[Build] Migrate to scikit-build-core (#939)

* cleanup

* init

* build first wheel that may not work

* build cython ext

* fix tvm build

* use sabi

* update rpath to support auditwheel

* pass editible build

* update ci

* fix warnings

* do not use ccache in self host runner

* test local uv cache

* test pip index

* update lib search to respect new lib location

* fix

* update ci

* enable cuda by default

* update src map

* fix

* fix

* fix

* Generate version with backend and git information at build time

* copy tvm_cython to wheels

* fix tvm lib search

* fmt

* remove unused

* auto detect ccache

* add back backend-related files

* remove jit cython adaptor to simplify code

* fmt

* fix ci

* ci fix 2

* ci fix 3

* workaround metal

* ci fix 4

* fmt

* fmt

* Revert "ci fix 4"

This reverts commit d1de8291c3e40927955f3ad3cf87a75c78813676.

* tmp

* fix metal

* trivial cleanup

* add detailed build-time version for cuda

* add back mlc

* Restore wheel info and other trivial updates

* update

* fix cuda

* upd

* fix metal ci

* test for ga build

* test for nvidia/cuda

* test ubuntu 20

* fix

* fix

* Do not use `uv build`

* fix

* fix

* log toolchain version

* merge wheel

* update

* debug

* fix

* update

* skip rocm

* update artifacts each

* fix

* fix

* add mac

* fix cache

* fix cache

* fix cache

* reset and add comment

* upd

* fix git version

* update deps

* trivial update

* use in-tree build dir and install to src to speedup editable build

* Revert "use in-tree build dir and install to src to speedup editable build"

This reverts commit 6ab87b05c5eed811210136b8dca4fc3677dd51f2.

* add build-dir

* update docs

* remove old scrips

* [1/n] cleanup scripts

* [Lint]: [pre-commit.ci] auto fixes [...]

* fix and update

* wait for tvm fix

* revert some tmp fix

* fix

* fix

* spell

* doc update

* test cibuildwheel

* fix and test macos on ci

* Update .github/workflows/dist.yml
Co-authored-by: Xuehai Pan <XuehaiPan@outlook.com>

* fix

* test ga event

* cleanup

* bump tvm to support api3

* test final version

* add cron

* Update .github/workflows/dist.yml
Co-authored-by: Xuehai Pan <XuehaiPan@outlook.com>

* fix

* test ccache for metal cibuildwheel

* test newer macos

* finish

---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xuehai Pan <XuehaiPan@outlook.com>

maint/scripts/pypi_distribution_tox.sh

-multi_python_version=("3.8" "3.9" "3.10" "3.11" "3.12")
-for python_version in "${multi_python_version[@]}"; do
-    echo "Installing Python ${python_version}..."
-    apt-get install -y python${python_version}
-done
-
-pip install -r requirements-build.txt
-
-# if dist and build directories exist, remove them
-if [ -d dist ]; then
-    rm -r dist
-fi
-
-# Build source distribution (disabled for now)
-# python setup.py sdist --formats=gztar,zip
-
-# Build wheels for different Python versions
-echo "Building wheels for multiple Python versions..."
-tox -e py38-pypi,py39-pypi,py310-pypi,py311-pypi,py312-pypi
-
-if [ $? -ne 0 ]; then
-    echo "Error: Failed to build the wheels."
-    exit 1
-else
-    echo "Wheels built successfully."
-fi
\ No newline at end of file

pyproject.toml

+[project]
+name = "tilelang"
+authors = [{name = "Tile-AI"}]
+maintainers = [{name = "Lei Wang", email = "leiwang1999@outlook.com"}]
+description = "A tile level programming language to generate high performance code."
+readme.file = "README.md"
+license = "MIT"
+keywords = ["BLAS", "CUDA", "HIP", "Code Generation", "TVM"]
+classifiers = [
+    "Environment :: GPU",
+    "Operating System :: POSIX :: Linux",
+    "Operating System :: OS Independent",
+    "Operating System :: MacOS",
+    "Programming Language :: Python :: 3.8",
+    "Programming Language :: Python :: 3.9",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Intended Audience :: Developers",
+    "Intended Audience :: Science/Research",
+    "Scientific/Engineering :: Artificial Intelligence",
+]
+
+readme.content-type = "text/markdown"
+requires-python = ">=3.8"
+
+dynamic = ["version"]
+
+# Somehow this does not work, hard-code for now
+# dynamic = ["version", "dependencies"]
+# [tool.setuptools.dynamic]
+# dependencies = {file = ["requirements.txt"]}
+dependencies = [
+    "numpy>=1.23.5",
+    "tqdm>=4.62.3",
+    "typing_extensions>=4.10.0",
+    "cloudpickle",
+    "ml_dtypes",
+    "psutil",
+    "torch",
+]
+
+[project.optional-dependencies]
+# mldtypes should be greater than 0.5.1
+# if you want to enable fp4
+fp4 = ["ml_dtypes>=0.5.1"]
+
 [build-system]
 requires = [
-    "build",
-    "cmake>=3.26",
-    "packaging",
-    "setuptools>=61",
-    "wheel",
-    "patchelf",
+    "setuptools>=63",
     "Cython>=3.0.0",
+    "scikit-build-core",
 ]
-build-backend = "setuptools.build_meta"
+build-backend = "scikit_build_core.build"
+
+[tool.scikit-build]
+wheel.py-api = "cp38"
+cmake.version = ">=3.26.1"
+build-dir = "build"
+
+# editable.rebuild = true
+
+# Include backend and git info in version
+metadata.version.provider = "version_provider"
+metadata.version.provider-path = "."
+experimental = true
+
+[tool.scikit-build.wheel.packages]
+tilelang = "tilelang"
+"tilelang/src" = "src"
+"tilelang/3rdparty" = "3rdparty"
+
+# TODO: we might want to not include these in wheel?
+"tilelang/benchmark" = "benchmark"
+"tilelang/examples" = "examples"
+"tilelang/testing" = "testing"
 
 [tool.yapf]
 based_on_style = "yapf"
@@ -67,3 +132,50 @@ ignore = [
 [tool.ruff.lint.per-file-ignores]
 "3rdparty/**/*" = ["ALL"]
 "examples/deepseek_v32/inference/**/*" = ["ALL"]
+
+[tool.cibuildwheel]
+archs = ["auto64"]
+# wait for tvm fix
+build = "cp38-*"
+
+[tool.cibuildwheel.macos]
+archs = ["arm64"]
+
+[tool.cibuildwheel.linux]
+# Pin to glibc 2.17 for x86 and 2.28 for aarch64 for now
+manylinux-x86_64-image = "manylinux2014"
+manylinux-aarch64-image = "manylinux_2_28"
+skip = "*-musllinux*"
+environment-pass = ["CUDA_VERSION"]
+repair-wheel-command = [
+  "auditwheel repair --exclude libcuda.so.1 --exclude /usr/local/cuda\\* -w {dest_dir} {wheel}",
+  "pipx run abi3audit --strict --report {wheel}",
+]
+
+# Install CUDA runtime and stub driver library
+# manylinux_2_28 uses gcc 14, which needs CUDA 12.8
+before-all = """
+set -eux
+
+case "$(uname -m)" in
+"x86_64")
+    yum-config-manager --add-repo https://developer.download.nvidia.cn/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
+    ;;
+"aarch64")
+    dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
+    ;;
+*)
+    exit 1
+    ;;
+esac
+
+# Assume CUDA_VERSION=xx.y
+v=${CUDA_VERSION:-12.4}
+v=${v:0:4}
+v=${v/./-}
+yum install -y cuda-minimal-build-${v} cuda-driver-devel-${v} cuda-nvrtc-devel-${v}
+"""
+
+[tool.cibuildwheel.linux.environment]
+# Equlivant to `source /opt/rh/gcc-toolset-12/enable`, safe when gcc-toolset-12 is not installed
+PATH = "/usr/local/cuda/bin:$PATH"

requirements-build.txt

-# Should be mirrored in pyproject.toml
-Cython>=3.0.0
-build
-cmake>=3.26
-packaging
-setuptools>=61
-torch
-wheel
-tox
-auditwheel
-patchelf
-ninja

requirements-dev.txt

-# lint requirements
--r requirements-lint.txt
-# build requirements
+# Requirements to run local build with `--no-build-isolation` or other developments
+
+Cython>=3.0.0
+build
 cmake>=3.26
-# runtime requirements
-cffi
-cpplint
-Cython
-docutils
-dtlib
-numpy>=1.23.5
-pytest>=6.2.4
-pytest_xdist>=2.2.1
-packaging>=21.0
-PyYAML
-tqdm>=4.62.3
-typing_extensions>=4.10.0
-requests
-cloudpickle
-ml_dtypes
-psutil
-scipy
+packaging
+setuptools>=61
 torch
-tabulate
 wheel
-setuptools
\ No newline at end of file
+tox
+ninja
+
+auditwheel; platform_system == 'Linux'
+patchelf; platform_system == 'Linux'
+delocate; platform_system == 'Darwin'

requirements.txt

 # runtime requirements
-Cython>=3.0.0
 numpy>=1.23.5
 tqdm>=4.62.3
 typing_extensions>=4.10.0
 cloudpickle
-# mldtypes should be greater than 0.5.1
-# if you want to enable fp4
 ml_dtypes
 psutil
 torch

tilelang/__init__.py

@@ -5,6 +5,10 @@ import ctypes
 import logging
 from tqdm import tqdm
 
+from importlib.metadata import version
+
+__version__ = version('tilelang')
+
 
 class TqdmLoggingHandler(logging.Handler):
     """Custom logging handler that directs log output to tqdm progress bar to avoid interference."""
@@ -57,9 +61,10 @@ from .env import enable_cache, disable_cache, is_cache_enabled  # noqa: F401
 from .env import env as env  # noqa: F401
 
 import tvm
-import tvm.base
+import tvm.base  # noqa: F401
 from tvm import DataType  # noqa: F401
 
+# Setup tvm search path before importing tvm
 from . import libinfo
 
 
@@ -71,8 +76,8 @@ def _load_tile_lang_lib():
     # pylint: disable=protected-access
     lib_name = "tilelang" if tvm.base._RUNTIME_ONLY else "tilelang_module"
     # pylint: enable=protected-access
-    lib_path = libinfo.find_lib_path(lib_name, optional=False)
-    return ctypes.CDLL(lib_path[0]), lib_path[0]
+    lib_path = libinfo.find_lib_path(lib_name)
+    return ctypes.CDLL(lib_path), lib_path
 
 
 # only load once here
@@ -101,8 +106,6 @@ from .transform import PassConfigKey  # noqa: F401
 
 from .engine import lower, register_cuda_postproc, register_hip_postproc  # noqa: F401
 
-from .version import __version__  # noqa: F401
-
 from .math import *  # noqa: F403
 
 from . import ir  # noqa: F401

tilelang/autotuner/tuner.py

@@ -30,7 +30,7 @@ from tilelang.autotuner.param import CompileArgs, ProfileArgs, AutotuneResult
 from tilelang.autotuner.capture import get_autotune_inputs
 from tilelang.utils.target import determine_target
 from tilelang.jit.param import _P, _RProg
-from tilelang.version import __version__
+from tilelang import __version__
 
 
 class TimeoutException(Exception):

tilelang/cache/kernel_cache.py

@@ -16,7 +16,7 @@ from tvm.tir import PrimFunc
 from tilelang.engine.param import KernelParam
 from tilelang import env
 from tilelang.jit import JITKernel
-from tilelang.version import __version__
+from tilelang import __version__
 
 KERNEL_PATH = "kernel.cu"
 WRAPPED_KERNEL_PATH = "wrapped_kernel.cu"

tilelang/env.py

@@ -4,6 +4,7 @@ import pathlib
 import logging
 import shutil
 import glob
+import site
 from dataclasses import dataclass
 from typing import Optional
 
@@ -19,6 +20,19 @@ TL_TEMPLATE_NOT_FOUND_MESSAGE = ("TileLang is not installed or found in the expe
 ", which may lead to compilation bugs when utilize tilelang backend."
 TVM_LIBRARY_NOT_FOUND_MESSAGE = ("TVM is not installed or found in the expected path")
 
+SITE_PACKAGES = site.getsitepackages()
+
+TL_LIBS = [os.path.join(i, 'tilelang/lib') for i in site.getsitepackages()]
+TL_LIBS = [i for i in TL_LIBS if os.path.exists(i)]
+
+TL_ROOT = os.path.dirname(os.path.abspath(__file__))
+
+DEV = False
+THIRD_PARTY_ROOT = os.path.join(TL_ROOT, '3rdparty')
+if not os.path.exists(THIRD_PARTY_ROOT):
+    DEV = True
+    THIRD_PARTY_ROOT = os.path.join(TL_ROOT, '..', '3rdparty')
+
 
 def _find_cuda_home() -> str:
     """Find the CUDA install path.
@@ -261,85 +275,51 @@ env = Environment()
 CUDA_HOME = env.CUDA_HOME
 ROCM_HOME = env.ROCM_HOME
 
+
+def prepend_pythonpath(path):
+    if not os.environ.get("PYTHONPATH", None):
+        os.environ["PYTHONPATH"] = path
+    else:
+        os.environ["PYTHONPATH"] = path + os.pathsep + os.environ["PYTHONPATH"]
+
+    sys.path.insert(0, path)
+
+
 # Initialize TVM paths
 if env.TVM_IMPORT_PYTHON_PATH is not None:
-    os.environ["PYTHONPATH"] = env.TVM_IMPORT_PYTHON_PATH + ":" + os.environ.get("PYTHONPATH", "")
-    sys.path.insert(0, env.TVM_IMPORT_PYTHON_PATH)
+    prepend_pythonpath(env.TVM_IMPORT_PYTHON_PATH)
 else:
-    install_tvm_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "3rdparty", "tvm")
-    if os.path.exists(install_tvm_path) and install_tvm_path not in sys.path:
-        os.environ["PYTHONPATH"] = (
-            install_tvm_path + "/python:" + os.environ.get("PYTHONPATH", ""))
-        sys.path.insert(0, install_tvm_path + "/python")
-        env.TVM_IMPORT_PYTHON_PATH = install_tvm_path + "/python"
-
-    develop_tvm_path = os.path.join(
-        os.path.dirname(os.path.abspath(__file__)), "..", "3rdparty", "tvm")
-    if os.path.exists(develop_tvm_path) and develop_tvm_path not in sys.path:
-        os.environ["PYTHONPATH"] = (
-            develop_tvm_path + "/python:" + os.environ.get("PYTHONPATH", ""))
-        sys.path.insert(0, develop_tvm_path + "/python")
-        env.TVM_IMPORT_PYTHON_PATH = develop_tvm_path + "/python"
-
-    develop_tvm_library_path = os.path.join(
-        os.path.dirname(os.path.abspath(__file__)), "..", "build", "tvm")
-    install_tvm_library_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "lib")
+    tvm_path = os.path.join(THIRD_PARTY_ROOT, "tvm")
+    assert os.path.exists(tvm_path), tvm_path
+    if tvm_path not in sys.path:
+        tvm_python_binding = os.path.join(tvm_path, 'python')
+        prepend_pythonpath(tvm_python_binding)
+        env.TVM_IMPORT_PYTHON_PATH = tvm_python_binding
+
     if os.environ.get("TVM_LIBRARY_PATH") is None:
-        if os.path.exists(develop_tvm_library_path):
-            os.environ["TVM_LIBRARY_PATH"] = develop_tvm_library_path
-        elif os.path.exists(install_tvm_library_path):
-            os.environ["TVM_LIBRARY_PATH"] = install_tvm_library_path
-        else:
-            logger.warning(TVM_LIBRARY_NOT_FOUND_MESSAGE)
-        # pip install build library path
-        lib_path = os.path.join(env.TILELANG_PACKAGE_PATH, "lib")
-        existing_path = os.environ.get("TVM_LIBRARY_PATH")
-        if existing_path:
-            os.environ["TVM_LIBRARY_PATH"] = f"{existing_path}:{lib_path}"
-        else:
-            os.environ["TVM_LIBRARY_PATH"] = lib_path
-        env.TVM_LIBRARY_PATH = os.environ.get("TVM_LIBRARY_PATH", None)
+        os.environ['TVM_LIBRARY_PATH'] = env.TVM_LIBRARY_PATH = os.pathsep.join(TL_LIBS)
 
 # Initialize CUTLASS paths
 if os.environ.get("TL_CUTLASS_PATH", None) is None:
-    install_cutlass_path = os.path.join(
-        os.path.dirname(os.path.abspath(__file__)), "3rdparty", "cutlass")
-    develop_cutlass_path = os.path.join(
-        os.path.dirname(os.path.abspath(__file__)), "..", "3rdparty", "cutlass")
-    if os.path.exists(install_cutlass_path):
-        os.environ["TL_CUTLASS_PATH"] = install_cutlass_path + "/include"
-        env.CUTLASS_INCLUDE_DIR = install_cutlass_path + "/include"
-    elif (os.path.exists(develop_cutlass_path) and develop_cutlass_path not in sys.path):
-        os.environ["TL_CUTLASS_PATH"] = develop_cutlass_path + "/include"
-        env.CUTLASS_INCLUDE_DIR = develop_cutlass_path + "/include"
+    cutlass_inc_path = os.path.join(THIRD_PARTY_ROOT, 'cutlass', 'include')
+    if os.path.exists(cutlass_inc_path):
+        os.environ["TL_CUTLASS_PATH"] = env.CUTLASS_INCLUDE_DIR = cutlass_inc_path
     else:
         logger.warning(CUTLASS_NOT_FOUND_MESSAGE)
 
 # Initialize COMPOSABLE_KERNEL paths
 if os.environ.get("TL_COMPOSABLE_KERNEL_PATH", None) is None:
-    install_ck_path = os.path.join(
-        os.path.dirname(os.path.abspath(__file__)), "3rdparty", "composable_kernel")
-    develop_ck_path = os.path.join(
-        os.path.dirname(os.path.abspath(__file__)), "..", "3rdparty", "composable_kernel")
-    if os.path.exists(install_ck_path):
-        os.environ["TL_COMPOSABLE_KERNEL_PATH"] = install_ck_path + "/include"
-        env.COMPOSABLE_KERNEL_INCLUDE_DIR = install_ck_path + "/include"
-    elif (os.path.exists(develop_ck_path) and develop_ck_path not in sys.path):
-        os.environ["TL_COMPOSABLE_KERNEL_PATH"] = develop_ck_path + "/include"
-        env.COMPOSABLE_KERNEL_INCLUDE_DIR = develop_ck_path + "/include"
+    ck_inc_path = os.path.join(THIRD_PARTY_ROOT, 'composable_kernel', 'include')
+    if os.path.exists(ck_inc_path):
+        os.environ["TL_COMPOSABLE_KERNEL_PATH"] = env.COMPOSABLE_KERNEL_INCLUDE_DIR = ck_inc_path
     else:
         logger.warning(COMPOSABLE_KERNEL_NOT_FOUND_MESSAGE)
 
 # Initialize TL_TEMPLATE_PATH
 if os.environ.get("TL_TEMPLATE_PATH", None) is None:
-    install_tl_template_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "src")
-    develop_tl_template_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "src")
-    if os.path.exists(install_tl_template_path):
-        os.environ["TL_TEMPLATE_PATH"] = install_tl_template_path
-        env.TILELANG_TEMPLATE_PATH = install_tl_template_path
-    elif (os.path.exists(develop_tl_template_path) and develop_tl_template_path not in sys.path):
-        os.environ["TL_TEMPLATE_PATH"] = develop_tl_template_path
-        env.TILELANG_TEMPLATE_PATH = develop_tl_template_path
+    tl_template_path = os.path.join(THIRD_PARTY_ROOT, "..", "src")
+    if os.path.exists(tl_template_path):
+        os.environ["TL_TEMPLATE_PATH"] = env.TILELANG_TEMPLATE_PATH = tl_template_path
     else:
         logger.warning(TL_TEMPLATE_NOT_FOUND_MESSAGE)
 

tilelang/jit/adapter/cython/adapter.py

 """The profiler and convert to torch utils"""
 
 import ctypes
-import fcntl
-import hashlib
 import logging
-import site
-import sys
-import sysconfig
 import torch
-import os
-from pathlib import Path
 
 from typing import List, Optional, Union, Callable, Dict, Tuple, Any
 from tilelang import tvm as tvm
@@ -25,155 +18,15 @@ from tilelang.jit.adapter.utils import is_cuda_target, is_hip_target, is_cpu_tar
 from tilelang.utils.target import determine_target
 from tilelang.utils.language import retrieve_func_from_module
 from tilelang.utils.tensor import map_torch_type
-from tilelang.contrib.cc import get_cplus_compiler, is_darwin
 
 logger = logging.getLogger(__name__)
 
-
-def get_cython_compiler() -> Optional[str]:
-    """Return the path to the Cython compiler.
-
-    Returns
-    -------
-    out: Optional[str]
-        The path to the Cython compiler, or None if none was found.
-    """
-
-    cython_names = ["cython", "cython3"]
-
-    # Check system PATH
-    dirs_in_path = list(os.get_exec_path())
-
-    # Add user site-packages bin directory
-    user_base = site.getuserbase()
-    if user_base:
-        user_bin = os.path.join(user_base, "bin")
-        if os.path.exists(user_bin):
-            dirs_in_path = [user_bin] + dirs_in_path
-
-    # If in a virtual environment, add its bin directory
-    if sys.prefix != sys.base_prefix:
-        venv_bin = os.path.join(sys.prefix, "bin")
-        if os.path.exists(venv_bin):
-            dirs_in_path = [venv_bin] + dirs_in_path
-
-    for cython_name in cython_names:
-        for d in dirs_in_path:
-            cython_path = os.path.join(d, cython_name)
-            if os.path.isfile(cython_path) and os.access(cython_path, os.X_OK):
-                return cython_path
-    return None
-
-
-# Add cache management functions at module level
-def get_cache_dir() -> Path:
-    """Get the cache directory for the current Python version."""
-    py_version = f"py{sys.version_info.major}{sys.version_info.minor}"
-    # current directory
-    current_dir = os.path.dirname(os.path.abspath(__file__))
-    cache_dir = Path(current_dir) / ".cycache" / py_version
-    cache_dir.mkdir(parents=True, exist_ok=True)
-    return cache_dir
-
-
-def get_cached_lib(source_code: str) -> Tuple[Optional[ctypes.CDLL], Path]:
-    """Try to load cached library or return None if not found."""
-    code_hash = hashlib.sha256(source_code.encode()).hexdigest()
-    cache_path = get_cache_dir() / f"{code_hash}.so"
-    lock_file = cache_path.with_suffix('.lock')
-    with open(lock_file, 'w') as lock:
-        fcntl.flock(lock.fileno(), fcntl.LOCK_EX)
-        try:
-            if cache_path.exists():
-                try:
-                    if cache_path.stat().st_size > 1024:
-                        return ctypes.CDLL(str(cache_path)), cache_path
-                    else:
-                        cache_path.unlink()  # remove the incomplete file
-                except Exception as e:
-                    logger.error(f"Failed to load cached library: {e}")
-                    return None, cache_path
-            return None, cache_path
-        finally:
-            fcntl.flock(lock.fileno(), fcntl.LOCK_UN)
-
-
-# read the cython_wrapper.pyx file
-current_dir = os.path.dirname(os.path.abspath(__file__))
-cython_wrapper_path = os.path.join(current_dir, "cython_wrapper.pyx")
-
-with open(cython_wrapper_path, "r") as f:
-    cython_wrapper_code = f.read()
-    cache_dir = get_cache_dir()
-    source_path = cache_dir / "cython_wrapper.cpp"
-    library_path = cache_dir / "cython_wrapper.so"
-    md5_path = cache_dir / "md5.txt"
-    code_hash = hashlib.sha256(cython_wrapper_code.encode()).hexdigest()
-    cache_path = cache_dir / f"{code_hash}.so"
-    lock_file = cache_path.with_suffix('.lock')
-
-    # Check if cached version exists and is valid
-    need_compile = True
-    if md5_path.exists() and library_path.exists():
-        with open(md5_path, "r") as f:
-            cached_hash = f.read().strip()
-            if cached_hash == code_hash:
-                logger.debug("Cython JIT adapter is up to date, no need to compile...")
-                need_compile = False
-            else:
-                logger.info("Cython JIT adapter is out of date, need to recompile...")
-    else:
-        logger.info("No cached version found for Cython JIT adapter, need to compile...")
-
-    if need_compile:
-        logger.info("Waiting for lock to compile Cython JIT adapter...")
-        with open(lock_file, 'w') as lock:
-            fcntl.flock(lock.fileno(), fcntl.LOCK_EX)
-            try:
-                # After acquiring the lock, check again if the file has been compiled by another process
-                if md5_path.exists() and library_path.exists():
-                    with open(md5_path, "r") as f:
-                        cached_hash = f.read().strip()
-                        if cached_hash == code_hash:
-                            logger.info(
-                                "Another process has already compiled the file, using it...")
-                            need_compile = False
-
-                if need_compile:
-                    logger.info("Compiling Cython JIT adapter...")
-                    temp_path = cache_dir / f"temp_{code_hash}.so"
-
-                    with open(md5_path, "w") as f:
-                        f.write(code_hash)
-
-                    # compile the cython_wrapper.pyx file into .cpp
-                    cython = get_cython_compiler()
-                    if cython is None:
-                        raise Exception("Cython is not installed, please install it first.")
-                    os.system(f"{cython} {cython_wrapper_path} --cplus -o {source_path}")
-                    python_include_path = sysconfig.get_path("include")
-                    cc = get_cplus_compiler()
-                    dynamic_flag = '-Wl,-undefined,dynamic_lookup' if is_darwin(
-                    ) else '-Wl,--unresolved-symbols=ignore-all'
-                    command = f"{cc} -shared -pthread -fPIC -fwrapv -O2 -Wall -fno-strict-aliasing {dynamic_flag} -I{python_include_path} {source_path} -o {temp_path}"
-                    os.system(command)
-
-                    # rename the temp file to the library file
-                    temp_path.rename(library_path)
-            except Exception as e:
-                if 'temp_path' in locals() and temp_path.exists():
-                    temp_path.unlink()
-                raise Exception(f"Failed to compile Cython JIT adapter: {e}") from e
-            finally:
-                if lock_file.exists():
-                    lock_file.unlink()
-
-    # add the .so file to the sys.path
-    cache_dir_str = str(cache_dir)
-    if cache_dir_str not in sys.path:
-        sys.path.append(cache_dir_str)
-
-from cython_wrapper import CythonKernelWrapper
+try:
+    # Load cython_wrapper.api3.so in env.py
+    from cython_wrapper import CythonKernelWrapper
+except ImportError:
+    # TODO: tolerance a build without cython backend
+    raise
 
 
 class CythonKernelAdapter(BaseKernelAdapter):

tilelang/libinfo.py

-"""Library information. This is a standalone file that can be used to get various info.
-Modified from: https://github.com/mlc-ai/mlc-llm/blob/main/python/mlc_llm/libinfo.py
-"""
-
-#! pylint: disable=protected-access
-import os
 import sys
+import os
 
-TILELANG_LIBRARY_PATH = os.environ.get("TILELANG_LIBRARY_PATH", None)
-
-
-def get_env_paths(env_var, splitter):
-    """Get path in env variable"""
-    if os.environ.get(env_var, None):
-        return [p.strip() for p in os.environ[env_var].split(splitter)]
-    return []
-
+from .env import TL_LIBS
 
-def get_dll_directories():
-    """Get extra tile lang dll directories"""
-    curr_dir = os.path.dirname(os.path.realpath(os.path.expanduser(__file__)))
-    source_dir = os.path.abspath(os.path.join(curr_dir, ".."))
-    dll_path = [
-        curr_dir,
-        os.path.join(source_dir, "build"),  # local build
-        os.path.join(source_dir, "build", "Release"),
-        os.path.join(curr_dir, "lib"),  # pypi build
-    ]
-    if TILELANG_LIBRARY_PATH:
-        dll_path.append(TILELANG_LIBRARY_PATH)
-    if "CONDA_PREFIX" in os.environ:
-        dll_path.append(os.path.join(os.environ["CONDA_PREFIX"], "lib"))
-    if sys.platform.startswith("linux") or sys.platform.startswith("freebsd"):
-        dll_path.extend(get_env_paths("LD_LIBRARY_PATH", ":"))
-    elif sys.platform.startswith("darwin"):
-        dll_path.extend(get_env_paths("DYLD_LIBRARY_PATH", ":"))
-    elif sys.platform.startswith("win32"):
-        dll_path.extend(get_env_paths("PATH", ";"))
-    return [os.path.abspath(p) for p in dll_path if os.path.isdir(p)]
 
-
-def find_lib_path(name, optional=False):
+def find_lib_path(name: str, py_ext=False):
     """Find tile lang library
 
     Parameters
@@ -50,7 +15,9 @@ def find_lib_path(name, optional=False):
     optional: boolean
         Whether the library is required
     """
-    if sys.platform.startswith("linux") or sys.platform.startswith("freebsd"):
+    if py_ext:
+        lib_name = f"{name}.abi3.so"
+    elif sys.platform.startswith("linux") or sys.platform.startswith("freebsd"):
         lib_name = f"lib{name}.so"
     elif sys.platform.startswith("win32"):
         lib_name = f"{name}.dll"
@@ -59,11 +26,11 @@ def find_lib_path(name, optional=False):
     else:
         lib_name = f"lib{name}.so"
 
-    dll_paths = get_dll_directories()
-    lib_dll_path = [os.path.join(p, lib_name) for p in dll_paths]
-    lib_found = [p for p in lib_dll_path if os.path.exists(p) and os.path.isfile(p)]
-    if not lib_found and not optional:
+    for lib_root in TL_LIBS:
+        lib_dll_path = os.path.join(lib_root, lib_name)
+        if os.path.exists(lib_dll_path) and os.path.isfile(lib_dll_path):
+            return lib_dll_path
+    else:
         message = (f"Cannot find libraries: {lib_name}\n" + "List of candidates:\n" +
-                   "\n".join(lib_dll_path))
+                   "\n".join(TL_LIBS))
         raise RuntimeError(message)
-    return lib_found

tilelang/version.py

-import os
-import subprocess
-from typing import Union
-
-# Get the absolute path of the current Python script's directory
-current_dir = os.path.dirname(os.path.abspath(__file__))
-
-# Get the absolute path of the project root directory (one level above the current directory)
-develop_project_root_dir = os.path.abspath(os.path.join(current_dir, ".."))
-installed_project_root_dir = os.path.abspath(os.path.join(current_dir))
-# Define the path to the VERSION file located in the project root directory
-develop_version_file_path = os.path.join(develop_project_root_dir, "VERSION")
-installed_version_file_path = os.path.join(installed_project_root_dir, "VERSION")
-
-if os.path.exists(develop_version_file_path):
-    version_file_path = develop_version_file_path
-elif os.path.exists(installed_version_file_path):
-    version_file_path = installed_version_file_path
-else:
-    raise FileNotFoundError("VERSION file not found in the project root directory")
-
-# Read and store the version information from the VERSION file
-# Use 'strip()' to remove any leading/trailing whitespace or newline characters
-with open(version_file_path, "r") as version_file:
-    __version__ = version_file.read().strip()
-
-
-def get_git_commit_id() -> Union[str, None]:
-    """Get the current git commit hash by running git in the current file's directory."""
-    try:
-        return subprocess.check_output(['git', 'rev-parse', 'HEAD'],
-                                       cwd=os.path.dirname(os.path.abspath(__file__)),
-                                       stderr=subprocess.DEVNULL,
-                                       encoding='utf-8').strip()
-    # FileNotFoundError is raised when git is not installed
-    except (subprocess.SubprocessError, FileNotFoundError):
-        return None
-
-
-# Append git commit hash to version if not already present
-# NOTE(lei): Although the local commit id cannot capture locally staged changes,
-# the local commit id can help mitigate issues caused by incorrect cache to some extent,
-# so it should still be kept.
-# Check WITH_COMMITID environment variable to control whether to include commit ID
-WITH_COMMITID = os.environ.get("WITH_COMMITID", "True").lower() == "true"
-if WITH_COMMITID and "+" not in __version__ and (commit_id := get_git_commit_id()):
-    # Use short commit ID (8 characters) for better compatibility
-    short_commit_id = commit_id[:8]
-    __version__ = f"{__version__}+{short_commit_id}"
-
-# Define the public API for the module
-__all__ = ["__version__"]

tox.ini

-[tox]
-envlist = py38,py39,py310,py311,py312
-isolated_build = False
-
-[testenv:py{38,39,310,311,312}]
-skip_install = false
-deps =
-    wheel
-    build
-setenv =
-    WITH_COMMITID = TRUE
-    PYTHON_EXECUTABLE = {envpython}
-    Python3_EXECUTABLE = {envpython}
-commands =
-    python -m build --wheel -o {toxinidir}/dist
-
-[testenv:py{38,39,310,311,312}-pypi]
-skip_install = false
-setenv =
-    PYPI_BUILD = TRUE
-    WITH_COMMITID = FALSE
-    PYTHON_EXECUTABLE = {envpython}
-    Python3_EXECUTABLE = {envpython}
-commands =
-    python setup.py bdist_wheel --plat-name=manylinux2014_x86_64
-
-[testenv:audit_manylinux2014]
-skip_install = true
-allowlist_externals =
-    bash
-deps =
-    auditwheel
-    patchelf
-commands =
-    bash -c 'auditwheel repair -L=/lib --exclude=/usr/local/cuda* --exclude=libcuda.so.1 --plat=manylinux2014_x86_64 dist/*'
-
-[testenv:py38]
-basepython = python3.8
-
-[testenv:py39]
-basepython = python3.9
-
-[testenv:py310]
-basepython = python3.10
-
-[testenv:py311]
-basepython = python3.11
-
-[testenv:py312]
-basepython = python3.12

version_provider.py

+from __future__ import annotations
+
+import os
+import platform
+import subprocess
+from typing import Optional
+from pathlib import Path
+
+ROOT = Path(__file__).parent
+
+base_version = (ROOT / 'VERSION').read_text().strip()
+
+
+def _read_cmake_bool(i: str | None, default=False):
+    if i is None:
+        return default
+    return i.lower() not in ('0', 'false', 'off', 'no', 'n', '')
+
+
+def get_git_commit_id() -> Optional[str]:
+    """Get the current git commit hash by running git in the current file's directory."""
+
+    r = subprocess.run(['git', 'rev-parse', 'HEAD'],
+                       cwd=ROOT,
+                       stdout=subprocess.PIPE,
+                       stderr=subprocess.PIPE,
+                       encoding='utf-8')
+    if r.returncode == 0:
+        return r.stdout.strip()
+    else:
+        return 'unknown'
+
+
+def dynamic_metadata(
+    field: str,
+    settings: dict[str, object] | None = None,
+) -> str:
+    assert field == 'version'
+
+    version = base_version
+
+    if not _read_cmake_bool(os.environ.get('NO_VERSION_LABEL')):
+        exts = []
+        backend = None
+        if _read_cmake_bool(os.environ.get('NO_TOOLCHAIN_VERSION')):
+            pass
+        elif platform.system() == 'Darwin':
+            # only on macosx_11_0_arm64, not necessary
+            # backend = 'metal'
+            pass
+        elif _read_cmake_bool(os.environ.get('USE_ROCM', '')):
+            backend = 'rocm'
+        elif 'USE_CUDA' in os.environ and not _read_cmake_bool(os.environ.get('USE_CUDA')):
+            backend = 'cpu'
+        else:  # cuda
+            # Read nvcc version from env.
+            # This is not exactly how it should be,
+            # but works for now if building in a nvidia/cuda image.
+            if cuda_version := os.environ.get('CUDA_VERSION'):
+                major, minor, *_ = cuda_version.split('.')
+                backend = f'cu{major}{minor}'
+            else:
+                backend = 'cuda'
+        if backend:
+            exts.append(backend)
+
+        if _read_cmake_bool(os.environ.get('NO_GIT_VERSION')):
+            pass
+        elif git_hash := get_git_commit_id():
+            exts.append(f'git{git_hash[:8]}')
+
+        if exts:
+            version += '+' + '.'.join(exts)
+
+    return version
+
+
+__all__ = ["dynamic_metadata"]