Files
wehub-resource-sync ec436095dd
Book-CI / test (macos-latest) (push) Waiting to run
Deploy / deploy (macos-latest) (push) Waiting to run
Deploy / deploy (ubuntu-latest) (push) Waiting to run
Deploy / deploy (windows-latest) (push) Waiting to run
Release to PyPI / Build & publish sglang-kt (push) Waiting to run
Release to PyPI / Build kt-kernel (Python 3.11) (push) Waiting to run
Release to PyPI / Build kt-kernel (Python 3.12) (push) Waiting to run
Release to PyPI / Publish kt-kernel to PyPI (push) Blocked by required conditions
Book-CI / test (ubuntu-latest) (push) Waiting to run
Book-CI / test (windows-latest) (push) Waiting to run
Release Fake Tag / publish (push) Waiting to run
chore: import upstream snapshot with attribution
2026-07-13 13:30:03 +08:00

284 lines
11 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:
# ── sglang-kt (must be on PyPI before users can pip install kt-kernel) ──
build-and-publish-sglang-kt:
name: Build & publish sglang-kt
runs-on: [self-hosted, linux, x64]
if: github.repository == 'kvcache-ai/ktransformers' && github.ref == 'refs/heads/main'
environment: prod
permissions:
id-token: write
contents: read
steps:
- name: Checkout repository
uses: actions/checkout@v4
with:
submodules: recursive
- name: Set up Python
uses: actions/setup-python@v4
with:
python-version: '3.12'
- name: Install build tools
run: |
python -m pip install --upgrade pip
pip install build wheel setuptools twine
- name: Build sglang-kt wheel
working-directory: third_party/sglang/python
run: |
KT_VERSION=$(python3 -c "exec(open('${{ github.workspace }}/version.py').read()); print(__version__)")
export SGLANG_KT_VERSION="$KT_VERSION"
echo "Building sglang-kt v${KT_VERSION} wheel..."
python -m build --wheel -v
ls dist/ | grep -q "sglang_kt" || (echo "ERROR: Wheel name does not contain sglang_kt" && exit 1)
- name: Publish sglang-kt 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 third_party/sglang/python/dist/*.whl
- name: Publish sglang-kt 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 third_party/sglang/python/dist/*.whl
# ── kt-kernel ──
build-kt-kernel:
name: Build kt-kernel (Python ${{ matrix.python-version }})
runs-on: [self-hosted, linux, x64, gpu]
strategy:
fail-fast: false
matrix:
python-version: ['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: |
# System packages (cmake/libhwloc-dev/pkg-config/libnuma-dev) are expected to be
# preinstalled on self-hosted runners. Skip apt-get to avoid sudo dependency.
for pkg in cmake pkg-config; do command -v $pkg >/dev/null || { echo "ERROR: $pkg missing on runner"; exit 1; }; done
python -m pip install --upgrade pip
pip install build wheel setuptools
pip install torch --index-url https://download.pytorch.org/whl/cu128
- name: Build kt-kernel wheel
working-directory: kt-kernel
env:
CPUINFER_BUILD_ALL_VARIANTS: '1'
CPUINFER_ENABLE_CPPTRACE: '0'
CPUINFER_USE_CUDA: '1'
CPUINFER_CUDA_ARCHS: '80;86;89;90;120'
CPUINFER_CUDA_STATIC_RUNTIME: '1'
CPUINFER_BUILD_TYPE: 'Release'
CPUINFER_PARALLEL: '4'
CPUINFER_FORCE_REBUILD: '1'
CUDA_HOME: '/usr/local/cuda-12.8'
run: |
echo "Building kt-kernel with:"
echo " - CUDA support (SM 80, 86, 89, 90, 120)"
echo " - CPU multi-variant (AMX, AVX512, AVX2)"
python -m build --wheel -v
- name: Verify wheel
working-directory: kt-kernel
run: |
echo "Generated wheel:"
ls -lh dist/
# Install and test
pip install dist/*.whl
python -c "import kt_kernel; print(f'✓ Version: {kt_kernel.__version__}')"
python -c "import kt_kernel; print(f'✓ CPU variant: {kt_kernel.__cpu_variant__}')"
# Verify CUDA support
python -c "
from kt_kernel import kt_kernel_ext
cpu_infer = kt_kernel_ext.CPUInfer(4)
methods = dir(cpu_infer)
has_cuda = 'submit_with_cuda_stream' in methods
print(f'✓ CUDA support: {has_cuda}')
"
# Verify CPU multi-variant support
echo "Checking CPU variants in wheel..."
python -m zipfile -l dist/*.whl | grep "_kt_kernel_ext_" || echo "Warning: No variant .so files found"
python -m zipfile -l dist/*.whl | grep "_kt_kernel_ext_amx.cpython" && echo "✓ AMX variant found" || echo "Note: AMX variant missing"
python -m zipfile -l dist/*.whl | grep "_kt_kernel_ext_avx512" && echo "✓ AVX512 variants found" || echo "Note: AVX512 variants missing"
python -m zipfile -l dist/*.whl | grep "_kt_kernel_ext_avx2.cpython" && echo "✓ AVX2 variant found" || echo "Note: AVX2 variant missing"
# Verify static linking (should NOT depend on libcudart.so).
# Use $RUNNER_TEMP (honors TMPDIR redirect to /mnt) — /tmp is the
# system disk on self-hosted runners and can be tight.
CHECK_DIR="${RUNNER_TEMP:-/tmp}/check"
rm -rf "$CHECK_DIR"
unzip -q dist/*.whl -d "$CHECK_DIR"
if ldd "$CHECK_DIR"/kt_kernel/*.so 2>/dev/null | grep -q "libcudart.so"; then
echo "ERROR: Dynamic cudart found, should be statically linked"
exit 1
else
echo "✓ CUDA runtime statically linked"
fi
- name: Repair wheel for manylinux
working-directory: kt-kernel
run: |
pip install auditwheel patchelf
mkdir -p wheelhouse
for wheel in dist/*.whl; do
auditwheel repair "$wheel" --plat manylinux_2_35_x86_64 --exclude libcuda.so.1 -w wheelhouse/
done
rm -f dist/*.whl && cp wheelhouse/*.whl dist/
- name: Upload artifact
uses: actions/upload-artifact@v4
with:
name: kt-kernel-wheels-py${{ matrix.python-version }}
path: kt-kernel/dist/*.whl
retention-days: 7
publish-pypi:
name: Publish kt-kernel to PyPI
needs: [build-and-publish-sglang-kt, build-kt-kernel]
runs-on: [self-hosted, linux, x64]
if: github.repository == 'kvcache-ai/ktransformers' && github.ref == 'refs/heads/main'
environment: prod
permissions:
id-token: write # For trusted publishing (OIDC)
contents: read
steps:
- name: Download all wheel artifacts
uses: actions/download-artifact@v4
with:
path: artifacts/
- name: Set up Python
uses: actions/setup-python@v4
with:
python-version: '3.12'
- name: Organize wheels into dist/
run: |
mkdir -p dist/
find artifacts/ -name "*.whl" -exec cp {} dist/ \;
echo "Wheels to publish:"
ls -lh dist/
- name: Get version from wheel
id: get_version
run: |
# Extract version from first wheel filename
wheel_name=$(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 "### Installation" >> $GITHUB_STEP_SUMMARY
echo '```bash' >> $GITHUB_STEP_SUMMARY
echo "pip install kt-kernel==${{ steps.get_version.outputs.VERSION }}" >> $GITHUB_STEP_SUMMARY
echo '```' >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "### Published Wheels" >> $GITHUB_STEP_SUMMARY
echo "Total: $(ls -1 dist/*.whl | wc -l) wheels (Python 3.10, 3.11, 3.12)" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "### Features" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "**CPU Multi-Variant Support:**" >> $GITHUB_STEP_SUMMARY
echo "- ✅ AMX (Intel Sapphire Rapids+, 2023)" >> $GITHUB_STEP_SUMMARY
echo "- ✅ AVX512 Base/VNNI/VBMI/BF16 (Intel Skylake-X/Ice Lake/Cascade Lake, 2017+)" >> $GITHUB_STEP_SUMMARY
echo "- ✅ AVX2 (Maximum compatibility, 2013+)" >> $GITHUB_STEP_SUMMARY
echo "- 🔧 Runtime CPU detection: Automatically selects optimal variant" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "**CUDA 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 "- 🔧 Works on CPU-only systems (CUDA features disabled gracefully)" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "**Requirements:**" >> $GITHUB_STEP_SUMMARY
echo "- Python 3.10, 3.11, or 3.12" >> $GITHUB_STEP_SUMMARY
echo "- Linux x86-64 (manylinux_2_17 compatible)" >> $GITHUB_STEP_SUMMARY
echo "- For CUDA features: NVIDIA driver with CUDA 11.8+ or 12.x support" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "PyPI link: https://pypi.org/project/kt-kernel/${{ steps.get_version.outputs.VERSION }}/" >> $GITHUB_STEP_SUMMARY