mirror of
https://github.com/kvcache-ai/ktransformers.git
synced 2026-03-14 18:37:23 +00:00
fix pypi cuda install (#1763)
This commit is contained in:
Jianwei Dong
134
.github/workflows/release-pypi.yml
vendored
134
.github/workflows/release-pypi.yml
vendored
@@ -89,18 +89,42 @@ jobs:
|
||||
pip install auditwheel patchelf
|
||||
echo "Repairing wheels for manylinux compatibility..."
|
||||
mkdir -p wheelhouse
|
||||
|
||||
for wheel in dist/*.whl; do
|
||||
echo "Processing $wheel..."
|
||||
auditwheel repair "$wheel" --plat manylinux_2_17_x86_64 -w wheelhouse/ || {
|
||||
echo "Warning: auditwheel repair failed, trying to rename platform tag..."
|
||||
# Fallback: rename the wheel file with manylinux tag
|
||||
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 "$wheel")
|
||||
new_name=$(echo "$wheel_name" | sed 's/linux_x86_64/manylinux_2_17_x86_64/')
|
||||
# Use # as sed delimiter to avoid conflict with /
|
||||
new_name=$(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/
|
||||
@@ -112,9 +136,87 @@ jobs:
|
||||
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_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 for SM 80, 86, 89, 90"
|
||||
python -m build --wheel -v
|
||||
|
||||
- name: Verify wheel
|
||||
working-directory: kt-kernel
|
||||
run: |
|
||||
ls -lh dist/
|
||||
# Check version suffix
|
||||
[[ $(ls dist/*.whl) == *"+cuda118"* ]] || (echo "ERROR: Missing +cuda118 suffix" && exit 1)
|
||||
|
||||
# Install and test
|
||||
pip install dist/*.whl
|
||||
python -c "import kt_kernel; print(f'Version: {kt_kernel.__version__}')"
|
||||
|
||||
# Verify static linking (should NOT depend on libcudart.so)
|
||||
unzip -q dist/*.whl -d /tmp/check
|
||||
! ldd /tmp/check/kt_kernel/*.so | grep -q "libcudart.so" || (echo "ERROR: Dynamic cudart found" && exit 1)
|
||||
echo "✓ CUDA runtime statically linked"
|
||||
|
||||
- name: Repair wheel for manylinux
|
||||
working-directory: kt-kernel
|
||||
run: |
|
||||
pip install auditwheel patchelf
|
||||
mkdir -p wheelhouse
|
||||
for wheel in dist/*.whl; do
|
||||
auditwheel repair "$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
|
||||
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
|
||||
@@ -186,11 +288,27 @@ jobs:
|
||||
echo "Total: $(ls -1 dist/*.whl | wc -l) wheels (3 Python versions: 3.10, 3.11, 3.12)" >> $GITHUB_STEP_SUMMARY
|
||||
echo "" >> $GITHUB_STEP_SUMMARY
|
||||
echo "### Features" >> $GITHUB_STEP_SUMMARY
|
||||
echo "**CPU-only build with multi-variant support:**" >> $GITHUB_STEP_SUMMARY
|
||||
echo "" >> $GITHUB_STEP_SUMMARY
|
||||
echo "**CPU wheels with 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 "**Runtime CPU detection:** Automatically selects the best variant for your CPU" >> $GITHUB_STEP_SUMMARY
|
||||
echo "**CUDA wheels with 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 "" >> $GITHUB_STEP_SUMMARY
|
||||
echo "**Installation:**" >> $GITHUB_STEP_SUMMARY
|
||||
echo '```bash' >> $GITHUB_STEP_SUMMARY
|
||||
echo "# CPU version" >> $GITHUB_STEP_SUMMARY
|
||||
echo "pip install kt-kernel==${{ steps.get_version.outputs.VERSION }}+cpu" >> $GITHUB_STEP_SUMMARY
|
||||
echo "" >> $GITHUB_STEP_SUMMARY
|
||||
echo "# CUDA version (requires NVIDIA driver with CUDA 11.8+ or 12.x support)" >> $GITHUB_STEP_SUMMARY
|
||||
echo "pip install kt-kernel==${{ steps.get_version.outputs.VERSION }}+cuda118" >> $GITHUB_STEP_SUMMARY
|
||||
echo '```' >> $GITHUB_STEP_SUMMARY
|
||||
echo "" >> $GITHUB_STEP_SUMMARY
|
||||
echo "PyPI link: https://pypi.org/project/kt-kernel/#history" >> $GITHUB_STEP_SUMMARY
|
||||
|
||||
Reference in New Issue
Block a user