add flash-attn v2.8.0

This commit is contained in:
Junya Morioka
2025-06-14 23:18:10 +09:00
parent 598eaf21ad
commit d1945566ef
+95 -95
View File
@@ -21,7 +21,7 @@ jobs:
| Flash-Attention | Python | PyTorch | CUDA | | Flash-Attention | Python | PyTorch | CUDA |
| --- | --- | --- | --- | | --- | --- | --- | --- |
| 2.7.4 | 3.10, 3.11, 3.12 | 2.7.1 | 12.8.1 |" | 2.8.0 | 3.10, 3.11, 3.12 | 2.4.1,2.5.1,2.6.0,2.7.1 | 12.1.1,12.4.1,12.8.1 |"
# ## Windows x86_64 # ## Windows x86_64
@@ -44,11 +44,11 @@ jobs:
strategy: strategy:
fail-fast: false fail-fast: false
matrix: matrix:
flash-attn-version: ["2.4.3", "2.5.9", "2.6.3"] flash-attn-version: ["2.8.0"]
python-version: ["3.10", "3.11", "3.12"] python-version: ["3.10", "3.11", "3.12"]
torch-version: ["2.7.1"] torch-version: ["2.4.1","2.5.1","2.6.0"]
# # https://developer.nvidia.com/cuda-toolkit-archive # # https://developer.nvidia.com/cuda-toolkit-archive
cuda-version: ["12.8.1"] cuda-version: ["12.1.1","12.4.1"]
exclude: exclude:
# torch < 2.2 does not support Python 3.12 # torch < 2.2 does not support Python 3.12
- python-version: "3.12" - python-version: "3.12"
@@ -158,9 +158,9 @@ jobs:
strategy: strategy:
fail-fast: false fail-fast: false
matrix: matrix:
flash-attn-version: ["2.7.4"] flash-attn-version: ["2.8.0"]
python-version: ["3.10", "3.11", "3.12"] python-version: ["3.10", "3.11", "3.12"]
torch-version: ["2.7.1"] torch-version: ["2.4.1","2.5.1","2.6.0","2.7.1"]
# https://developer.nvidia.com/cuda-toolkit-archive # https://developer.nvidia.com/cuda-toolkit-archive
cuda-version: ["12.8.1"] cuda-version: ["12.8.1"]
exclude: exclude:
@@ -284,104 +284,104 @@ jobs:
run: | run: |
sudo rm -rf /opt/hostedtoolcache/Python/${{ matrix.python-version }}* sudo rm -rf /opt/hostedtoolcache/Python/${{ matrix.python-version }}*
build_windows_wheels: # build_windows_wheels:
name: Build wheels and Upload (Windows x86_64, GitHub hosted runner) # name: Build wheels and Upload (Windows x86_64, GitHub hosted runner)
runs-on: windows-latest # runs-on: windows-latest
timeout-minutes: 1000 # timeout-minutes: 1000
env: # env:
MAX_JOBS: 2 # MAX_JOBS: 2
NVCC_THREADS: 2 # NVCC_THREADS: 2
strategy: # strategy:
fail-fast: false # fail-fast: false
matrix: # matrix:
flash-attn-version: ["2.5.9", "2.6.3", "2.7.4"] # flash-attn-version: ["2.5.9", "2.6.3", "2.7.4"]
python-version: ["3.10", "3.11", "3.12"] # python-version: ["3.10", "3.11", "3.12"]
torch-version: ["2.4.1", "2.5.1", "2.6.0", "2.7.1"] # torch-version: ["2.4.1", "2.5.1", "2.6.0", "2.7.1"]
# https://developer.nvidia.com/cuda-toolkit-archive # # https://developer.nvidia.com/cuda-toolkit-archive
# CUDA 11.8 cannot build flash-attn. # # CUDA 11.8 cannot build flash-attn.
# https://github.com/Dao-AILab/flash-attention/issues/595 # # https://github.com/Dao-AILab/flash-attention/issues/595
cuda-version: ["12.4.1", "12.8.1"] # cuda-version: ["12.4.1", "12.8.1"]
exclude: # exclude:
# torch < 2.2 does not support Python 3.12 # # torch < 2.2 does not support Python 3.12
- python-version: "3.12" # - python-version: "3.12"
torch-version: "2.0.1" # torch-version: "2.0.1"
- python-version: "3.12" # - python-version: "3.12"
torch-version: "2.1.2" # torch-version: "2.1.2"
# torch 2.0.1 does not support CUDA 12.x # # torch 2.0.1 does not support CUDA 12.x
- torch-version: "2.0.1" # - torch-version: "2.0.1"
cuda-version: "12.1.1" # cuda-version: "12.1.1"
- torch-version: "2.0.1" # - torch-version: "2.0.1"
cuda-version: "12.4.1" # cuda-version: "12.4.1"
- torch-version: "2.0.1" # - torch-version: "2.0.1"
cuda-version: "12.6.3" # cuda-version: "12.6.3"
- torch-version: "2.0.1" # - torch-version: "2.0.1"
cuda-version: "12.8.1" # cuda-version: "12.8.1"
# torch 2.7.0 does not support CUDA 12.4 # # torch 2.7.0 does not support CUDA 12.4
- torch-version: "2.7.0" # - torch-version: "2.7.0"
cuda-version: "12.4.1" # cuda-version: "12.4.1"
steps: # steps:
- uses: actions/checkout@v4 # - uses: actions/checkout@v4
- name: Enable Git long paths # - name: Enable Git long paths
shell: pwsh # shell: pwsh
run: git config --system core.longpaths true # run: git config --system core.longpaths true
- uses: actions/setup-python@v5 # - uses: actions/setup-python@v5
with: # with:
python-version: ${{ matrix.python-version }} # python-version: ${{ matrix.python-version }}
- uses: Jimver/cuda-toolkit@v0.2.24 # - uses: Jimver/cuda-toolkit@v0.2.24
with: # with:
cuda: ${{ matrix.cuda-version }} # cuda: ${{ matrix.cuda-version }}
method: "network" # method: "network"
use-github-cache: false # use-github-cache: false
use-local-cache: false # use-local-cache: false
- name: Install VS2022 BuildTools # - name: Install VS2022 BuildTools
shell: pwsh # shell: pwsh
run: | # run: |
choco install -y visualstudio2022buildtools ` # choco install -y visualstudio2022buildtools `
--version=117.14.1 ` # --version=117.14.1 `
--params "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64" # --params "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64"
- name: Install build dependencies # - name: Install build dependencies
shell: pwsh # shell: pwsh
run: | # run: |
pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil ninja # pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil ninja
- name: Build wheels # - name: Build wheels
shell: pwsh # shell: pwsh
run: | # run: |
.\build_windows.ps1 -FlashAttnVersion "${{ matrix.flash-attn-version }}" -PythonVersion "${{ matrix.python-version }}" -TorchVersion "${{ matrix.torch-version }}" -CudaVersion "${{ matrix.cuda-version }}" # .\build_windows.ps1 -FlashAttnVersion "${{ 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 } # $wheelName = Get-ChildItem -Path "flash-attention\dist\*.whl" | Select-Object -First 1 | ForEach-Object { $_.Name }
echo "wheel_name=$wheelName" >> $env:GITHUB_ENV # echo "wheel_name=$wheelName" >> $env:GITHUB_ENV
- name: Install Test # - name: Install Test
shell: pwsh # shell: pwsh
run: | # run: |
pip install flash-attention/dist/$env:wheel_name # pip install flash-attention/dist/$env:wheel_name
python -c "import flash_attn; print(flash_attn.__version__)" # python -c "import flash_attn; print(flash_attn.__version__)"
- name: Upload Release Asset # - name: Upload Release Asset
shell: pwsh # shell: pwsh
env: # env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} # GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
run: | # run: |
$tag_name = "${env:GITHUB_REF}".Replace("refs/tags/", "") # $tag_name = "${env:GITHUB_REF}".Replace("refs/tags/", "")
$wheel_path = "flash-attention/dist/$env:wheel_name" # $wheel_path = "flash-attention/dist/$env:wheel_name"
# Check if the file exists # # Check if the file exists
if (-not (Test-Path $wheel_path)) { # if (-not (Test-Path $wheel_path)) {
$tag_name = "${env:GITHUB_REF}".Replace("refs/tags/", "") # $tag_name = "${env:GITHUB_REF}".Replace("refs/tags/", "")
$wheel_path = "flash-attention/dist/$env:wheel_name" # $wheel_path = "flash-attention/dist/$env:wheel_name"
# Check if the file exists # # Check if the file exists
if (-not (Test-Path $wheel_path)) { # if (-not (Test-Path $wheel_path)) {
Write-Host "Error: Wheel file not found at $wheel_path" # Write-Host "Error: Wheel file not found at $wheel_path"
exit 1 # exit 1
} # }
# Upload the release asset using GitHub CLI # # Upload the release asset using GitHub CLI
gh release upload "$tag_name" "$wheel_path" --clobber # gh release upload "$tag_name" "$wheel_path" --clobber
Write-Host "Successfully uploaded $env:wheel_name to release $tag_name" # Write-Host "Successfully uploaded $env:wheel_name to release $tag_name"