cache:
          keys:
             - cache-key-pyenv-3-9-7-v1

       - run:
           name: Setup pyenv
           # We used to use the following commands to update pyenv.
           #   git clone https://github.com/pyenv/pyenv-update.git $(pyenv root)/plugins/pyenv-update
           #   pyenv update
           # However, it is not deterministic since pyenv is being updated.
           # It is now fixed to a version. (v2.3.0 is broken since it cause bash to fail when it try to do "eval $(pyenv init -)")
           #
           # We use "-sf" to skip already installed version after cache restoring the pyenv dir. This should avoid redoing
           # downloading and installing. The key name has a "v1" in the end so that we can debug and try different cache objects
           # while doing development.
           command: |
             cd /opt/circleci/.pyenv/
             git remote update
             git checkout v2.2.0
             pyenv install -sf <<parameters.version>>
             pyenv global <<parameters.version>>

       - save_cache:
          paths:
             - /opt/circleci/.pyenv
          key: cache-key-pyenv-3-9-7-v1

# -------------------------------------------------------------------------------------
# Jobs to run
# -------------------------------------------------------------------------------------

jobs:
  cpu_tests_py38:
    <<: *cpu_py38

    working_directory: ~/fairscale

    steps:
      - checkout
      - <<: *check_test_list
      - <<: *setup_venv

      # Do this first to test repo dependencies. Basic import should work after this
      # installation. See issue #1042 for an example.
      - <<: *install_repo

      # Cache the venv directory that contains dependencies
      - restore_cache:
          keys:
            - cache-key-cpu-py38-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}
      - <<: *install_dep_pytorch_stable

      - save_cache:
          paths:
            - ~/venv
          key: cache-key-cpu-py38-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *run_unittests
      - <<: *run_doc_build

      - store_test_results:
          path: test-results

  cpu_tests_py39:
    <<: *cpu_py39

    working_directory: ~/fairscale

    steps:
      - checkout
      - <<: *check_test_list
      - <<: *setup_venv

      # Do this first to test repo dependencies. Basic import should work after this
      # installation. See issue #1042 for an example.
      - <<: *install_repo

      # Cache the venv directory that contains dependencies
      - restore_cache:
          keys:
            - cache-key-cpu-py39-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *install_dep_pytorch_stable

      - save_cache:
          paths:
            - ~/venv
          key: cache-key-cpu-py39-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *run_unittests
      - <<: *run_doc_build

      - store_test_results:
          path: test-results

  cpu_tests_py310:
    <<: *cpu_py310

    working_directory: ~/fairscale

    steps:
      - checkout
      - <<: *check_test_list
      - <<: *setup_venv

      # Do this first to test repo dependencies. Basic import should work after this
      # installation. See issue #1042 for an example.
      - <<: *install_repo

      # Cache the venv directory that contains dependencies
      - restore_cache:
          keys:
            - cache-key-cpu-py310-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *install_dep_pytorch_stable

      - save_cache:
          paths:
            - ~/venv
          key: cache-key-cpu-py310-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *run_unittests
      - <<: *run_doc_build

      - store_test_results:
          path: test-results

  gpu_tests_lts:
    parameters:
      test_list_file:
        type: string
        default: "/dev/non_exist"

    <<: *gpu_cu_11_2_small_multi

    working_directory: ~/fairscale

    steps:
      - checkout

      - run: nvidia-smi

      # Run this to make sure we use python3 from the system.
      - setup_pyenv:
          version: 3.9.7

      - <<: *setup_venv

      # Cache the venv directory that contains dependencies
      - restore_cache:
          keys:
            - cache-key-py-3-9-7-gpu-torch-1-8-2-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *install_dep_pytorch_lts

      - save_cache:
          paths:
            - ~/venv
          key: cache-key-py-3-9-7-gpu-torch-1-8-2-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *install_repo

      - run_unittests_from_list:
          test_list_file: <<parameters.test_list_file>>

      - store_test_results:
          path: test-results

      # Disabled codecov since https://codecov.io/bash seems to be down.
      #- <<: *upload_coverage

  gpu_tests_stable:
    parameters:
      test_list_file:
        type: string
        default: "/dev/non_exist"

    <<: *gpu_cu_11_2_small_multi

    working_directory: ~/fairscale

    steps:
      - checkout

      - run: nvidia-smi

      # Run this to make sure we use python3 from the system.
      - setup_pyenv:
          version: 3.9.7

      - <<: *setup_venv

      # Cache the venv directory that contains dependencies
      - restore_cache:
          keys:
            - cache-key-py-3-9-7-gpu-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *install_dep_pytorch_stable

      - save_cache:
          paths:
            - ~/venv
          key: cache-key-py-3-9-7-gpu-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *install_repo

      - run_unittests_from_list:
          test_list_file: <<parameters.test_list_file>>

      - store_test_results:
          path: test-results

  gpu_tests_pytorch_nightly:
    parameters:
      test_list_file:
        type: string
        default: "/dev/non_exist"

    <<: *gpu_cu_11_2_medium_multi

    working_directory: ~/fairscale

    steps:
      - checkout

      - run: nvidia-smi

      # Run this to make sure we use python3 from the system.
      - setup_pyenv:
          version: 3.9.7

      - <<: *setup_venv

      # Cache the venv directory that contains dependencies
      - restore_cache:
          keys:
            - cache-key-py-3-9-7-gpu-torch-1-12-0424-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *install_dep_pytorch_nightly

      - save_cache:
          paths:
            - ~/venv
          key: cache-key-py-3-9-7-gpu-torch-1-12-0424-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *install_repo

      - run_unittests_from_list:
          test_list_file: <<parameters.test_list_file>>

      - store_test_results:
          path: test-results

  benchmarks_1:
    <<: *gpu_cu_11_2_small_multi

    working_directory: ~/fairscale

    steps:
      - checkout

      - run: nvidia-smi

      - setup_pyenv:
          version: 3.9.7

      - <<: *setup_venv

      # Cache the venv directory that contains dependencies
      - restore_cache:
          keys:
            - cache-key-py-3-9-7-benchmarks-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      # Cache the MNIST directory that contains benchmark data
      - restore_cache:
          keys:
            - cache-key-benchmark-MNIST-{{.Environment.CACHE_VERSION}}-{{checksum "benchmarks/datasets/mnist.py"}}

      - <<: *install_dep_pytorch_stable

      - save_cache:
          paths:
            - ~/venv
          key: cache-key-py-3-9-7-benchmarks-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *install_repo

      - <<: *run_pipe_benchmark

      - <<: *run_offload_benchmark

      - <<: *run_oss_amp

      - <<: *run_oss_for_each

      - <<: *run_oss_gloo

      - <<: *run_fsdp_benchmark

      - save_cache:
          paths:
            - /tmp/MNIST
          key: cache-key-benchmark-MNIST-{{.Environment.CACHE_VERSION}}-{{checksum "benchmarks/datasets/mnist.py"}}

  benchmarks_2:
    <<: *gpu_cu_11_2_medium_multi

    working_directory: ~/fairscale

    steps:
      - checkout

      - run: nvidia-smi

      - setup_pyenv:
          version: 3.9.7

      - <<: *setup_venv

      # Do this first to test an issue like #1042. We keep benchmark_1 doing it
      # after requirements-dev.txt, but change benchmark_2 to do this earlier to
      # ensure both cases are tested.
      - <<: *install_repo

      # Cache the venv directory that contains dependencies
      - restore_cache:
          keys:
            - cache-key-py-3-9-7-benchmarks-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}


      # Cache the MNIST directory that contains benchmark data
      - restore_cache:
          keys:
            - cache-key-benchmark-MNIST-{{.Environment.CACHE_VERSION}}-{{checksum "benchmarks/datasets/mnist.py"}}

      - <<: *install_dep_pytorch_stable

      - save_cache:
          paths:
            - ~/venv
          key: cache-key-py-3-9-7-benchmarks-torch-1-11-0-{{.Environment.CACHE_VERSION}}-{{checksum "setup.py"}}-{{checksum "requirements-dev.txt"}}

      - <<: *run_oss_benchmark

      - save_cache:
          paths:
            - /tmp/MNIST
          key: cache-key-benchmark-MNIST-{{.Environment.CACHE_VERSION}}-{{checksum "benchmarks/datasets/mnist.py"}}


workflows:
  version: 2
  build:
    jobs:
      - cpu_tests_py38
      - cpu_tests_py39
      - cpu_tests_py310
      - gpu_tests_lts:
          test_list_file: tests/ci_test_list_1.txt
      - gpu_tests_stable:
          test_list_file: tests/ci_test_list_1.txt
      - gpu_tests_pytorch_nightly:
          test_list_file: tests/ci_test_list_1.txt
      - gpu_tests_lts:
          test_list_file: tests/ci_test_list_2.txt
      - gpu_tests_stable:
          test_list_file: tests/ci_test_list_2.txt
      - gpu_tests_pytorch_nightly:
          test_list_file: tests/ci_test_list_2.txt
      - gpu_tests_lts:
          test_list_file: tests/ci_test_list_3.txt
      - gpu_tests_stable:
          test_list_file: tests/ci_test_list_3.txt
      - gpu_tests_pytorch_nightly:
          test_list_file: tests/ci_test_list_3.txt
      - benchmarks_1
      - benchmarks_2