mirror of
https://github.com/BillyOutlast/flash-attention-prebuild-wheels-rocm.git
synced 2026-07-01 01:37:53 -04:00
[WIP] refactor windows build
This commit is contained in:
Junya Morioka
+183
-203
@@ -32,112 +32,110 @@ jobs:
|
||||
# #########################################################
|
||||
# Build wheels with GitHub hosted runner
|
||||
# #########################################################
|
||||
# build_wheels:
|
||||
# name: Build wheels and Upload (Linux x86_64, GitHub hosted runner)
|
||||
# needs: create_releases
|
||||
# runs-on: ubuntu-22.04
|
||||
# env:
|
||||
# DEBIAN_FRONTEND: noninteractive
|
||||
# TERM: xterm-256color
|
||||
# timeout-minutes: 1000
|
||||
# strategy:
|
||||
# fail-fast: false
|
||||
# matrix:
|
||||
# flash-attn-version: ["2.4.3", "2.5.9", "2.6.3"]
|
||||
# python-version: ["3.10", "3.11", "3.12"]
|
||||
# torch-version: ["2.7.1"]
|
||||
# # # https://developer.nvidia.com/cuda-toolkit-archive
|
||||
# cuda-version: ["12.8.1"]
|
||||
# exclude:
|
||||
# # torch < 2.2 does not support Python 3.12
|
||||
# - python-version: "3.12"
|
||||
# torch-version: "2.0.1"
|
||||
# - python-version: "3.12"
|
||||
# torch-version: "2.1.2"
|
||||
# # torch 2.0.1 does not support CUDA 12.x
|
||||
# - torch-version: "2.0.1"
|
||||
# cuda-version: "12.1.1"
|
||||
# - torch-version: "2.0.1"
|
||||
# cuda-version: "12.4.1"
|
||||
# - torch-version: "2.0.1"
|
||||
# cuda-version: "12.6.3"
|
||||
# - torch-version: "2.0.1"
|
||||
# cuda-version: "12.8.1"
|
||||
# # torch 2.7.0 does not support CUDA 12.4
|
||||
# - torch-version: "2.7.0"
|
||||
# cuda-version: "12.4.1"
|
||||
# steps:
|
||||
# - uses: actions/checkout@v4
|
||||
build_wheels:
|
||||
name: Build wheels and Upload (Linux x86_64, GitHub hosted runner)
|
||||
needs: create_releases
|
||||
runs-on: ubuntu-22.04
|
||||
env:
|
||||
DEBIAN_FRONTEND: noninteractive
|
||||
TERM: xterm-256color
|
||||
MAX_JOBS: 2
|
||||
NVCC_THREADS: 2
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
flash-attn-version: ["2.4.3", "2.5.9", "2.6.3"]
|
||||
python-version: ["3.10", "3.11", "3.12"]
|
||||
torch-version: ["2.7.1"]
|
||||
# # https://developer.nvidia.com/cuda-toolkit-archive
|
||||
cuda-version: ["12.8.1"]
|
||||
exclude:
|
||||
# torch < 2.2 does not support Python 3.12
|
||||
- python-version: "3.12"
|
||||
torch-version: "2.0.1"
|
||||
- python-version: "3.12"
|
||||
torch-version: "2.1.2"
|
||||
# torch 2.0.1 does not support CUDA 12.x
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.1.1"
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.4.1"
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.6.3"
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.8.1"
|
||||
# torch 2.7.0 does not support CUDA 12.4
|
||||
- torch-version: "2.7.0"
|
||||
cuda-version: "12.4.1"
|
||||
steps:
|
||||
- uses: actions/checkout@v4
|
||||
|
||||
# - name: Maximize build space
|
||||
# run: |
|
||||
# df -h
|
||||
# echo "-----------------------------"
|
||||
# sudo rm -rf /usr/share/dotnet
|
||||
# sudo rm -rf /usr/local/lib/android
|
||||
# sudo rm -rf /opt/ghc
|
||||
# sudo rm -rf /opt/hostedtoolcache/CodeQL
|
||||
# df -h
|
||||
- name: Maximize build space
|
||||
run: |
|
||||
df -h
|
||||
echo "-----------------------------"
|
||||
sudo rm -rf /usr/share/dotnet
|
||||
sudo rm -rf /usr/local/lib/android
|
||||
sudo rm -rf /opt/ghc
|
||||
sudo rm -rf /opt/hostedtoolcache/CodeQL
|
||||
df -h
|
||||
|
||||
# - name: Set Swap Space
|
||||
# uses: pierotofy/set-swap-space@master
|
||||
# with:
|
||||
# swap-size-gb: 48
|
||||
- name: Set Swap Space
|
||||
uses: pierotofy/set-swap-space@master
|
||||
with:
|
||||
swap-size-gb: 48
|
||||
|
||||
# - uses: actions/setup-python@v5
|
||||
# with:
|
||||
# python-version: ${{ matrix.python-version }}
|
||||
- uses: actions/setup-python@v5
|
||||
with:
|
||||
python-version: ${{ matrix.python-version }}
|
||||
|
||||
# - uses: Jimver/cuda-toolkit@master
|
||||
# with:
|
||||
# cuda: ${{ matrix.cuda-version }}
|
||||
# sub-packages: '["nvcc", "toolkit"]'
|
||||
# method: "network"
|
||||
# use-github-cache: false
|
||||
# use-local-cache: false
|
||||
- uses: Jimver/cuda-toolkit@master
|
||||
with:
|
||||
cuda: ${{ matrix.cuda-version }}
|
||||
sub-packages: '["nvcc", "toolkit"]'
|
||||
method: "network"
|
||||
use-github-cache: false
|
||||
use-local-cache: false
|
||||
|
||||
# - name: Install build dependencies
|
||||
# run: |
|
||||
# sudo apt install -y ninja-build clang
|
||||
# pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil
|
||||
- name: Install build dependencies
|
||||
run: |
|
||||
sudo apt install -y ninja-build clang
|
||||
pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil
|
||||
|
||||
# - name: Set environment variables
|
||||
# run: |
|
||||
# export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
|
||||
# export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
|
||||
# export MAX_JOBS=2
|
||||
# export NVCC_THREADS=2
|
||||
- name: Set environment variables
|
||||
run: |
|
||||
export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
|
||||
export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
|
||||
|
||||
# - name: Build wheels
|
||||
# timeout-minutes: 800
|
||||
# run: |
|
||||
# chmod +x build_linux.sh
|
||||
# ./build_linux.sh ${{ matrix.flash-attn-version }} ${{ matrix.python-version }} ${{ matrix.torch-version }} ${{ matrix.cuda-version }}
|
||||
# wheel_name=$(basename $(ls flash-attention/dist/*.whl | head -n 1))
|
||||
# echo "wheel_name=$wheel_name" >> $GITHUB_ENV
|
||||
- name: Build wheels
|
||||
run: |
|
||||
chmod +x build_linux.sh
|
||||
./build_linux.sh ${{ matrix.flash-attn-version }} ${{ matrix.python-version }} ${{ matrix.torch-version }} ${{ matrix.cuda-version }}
|
||||
wheel_name=$(basename $(ls flash-attention/dist/*.whl | head -n 1))
|
||||
echo "wheel_name=$wheel_name" >> $GITHUB_ENV
|
||||
|
||||
# - name: Install Test
|
||||
# run: |
|
||||
# pip install flash-attention/dist/${{ env.wheel_name }}
|
||||
# python -c "import flash_attn; print(flash_attn.__version__)"
|
||||
- name: Install Test
|
||||
run: |
|
||||
pip install flash-attention/dist/${{ env.wheel_name }}
|
||||
python -c "import flash_attn; print(flash_attn.__version__)"
|
||||
|
||||
# - name: Upload Release Asset
|
||||
# env:
|
||||
# GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
|
||||
# run: |
|
||||
# tag_name=${{ github.ref_name }}
|
||||
# wheel_path="flash-attention/dist/${{ env.wheel_name }}"
|
||||
- name: Upload Release Asset
|
||||
env:
|
||||
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
|
||||
run: |
|
||||
tag_name=${{ github.ref_name }}
|
||||
wheel_path="flash-attention/dist/${{ env.wheel_name }}"
|
||||
|
||||
# # Check if the file exists
|
||||
# if [ ! -f "$wheel_path" ]; then
|
||||
# echo "Error: Wheel file not found at $wheel_path"
|
||||
# exit 1
|
||||
# fi
|
||||
# Check if the file exists
|
||||
if [ ! -f "$wheel_path" ]; then
|
||||
echo "Error: Wheel file not found at $wheel_path"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
# # Upload the release asset using GitHub CLI
|
||||
# gh release upload "$tag_name" "$wheel_path" --clobber
|
||||
# Upload the release asset using GitHub CLI
|
||||
gh release upload "$tag_name" "$wheel_path" --clobber
|
||||
|
||||
# echo "Successfully uploaded $wheel_name to release $tag_name"
|
||||
echo "Successfully uploaded $wheel_name to release $tag_name"
|
||||
|
||||
# #########################################################
|
||||
# Build wheels with self-hosted runner
|
||||
@@ -154,6 +152,8 @@ jobs:
|
||||
env:
|
||||
DEBIAN_FRONTEND: noninteractive
|
||||
TERM: xterm-256color
|
||||
MAX_JOBS: 2
|
||||
NVCC_THREADS: 2
|
||||
timeout-minutes: 1000
|
||||
strategy:
|
||||
fail-fast: false
|
||||
@@ -243,8 +243,6 @@ jobs:
|
||||
run: |
|
||||
export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
|
||||
export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
|
||||
export MAX_JOBS=2
|
||||
export NVCC_THREADS=2
|
||||
|
||||
- name: Build wheels
|
||||
timeout-minutes: 800
|
||||
@@ -286,122 +284,104 @@ jobs:
|
||||
run: |
|
||||
sudo rm -rf /opt/hostedtoolcache/Python/${{ matrix.python-version }}*
|
||||
|
||||
# build_windows_wheels:
|
||||
# name: Build wheels and Upload (Windows x86_64, GitHub hosted runner)
|
||||
# runs-on: windows-latest
|
||||
# timeout-minutes: 1000
|
||||
# strategy:
|
||||
# fail-fast: false
|
||||
# matrix:
|
||||
# flash-attn-version: ["2.5.9", "2.6.3", "2.7.4"]
|
||||
# python-version: ["3.10", "3.11", "3.12"]
|
||||
# torch-version: ["2.4.1", "2.5.1", "2.6.0", "2.7.1"]
|
||||
# # https://developer.nvidia.com/cuda-toolkit-archive
|
||||
# # CUDA 11.8 cannot build flash-attn.
|
||||
# # https://github.com/Dao-AILab/flash-attention/issues/595
|
||||
# cuda-version: ["12.4.1", "12.8.1"]
|
||||
# steps:
|
||||
# - uses: actions/checkout@v4
|
||||
build_windows_wheels:
|
||||
name: Build wheels and Upload (Windows x86_64, GitHub hosted runner)
|
||||
runs-on: windows-latest
|
||||
timeout-minutes: 1000
|
||||
env:
|
||||
MAX_JOBS: 2
|
||||
NVCC_THREADS: 2
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
flash-attn-version: ["2.5.9", "2.6.3", "2.7.4"]
|
||||
python-version: ["3.10", "3.11", "3.12"]
|
||||
torch-version: ["2.4.1", "2.5.1", "2.6.0", "2.7.1"]
|
||||
# https://developer.nvidia.com/cuda-toolkit-archive
|
||||
# CUDA 11.8 cannot build flash-attn.
|
||||
# https://github.com/Dao-AILab/flash-attention/issues/595
|
||||
cuda-version: ["12.4.1", "12.8.1"]
|
||||
exclude:
|
||||
# torch < 2.2 does not support Python 3.12
|
||||
- python-version: "3.12"
|
||||
torch-version: "2.0.1"
|
||||
- python-version: "3.12"
|
||||
torch-version: "2.1.2"
|
||||
# torch 2.0.1 does not support CUDA 12.x
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.1.1"
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.4.1"
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.6.3"
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.8.1"
|
||||
# torch 2.7.0 does not support CUDA 12.4
|
||||
- torch-version: "2.7.0"
|
||||
cuda-version: "12.4.1"
|
||||
steps:
|
||||
- uses: actions/checkout@v4
|
||||
|
||||
# - name: Enable Git long paths
|
||||
# run: git config --system core.longpaths true
|
||||
# shell: pwsh
|
||||
- name: Enable Git long paths
|
||||
shell: pwsh
|
||||
run: git config --system core.longpaths true
|
||||
|
||||
# - name: Install VS2022 BuildTools
|
||||
# run: |
|
||||
# choco install -y visualstudio2022buildtools `
|
||||
# --version=117.14.1 `
|
||||
# --params "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64"
|
||||
# shell: pwsh
|
||||
- uses: actions/setup-python@v5
|
||||
with:
|
||||
python-version: ${{ matrix.python-version }}
|
||||
|
||||
# - uses: actions/setup-python@v5
|
||||
# with:
|
||||
# python-version: ${{ matrix.python-version }}
|
||||
- uses: Jimver/cuda-toolkit@v0.2.24
|
||||
with:
|
||||
cuda: ${{ matrix.cuda-version }}
|
||||
method: "network"
|
||||
use-github-cache: false
|
||||
use-local-cache: false
|
||||
|
||||
# - uses: Jimver/cuda-toolkit@v0.2.24
|
||||
# with:
|
||||
# cuda: ${{ matrix.cuda-version }}
|
||||
# method: "network"
|
||||
# use-github-cache: false
|
||||
# use-local-cache: false
|
||||
- name: Install VS2022 BuildTools
|
||||
shell: pwsh
|
||||
run: |
|
||||
choco install -y visualstudio2022buildtools `
|
||||
--version=117.14.1 `
|
||||
--params "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64"
|
||||
|
||||
# - name: Set CUDA and PyTorch versions
|
||||
# run: |
|
||||
# $cudaVersion = "${{ matrix.cuda-version }}" -replace '\.', ''
|
||||
# $cudaVersion = $cudaVersion.Substring(0, [Math]::Min(3, $cudaVersion.Length))
|
||||
# echo "MATRIX_CUDA_VERSION=$cudaVersion" >> $env:GITHUB_ENV
|
||||
# $torchVersion = "${{ matrix.torch-version }}" -replace '^(\d+\.\d+).*', '$1'
|
||||
# echo "MATRIX_TORCH_VERSION=$torchVersion" >> $env:GITHUB_ENV
|
||||
# shell: pwsh
|
||||
- name: Install build dependencies
|
||||
shell: pwsh
|
||||
run: |
|
||||
pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil ninja
|
||||
|
||||
# - name: Install build dependencies
|
||||
# run: |
|
||||
# pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil ninja
|
||||
# shell: pwsh
|
||||
- name: Build wheels
|
||||
shell: pwsh
|
||||
run: |
|
||||
.\build_windows.ps1 -FlashAttnVersion "${{ matrix.flash-attn-version }}" -PythonVersion "${{ matrix.python-version }}" -TorchVersion "${{ matrix.torch-version }}" -CudaVersion "${{ matrix.cuda-version }}"
|
||||
$wheelName = Get-ChildItem -Path "flash-attention\dist\*.whl" | Select-Object -First 1 | ForEach-Object { $_.Name }
|
||||
echo "wheel_name=$wheelName" >> $env:GITHUB_ENV
|
||||
|
||||
# - name: Install PyTorch ${{ matrix.torch-version }}+cu${{ matrix.cuda-version }}
|
||||
# run: |
|
||||
# $env:TORCH_CUDA_VERSION = python -c "from os import environ as env; support_cuda_versions = { '2.1': [121], '2.2': [121], '2.3': [121], '2.4': [121, 124], '2.5': [121, 124], '2.6': [124, 126], '2.7': [126, 128], '2.8': [128], }; target_cuda_versions = support_cuda_versions[env['MATRIX_TORCH_VERSION']]; cuda_version = int(env['MATRIX_CUDA_VERSION']); closest_version = min(target_cuda_versions, key=lambda x: abs(x - cuda_version)); print(closest_version)"
|
||||
- name: Install Test
|
||||
shell: pwsh
|
||||
run: |
|
||||
pip install flash-attention/dist/$env:wheel_name
|
||||
python -c "import flash_attn; print(flash_attn.__version__)"
|
||||
|
||||
# if ("${{ matrix.torch-version }}" -like "*dev*") {
|
||||
# pip install --pre torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/nightly/cu$env:TORCH_CUDA_VERSION
|
||||
# } else {
|
||||
# pip install --no-cache-dir torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/cu$env:TORCH_CUDA_VERSION
|
||||
# }
|
||||
- name: Upload Release Asset
|
||||
shell: pwsh
|
||||
env:
|
||||
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
|
||||
run: |
|
||||
$tag_name = "${env:GITHUB_REF}".Replace("refs/tags/", "")
|
||||
$wheel_path = "flash-attention/dist/$env:wheel_name"
|
||||
|
||||
# nvcc --version
|
||||
# python -V
|
||||
# python -c "import torch; print('PyTorch:', torch.__version__)"
|
||||
# python -c "import torch; print('CUDA:', torch.version.cuda)"
|
||||
# python -c "from torch.utils import cpp_extension; print(cpp_extension.CUDA_HOME)"
|
||||
# shell: pwsh
|
||||
# Check if the file exists
|
||||
if (-not (Test-Path $wheel_path)) {
|
||||
$tag_name = "${env:GITHUB_REF}".Replace("refs/tags/", "")
|
||||
$wheel_path = "flash-attention/dist/$env:wheel_name"
|
||||
|
||||
# - name: Checkout flash-attn
|
||||
# run: |
|
||||
# git clone https://github.com/Dao-AILab/flash-attention.git -b "v${{ matrix.flash-attn-version }}"
|
||||
# shell: pwsh
|
||||
# Check if the file exists
|
||||
if (-not (Test-Path $wheel_path)) {
|
||||
Write-Host "Error: Wheel file not found at $wheel_path"
|
||||
exit 1
|
||||
}
|
||||
|
||||
# - name: Build wheels
|
||||
# timeout-minutes: 800
|
||||
# run: |
|
||||
# Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
|
||||
# Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools' -DevCmdArguments '-arch=x64 -host_arch=x64'
|
||||
# $env:DISTUTILS_USE_SDK = 1
|
||||
# $env:BUILD_TARGET = "cuda"
|
||||
# Upload the release asset using GitHub CLI
|
||||
gh release upload "$tag_name" "$wheel_path" --clobber
|
||||
|
||||
# $env:MAX_JOBS = "2"
|
||||
# $env:NVCC_THREADS = "2"
|
||||
# $env:FLASH_ATTENTION_FORCE_BUILD = "TRUE"
|
||||
|
||||
# cd flash-attention
|
||||
# python setup.py bdist_wheel --dist-dir=dist
|
||||
# $baseWheelName = Get-ChildItem -Path "dist\*.whl" | Select-Object -First 1 | ForEach-Object { $_.Name }
|
||||
# $wheelName = $baseWheelName.Replace("${{ matrix.flash-attn-version }}", "${{ matrix.flash-attn-version }}+cu$env:MATRIX_CUDA_VERSION" + "torch$env:MATRIX_TORCH_VERSION")
|
||||
# Move-Item "dist\$baseWheelName" "dist\$wheelName"
|
||||
# echo "wheel_name=$wheelName" >> $env:GITHUB_ENV
|
||||
# shell: pwsh
|
||||
|
||||
# - name: Install Test
|
||||
# run: |
|
||||
# pip install flash-attention/dist/$env:wheel_name
|
||||
# python -c "import flash_attn; print(flash_attn.__version__)"
|
||||
# shell: pwsh
|
||||
|
||||
# - name: Upload Release Asset
|
||||
# env:
|
||||
# GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
|
||||
# run: |
|
||||
# $tag_name = "${env:GITHUB_REF}".Replace("refs/tags/", "")
|
||||
# $wheel_path = "flash-attention/dist/$env:wheel_name"
|
||||
|
||||
# # Check if the file exists
|
||||
# if (-not (Test-Path $wheel_path)) {
|
||||
# Write-Host "Error: Wheel file not found at $wheel_path"
|
||||
# exit 1
|
||||
# }
|
||||
|
||||
# # Upload the release asset using GitHub CLI
|
||||
# gh release upload "$tag_name" "$wheel_path" --clobber
|
||||
|
||||
# Write-Host "Successfully uploaded $env:wheel_name to release $tag_name"
|
||||
# shell: pwsh
|
||||
Write-Host "Successfully uploaded $env:wheel_name to release $tag_name"
|
||||
|
||||
@@ -1,118 +1,59 @@
|
||||
# Ref: https://github.com/Foul-Tarnished/flash-attention/blob/main/.github/workflows/build-wheels.yml
|
||||
|
||||
name: "Test Windows build"
|
||||
on:
|
||||
workflow_dispatch:
|
||||
|
||||
jobs:
|
||||
build_wheels:
|
||||
name: Build wheels and Upload
|
||||
build_windows_wheels:
|
||||
name: Build wheels and Upload (Windows x86_64, GitHub hosted runner)
|
||||
runs-on: windows-latest
|
||||
timeout-minutes: 1000
|
||||
env:
|
||||
MAX_JOBS: 2
|
||||
NVCC_THREADS: 2
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
flash-attn-version: ["2.6.3"]
|
||||
python-version: ["3.11"]
|
||||
torch-version: ["2.6.0"]
|
||||
# https://developer.nvidia.com/cuda-toolkit-archive
|
||||
cuda-version: ["12.6.3"]
|
||||
exclude:
|
||||
# torch < 2.2 does not support Python 3.12
|
||||
- python-version: "3.12"
|
||||
torch-version: "2.0.1"
|
||||
- python-version: "3.12"
|
||||
torch-version: "2.1.2"
|
||||
# torch 2.0.1 does not support CUDA 12.x
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.1.1"
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.4.1"
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.6.3"
|
||||
- torch-version: "2.0.1"
|
||||
cuda-version: "12.8.1"
|
||||
# torch 2.7.0 does not support CUDA 12.4
|
||||
- torch-version: "2.7.0"
|
||||
cuda-version: "12.4.1"
|
||||
torch-version: ["2.7.1"]
|
||||
cuda-version: ["12.4.1"]
|
||||
steps:
|
||||
- uses: actions/checkout@v4
|
||||
|
||||
- name: Enable Git long paths
|
||||
shell: pwsh
|
||||
run: git config --system core.longpaths true
|
||||
shell: pwsh
|
||||
|
||||
- name: Install VS2022 BuildTools 17.9.7
|
||||
run: choco install -y visualstudio2022buildtools --version=117.9.7.0 --params "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64 --installChannelUri https://aka.ms/vs/17/release/180911598_-255012421/channel"
|
||||
shell: pwsh
|
||||
|
||||
- uses: actions/setup-python@v5
|
||||
with:
|
||||
python-version: ${{ matrix.python-version }}
|
||||
|
||||
- uses: Jimver/cuda-toolkit@master
|
||||
- uses: Jimver/cuda-toolkit@v0.2.24
|
||||
with:
|
||||
cuda: ${{ matrix.cuda-version }}
|
||||
method: "network"
|
||||
use-github-cache: false
|
||||
use-local-cache: false
|
||||
|
||||
- name: Set CUDA and PyTorch versions
|
||||
run: |
|
||||
$cudaVersion = "${{ matrix.cuda-version }}" -replace '\.', ''
|
||||
$cudaVersion = $cudaVersion.Substring(0, [Math]::Min(3, $cudaVersion.Length))
|
||||
echo "MATRIX_CUDA_VERSION=$cudaVersion" >> $env:GITHUB_ENV
|
||||
$torchVersion = "${{ matrix.torch-version }}" -replace '^(\d+\.\d+).*', '$1'
|
||||
echo "MATRIX_TORCH_VERSION=$torchVersion" >> $env:GITHUB_ENV
|
||||
- name: Install VS2022 BuildTools
|
||||
shell: pwsh
|
||||
run: |
|
||||
choco install -y visualstudio2022buildtools `
|
||||
--version=117.14.1 `
|
||||
--params "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64"
|
||||
|
||||
- name: Install build dependencies
|
||||
shell: pwsh
|
||||
run: |
|
||||
pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil ninja
|
||||
shell: pwsh
|
||||
|
||||
- name: Install PyTorch ${{ matrix.torch-version }}+cu${{ matrix.cuda-version }}
|
||||
run: |
|
||||
$env:TORCH_CUDA_VERSION = python -c "from os import environ as env; support_cuda_versions = { '2.0': [117, 118], '2.1': [118, 121], '2.2': [118, 121], '2.3': [118, 121], '2.4': [118, 121, 124], '2.5': [118, 121, 124], '2.6': [118, 124, 126], '2.7': [118, 126, 128], '2.8': [128], }; target_cuda_versions = support_cuda_versions[env['MATRIX_TORCH_VERSION']]; cuda_version = int(env['MATRIX_CUDA_VERSION']); closest_version = min(target_cuda_versions, key=lambda x: abs(x - cuda_version)); print(closest_version)"
|
||||
|
||||
if ("${{ matrix.torch-version }}" -like "*dev*") {
|
||||
pip install --pre torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/nightly/cu$env:TORCH_CUDA_VERSION
|
||||
} else {
|
||||
pip install --no-cache-dir torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/cu$env:TORCH_CUDA_VERSION
|
||||
}
|
||||
|
||||
nvcc --version
|
||||
python -V
|
||||
python -c "import torch; print('PyTorch:', torch.__version__)"
|
||||
python -c "import torch; print('CUDA:', torch.version.cuda)"
|
||||
python -c "from torch.utils import cpp_extension; print(cpp_extension.CUDA_HOME)"
|
||||
shell: pwsh
|
||||
|
||||
- name: Checkout flash-attn
|
||||
run: |
|
||||
git clone https://github.com/Dao-AILab/flash-attention.git -b "v${{ matrix.flash-attn-version }}"
|
||||
shell: pwsh
|
||||
|
||||
- name: Build wheels
|
||||
timeout-minutes: 800
|
||||
run: |
|
||||
Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
|
||||
Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools' -DevCmdArguments '-arch=x64 -host_arch=x64'
|
||||
$env:DISTUTILS_USE_SDK = 1
|
||||
$env:BUILD_TARGET = "cuda"
|
||||
|
||||
$env:MAX_JOBS = "2"
|
||||
$env:NVCC_THREADS = "2"
|
||||
$env:FLASH_ATTENTION_FORCE_BUILD = "TRUE"
|
||||
|
||||
cd flash-attention
|
||||
python setup.py bdist_wheel --dist-dir=dist
|
||||
$baseWheelName = Get-ChildItem -Path "dist\*.whl" | Select-Object -First 1 | ForEach-Object { $_.Name }
|
||||
$wheelName = $baseWheelName.Replace("${{ matrix.flash-attn-version }}", "${{ matrix.flash-attn-version }}+cu$env:MATRIX_CUDA_VERSION" + "torch$env:MATRIX_TORCH_VERSION")
|
||||
Move-Item "dist\$baseWheelName" "dist\$wheelName"
|
||||
echo "wheel_name=$wheelName" >> $env:GITHUB_ENV
|
||||
shell: pwsh
|
||||
run: |
|
||||
.\build_windows.ps1 -FlashAttnVersion "${{ matrix.flash-attn-version }}" -PythonVersion "${{ matrix.python-version }}" -TorchVersion "${{ matrix.torch-version }}" -CudaVersion "${{ matrix.cuda-version }}"
|
||||
$wheelName = Get-ChildItem -Path "flash-attention\dist\*.whl" | Select-Object -First 1 | ForEach-Object { $_.Name }
|
||||
echo "wheel_name=$wheelName" >> $env:GITHUB_ENV
|
||||
|
||||
- name: Install Test
|
||||
shell: pwsh
|
||||
run: |
|
||||
pip install flash-attention/dist/$env:wheel_name
|
||||
python -c "import flash_attn; print(flash_attn.__version__)"
|
||||
shell: pwsh
|
||||
|
||||
@@ -0,0 +1,71 @@
|
||||
param(
|
||||
[Parameter(Mandatory=$true)]
|
||||
[string]$FlashAttnVersion,
|
||||
|
||||
[Parameter(Mandatory=$true)]
|
||||
[string]$PythonVersion,
|
||||
|
||||
[Parameter(Mandatory=$true)]
|
||||
[string]$TorchVersion,
|
||||
|
||||
[Parameter(Mandatory=$true)]
|
||||
[string]$CudaVersion
|
||||
)
|
||||
|
||||
# Error handling
|
||||
$ErrorActionPreference = "Stop"
|
||||
|
||||
Write-Host "Building Flash Attention with parameters:"
|
||||
Write-Host " Flash-Attention: $FlashAttnVersion"
|
||||
Write-Host " Python: $PythonVersion"
|
||||
Write-Host " PyTorch: $TorchVersion"
|
||||
Write-Host " CUDA: $CudaVersion"
|
||||
|
||||
# Set CUDA and PyTorch versions
|
||||
$CudaVersionClean = $CudaVersion -replace '\.', ''
|
||||
$MatrixCudaVersion = $CudaVersionClean.Substring(0, [Math]::Min(3, $CudaVersionClean.Length))
|
||||
$MatrixTorchVersion = $TorchVersion -replace '^(\d+\.\d+).*', '$1'
|
||||
|
||||
Write-Host "Derived versions:"
|
||||
Write-Host " CUDA Matrix: $MatrixCudaVersion"
|
||||
Write-Host " Torch Matrix: $MatrixTorchVersion"
|
||||
|
||||
# Install PyTorch
|
||||
Write-Host "Installing PyTorch $TorchVersion+cu$CudaVersion..."
|
||||
$env:TORCH_CUDA_VERSION = python -c "from os import environ as env; support_cuda_versions = { '2.1': [121], '2.2': [121], '2.3': [121], '2.4': [121, 124], '2.5': [121, 124], '2.6': [124, 126], '2.7': [126, 128], '2.8': [128], }; target_cuda_versions = support_cuda_versions['$MatrixTorchVersion']; cuda_version = int('$MatrixCudaVersion'); closest_version = min(target_cuda_versions, key=lambda x: abs(x - cuda_version)); print(closest_version)"
|
||||
|
||||
if ($TorchVersion -like "*dev*") {
|
||||
pip install --pre torch=$TorchVersion --index-url https://download.pytorch.org/whl/nightly/cu$env:TORCH_CUDA_VERSION
|
||||
} else {
|
||||
pip install --no-cache-dir torch=$TorchVersion --index-url https://download.pytorch.org/whl/cu$env:TORCH_CUDA_VERSION
|
||||
}
|
||||
|
||||
# Verify installation
|
||||
Write-Host "Verifying installations..."
|
||||
nvcc --version
|
||||
python -V
|
||||
python -c "import torch; print('PyTorch:', torch.__version__)"
|
||||
python -c "import torch; print('CUDA:', torch.version.cuda)"
|
||||
python -c "from torch.utils import cpp_extension; print(cpp_extension.CUDA_HOME)"
|
||||
|
||||
# Checkout flash-attn
|
||||
Write-Host "Checking out flash-attention v$FlashAttnVersion..."
|
||||
git clone https://github.com/Dao-AILab/flash-attention.git -b "v$FlashAttnVersion"
|
||||
|
||||
# Build wheels
|
||||
Write-Host "Building wheels..."
|
||||
Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
|
||||
Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools' -DevCmdArguments '-arch=x64 -host_arch=x64'
|
||||
$env:DISTUTILS_USE_SDK = 1
|
||||
$env:BUILD_TARGET = "cuda"
|
||||
# Use environment variables from workflow if available, otherwise use defaults
|
||||
if (-not $env:MAX_JOBS) { $env:MAX_JOBS = "2" }
|
||||
if (-not $env:NVCC_THREADS) { $env:NVCC_THREADS = "2" }
|
||||
$env:FLASH_ATTENTION_FORCE_BUILD = "TRUE"
|
||||
|
||||
cd flash-attention
|
||||
python setup.py bdist_wheel --dist-dir=dist
|
||||
$baseWheelName = Get-ChildItem -Path "dist\*.whl" | Select-Object -First 1 | ForEach-Object { $_.Name }
|
||||
$wheelName = $baseWheelName.Replace($FlashAttnVersion, "$FlashAttnVersion+cu$MatrixCudaVersion" + "torch$MatrixTorchVersion")
|
||||
Move-Item "dist\$baseWheelName" "dist\$wheelName"
|
||||
Write-Host "Built wheel: $wheelName"
|
||||
Reference in New Issue
Block a user