Files
ktransformers/.github/workflows/release-pypi.yml
2025-12-29 12:42:06 +08:00

319 lines
12 KiB
YAML

name: Release to PyPI
on:
push:
branches:
- main
paths:
- "version.py"
workflow_dispatch:
inputs:
test_pypi:
description: 'Publish to TestPyPI instead of PyPI (for testing)'
required: false
default: 'false'
type: choice
options:
- 'true'
- 'false'
permissions:
contents: read
jobs:
build-kt-kernel:
name: Build kt-kernel CPU-only (Python ${{ matrix.python-version }})
runs-on: [self-hosted, linux, x64]
strategy:
fail-fast: false
matrix:
python-version: ['3.10', '3.11', '3.12']
steps:
- name: Checkout repository
uses: actions/checkout@v4
with:
submodules: recursive
- name: Set up Python ${{ matrix.python-version }}
uses: actions/setup-python@v4
with:
python-version: ${{ matrix.python-version }}
- name: Install system dependencies
run: |
apt-get update
apt-get install -y cmake libhwloc-dev pkg-config libnuma-dev
- name: Install Python build tools
run: |
python -m pip install --upgrade pip
pip install build wheel setuptools
- name: Build kt-kernel wheel (CPU-only, multi-variant)
working-directory: kt-kernel
env:
CPUINFER_BUILD_ALL_VARIANTS: '1'
CPUINFER_USE_CUDA: '0'
CPUINFER_BUILD_TYPE: 'Release'
CPUINFER_PARALLEL: '4'
CPUINFER_FORCE_REBUILD: '1'
run: |
echo "Building kt-kernel CPU-only with all CPU variants (AMX, AVX512, AVX2)"
python -m build --wheel -v
- name: List generated wheels
working-directory: kt-kernel
run: |
echo "Generated wheels:"
ls -lh dist/
- name: Test wheel import
working-directory: kt-kernel
run: |
pip install dist/*.whl
python -c "import kt_kernel; print('✓ Import successful'); print(f'CPU variant detected: {kt_kernel.__cpu_variant__}'); print(f'Version: {kt_kernel.__version__}')"
- name: Verify wheel contains all variants
working-directory: kt-kernel
run: |
echo "Checking wheel contents for CPU variants..."
python -m zipfile -l dist/*.whl | grep "_kt_kernel_ext_" || echo "ERROR: No variant .so files found!"
python -m zipfile -l dist/*.whl | grep "_kt_kernel_ext_amx.cpython" && echo "✓ AMX variant found" || echo "✗ AMX variant missing"
python -m zipfile -l dist/*.whl | grep "_kt_kernel_ext_avx512.cpython" && echo "✓ AVX512 variant found" || echo "✗ AVX512 variant missing"
python -m zipfile -l dist/*.whl | grep "_kt_kernel_ext_avx2.cpython" && echo "✓ AVX2 variant found" || echo "✗ AVX2 variant missing"
- name: Repair wheel for manylinux compatibility
working-directory: kt-kernel
run: |
pip install auditwheel patchelf
echo "Repairing wheels for manylinux compatibility..."
mkdir -p wheelhouse
for wheel in dist/*.whl; do
echo "Processing $wheel..."
success=0
# Try different manylinux versions (newest to oldest)
for plat in manylinux_2_31_x86_64 manylinux_2_28_x86_64 manylinux_2_17_x86_64; do
echo " Trying $plat..."
if auditwheel repair "$wheel" --plat "$plat" -w wheelhouse/ 2>&1; then
echo " ✓ Successfully repaired with $plat"
success=1
break
fi
done
# If all auditwheel attempts failed, use rename fallback
if [ $success -eq 0 ]; then
echo " Warning: auditwheel repair failed, using rename fallback..."
wheel_name=$(basename "$wheel")
# Use # as sed delimiter to avoid conflict with /
new_name=$(echo "$wheel_name" | sed 's#linux_x86_64#manylinux_2_17_x86_64#')
cp "$wheel" "wheelhouse/$new_name"
echo " ✓ Renamed to $new_name"
fi
done
echo "Repaired wheels:"
ls -lh wheelhouse/
# Verify all wheels contain 3 CPU variants
echo "Verifying CPU variants in repaired wheels..."
for wheel in wheelhouse/*.whl; do
echo "Checking $(basename $wheel):"
python -m zipfile -l "$wheel" | grep "\.so" | grep -E "(amx|avx512|avx2)"
done
# Replace original wheels with repaired ones
rm -f dist/*.whl
cp wheelhouse/*.whl dist/
- name: Upload wheel artifact
uses: actions/upload-artifact@v4
with:
name: kt-kernel-wheels-py${{ matrix.python-version }}
path: kt-kernel/dist/*.whl
retention-days: 7
build-kt-kernel-cuda:
name: Build kt-kernel CUDA (Python ${{ matrix.python-version }})
runs-on: [self-hosted, linux, x64, gpu]
strategy:
fail-fast: false
matrix:
python-version: ['3.10', '3.11', '3.12']
steps:
- name: Checkout repository
uses: actions/checkout@v4
with:
submodules: recursive
- name: Set up Python ${{ matrix.python-version }}
uses: actions/setup-python@v4
with:
python-version: ${{ matrix.python-version }}
- name: Verify CUDA availability
run: |
nvidia-smi || (echo "ERROR: GPU not available" && exit 1)
nvcc --version || (echo "ERROR: CUDA toolkit not found" && exit 1)
- name: Install dependencies
run: |
apt-get update && apt-get install -y cmake libhwloc-dev pkg-config libnuma-dev
python -m pip install --upgrade pip
pip install build wheel setuptools torch --index-url https://download.pytorch.org/whl/cu118
- name: Build CUDA wheel
working-directory: kt-kernel
env:
CPUINFER_BUILD_ALL_VARIANTS: '1'
CPUINFER_USE_CUDA: '1'
CPUINFER_CUDA_ARCHS: '80;86;89;90'
CPUINFER_CUDA_STATIC_RUNTIME: '1'
CPUINFER_BUILD_TYPE: 'Release'
CPUINFER_PARALLEL: '4'
CPUINFER_FORCE_REBUILD: '1'
CUDA_HOME: '/usr/local/cuda-11.8'
run: |
echo "Building CUDA wheel with multi-CPU-variant support (AMX, AVX512, AVX2)"
echo "CUDA architectures for GPU sync: SM 80, 86, 89, 90"
python -m build --wheel -v
- name: Verify wheel
working-directory: kt-kernel
run: |
ls -lh dist/
# Check version suffix
[[ $(ls dist/*.whl) == *"+cuda118"* ]] || (echo "ERROR: Missing +cuda118 suffix" && exit 1)
# Install and test
pip install dist/*.whl
python -c "import kt_kernel; print(f'Version: {kt_kernel.__version__}')"
# Verify static linking (should NOT depend on libcudart.so)
rm -rf /tmp/check
unzip -q dist/*.whl -d /tmp/check
! ldd /tmp/check/kt_kernel/*.so | grep -q "libcudart.so" || (echo "ERROR: Dynamic cudart found" && exit 1)
echo "✓ CUDA runtime statically linked"
- name: Repair wheel for manylinux
working-directory: kt-kernel
run: |
pip install auditwheel patchelf
mkdir -p wheelhouse
for wheel in dist/*.whl; do
auditwheel repair "$wheel" --plat manylinux_2_17_x86_64 --exclude libcuda.so.1 -w wheelhouse/ || \
cp "$wheel" wheelhouse/$(basename "$wheel" | sed 's/linux_x86_64/manylinux_2_17_x86_64/')
done
rm -f dist/*.whl && cp wheelhouse/*.whl dist/
- name: Upload artifact
uses: actions/upload-artifact@v4
with:
name: kt-kernel-cuda-wheels-py${{ matrix.python-version }}
path: kt-kernel/dist/*.whl
retention-days: 7
publish-pypi:
name: Publish to PyPI
needs: [build-kt-kernel, build-kt-kernel-cuda]
runs-on: [self-hosted, linux, x64]
if: github.repository == 'kvcache-ai/ktransformers' && github.ref == 'refs/heads/main'
environment: prod
permissions:
id-token: write # For trusted publishing (OIDC)
contents: read
steps:
- name: Download all wheel artifacts
uses: actions/download-artifact@v4
with:
path: artifacts/
- name: Organize wheels into dist/
run: |
mkdir -p dist/
find artifacts/ -name "*.whl" -exec cp {} dist/ \;
echo "Wheels to publish:"
ls -lh dist/
- name: Get version from wheel
id: get_version
run: |
# Extract version from first wheel filename
wheel_name=$(ls dist/*.whl | head -1 | xargs basename)
# Extract version (format: kt_kernel-X.Y.Z-...)
version=$(echo "$wheel_name" | sed 's/kt_kernel-\([0-9.]*\)-.*/\1/')
echo "VERSION=$version" >> $GITHUB_OUTPUT
echo "Publishing version: $version"
- name: Install twine
run: |
python -m pip install --upgrade pip
pip install twine
- name: Publish to TestPyPI (if requested)
if: github.event.inputs.test_pypi == 'true'
env:
TWINE_USERNAME: __token__
TWINE_PASSWORD: ${{ secrets.TEST_PYPI_API_TOKEN }}
run: |
python -m twine upload \
--repository testpypi \
--skip-existing \
--verbose \
dist/*.whl
- name: Publish to PyPI
if: github.event.inputs.test_pypi != 'true'
env:
TWINE_USERNAME: __token__
TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }}
run: |
python -m twine upload \
--skip-existing \
--verbose \
dist/*.whl
- name: Create release summary
run: |
echo "## 🎉 kt-kernel v${{ steps.get_version.outputs.VERSION }} Published to PyPI" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "### Published Packages" >> $GITHUB_STEP_SUMMARY
echo "- **kt-kernel** (CPU-only)" >> $GITHUB_STEP_SUMMARY
echo "- **kt-kernel-cuda** (CUDA support)" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "Total wheels: $(ls -1 dist/*.whl | wc -l) (3 Python versions: 3.10, 3.11, 3.12)" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "### Installation" >> $GITHUB_STEP_SUMMARY
echo '```bash' >> $GITHUB_STEP_SUMMARY
echo "# CPU version (AMX/AVX512/AVX2 multi-variant)" >> $GITHUB_STEP_SUMMARY
echo "pip install kt-kernel==${{ steps.get_version.outputs.VERSION }}" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "# CUDA version (requires NVIDIA driver with CUDA 11.8+ or 12.x support)" >> $GITHUB_STEP_SUMMARY
echo "pip install kt-kernel-cuda==${{ steps.get_version.outputs.VERSION }}" >> $GITHUB_STEP_SUMMARY
echo '```' >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "### Features" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "**kt-kernel (CPU) - Multi-variant support:**" >> $GITHUB_STEP_SUMMARY
echo "- ✅ AMX (Intel Sapphire Rapids+)" >> $GITHUB_STEP_SUMMARY
echo "- ✅ AVX512 (Intel Skylake-X/Ice Lake/Cascade Lake)" >> $GITHUB_STEP_SUMMARY
echo "- ✅ AVX2 (Maximum compatibility)" >> $GITHUB_STEP_SUMMARY
echo "- 🔧 Runtime CPU detection: Automatically selects optimal variant" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "**kt-kernel-cuda (CUDA) - Multi-architecture support:**" >> $GITHUB_STEP_SUMMARY
echo "- ✅ SM 80 (Ampere: A100, RTX 3000 series)" >> $GITHUB_STEP_SUMMARY
echo "- ✅ SM 86 (Ampere: RTX 3060-3090)" >> $GITHUB_STEP_SUMMARY
echo "- ✅ SM 89 (Ada Lovelace: RTX 4000 series)" >> $GITHUB_STEP_SUMMARY
echo "- ✅ SM 90 (Hopper: H100)" >> $GITHUB_STEP_SUMMARY
echo "- 🔧 Static CUDA runtime: Compatible with CUDA 11.8+ and 12.x drivers" >> $GITHUB_STEP_SUMMARY
echo "- 🔧 Includes multi-variant CPU code (AMX/AVX512/AVX2)" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "### Links" >> $GITHUB_STEP_SUMMARY
echo "- CPU package: https://pypi.org/project/kt-kernel/${{ steps.get_version.outputs.VERSION }}/" >> $GITHUB_STEP_SUMMARY
echo "- CUDA package: https://pypi.org/project/kt-kernel-cuda/${{ steps.get_version.outputs.VERSION }}/" >> $GITHUB_STEP_SUMMARY