update workflow

This commit is contained in:
Junya Morioka
2025-06-15 21:19:23 +09:00
parent dedf5eb8d5
commit e15e8237d4
5 changed files with 143 additions and 373 deletions
+16 -13
View File
@@ -39,16 +39,14 @@ jobs:
env:
DEBIAN_FRONTEND: noninteractive
TERM: xterm-256color
MAX_JOBS: 2
NVCC_THREADS: 2
strategy:
fail-fast: false
matrix:
flash-attn-version: ["2.8.0"]
python-version: ["3.10", "3.11", "3.12"]
torch-version: ["2.4.1","2.5.1","2.6.0"]
torch-version: ["2.4.1", "2.5.1", "2.6.0"]
# # https://developer.nvidia.com/cuda-toolkit-archive
cuda-version: ["12.1.1","12.4.1"]
cuda-version: ["12.1.1", "12.4.1"]
exclude:
# torch < 2.2 does not support Python 3.12
- python-version: "3.12"
@@ -111,6 +109,9 @@ jobs:
export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
- name: Build wheels
env:
MAX_JOBS: 2
NVCC_THREADS: 2
run: |
chmod +x build_linux.sh
./build_linux.sh ${{ matrix.flash-attn-version }} ${{ matrix.python-version }} ${{ matrix.torch-version }} ${{ matrix.cuda-version }}
@@ -119,7 +120,7 @@ jobs:
- name: Install Test
run: |
pip install flash-attention/dist/${{ env.wheel_name }}
pip install --no-cache-dir --force-reinstall flash-attention/dist/${{ env.wheel_name }}
python -c "import flash_attn; print(flash_attn.__version__)"
- name: Upload Release Asset
@@ -155,15 +156,13 @@ jobs:
env:
DEBIAN_FRONTEND: noninteractive
TERM: xterm-256color
MAX_JOBS: 2
NVCC_THREADS: 2
timeout-minutes: 1000
timeout-minutes: 2000
strategy:
fail-fast: false
matrix:
flash-attn-version: ["2.8.0"]
python-version: ["3.10", "3.11", "3.12"]
torch-version: ["2.4.1","2.5.1","2.6.0","2.7.1"]
torch-version: ["2.4.1", "2.5.1", "2.6.0", "2.7.1"]
# https://developer.nvidia.com/cuda-toolkit-archive
cuda-version: ["12.8.1"]
exclude:
@@ -251,8 +250,11 @@ jobs:
export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
- name: Build wheels
timeout-minutes: 800
timeout-minutes: 1200
shell: bash
env:
MAX_JOBS: 4
NVCC_THREADS: 3
run: |
chmod +x build_linux.sh
./build_linux.sh ${{ matrix.flash-attn-version }} ${{ matrix.python-version }} ${{ matrix.torch-version }} ${{ matrix.cuda-version }}
@@ -262,7 +264,7 @@ jobs:
- name: Install Test
shell: bash
run: |
pip install flash-attention/dist/${{ env.wheel_name }}
pip install --no-cache-dir --force-reinstall flash-attention/dist/${{ env.wheel_name }}
python -c "import flash_attn; print(flash_attn.__version__)"
- name: Upload Release Asset
@@ -288,7 +290,8 @@ jobs:
shell: bash
if: always()
run: |
sudo rm -rf /opt/hostedtoolcache/Python/${{ matrix.python-version }}*
rm -rf /opt/hostedtoolcache/Python
rm -rf ~/.cache/pip
# build_windows_wheels:
# name: Build wheels and Upload (Windows x86_64, GitHub hosted runner)
@@ -365,7 +368,7 @@ jobs:
# - name: Install Test
# shell: pwsh
# run: |
# pip install flash-attention/dist/$env:wheel_name
# pip install --no-cache-dir --force-reinstall flash-attention/dist/$env:wheel_name
# python -c "import flash_attn; print(flash_attn.__version__)"
# - name: Upload Release Asset
@@ -0,0 +1,118 @@
name: Build wheels and upload to GitHub Releases
on:
push:
tags:
- "v*"
jobs:
# #########################################################
# Build wheels with self-hosted runner
# #########################################################
build_wheels_self_hosted:
name: Build wheels and Upload (Linux x86_64, self-hosted runner)
runs-on: self-hosted
container:
image: ubuntu:22.04
defaults:
run:
shell: bash
env:
DEBIAN_FRONTEND: noninteractive
TERM: xterm-256color
timeout-minutes: 2000
strategy:
fail-fast: false
matrix:
flash-attn-version: ["2.8.0"]
python-version: ["3.11"]
torch-version: ["2.7.1"]
# https://developer.nvidia.com/cuda-toolkit-archive
cuda-version: ["12.8.1"]
steps:
- name: Install tools
shell: bash
run: |
apt-get update && apt-get install -y --no-install-recommends \
curl \
ca-certificates \
sudo \
software-properties-common \
wget \
unzip \
zip \
git \
build-essential \
gcc \
g++ \
clang \
ninja-build \
keyboard-configuration
- name: Install gh
shell: bash
run: |
sudo mkdir -p -m 755 /etc/apt/keyrings
out=$(mktemp)
wget -nv -O$out https://cli.github.com/packages/githubcli-archive-keyring.gpg
cat $out | sudo tee /etc/apt/keyrings/githubcli-archive-keyring.gpg > /dev/null
sudo chmod go+r /etc/apt/keyrings/githubcli-archive-keyring.gpg
echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" | sudo tee /etc/apt/sources.list.d/github-cli.list > /dev/null
sudo apt update
sudo apt install gh -y
- uses: actions/checkout@v4
- name: Configure Git safe directory
shell: bash
run: |
git config --global --add safe.directory $(pwd)
- uses: actions/setup-python@v5
with:
python-version: ${{ matrix.python-version }}
- uses: Jimver/cuda-toolkit@master
with:
cuda: ${{ matrix.cuda-version }}
sub-packages: '["nvcc", "toolkit"]'
method: "network"
use-github-cache: false
use-local-cache: false
- name: Install build dependencies
shell: bash
run: |
sudo apt install -y ninja-build clang
pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil
- name: Set environment variables
shell: bash
run: |
export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
- name: Build wheels
timeout-minutes: 1200
shell: bash
env:
MAX_JOBS: 4
NVCC_THREADS: 3
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" >> $GITHUB_ENV
- name: Install Test
shell: bash
run: |
pip install --no-cache-dir --force-reinstall flash-attention/dist/${{ env.wheel_name }}
python -c "import flash_attn; print(flash_attn.__version__)"
- name: Clean up
shell: bash
if: always()
run: |
rm -rf /opt/hostedtoolcache/Python
rm -rf ~/.cache/pip
-130
View File
@@ -1,130 +0,0 @@
name: "Test self-hosted runner"
on:
workflow_dispatch:
jobs:
docker:
runs-on:
- self-hosted
timeout-minutes: 10
steps:
- run: ls
- run: pwd
containers-test:
runs-on:
- self-hosted
container:
image: ubuntu:22.04
timeout-minutes: 10
steps:
- run: ls
- run: pwd
test-build:
runs-on:
- self-hosted
timeout-minutes: 1000
container:
image: ubuntu:22.04
env:
DEBIAN_FRONTEND: noninteractive
TERM: xterm-256color
strategy:
fail-fast: false
matrix:
flash-attn-version: ["2.4.3"]
python-version: ["3.12"]
torch-version: ["2.7.0"]
# https://developer.nvidia.com/cuda-toolkit-archive
cuda-version: ["11.8.0", "12.8.1"]
steps:
- name: Install tools
run: |
apt-get update && apt-get install -y --no-install-recommends \
curl \
ca-certificates \
sudo \
software-properties-common \
wget \
unzip \
zip \
git \
build-essential \
gcc \
g++ \
clang \
ninja-build \
keyboard-configuration
- uses: actions/checkout@v4
- uses: actions/setup-python@v5
with:
python-version: ${{ matrix.python-version }}
- uses: Jimver/cuda-toolkit@master
with:
cuda: ${{ matrix.cuda-version }}
linux-local-args: '["--toolkit"]'
method: "network"
env:
DEBIAN_FRONTEND: noninteractive
- name: Set CUDA and PyTorch versions
run: |
echo "MATRIX_CUDA_VERSION=$(echo ${{ matrix.cuda-version }} | awk -F \. {'print $1 $2'})" >> $GITHUB_ENV
echo "MATRIX_TORCH_VERSION=$(echo ${{ matrix.torch-version }} | awk -F \. {'print $1 "." $2'})" >> $GITHUB_ENV
echo "CACHE_KEY=cuda-ext-${{ matrix.flash-attn-version }}-py${{ matrix.python-version }}-torch${{ matrix.torch-version }}-cuda${{ matrix.cuda-version }}" >> $GITHUB_ENV
- name: Install build dependencies
run: |
pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil
- name: Install PyTorch ${{ matrix.torch-version }}+cu${{ matrix.cuda-version }}
run: |
export TORCH_CUDA_VERSION=$(python -c "from os import environ as env; \
support_cuda_versions = { \
'2.0': [117, 118], \
'2.1': [118, 121], \
'2.2': [118, 121], \
'2.3': [118, 121], \
'2.4': [118, 121, 124], \
'2.5': [118, 121, 124], \
'2.6': [118, 124, 126], \
'2.7': [118, 126, 128], \
}; \
target_cuda_versions = support_cuda_versions[env['MATRIX_TORCH_VERSION']]; \
cuda_version = int(env['MATRIX_CUDA_VERSION']); \
closest_version = min(target_cuda_versions, key=lambda x: abs(x - cuda_version)); \
print(closest_version) \
")
pip install --no-cache-dir torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
nvcc --version
python -V
python -c "import torch; print('PyTorch:', torch.__version__)"
python -c "import torch; print('CUDA:', torch.version.cuda)"
python -c "from torch.utils import cpp_extension; print(cpp_extension.CUDA_HOME)"
- name: Checkout flash-attn
run: |
git clone https://github.com/Dao-AILab/flash-attention.git -b "v${{ matrix.flash-attn-version }}"
cd flash-attention
- name: Build wheels
timeout-minutes: 800
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
export FLASH_ATTENTION_FORCE_BUILD="TRUE"
cd flash-attention
python setup.py bdist_wheel --dist-dir=dist
base_wheel_name=$(basename $(ls dist/*.whl | head -n 1))
wheel_name=$(echo $base_wheel_name | sed "s/${{ matrix.flash-attn-version }}/${{ matrix.flash-attn-version }}+cu${{ env.MATRIX_CUDA_VERSION }}torch${{ env.MATRIX_TORCH_VERSION }}/")
mv dist/$base_wheel_name dist/$wheel_name
echo "wheel_name=$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__)"
-224
View File
@@ -1,224 +0,0 @@
name: Test workflow
on:
workflow_dispatch:
jobs:
# #########################################################
# Build wheels with GitHub hosted runner
# #########################################################
build_wheels:
name: Build wheels
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: ["2.8.0.dev20250523"]
# # 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: 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: Jimver/cuda-toolkit@master
with:
cuda: ${{ matrix.cuda-version }}
sub-packages: '["nvcc", "toolkit"]'
method: "network"
- name: Install build dependencies
run: |
sudo apt install -y ninja-build clang
pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil
- name: Set environment variables
run: |
export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
export MAX_JOBS=2
export NVCC_THREADS=2
export FLASH_ATTENTION_FORCE_BUILD=TRUE
- 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" >> $GITHUB_ENV
- name: Install Test
run: |
pip install flash-attention/dist/${{ env.wheel_name }}
python -c "import flash_attn; print(flash_attn.__version__)"
# #########################################################
# Build wheels with self-hosted runner
# #########################################################
# build_wheels_self_hosted:
# name: Build wheels
# runs-on: self-hosted
# container:
# image: ubuntu:22.04
# defaults:
# run:
# shell: bash
# env:
# DEBIAN_FRONTEND: noninteractive
# TERM: xterm-256color
# timeout-minutes: 1000
# strategy:
# fail-fast: false
# matrix:
# flash-attn-version: ["2.7.4"]
# python-version: ["3.10", "3.11", "3.12"]
# torch-version: ["2.8.0.dev20250523"]
# # 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:
# - name: Install tools
# shell: bash
# run: |
# apt-get update && apt-get install -y --no-install-recommends \
# curl \
# ca-certificates \
# sudo \
# software-properties-common \
# wget \
# unzip \
# zip \
# git \
# build-essential \
# gcc \
# g++ \
# clang \
# ninja-build \
# keyboard-configuration
# - uses: actions/checkout@v4
# - uses: actions/setup-python@v5
# with:
# python-version: ${{ matrix.python-version }}
# - uses: Jimver/cuda-toolkit@master
# env:
# DEBIAN_FRONTEND: noninteractive
# with:
# cuda: ${{ matrix.cuda-version }}
# sub-packages: '["nvcc", "toolkit"]'
# method: "network"
# - name: Install build dependencies
# run: |
# pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil
# - name: Set environment variables
# run: |
# export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
# export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
# export MAX_JOBS=2
# export NVCC_THREADS=2
# export FLASH_ATTENTION_FORCE_BUILD=TRUE
# - 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" >> $GITHUB_ENV
# - name: Install Test
# run: |
# pip install flash-attention/dist/${{ env.wheel_name }}
# python -c "import flash_attn; print(flash_attn.__version__)"
# - name: Get the tag version
# id: extract_branch
# shell: bash
# run: echo "branch=${GITHUB_REF#refs/tags/}" >> $GITHUB_OUTPUT
# - name: Get Release with Tag
# id: get_release
# uses: joutvhu/get-release@v1
# env:
# GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
# with:
# tag_name: ${{ steps.extract_branch.outputs.branch }}
# - name: Upload Release Asset
# uses: actions/upload-release-asset@v1
# env:
# GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
# with:
# upload_url: ${{ steps.get_release.outputs.upload_url }}
# asset_path: flash-attention/dist/${{ env.wheel_name }}
# asset_name: ${{ env.wheel_name }}
# asset_content_type: application/*
# - name: Clean up
# if: always()
# run: |
# sudo rm -rf /opt/hostedtoolcache/Python/${{ matrix.python-version }}*
+9 -6
View File
@@ -24,7 +24,7 @@ echo " Torch Matrix: $MATRIX_TORCH_VERSION"
# Install PyTorch
echo "Installing PyTorch $TORCH_VERSION+cu$CUDA_VERSION..."
TORCH_CUDA_VERSION=$(python -c "from os import environ as env; \
TORCH_CUDA_VERSION=$(python -c "\
support_cuda_versions = { \
'2.0': [117, 118], \
'2.1': [118, 121], \
@@ -36,18 +36,21 @@ TORCH_CUDA_VERSION=$(python -c "from os import environ as env; \
'2.7': [118, 126, 128], \
'2.8': [128], \
}; \
matrix_cuda_version = '$MATRIX_CUDA_VERSION'; \
cuda_version = int('$MATRIX_CUDA_VERSION'); \
matrix_torch_version = '$MATRIX_TORCH_VERSION'; \
target_cuda_versions = support_cuda_versions[matrix_torch_version]; \
cuda_version = int(matrix_cuda_version); \
closest_version = min(target_cuda_versions, key=lambda x: abs(x - cuda_version)); \
target_cuda_versions = [v for v in target_cuda_versions if str(v)[:2] == str(cuda_version)[:2]]; \
if len(target_cuda_versions) == 0: \
closest_version = support_cuda_versions[matrix_torch_version][-1]; \
else: \
closest_version = min(target_cuda_versions, key=lambda x: abs(x - cuda_version)); \
print(closest_version) \
")
if [[ $TORCH_VERSION == *"dev"* ]]; then
pip install --pre torch==$TORCH_VERSION --index-url https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}
pip install --force-reinstall --no-cache-dir --pre torch==$TORCH_VERSION --index-url https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}
else
pip install --no-cache-dir torch==$TORCH_VERSION --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
pip install --force-reinstall --no-cache-dir torch==$TORCH_VERSION --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
fi
# Verify installation