tilelang/pyproject.toml

[project]
name = "tilelang"
description = "A tile level programming language to generate high performance code."
readme = "README.md"
requires-python = ">=3.10"
authors = [{ name = "TileLang Contributors" }, { name = "Tile-AI" }]
maintainers = [{ name = "Lei Wang", email = "leiwang1999@outlook.com" }]
license = "MIT"
keywords = ["BLAS", "CUDA", "HIP", "Code Generation", "TVM"]
classifiers = [
    "Development Status :: 4 - Beta",
    "Environment :: GPU",
    "Operating System :: Microsoft :: Windows",
    "Operating System :: POSIX :: Linux",
    "Operating System :: MacOS",
    "Programming Language :: C++",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Programming Language :: Python :: 3.14",
    "Programming Language :: Python :: Implementation :: CPython",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
dynamic = ["version"]
dependencies = [
    "apache-tvm-ffi>=0.1.11,<0.1.13",
    # torch-c-dlpack-ext provides prebuilt torch extensions.
    # Without it, TVM FFI may require JIT compilation on first import.
    "torch-c-dlpack-ext; python_version < '3.14'",
    "cloudpickle",
    "ml-dtypes",
    "numpy>=1.23.5",
    "psutil",
    "torch",
    "torch>=2.4; platform_system == 'Darwin'",
    # jit-compiling a torch extension needs setuptools
    "setuptools; platform_system == 'Darwin'",
    "tqdm>=4.62.3",
    "typing-extensions>=4.10.0",
    "z3-solver>=4.13.0,<4.15.5",
]

[project.optional-dependencies]
# mldtypes should be greater than 0.5.1
# if you want to enable fp4
fp4 = ["ml-dtypes>=0.5.1"]
# if you want to enable layout inference visualization
vis = ["matplotlib"]
# if you want to build with CUDA NVCC support
nvcc = [
    "torch",
    "nvidia-cuda-nvcc>=13.0.48",
    "nvidia-cuda-cccl>=13.0.50",
    # On Windows, ship nvrtc through the runtime deps so `pip install tilelang`
    # works without a host CUDA install (TileLang JITs CUDA kernels via nvrtc).
    "nvidia-cuda-nvrtc>=13; platform_system == 'Windows'",
    # cuRAND is required when kernels use tilelang.cuda.language.random
    "nvidia-curand>=10.4; platform_system == 'Windows'",
]

[[tool.uv.index]]
name = "pytorch-cu130"
url = "https://download.pytorch.org/whl/cu130"
explicit = true

[tool.uv.sources]
torch = [
    { index = "pytorch-cu130", marker = "sys_platform == 'win32'", extra = "nvcc" },
]

[build-system]
requires = [
    "cython>=3.1.0",
    "scikit-build-core",
    "z3-solver>=4.13.0,<4.15.5",
    # Not for auditwheel, explicitly add patchelf for repairing libz3.so.
    # See tvm's CMakeLists.txt for more information.
    "patchelf>=0.17.2; platform_system == 'Linux'",
    # Under PEP 517 build isolation on Windows, provide nvcc and CUDA headers so
    # CMake's find_package(CUDAToolkit) succeeds without a host CUDA install.
    "nvidia-cuda-nvcc>=13; platform_system == 'Windows'",
    "nvidia-cuda-cccl>=13; platform_system == 'Windows'",
    "nvidia-cuda-nvrtc>=13; platform_system == 'Windows'",
    # CMake's MSVC helpers in FindPipCUDAToolkit.cmake only activate under the
    # Ninja generator; require ninja on Windows so scikit-build-core can pick
    # it up (paired with the override below that passes -G Ninja).
    "ninja>=1.11; platform_system == 'Windows'",
]
build-backend = "scikit_build_core.build"

[tool.scikit-build]
wheel.py-api = "cp39"
cmake.version = ">=3.26.1"
build-dir = "build"

# Include backend and git info in version
metadata.version.provider = "version_provider"
metadata.version.provider-path = "."
experimental = true

# On Windows, force the Ninja generator so the project's MSVC environment
# activation path runs (it is guarded on `CMAKE_GENERATOR MATCHES "Ninja"`).
# Linux and macOS keep the scikit-build-core default (Makefiles).
[[tool.scikit-build.overrides]]
if.platform-system = "win32"
cmake.args = ["-G", "Ninja"]

# editable.rebuild = true

# build.verbose = true
# logging.level = "DEBUG"

[tool.scikit-build.sdist]
include = [
    "./VERSION",
    ".git_commit.txt",
    "./LICENSE",
    "THIRDPARTYNOTICES.txt",
    "version_provider.py",
    "requirements*.txt",
    "tilelang/jit/adapter/cython/cython_wrapper.pyx",
    "tilelang/jit/adapter/pto/cython_wrapper.pyx",
    "CMakeLists.txt",
    "src/**",
    "cmake/**",
    # The vendored 3rdparty contents in sdist should be same as wheel.
    # Need full TVM to build from source.
    "3rdparty/tvm",
    # CUTLASS
    "3rdparty/cutlass/include",
    "3rdparty/cutlass/tools",
    # Vendored HIP headers (build-time only) so source builds work on hosts
    # without a ROCm install. Not mapped into the wheel at runtime.
    "3rdparty/hip-headers/include",
    "testing/**",
    "examples/**",
]
exclude = [
    ".git",
    ".github",
    "**/.git",
    "**/.github",
    "3rdparty/**",
    "build",
]

[tool.scikit-build.wheel.packages]
tilelang = "tilelang"
"tilelang/src" = "src"
# NOTE: The mapping below places the contents of '3rdparty' inside 'tilelang/3rdparty' in the wheel.
# This is necessary to find TVM shared libraries at runtime.
# The vendored 3rdparty contents in wheel should be same as sdist.
# TVM
"tilelang/3rdparty/tvm/src" = "3rdparty/tvm/src"
"tilelang/3rdparty/tvm/python" = "3rdparty/tvm/python"
"tilelang/3rdparty/tvm/include" = "3rdparty/tvm/include"
"tilelang/3rdparty/tvm/version.py" = "3rdparty/tvm/version.py"
# CUTLASS
"tilelang/3rdparty/cutlass/include" = "3rdparty/cutlass/include"
"tilelang/3rdparty/cutlass/tools" = "3rdparty/cutlass/tools"



[tool.codespell]
ignore-words = "docs/spelling_wordlist.txt"
skip = [
    "build",
    "3rdparty",
    "dist",
    ".venv"
]

[tool.ruff]
target-version = "py310"
line-length = 140
output-format = "full"

exclude = [
    "3rdparty",
    "examples/deepseek_v32/inference",
]

[tool.ruff.format]
quote-style = "double"
indent-style = "space"
skip-magic-trailing-comma = false
line-ending = "auto"
docstring-code-format = false
docstring-code-line-length = "dynamic"

[tool.ruff.lint.per-file-ignores]
# Do not upgrade type hint in testing and examples.
# See https://github.com/tile-ai/tilelang/issues/1079 for more information.
"testing/**.py" = ["UP", "FA"]
"examples/**.py" = ["UP", "FA"]

[tool.ruff.lint]
select = [
    # pycodestyle
    "E", "W",
    # Pyflakes
    "F",
    # pyupgrade
    "UP", "FA",
    # flake8-bugbear
    "B",
    # flake8-simplify
    "SIM",
    # isort
    # "I",
]
ignore = [
    # Module level import not at top of file
    "E402",
    # star imports
    "F405", "F403",
    # ambiguous name
    "E741",
    # line too long
    "E501",
    # if-else-block instead of ternary
    "SIM108",
    # key in dict.keys()
    "SIM118",
    # open file w.o. ctx manager
    "SIM115",
    # memory leaks
    "B019",
    # zip without explicit strict
    "B905",
    # No such file or directory
    "E902",
]

[tool.pytest.ini_options]
verbosity_assertions = 3
filterwarnings = ["always"]
markers = [
    "perf: benchmark and performance API tests skipped by default",
    "slow: long-running correctness sweeps",
    # Feature markers emitted by vendored TVM's tvm.testing.utils requires_*
    # decorators (see 3rdparty/tvm/python/tvm/testing/utils.py). TVM registers
    # these via its own conftest, but tilelang only imports the decorators, so
    # they must be registered here to avoid PytestUnknownMarkWarning.
    "pto: tests for the PTO backend",
    "cpu: Mark a test as running on CPU",
    "gpu: Mark a test as using a GPU",
    "multi_gpu: Mark a test as using multiple GPUs",
    "cuda: Mark a test as using CUDA",
    "llvm: Mark a test as using LLVM",
    "metal: Mark a test as using Metal",
    "rocm: Mark a test as using ROCm",
]

[tool.cibuildwheel]
archs = ["auto64"]
skip = "*musllinux*"
build-frontend = "build"
environment = { PYTHONDEVMODE = "1", PYTHONUNBUFFERED = "1" }
environment-pass = [
    "CUDA_VERSION",
    "NO_VERSION_LABEL",
    "NO_TOOLCHAIN_VERSION",
    "NO_GIT_VERSION",
    "COLUMNS",
    "CMAKE_GENERATOR",
    "CMAKE_BUILD_PARALLEL_LEVEL",
    "FORCE_COLOR",
    "CLICOLOR_FORCE",
]
before-build = "env -0 | sort -z | tr '\\0' '\\n'"
windows.before-build = "set"
test-command = [
    'python -c "import tilelang; print(tilelang.__version__)"',
]

[tool.cibuildwheel.linux]
environment.PYTHONDEVMODE = "1"
environment.PYTHONUNBUFFERED = "1"
environment.PATH = "/usr/local/cuda/bin:$PATH"
environment.LD_LIBRARY_PATH = "/usr/local/cuda/lib64:/usr/local/cuda/lib64/stubs:$LD_LIBRARY_PATH"
# Build a wheel with CUDA, ROCm, and Ascend support. ROCm uses vendored HIP
# headers and Ascend uses local ACL declarations; both link to runtime stubs,
# so the manylinux container only needs the CUDA toolkit. Execution still
# requires the selected backend's real runtime and device.
environment.USE_CUDA = "ON"
environment.USE_ROCM = "ON"
environment.USE_ASCEND = "ON"
manylinux-x86_64-image  = "manylinux_2_28"  # AlmaLinux 8
manylinux-aarch64-image = "manylinux_2_34"  # Z3 requires
# Install CUDA runtime and stub driver library
# manylinux_2_28 uses gcc 14, which needs CUDA >=12.8
before-all = """
set -eux

cat /etc/*-release
uname -a

case "$(uname -m)" in
    "x86_64")
        DEFAULT_CUDA_VERSION="13.0"
        dnf config-manager --add-repo https://developer.download.nvidia.cn/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
        ;;
    "aarch64")
        DEFAULT_CUDA_VERSION="13.0"
        dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
        ;;
    *)
        exit 1
        ;;
esac

cudaver="$(echo "${CUDA_VERSION:-$DEFAULT_CUDA_VERSION}" | cut -d '.' -f-2)"
v="${cudaver//./-}"
dnf install -y "cuda-minimal-build-${v}" "cuda-driver-devel-${v}" "cuda-nvrtc-devel-${v}"
dnf clean all
"""
# Do not bundle libtvm_ffi.so and libz3.so* because they are shipped by the dependency packages.
repair-wheel-command = "auditwheel -v repair --exclude libtvm_ffi.so --exclude 'libz3.so*' --exclude libcuda.so.1 --exclude '/usr/local/cuda*' -w {dest_dir} {wheel}"

[tool.cibuildwheel.macos]
repair-wheel-command = "delocate-wheel --verbose --ignore-missing-dependencies --no-sanitize-rpaths --require-archs {delocate_archs} -w {dest_dir} -v {wheel}"

[[tool.cibuildwheel.overrides]]
select = "*linux*x86_64*"
# x86_64 runners in GitHub Actions have limited storage,
# pre-install torch without caching to reduce disk usage during install tilelang.
before-test = [
    "pip install torch --no-cache-dir",
]