name: build-gpu-wheels

permissions:

contents: read

on:

workflow_dispatch:

inputs:

link_mode:

description: 'Linking mode'

required: true

default: 'dynamic'

type: choice

options:

- dynamic

- static

cuda_linux:

description: 'Build CUDA 12 wheels (Linux x86_64)'

required: false

default: false

type: boolean

rocm_linux:

description: 'Build ROCm wheels (Linux x86_64)'

required: false

default: false

type: boolean

sycl_linux:

description: 'Build SYCL wheels (Linux x86_64)'

required: false

default: false

type: boolean

vulkan_linux:

description: 'Build Vulkan wheels (Linux x86_64)'

required: false

default: false

type: boolean

cuda_windows:

description: 'Build CUDA 12 wheels (Windows x86_64)'

required: false

default: false

type: boolean

vulkan_windows:

description: 'Build Vulkan wheels (Windows x86_64)'

required: false

default: false

type: boolean

vulkan_macos_intel:

description: 'Build Vulkan wheels (MacOS Intel)'

required: false

default: false

type: boolean

upload_release:

description: 'Upload wheels to GitHub Release'

required: false

default: false

type: boolean

publish_index:

description: 'Publish to GitHub Pages wheel index'

required: false

default: false

type: boolean

concurrency:

group: ${{ github.workflow }}-${{ github.ref }}-${{ inputs.link_mode }}

cancel-in-progress: true

env:

CIBW_BUILD_VERBOSITY: 1

CIBW_BUILD: "cp310-* cp311-* cp312-* cp313-* cp314-*"

CIBW_SKIP: "*-musllinux_* *-win32 *-manylinux_i686"

FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"

jobs:

# ---------------------------------------------------------------------------

# CUDA wheels (Linux only, CUDA 12.4)

# ---------------------------------------------------------------------------

build_cuda:

name: CUDA 12.4 - Linux x86_64 (${{ inputs.link_mode }})

if: inputs.cuda_linux

runs-on: ubuntu-22.04

steps:

- name: Checkout repository

uses: actions/checkout@v6

with:

fetch-depth: 0

- name: Free disk space

run: |

sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc

df -h

- name: Cache thirdparty deps

id: cache-deps

uses: actions/cache@v5

with:

path: thirdparty/

key: deps-cuda-${{ inputs.link_mode }}-${{ hashFiles('scripts/manage.py', '.github/workflows/build-gpu-wheels.yml') }}

- name: Set up Python

uses: actions/setup-python@v6

with:

python-version: '3.11'

- name: Install cibuildwheel

run: pip install cibuildwheel==3.4.0

- name: Set CCACHE_DIR to workspace

run: echo "CCACHE_DIR=${{ github.workspace }}/.ccache" >> "$GITHUB_ENV"

- name: ccache

uses: hendrikmuhs/[email protected]

with:

key: ccache-cuda-${{ inputs.link_mode }}

evict-old-files: 1d

- name: Build CUDA wheels

run: python -m cibuildwheel --output-dir wheelhouse

env:

CIBW_ARCHS_LINUX: x86_64

CIBW_MANYLINUX_X86_64_IMAGE: manylinux2014

CIBW_BEFORE_ALL_LINUX: >

yum install -y git yum-utils epel-release &&

yum install -y ccache &&

yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo &&

yum install -y cuda-nvcc-12-4 cuda-cudart-devel-12-4 cuda-driver-devel-12-4 libcublas-devel-12-4 &&

ln -s /usr/local/cuda-12.4 /usr/local/cuda &&

if [ -f /project/thirdparty/.deps-complete ]; then echo "=== Deps cached, skipping build ==="; exit 0; fi &&

CI_PYTHON=$(ls -d /opt/python/cp3*/bin | head -1) &&

$CI_PYTHON/pip install cmake ninja &&

PATH=/usr/local/cuda/bin:$CI_PYTHON:$PATH

CCACHE_DIR=/project/.ccache

CCACHE_DEPEND=1

CMAKE_C_COMPILER_LAUNCHER=ccache

CMAKE_CXX_COMPILER_LAUNCHER=ccache

CMAKE_CUDA_COMPILER_LAUNCHER=ccache

CMAKE_CUDA_ARCHITECTURES="75"

GGML_NATIVE=OFF GGML_CUDA=1 GGML_METAL=0

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

$CI_PYTHON/python scripts/manage.py build --all --deps-only --no-sd-examples

${{ inputs.link_mode == 'dynamic' && '--dynamic' || '' }} &&

touch /project/thirdparty/.deps-complete

CIBW_BEFORE_BUILD_LINUX: >

sed -i 's/^name = "cyllama"/name = "cyllama-cuda12"/' {project}/pyproject.toml &&

grep -q '^name = "cyllama-cuda12"' {project}/pyproject.toml || { echo "ERROR: sed rename to cyllama-cuda12 failed"; exit 1; } &&

cd {project} && python scripts/manage.py write_build_config

CIBW_ENVIRONMENT_LINUX: >

GGML_CUDA=1

GGML_METAL=0

CUDA_PATH=/usr/local/cuda

CMAKE_CUDA_ARCHITECTURES="75"

GGML_NATIVE=OFF

PATH=/usr/local/cuda/bin:$PATH

LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

CFLAGS=-s

CXXFLAGS=-s

CCACHE_DIR=/project/.ccache

CCACHE_DEPEND=1

CMAKE_C_COMPILER_LAUNCHER=ccache

CMAKE_CXX_COMPILER_LAUNCHER=ccache

CMAKE_CUDA_COMPILER_LAUNCHER=ccache

WITH_DYLIB=${{ inputs.link_mode == 'dynamic' && '1' || '0' }}

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

CIBW_TEST_COMMAND: ""

CIBW_REPAIR_WHEEL_COMMAND_LINUX: >

test -d {project}/thirdparty/llama.cpp/dynamic || { echo "ERROR: dynamic lib dir missing: {project}/thirdparty/llama.cpp/dynamic"; exit 1; } &&

LD_LIBRARY_PATH={project}/thirdparty/llama.cpp/dynamic:$LD_LIBRARY_PATH

auditwheel repair -w {dest_dir} {wheel}

--plat manylinux_2_35_x86_64

--exclude libcuda.so.1

--exclude libcudart.so.12

--exclude libcublas.so.12

--exclude libcublasLt.so.12

--exclude libgomp.so.1

- name: List built wheels

run: ls -lah wheelhouse/

- name: Upload wheels

uses: actions/upload-artifact@v7

with:

name: wheels-cuda-linux-x86_64

path: ./wheelhouse/*.whl

# ---------------------------------------------------------------------------

# ROCm wheels (Linux only, ROCm 6.3)

# ---------------------------------------------------------------------------

build_rocm:

name: ROCm 6.3 - Linux x86_64 (${{ inputs.link_mode }})

if: inputs.rocm_linux

runs-on: ubuntu-22.04

steps:

- name: Checkout repository

uses: actions/checkout@v6

with:

fetch-depth: 0

- name: Free disk space

run: |

sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc

df -h

- name: Cache thirdparty deps

id: cache-deps

uses: actions/cache@v5

with:

path: thirdparty/

key: deps-rocm-${{ inputs.link_mode }}-${{ hashFiles('scripts/manage.py', '.github/workflows/build-gpu-wheels.yml') }}

- name: Set up Python

uses: actions/setup-python@v6

with:

python-version: '3.11'

- name: Install cibuildwheel

run: pip install cibuildwheel==3.4.0

- name: Set CCACHE_DIR to workspace

run: echo "CCACHE_DIR=${{ github.workspace }}/.ccache" >> "$GITHUB_ENV"

- name: ccache

uses: hendrikmuhs/[email protected]

with:

key: ccache-rocm-${{ inputs.link_mode }}

evict-old-files: 1d

- name: Build ROCm wheels

run: python -m cibuildwheel --output-dir wheelhouse

env:

CIBW_ARCHS_LINUX: x86_64

CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28

CIBW_BEFORE_ALL_LINUX: >

dnf install -y git epel-release &&

dnf install -y ccache &&

dnf install -y https://repo.radeon.com/amdgpu-install/6.3.3/rhel/8.10/amdgpu-install-6.3.60303-1.el8.noarch.rpm &&

dnf install -y rocm-hip-runtime-devel hipblas-devel rocblas-devel libstdc++-static &&

ls -d /opt/rocm-* | head -1 | xargs -I{} ln -sf {} /opt/rocm &&

ln -sf /usr/lib/gcc/x86_64-redhat-linux/8/libstdc++.a /opt/rocm/lib/libstdc++.a &&

if [ -f /project/thirdparty/.deps-complete ]; then echo "=== Deps cached, skipping build ==="; exit 0; fi &&

CI_PYTHON=$(ls -d /opt/python/cp3*/bin | head -1) &&

$CI_PYTHON/pip install cmake ninja &&

PATH=/opt/rocm/bin:$CI_PYTHON:$PATH

LIBRARY_PATH=/usr/lib/gcc/x86_64-redhat-linux/8:$LIBRARY_PATH

CCACHE_DIR=/project/.ccache

CCACHE_DEPEND=1

CMAKE_C_COMPILER_LAUNCHER=ccache

CMAKE_CXX_COMPILER_LAUNCHER=ccache

CMAKE_HIP_COMPILER_LAUNCHER=ccache

CMAKE_HIP_ARCHITECTURES="gfx90a;gfx942;gfx1100"

HIP_PATH=/opt/rocm

GGML_NATIVE=OFF GGML_HIP=1 GGML_METAL=0

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

$CI_PYTHON/python scripts/manage.py build --all --deps-only --no-sd-examples

${{ inputs.link_mode == 'dynamic' && '--dynamic' || '' }} &&

touch /project/thirdparty/.deps-complete

CIBW_BEFORE_BUILD_LINUX: >

sed -i 's/^name = "cyllama"/name = "cyllama-rocm"/' {project}/pyproject.toml &&

grep -q '^name = "cyllama-rocm"' {project}/pyproject.toml || { echo "ERROR: sed rename to cyllama-rocm failed"; exit 1; } &&

cd {project} && python scripts/manage.py write_build_config

CIBW_ENVIRONMENT_LINUX: >

GGML_NATIVE=OFF

GGML_HIP=1

GGML_METAL=0

HIP_PATH=/opt/rocm

CMAKE_HIP_ARCHITECTURES="gfx90a;gfx942;gfx1100"

PATH=/opt/rocm/bin:$PATH

LD_LIBRARY_PATH=/opt/rocm/lib:$LD_LIBRARY_PATH

LIBRARY_PATH=/usr/lib/gcc/x86_64-redhat-linux/8:$LIBRARY_PATH

CFLAGS=-s

CXXFLAGS=-s

CCACHE_DIR=/project/.ccache

CCACHE_DEPEND=1

CMAKE_C_COMPILER_LAUNCHER=ccache

CMAKE_CXX_COMPILER_LAUNCHER=ccache

CMAKE_HIP_COMPILER_LAUNCHER=ccache

WITH_DYLIB=${{ inputs.link_mode == 'dynamic' && '1' || '0' }}

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

CIBW_TEST_COMMAND: ""

CIBW_REPAIR_WHEEL_COMMAND_LINUX: >

test -d {project}/thirdparty/llama.cpp/dynamic || { echo "ERROR: dynamic lib dir missing: {project}/thirdparty/llama.cpp/dynamic"; exit 1; } &&

LD_LIBRARY_PATH={project}/thirdparty/llama.cpp/dynamic:$LD_LIBRARY_PATH

auditwheel repair -w {dest_dir} {wheel}

--plat manylinux_2_35_x86_64

--exclude libamdhip64.so.6

--exclude libhipblas.so.2

--exclude librocblas.so.4

--exclude libhsa-runtime64.so.1

--exclude librocsolver.so.0

--exclude libhipblaslt.so.0

--exclude libamd_comgr.so.2

--exclude librocprofiler-register.so.0

--exclude libgomp.so.1

- name: List built wheels

run: ls -lah wheelhouse/

- name: Upload wheels

uses: actions/upload-artifact@v7

with:

name: wheels-rocm-linux-x86_64

path: ./wheelhouse/*.whl

# ---------------------------------------------------------------------------

# SYCL wheels (Linux only, Intel oneAPI)

# ---------------------------------------------------------------------------

build_sycl:

name: SYCL - Linux x86_64 (${{ inputs.link_mode }})

if: inputs.sycl_linux

runs-on: ubuntu-22.04

steps:

- name: Checkout repository

uses: actions/checkout@v6

with:

fetch-depth: 0

- name: Cache thirdparty deps

id: cache-deps

uses: actions/cache@v5

with:

path: thirdparty/

key: deps-sycl-${{ inputs.link_mode }}-${{ hashFiles('scripts/manage.py', '.github/workflows/build-gpu-wheels.yml') }}

- name: Set up Python

uses: actions/setup-python@v6

with:

python-version: '3.11'

- name: Install cibuildwheel

run: pip install cibuildwheel==3.4.0

- name: Set CCACHE_DIR to workspace

run: echo "CCACHE_DIR=${{ github.workspace }}/.ccache" >> "$GITHUB_ENV"

- name: ccache

uses: hendrikmuhs/[email protected]

with:

key: ccache-sycl-${{ inputs.link_mode }}

evict-old-files: 1d

- name: Build SYCL wheels

run: python -m cibuildwheel --output-dir wheelhouse

env:

CIBW_ARCHS_LINUX: x86_64

CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28

CIBW_BEFORE_ALL_LINUX: >

dnf install -y git epel-release &&

dnf install -y ccache &&

rpm --import https://yum.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB &&

dnf config-manager --add-repo https://yum.repos.intel.com/oneapi &&

dnf install -y intel-oneapi-dpcpp-cpp-2025.3 intel-oneapi-mkl-devel-2025.3 &&

if [ -f /project/thirdparty/.deps-complete ]; then echo "=== Deps cached, skipping build ==="; exit 0; fi &&

CI_PYTHON=$(ls -d /opt/python/cp3*/bin | head -1) &&

$CI_PYTHON/pip install cmake ninja &&

source /opt/intel/oneapi/setvars.sh &&

export CC=icx CXX=icpx &&

PATH=$CI_PYTHON:$PATH

CCACHE_DIR=/project/.ccache

CMAKE_C_COMPILER_LAUNCHER=ccache

CMAKE_CXX_COMPILER_LAUNCHER=ccache

GGML_NATIVE=OFF GGML_SYCL=1 GGML_METAL=0

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

$CI_PYTHON/python scripts/manage.py build --all --deps-only --no-sd-examples

${{ inputs.link_mode == 'dynamic' && '--dynamic' || '' }} &&

touch /project/thirdparty/.deps-complete

CIBW_BEFORE_BUILD_LINUX: >

sed -i 's/^name = "cyllama"/name = "cyllama-sycl"/' {project}/pyproject.toml &&

grep -q '^name = "cyllama-sycl"' {project}/pyproject.toml || { echo "ERROR: sed rename to cyllama-sycl failed"; exit 1; } &&

cd {project} && python scripts/manage.py write_build_config

CIBW_ENVIRONMENT_LINUX: >

GGML_NATIVE=OFF

GGML_SYCL=1

GGML_METAL=0

CC=icx

CXX=icpx

PATH=/opt/intel/oneapi/compiler/latest/bin:$PATH

LD_LIBRARY_PATH=/opt/intel/oneapi/compiler/latest/lib:/opt/intel/oneapi/mkl/latest/lib:$LD_LIBRARY_PATH

CFLAGS=-s

CXXFLAGS=-s

CCACHE_DIR=/project/.ccache

CMAKE_C_COMPILER_LAUNCHER=ccache

CMAKE_CXX_COMPILER_LAUNCHER=ccache

WITH_DYLIB=${{ inputs.link_mode == 'dynamic' && '1' || '0' }}

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

CIBW_TEST_COMMAND: ""

CIBW_REPAIR_WHEEL_COMMAND_LINUX: >

test -d {project}/thirdparty/llama.cpp/dynamic || { echo "ERROR: dynamic lib dir missing: {project}/thirdparty/llama.cpp/dynamic"; exit 1; } &&

LD_LIBRARY_PATH={project}/thirdparty/llama.cpp/dynamic:$LD_LIBRARY_PATH

auditwheel repair -w {dest_dir} {wheel}

--plat manylinux_2_35_x86_64

--exclude libsycl.so.8

--exclude libOpenCL.so.1

--exclude libsvml.so

--exclude libimf.so

--exclude libintlc.so.5

--exclude libtbb.so.12

--exclude libgomp.so.1

- name: List built wheels

run: ls -lah wheelhouse/

- name: Upload wheels

uses: actions/upload-artifact@v7

with:

name: wheels-sycl-linux-x86_64

path: ./wheelhouse/*.whl

# ---------------------------------------------------------------------------

# Vulkan wheels (Linux)

# ---------------------------------------------------------------------------

build_vulkan:

name: Vulkan - Linux x86_64 (${{ inputs.link_mode }})

if: inputs.vulkan_linux

runs-on: ubuntu-22.04

steps:

- name: Checkout repository

uses: actions/checkout@v6

with:

fetch-depth: 0

clean: true

- name: Cache thirdparty deps

id: cache-deps

uses: actions/cache@v5

with:

path: thirdparty/

key: deps-vulkan-${{ inputs.link_mode }}-${{ hashFiles('scripts/manage.py', '.github/workflows/build-gpu-wheels.yml') }}

- name: Set up Python

uses: actions/setup-python@v6

with:

python-version: '3.11'

- name: Install cibuildwheel

run: pip install cibuildwheel==3.4.0

- name: Set CCACHE_DIR to workspace

run: echo "CCACHE_DIR=${{ github.workspace }}/.ccache" >> "$GITHUB_ENV"

- name: ccache

uses: hendrikmuhs/[email protected]

with:

key: ccache-vulkan-${{ inputs.link_mode }}

evict-old-files: 1d

- name: Build Vulkan wheels

run: python -m cibuildwheel --output-dir wheelhouse

env:

CIBW_ARCHS_LINUX: x86_64

CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28

CIBW_BEFORE_ALL_LINUX: >

dnf install -y git epel-release &&

dnf install -y ccache &&

dnf install -y vulkan-headers vulkan-loader-devel &&

if [ -f /project/thirdparty/.deps-complete ]; then echo "=== Deps cached, skipping build ==="; exit 0; fi &&

git clone --depth 1 https://github.com/google/shaderc.git /tmp/shaderc &&

cd /tmp/shaderc && python3 utils/git-sync-deps &&

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DSHADERC_SKIP_TESTS=ON -DSHADERC_SKIP_EXAMPLES=ON -DSHADERC_SKIP_COPYRIGHT_CHECK=ON &&

cmake --build build --target glslc_exe --parallel 4 &&

cp build/glslc/glslc /usr/local/bin/glslc &&

cmake -S /tmp/shaderc/third_party/spirv-headers -B /tmp/spirv-headers-build -DCMAKE_INSTALL_PREFIX=/usr/local -DSPIRV_HEADERS_SKIP_EXAMPLES=ON -DSPIRV_HEADERS_SKIP_INSTALL=OFF &&

cmake --install /tmp/spirv-headers-build &&

cd /project && rm -rf /tmp/shaderc /tmp/spirv-headers-build &&

CI_PYTHON=$(ls -d /opt/python/cp3*/bin | head -1) &&

$CI_PYTHON/pip install cmake ninja &&

PATH=$CI_PYTHON:$PATH

CCACHE_DIR=/project/.ccache

CMAKE_C_COMPILER_LAUNCHER=ccache

CMAKE_CXX_COMPILER_LAUNCHER=ccache

GGML_NATIVE=OFF GGML_VULKAN=1 GGML_METAL=0

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

$CI_PYTHON/python scripts/manage.py build --all --deps-only --no-sd-examples

${{ inputs.link_mode == 'dynamic' && '--dynamic' || '' }} &&

touch /project/thirdparty/.deps-complete

CIBW_BEFORE_BUILD_LINUX: >

sed -i 's/^name = "cyllama"/name = "cyllama-vulkan"/' {project}/pyproject.toml &&

grep -q '^name = "cyllama-vulkan"' {project}/pyproject.toml || { echo "ERROR: sed rename to cyllama-vulkan failed"; exit 1; } &&

cd {project} && python scripts/manage.py write_build_config

CIBW_ENVIRONMENT_LINUX: >

GGML_NATIVE=OFF

GGML_VULKAN=1

GGML_METAL=0

CFLAGS=-s

CXXFLAGS=-s

CCACHE_DIR=/project/.ccache

CMAKE_C_COMPILER_LAUNCHER=ccache

CMAKE_CXX_COMPILER_LAUNCHER=ccache

WITH_DYLIB=${{ inputs.link_mode == 'dynamic' && '1' || '0' }}

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

CIBW_TEST_COMMAND: ""

CIBW_REPAIR_WHEEL_COMMAND_LINUX: >

test -d {project}/thirdparty/llama.cpp/dynamic || { echo "ERROR: dynamic lib dir missing: {project}/thirdparty/llama.cpp/dynamic"; exit 1; } &&

LD_LIBRARY_PATH={project}/thirdparty/llama.cpp/dynamic:$LD_LIBRARY_PATH

auditwheel repair -w {dest_dir} {wheel}

--plat manylinux_2_35_x86_64

--exclude libvulkan.so.1

--exclude libgomp.so.1

- name: List built wheels

run: ls -lah wheelhouse/

- name: Upload wheels

uses: actions/upload-artifact@v7

with:

name: wheels-vulkan-linux-x86_64

path: ./wheelhouse/*.whl

# ---------------------------------------------------------------------------

# CUDA 12.4 - Windows (x86_64)

# Uses NVIDIA CUDA Toolkit 12.4. delvewheel handles wheel repair.

# ---------------------------------------------------------------------------

build_cuda_windows:

name: CUDA 12.4 - Windows x86_64 (${{ inputs.link_mode }})

if: inputs.cuda_windows

runs-on: windows-latest

steps:

- name: Checkout repository

uses: actions/checkout@v6

with:

fetch-depth: 0

- name: Cache thirdparty deps

id: cache-deps

uses: actions/cache@v5

with:

path: thirdparty/

key: deps-cuda-windows-${{ inputs.link_mode }}-${{ hashFiles('scripts/manage.py', '.github/workflows/build-gpu-wheels.yml') }}

- name: Install CUDA Toolkit 12.4

uses: Jimver/[email protected]

id: cuda-toolkit

with:

cuda: '12.4.0'

method: 'network'

sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "visual_studio_integration"]'

- name: Verify CUDA installation

shell: bash

run: |

echo "CUDA_PATH=$CUDA_PATH"

nvcc --version

ls "$CUDA_PATH/lib/x64/" | head -20

- name: Set up Python

uses: actions/setup-python@v6

with:

python-version: '3.11'

- name: Install cibuildwheel

run: pip install cibuildwheel==3.4.0

- name: Build CUDA wheels

run: python -m cibuildwheel --output-dir wheelhouse

env:

CIBW_ARCHS_WINDOWS: AMD64

CIBW_BEFORE_BUILD_WINDOWS: >

pip install delvewheel &&

python -c "import os; t=open('pyproject.toml').read(); open('pyproject.toml','w').write(t.replace('name = \"cyllama\"','name = \"cyllama-cuda12\"',1))" &&

python -c "t=open('pyproject.toml').read(); assert 'name = \"cyllama-cuda12\"' in t, 'rename failed'" &&

cd {project} &&

python scripts/manage.py build --all --deps-only --no-sd-examples

${{ inputs.link_mode == 'dynamic' && '--dynamic' || '' }} &&

python scripts/manage.py write_build_config

CIBW_ENVIRONMENT_WINDOWS: >

GGML_CUDA=1

GGML_METAL=0

GGML_NATIVE=OFF

CMAKE_CUDA_ARCHITECTURES="75"

WITH_DYLIB=${{ inputs.link_mode == 'dynamic' && '1' || '0' }}

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

CIBW_TEST_COMMAND: ""

CIBW_REPAIR_WHEEL_COMMAND_WINDOWS: >

delvewheel repair -w {dest_dir} {wheel}

--add-path {project}\thirdparty\llama.cpp\dynamic

--add-path {project}\thirdparty\whisper.cpp\dynamic

--add-path {project}\thirdparty\stable-diffusion.cpp\dynamic

--include ggml-cuda.dll

--no-dll nvcuda.dll

--no-dll cudart64_12.dll

--no-dll cublas64_12.dll

--no-dll cublasLt64_12.dll

- name: List built wheels

shell: bash

run: ls -lah wheelhouse/

- name: Upload wheels

uses: actions/upload-artifact@v7

with:

name: wheels-cuda-windows-x86_64

path: ./wheelhouse/*.whl

# ---------------------------------------------------------------------------

# Vulkan - Windows (x86_64)

# Uses the LunarG Vulkan SDK. delvewheel handles wheel repair.

# ---------------------------------------------------------------------------

build_vulkan_windows:

name: Vulkan - Windows x86_64 (${{ inputs.link_mode }})

if: inputs.vulkan_windows

runs-on: windows-latest

steps:

- name: Checkout repository

uses: actions/checkout@v6

with:

fetch-depth: 0

- name: Cache thirdparty deps

id: cache-deps

uses: actions/cache@v5

with:

path: thirdparty/

key: deps-vulkan-windows-${{ inputs.link_mode }}-${{ hashFiles('scripts/manage.py', '.github/workflows/build-gpu-wheels.yml') }}

- name: Install Vulkan SDK

shell: pwsh

run: |

$VK_LATEST = (Invoke-RestMethod -Uri "https://vulkan.lunarg.com/sdk/latest/windows.json").windows

Write-Host "Latest Vulkan SDK: ${VK_LATEST}"

$VK_URL = "https://sdk.lunarg.com/sdk/download/${VK_LATEST}/windows/vulkan-sdk.exe"

Write-Host "Downloading Vulkan SDK from ${VK_URL}..."

Invoke-WebRequest -Uri $VK_URL -OutFile vulkan-sdk.exe

Write-Host "Installing Vulkan SDK..."

Start-Process -FilePath .\vulkan-sdk.exe -ArgumentList "--accept-licenses --default-answer --confirm-command install" -Wait -NoNewWindow

$VK_PATH = "C:\VulkanSDK\${VK_LATEST}"

Write-Host "VULKAN_SDK=${VK_PATH}"

echo "VULKAN_SDK=${VK_PATH}" >> $env:GITHUB_ENV

echo "${VK_PATH}\Bin" >> $env:GITHUB_PATH

& "${VK_PATH}\Bin\glslc.exe" --version

- name: Set up Python

uses: actions/setup-python@v6

with:

python-version: '3.11'

- name: Install cibuildwheel

run: pip install cibuildwheel==3.4.0

- name: Build Vulkan wheels

run: python -m cibuildwheel --output-dir wheelhouse

env:

CIBW_ARCHS_WINDOWS: AMD64

CIBW_BEFORE_BUILD_WINDOWS: >

pip install delvewheel &&

python -c "import os; t=open('pyproject.toml').read(); open('pyproject.toml','w').write(t.replace('name = \"cyllama\"','name = \"cyllama-vulkan\"',1))" &&

python -c "t=open('pyproject.toml').read(); assert 'name = \"cyllama-vulkan\"' in t, 'rename failed'" &&

cd {project} &&

python scripts/manage.py build --all --deps-only --no-sd-examples

${{ inputs.link_mode == 'dynamic' && '--dynamic' || '' }} &&

python scripts/manage.py write_build_config

CIBW_ENVIRONMENT_WINDOWS: >

GGML_VULKAN=1

GGML_METAL=0

GGML_NATIVE=OFF

WITH_DYLIB=${{ inputs.link_mode == 'dynamic' && '1' || '0' }}

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

CIBW_TEST_COMMAND: ""

CIBW_REPAIR_WHEEL_COMMAND_WINDOWS: >

delvewheel repair -w {dest_dir} {wheel}

--add-path {project}\thirdparty\llama.cpp\dynamic

--add-path {project}\thirdparty\whisper.cpp\dynamic

--add-path {project}\thirdparty\stable-diffusion.cpp\dynamic

--include ggml-vulkan.dll

- name: List built wheels

shell: bash

run: ls -lah wheelhouse/

- name: Upload wheels

uses: actions/upload-artifact@v7

with:

name: wheels-vulkan-windows-x86_64

path: ./wheelhouse/*.whl

# ---------------------------------------------------------------------------

# Vulkan - macOS Intel (x86_64)

# Uses MoltenVK (Vulkan over Metal). The Vulkan SDK provides the loader,

# headers, and glslc. delocate handles wheel repair.

# ---------------------------------------------------------------------------

build_vulkan_macos_intel:

name: Vulkan - macOS Intel x86_64 (${{ inputs.link_mode }})

if: inputs.vulkan_macos_intel

runs-on: macos-15-intel

steps:

- name: Checkout repository

uses: actions/checkout@v6

with:

fetch-depth: 0

- name: Cache thirdparty deps

id: cache-deps

uses: actions/cache@v5

with:

path: thirdparty/

key: deps-vulkan-macos-intel-${{ inputs.link_mode }}-${{ hashFiles('scripts/manage.py', '.github/workflows/build-gpu-wheels.yml') }}

- name: Install Vulkan SDK

run: |

brew install vulkan-headers vulkan-loader molten-vk shaderc spirv-headers

SHADERC_BIN="$(brew --prefix shaderc)/bin"

VULKAN_PREFIX="$(brew --prefix vulkan-loader)"

echo "PATH=${SHADERC_BIN}:${PATH}" >> "$GITHUB_ENV"

echo "VULKAN_SDK=${VULKAN_PREFIX}" >> "$GITHUB_ENV"

echo "CMAKE_PREFIX_PATH=$(brew --prefix vulkan-headers):$(brew --prefix vulkan-loader):$(brew --prefix molten-vk):$(brew --prefix spirv-headers)" >> "$GITHUB_ENV"

${SHADERC_BIN}/glslc --version

- name: Set up Python

uses: actions/setup-python@v6

with:

python-version: '3.11'

- name: Install cibuildwheel

run: pip install cibuildwheel==3.4.0

- name: ccache

uses: hendrikmuhs/[email protected]

with:

key: ccache-vulkan-macos-intel-${{ inputs.link_mode }}

evict-old-files: 1d

- name: Build Vulkan wheels

run: python -m cibuildwheel --output-dir wheelhouse

env:

CIBW_ARCHS_MACOS: x86_64

CIBW_BEFORE_BUILD_MACOS: >

pip install delocate wheel &&

sed -i '' 's/^name = "cyllama"/name = "cyllama-vulkan"/' {project}/pyproject.toml &&

grep -q '^name = "cyllama-vulkan"' {project}/pyproject.toml || { echo "ERROR: sed rename to cyllama-vulkan failed"; exit 1; } &&

cd {project} &&

python scripts/manage.py build --all --deps-only --no-sd-examples

${{ inputs.link_mode == 'dynamic' && '--dynamic' || '' }} &&

python scripts/manage.py write_build_config

CIBW_ENVIRONMENT_MACOS: >

MACOSX_DEPLOYMENT_TARGET=11.0

GGML_VULKAN=1

GGML_METAL=0

GGML_NATIVE=OFF

CMAKE_C_COMPILER_LAUNCHER=ccache

CMAKE_CXX_COMPILER_LAUNCHER=ccache

WITH_DYLIB=${{ inputs.link_mode == 'dynamic' && '1' || '0' }}

SD_USE_VENDORED_GGML=${{ inputs.link_mode == 'dynamic' && '0' || '1' }}

CIBW_TEST_COMMAND: ""

CIBW_REPAIR_WHEEL_COMMAND_MACOS: >

delocate-wheel --require-archs {delocate_archs} -w {dest_dir} -v {wheel}

--exclude libvulkan

--exclude libMoltenVK &&

python {project}/scripts/manage.py fix_macos_vulkan_wheel {dest_dir}

- name: List built wheels

run: ls -lah wheelhouse/

- name: Upload wheels

uses: actions/upload-artifact@v7

with:

name: wheels-vulkan-macos-intel-x86_64

path: ./wheelhouse/*.whl

# ---------------------------------------------------------------------------

# Import smoke tests (no GPU required)

# ---------------------------------------------------------------------------

smoke_test:

name: Smoke test - ${{ matrix.backend }}

needs: [build_cuda, build_rocm, build_sycl, build_vulkan, build_cuda_windows, build_vulkan_windows, build_vulkan_macos_intel]

if: always() && !cancelled() && (needs.build_cuda.result == 'success' || needs.build_rocm.result == 'success' || needs.build_sycl.result == 'success' || needs.build_vulkan.result == 'success' || needs.build_cuda_windows.result == 'success' || needs.build_vulkan_windows.result == 'success' || needs.build_vulkan_macos_intel.result == 'success')

runs-on: ${{ matrix.runs-on }}

strategy:

fail-fast: false

matrix:

include:

- backend: cuda

runs-on: ubuntu-latest

artifact: wheels-cuda-linux-x86_64

condition: inputs.cuda_linux

- backend: rocm

runs-on: ubuntu-latest

artifact: wheels-rocm-linux-x86_64

condition: inputs.rocm_linux

- backend: sycl

runs-on: ubuntu-latest

artifact: wheels-sycl-linux-x86_64

condition: inputs.sycl_linux

- backend: vulkan

runs-on: ubuntu-latest

artifact: wheels-vulkan-linux-x86_64

condition: inputs.vulkan_linux

- backend: cuda-windows

runs-on: windows-latest

artifact: wheels-cuda-windows-x86_64

condition: inputs.cuda_windows

- backend: vulkan-windows

runs-on: windows-latest

artifact: wheels-vulkan-windows-x86_64

condition: inputs.vulkan_windows

- backend: vulkan-macos-intel

runs-on: macos-15-intel

artifact: wheels-vulkan-macos-intel-x86_64

condition: inputs.vulkan_macos_intel

steps:

- name: Check if backend was built

id: check

shell: bash

run: |

if [ "${{ matrix.condition }}" != "true" ]; then

echo "skip=true" >> "$GITHUB_OUTPUT"

else

echo "skip=false" >> "$GITHUB_OUTPUT"

fi

- uses: actions/setup-python@v6

if: steps.check.outputs.skip != 'true'

with:

python-version: "3.12"

- name: Download wheel artifacts

if: steps.check.outputs.skip != 'true'

uses: actions/download-artifact@v8

with:

name: ${{ matrix.artifact }}

path: wheels

- name: Find matching wheel

if: steps.check.outputs.skip != 'true'

id: find-wheel

shell: bash

run: |

WHEEL=$(find wheels -name "*cp312*" | head -1)

if [ -z "$WHEEL" ]; then

WHEEL=$(find wheels -name "*.whl" | head -1)

fi

echo "wheel=$WHEEL" >> "$GITHUB_OUTPUT"

echo "Found wheel: $WHEEL"

- name: Install wheel in clean venv

if: steps.check.outputs.skip != 'true'

shell: bash

run: |

python -m venv smoke-venv

if [ "$RUNNER_OS" = "Windows" ]; then

source smoke-venv/Scripts/activate

else

source smoke-venv/bin/activate

fi

pip install "${{ steps.find-wheel.outputs.wheel }}"

# -------------------------------------------------------------------

# Windows-only: link-test every bundled ggml-*.dll using the same

# LoadLibraryW(path, NULL, 0) semantics that ggml_backend_load()

# uses in C. Catches the class of bug where delvewheel-bundled

# siblings (e.g. vulkan-1-<hash>.dll) are unreachable because the

# libs dir is not on PATH — a failure mode invisible to GPU-less

# runners since `cyllama info` still exits 0 when a backend DLL

# fails to load. Independent of GPU/ICD presence.

# -------------------------------------------------------------------

- name: Smoke test - DLL link (Windows)

if: steps.check.outputs.skip != 'true' && runner.os == 'Windows'

shell: bash

run: |

source smoke-venv/Scripts/activate

python -c "

import ctypes, glob, os, site, sys

failed = []

for sp in site.getsitepackages():

for libs in glob.glob(os.path.join(sp, 'cyllama*.libs')):

for dll in sorted(glob.glob(os.path.join(libs, 'ggml-*.dll'))):

try:

ctypes.WinDLL(dll)

except OSError as e:

failed.append((dll, str(e)))

else:

print(f'OK {dll}')

if failed:

for d, e in failed:

print(f'FAIL {d}: {e}', file=sys.stderr)

sys.exit(1)

"

- name: Smoke test - imports

if: steps.check.outputs.skip != 'true'

shell: bash

run: |

if [ "$RUNNER_OS" = "Windows" ]; then

source smoke-venv/Scripts/activate

else

source smoke-venv/bin/activate

fi

python -c "

import cyllama

print(f'cyllama {cyllama.__version__} imported successfully')

# Core API imports

from cyllama import LLM, complete, chat, AsyncLLM

from cyllama import batch_generate, BatchGenerator

from cyllama import estimate_gpu_layers, estimate_memory_usage

from cyllama import GenerationConfig, Response

# Cython extension imports

from cyllama.llama import llama_cpp

print('llama_cpp extension loaded')

# Agent framework

from cyllama.agents import ReActAgent, ConstrainedAgent, ContractAgent, tool

print('agents loaded')

# Integrations

from cyllama.integrations.openai_compat import OpenAICompatibleClient

print('integrations loaded')

# Optional extensions

try:

from cyllama.whisper import whisper_cpp

print('whisper_cpp extension loaded')

except ImportError as e:

print(f'whisper_cpp not available: {e}')

try:

from cyllama.sd import stable_diffusion

print('stable_diffusion extension loaded')

except ImportError as e:

print(f'stable_diffusion not available: {e}')

# Backend info

try:

from cyllama import _backend

print(f'Backend: cuda={getattr(_backend, \"cuda\", False)}, '

f'vulkan={getattr(_backend, \"vulkan\", False)}, '

f'hip={getattr(_backend, \"hip\", False)}, '

f'sycl={getattr(_backend, \"sycl\", False)}')

except ImportError:

print('No _backend module')

print('All imports passed (${{ matrix.backend }})')

"

# -------------------------------------------------------------------

# cyllama info -- validates build_config.json is present in the wheel

# and reports which backend flags it was built with. Catches the

# class of bug behind "build_config.json missing from wheels" (0.2.10)

# and "build_config.json shipped with empty versions" (0.2.10).

# -------------------------------------------------------------------

- name: Smoke test - cyllama info

if: steps.check.outputs.skip != 'true'

shell: bash

run: |

if [ "$RUNNER_OS" = "Windows" ]; then

source smoke-venv/Scripts/activate

else

source smoke-venv/bin/activate

fi

python -m cyllama info

# -------------------------------------------------------------------

# SYCL-only: install Intel's OpenCL CPU runtime so the SYCL dispatch

# stack has a compliant device on a GPU-less runner, then run a

# 1-token decode. Exercises the full pipeline (Cython wrappers ->

# libllama -> libggml-sycl -> SYCL runtime -> OpenCL CPU device ->

# sampler -> teardown) on free hardware. Catches ABI/symbol/teardown

# regressions that bare imports miss. Best-effort: if the Intel

# runtime install fails (repo mirror issue, etc.) we log and move on

# rather than breaking the whole GPU-wheels workflow.

# -------------------------------------------------------------------

- name: Install Intel OpenCL CPU runtime

if: steps.check.outputs.skip != 'true' && matrix.backend == 'sycl'

continue-on-error: true

run: |

wget -qO - https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB \

| sudo gpg --dearmor -o /usr/share/keyrings/oneapi-archive-keyring.gpg

echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" \

| sudo tee /etc/apt/sources.list.d/oneAPI.list

sudo apt-get update

sudo apt-get install -y --no-install-recommends \

intel-oneapi-runtime-opencl \

intel-oneapi-runtime-dpcpp-cpp

- name: Cache test model (SYCL)

if: steps.check.outputs.skip != 'true' && matrix.backend == 'sycl'

id: cache-model-sycl

uses: actions/cache@v5

with:

path: models/Llama-3.2-1B-Instruct-Q8_0.gguf

key: test-model-llama-3.2-1b-q8

- name: Download test model (SYCL)

if: steps.check.outputs.skip != 'true' && matrix.backend == 'sycl' && steps.cache-model-sycl.outputs.cache-hit != 'true'

run: |

mkdir -p models

curl -sL https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q8_0.gguf \

-o models/Llama-3.2-1B-Instruct-Q8_0.gguf

- name: Smoke test - SYCL 1-token CPU inference

if: steps.check.outputs.skip != 'true' && matrix.backend == 'sycl'

continue-on-error: true

env:

# Force the SYCL runtime onto the OpenCL CPU device the Intel

# runtime just installed (no GPU on ubuntu-latest).