Use large runners on Windows to build with FA

facebookresearch · Oct 20, 2023 · c97990b · c97990b
1 parent de59cfc
commit c97990b
Show file tree

Hide file tree

Showing 3 changed files with 27 additions and 27 deletions.
diff --git a/.github/actions/setup-windows-runner/action.yml b/.github/actions/setup-windows-runner/action.yml
@@ -27,13 +27,6 @@ runs:
     - name: Setup MSVC
       uses: ilammy/msvc-dev-cmd@v1
 
-    - name: Configure Pagefile
-      # windows runners will OOM with many CUDA architectures
-      # we cheat here with a page file
-      uses: al-cheb/configure-pagefile-action@v1.3
-      with:
-        minimum-size: 8GB
-
     # really unfortunate: https://github.com/ilammy/msvc-dev-cmd#name-conflicts-with-shell-bash
     - name: Remove link.exe
       shell: bash

diff --git a/.github/workflows/wheels.yml b/.github/workflows/wheels.yml
@@ -19,8 +19,8 @@ jobs:
       fail-fast: false
       matrix:
         os:
-          - 4-core-ubuntu
-          - windows-2019
+          - 8-core-ubuntu
+          - windows-8-core
         python:
           - "3.8"
           - "3.9"
@@ -39,7 +39,7 @@ jobs:
           - torch_version: "2.1.0"
             publish: true
           # Publish source distribution only from this runner
-          - os: 4-core-ubuntu
+          - os: 8-core-ubuntu
             python: "3.10"
             torch_version: "2.1.0"
             sdist: true
@@ -59,4 +59,3 @@ jobs:
       twine_username: __token__
     secrets:
       twine_password: ${{ secrets.PYPI_TOKEN }}
-
diff --git a/.github/workflows/wheels_reusable.yml b/.github/workflows/wheels_reusable.yml
@@ -39,7 +39,7 @@ on:
 env:
   # you need at least cuda 5.0 for some of the stuff compiled here.
   TORCH_CUDA_ARCH_LIST: "5.0+PTX 6.0 6.1 7.0 7.5 8.0+PTX"
-  MAX_JOBS: 1
+  MAX_JOBS: 4
   DISTUTILS_USE_SDK: 1 # otherwise distutils will complain on windows about multiple versions of msvc
   XFORMERS_BUILD_TYPE: "Release"
   TWINE_USERNAME: __token__
@@ -123,16 +123,24 @@ jobs:
           echo "${{ inputs.pypirc }}" > ~/.pypirc
           cat ~/.pypirc
 
+      - name: Create venv
+        run: |
+          $PY -m venv venv
+          . ./venv/bin/activate
+          which pip
+          pip install packaging ninja
+          which ninja
       - name: Define version
         id: xformers_version
         run: |
           set -Eeuo pipefail
+          . ./venv/bin/activate
           git config --global --add safe.directory "*"
-          $PY -m pip install packaging
-          version=`$PY packaging/compute_wheel_version.py`
+          version=`python packaging/compute_wheel_version.py`
           echo $version > version.txt
           echo "BUILD_VERSION=$version${{ steps.cuda_info.outputs.CUDA_VERSION_SUFFIX }}" >> ${GITHUB_ENV}
           echo "BUILD_VERSION=$version${{ steps.cuda_info.outputs.CUDA_VERSION_SUFFIX }}" >> ${GITHUB_OUTPUT}
+          which ninja
           cat ${GITHUB_ENV}
       - run: echo "xformers-${BUILD_VERSION}"
       - run: echo "release version (will upload to PyTorch)"
@@ -152,14 +160,6 @@ jobs:
           ./aws/install
           aws --version
 
-      # TODO: This is needed because github-actions takes more than 6 hours (!)
-      # to build Flash-Attention on windows, and times out.
-      # We need to use bigger gh-action workers...
-      - if: runner.os == 'Windows'
-        name: (Windows) Disable Flash-Attention build
-        run: |
-          echo "XFORMERS_DISABLE_FLASH_ATTN=1" >> ${GITHUB_ENV}
-
       - if: runner.os == 'Windows'
         name: (Windows) Setup Runner
         uses: ./.github/actions/setup-windows-runner
@@ -168,7 +168,9 @@ jobs:
           python: ${{ inputs.python }}
 
       - name: Install dependencies
-        run: $PY -m pip install wheel setuptools twine -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu${{ inputs.cuda_short_version }}
+        run: |
+          . ./venv/bin/activate
+          pip install wheel setuptools twine -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu${{ inputs.cuda_short_version }}
 
       - if: runner.os == 'Linux'
         name: (Linux) install cuda
@@ -181,17 +183,20 @@ jobs:
       - name: Create source distribution
         if: inputs.sdist
         run: |
+          . ./venv/bin/activate
           rm -rf dist/
           (cd third_party/flash-attention && \
             git describe --tags --always > version.txt && \
             echo "Flash-Attention version" && \
             cat version.txt
           )
 
-          $PY setup.py sdist -d sdist/
+          python setup.py sdist -d sdist/
 
       - name: Build wheel
-        run: $PY setup.py bdist_wheel -d dist/ -k $PLAT_ARG
+        run: |
+          . ./venv/bin/activate
+          python setup.py bdist_wheel -d dist/ -k $PLAT_ARG
         env:
           PLAT_ARG: ${{ contains(inputs.os, 'ubuntu') && '--plat-name manylinux2014_x86_64' || '' }}
 
@@ -204,15 +209,18 @@ jobs:
 
       - name: Upload wheel to PyPi
         if: inputs.publish && steps.cuda_info.outputs.PUBLISH_PYPI == '1'
-        run: $PY -m twine upload dist/*.whl
+        run: |
+          . ./venv/bin/activate
+          python -m twine upload dist/*.whl
         env:
           TWINE_USERNAME: ${{ inputs.twine_username }}
           TWINE_PASSWORD: ${{ secrets.twine_password }}
 
       - name: Upload source distribution to PyPi
         if: inputs.publish && steps.cuda_info.outputs.PUBLISH_PYPI == '1' && inputs.sdist
         run: |
-          $PY -m twine upload sdist/*
+          . ./venv/bin/activate
+          python -m twine upload sdist/*
         env:
           TWINE_USERNAME: ${{ inputs.twine_username }}
           TWINE_PASSWORD: ${{ secrets.twine_password }}