setup.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# Copyright 2024, 2026 Arm Limited and/or its affiliates.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

# Part of this code is from pybind11 cmake_example, so attach the license below.
# That project has since dropped setup.py, so this points at the last revision
# that still had it instead of at a branch.
# https://github.com/pybind/cmake_example/blob/7a94877f581a14de4de1a096fb053a55fc2a66bf/setup.py

# Copyright (c) 2016 The Pybind Development Team, All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
#    list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright notice,
#    this list of conditions and the following disclaimer in the documentation
#    and/or other materials provided with the distribution.
#
# 3. Neither the name of the copyright holder nor the names of its contributors
#    may be used to endorse or promote products derived from this software
#    without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
# ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
# WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
# You are under no obligation whatsoever to provide any bug fixes, patches, or
# upgrades to the features, functionality or performance of the source code
# ("Enhancements") to anyone; however, if you choose to make your Enhancements
# available either publicly, or directly to the author of this software, without
# imposing a separate written license agreement for such Enhancements, then you
# hereby grant the following license: a non-exclusive, royalty-free perpetual
# license to install, use, modify, prepare derivative works, incorporate into
# other computer software, distribute, and sublicense such enhancements or
# derivative works thereof, in binary and source code form.

import ast
import contextlib
import functools

# Import this before distutils so that setuptools can intercept the distuils
# imports.
import importlib.util
import logging
import os
import re
import shlex
import shutil
import site
import stat
import subprocess
import sys
from distutils import log  # type: ignore[import-not-found]
from distutils.sysconfig import get_python_lib  # type: ignore[import-not-found]
from pathlib import Path, PurePosixPath
from typing import Dict, FrozenSet, List, Optional, Set, Tuple

# Clean dynamic import using importlib
_install_utils_path = Path(__file__).parent / "install_utils.py"
_spec = importlib.util.spec_from_file_location("install_utils", _install_utils_path)
if _spec is None:
    raise ImportError(f"Could not create module spec for {_install_utils_path}")
install_utils = importlib.util.module_from_spec(_spec)
if _spec.loader is None:
    raise ImportError(f"Module spec has no loader for {_install_utils_path}")
_spec.loader.exec_module(install_utils)

from setuptools import Distribution, Extension, find_namespace_packages, setup
from setuptools.command.build import build
from setuptools.command.build_ext import build_ext
from setuptools.command.build_py import build_py

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [%(levelname)s] %(message)s",
)

# Headers swept in by a directory copy that a consumer of the wheel cannot use, because each needs
# something the wheel does not carry. Publishing one is worse than leaving it out: the failure arrives in
# someone else's project rather than here.
#
# Matched on the path ending, not the bare file name, so an entry names one specific header rather than
# every file that happens to share its name. Anything under a directory beginning with "test" is already
# dropped separately, so those need no entry here.
#
# Only headers that nothing else the wheel installs includes belong here. A header other shipped headers
# pull in must keep shipping even when it cannot be compiled on its own.
_UNSHIPPABLE_HEADERS = frozenset(
    {
        # Needs a header generated when the schema is compiled, which in turn needs the FlatBuffers C++
        # headers. Those are a third-party library this wheel does not vendor.
        "runtime/executor/tensor_parser.h",
        # Reads processor details through cpuinfo, whose headers the wheel does not publish.
        "extension/threadpool/cpuinfo_utils.h",
        # Holds a pthreadpool member by value, so it needs that library's header, which the wheel does not
        # publish either. The component it belongs to is a link dependency the runtime carries, not
        # something a consumer includes.
        "extension/threadpool/threadpool.h",
        # Declares CPUCachingAllocator, whose implementation is in a component no shipped library links,
        # so including it compiles and then fails at link time with an undefined reference.
        "extension/memory_allocator/cpu_caching_malloc_allocator.h",
        # Declares BundledModule, which is built only for the Python bindings, so its implementation is in
        # the Python extension. A C++ application cannot link that, and building the source instead needs
        # bundled-program headers the wheel does not publish.
        "extension/module/bundled_module.h",
        # Declares FileDescriptorDataLoader, whose implementation is in no CMake target at all, so no
        # shipped library defines it. Including it compiles and then fails at link time.
        "extension/data_loader/file_descriptor_data_loader.h",
    }
)

try:
    from tools.cmake.cmake_cache import CMakeCache
except ImportError:
    sys.path.insert(
        0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "tools", "cmake")
    )
    from cmake_cache import CMakeCache  # type: ignore[no-redef, import-not-found]


def _is_macos() -> bool:
    return sys.platform == "darwin"


def _is_windows() -> bool:
    return sys.platform == "win32"


def _is_env_flag_enabled(name: str) -> bool:
    return os.environ.get(name, "").strip().upper() in {"1", "ON", "TRUE", "YES"}


def _is_minimal_build() -> bool:
    return _is_env_flag_enabled("EXECUTORCH_BUILD_MINIMAL")


def _minimal_cmake_flags() -> List[str]:
    return [
        "-DEXECUTORCH_BUILD_COREML=OFF",
        "-DEXECUTORCH_BUILD_CUDA=OFF",
        "-DEXECUTORCH_BUILD_DEVTOOLS=OFF",
        "-DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=OFF",
        "-DEXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR=OFF",
        "-DEXECUTORCH_BUILD_EXTENSION_LLM=OFF",
        "-DEXECUTORCH_BUILD_EXTENSION_LLM_RUNNER=OFF",
        "-DEXECUTORCH_BUILD_EXTENSION_MODULE=OFF",
        "-DEXECUTORCH_BUILD_EXTENSION_NAMED_DATA_MAP=OFF",
        "-DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=OFF",
        "-DEXECUTORCH_BUILD_EXTENSION_TENSOR=OFF",
        "-DEXECUTORCH_BUILD_EXTENSION_TRAINING=OFF",
        "-DEXECUTORCH_BUILD_KERNELS_CUSTOM_AOT=OFF",
        "-DEXECUTORCH_BUILD_KERNELS_LLM=OFF",
        "-DEXECUTORCH_BUILD_KERNELS_LLM_AOT=OFF",
        "-DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=OFF",
        "-DEXECUTORCH_BUILD_KERNELS_QUANTIZED=OFF",
        "-DEXECUTORCH_BUILD_KERNELS_QUANTIZED_AOT=OFF",
        "-DEXECUTORCH_BUILD_KERNELS_TORCHAO=OFF",
        "-DEXECUTORCH_BUILD_MLX=OFF",
        "-DEXECUTORCH_BUILD_OPENVINO=OFF",
        "-DEXECUTORCH_BUILD_PORTABLE_OPS=OFF",
        "-DEXECUTORCH_BUILD_PYBIND=OFF",
        "-DEXECUTORCH_BUILD_QNN=OFF",
        "-DEXECUTORCH_BUILD_TESTS=OFF",
        "-DEXECUTORCH_BUILD_VULKAN=OFF",
        "-DEXECUTORCH_BUILD_XNNPACK=OFF",
        "-DEXECUTORCH_BUILD_CMSIS_NN_PYBINDS=OFF",
    ]


_VENDORED_DIR_NAMES = frozenset({"third-party", "third_party"})

# Used only when .gitmodules cannot be read, as in a source distribution. A test keeps it in step.
_VENDORED_SUBMODULE_FALLBACK = (
    "backends/cadence/utils/FACTO",
    "extension/llm/tokenizers",
)


@functools.lru_cache(maxsize=None)
def _vendored_prefixes() -> Tuple[str, ...]:
    """Source-tree prefixes holding code from another repository.

    Two shapes reach the wheel. Most vendored code sits in a directory named third-party,
    which the name above covers wherever it appears. The rest are git submodules checked out
    under an ordinary name, so they can only be recognized by asking git what they are.

    None of them are importable from where they sit. FACTO is pure Python but its nested copy
    cannot satisfy backends/cadence/utils/facto_util.py, which imports the top level facto.specdb,
    and the tokenizers ship separately as pytorch-tokenizers in the dependency list. The rest,
    XNNPACK and the Vulkan headers among them, are C++ sources that the wheel has no use for once
    the libraries are built.

    Read through git rather than by scanning the file, so only real submodule entries count.
    A hand-rolled reader accepts a `path` line from any section, and one stray line elsewhere
    in the file would drop a first-party package from the wheel with nothing to warn about.

    Submodules at the repository root are skipped. Those are build tooling, never copied into
    the package, and carrying a bare single-word name here would make the match below drop any
    directory that happened to share it.
    """
    root = Path(__file__).parent
    if not (root / ".gitmodules").is_file():
        # A source distribution carries no .gitmodules, so nothing can be read there. Fall
        # back to the directories the vendored trees occupy, or the exclusion would quietly
        # do half its job and those files would ship again.
        return _VENDORED_SUBMODULE_FALLBACK
    try:
        listed = subprocess.run(
            [
                "git",
                "config",
                "-z",
                "-f",
                ".gitmodules",
                "--get-regexp",
                r"^submodule\..*\.path$",
            ],
            cwd=root,
            capture_output=True,
            text=True,
            check=False,
        )
    except OSError:
        # No git on PATH, so fall back for the same reason as above.
        return _VENDORED_SUBMODULE_FALLBACK

    if listed.returncode or not listed.stdout.strip():
        # git ran and told us nothing useful, which happens when the file has a bad section
        # header or conflict markers in it. Reading that as "no submodules" would turn the
        # exclusion off without a word, so fall back rather than trust an empty answer.
        return _VENDORED_SUBMODULE_FALLBACK

    prefixes = []
    # -z separates each record with NUL and its key from its value with a newline, so neither a
    # name nor a path containing a space can be misread. Splitting the default space-separated
    # output cannot do that: "submodule.a b.path c d/e" is ambiguous either way round.
    for record in listed.stdout.split("\0"):
        if not record:
            continue
        _, separator, value = record.partition("\n")
        if not separator:
            continue
        # Normalize, because git accepts a trailing slash, a ./ prefix and doubled
        # separators as the same path, and the raw text would stop matching the real
        # directory.
        parts = Path(value).parts
        if len(parts) < 2 or any(part in _VENDORED_DIR_NAMES for part in parts):
            continue
        prefixes.append("/".join(parts))
    return tuple(sorted(prefixes))


def _is_vendored_path(path: str) -> bool:
    """Whether a source-tree path holds code from another repository."""
    parts = Path(path).parts
    if any(part in _VENDORED_DIR_NAMES for part in parts):
        return True
    # A submodule path is relative to the repository root, while a path here may be relative
    # to src/executorch or carry a src/executorch prefix, so match on any suffix boundary.
    # Whole-component match: the prefix must be the entire path, or sit at its start, end, or
    # middle bounded by separators. Substring matching would let a directory whose name merely
    # begins with a prefix be dropped.
    posix = "/".join(parts)
    return any(
        posix == prefix
        or posix.startswith(f"{prefix}/")
        or posix.endswith(f"/{prefix}")
        or f"/{prefix}/" in posix
        for prefix in _vendored_prefixes()
    )


def _minimal_packages() -> List[str]:
    return sorted(
        find_namespace_packages(
            # Anchored on this file, not the working directory, so the list does not change
            # with where the build was started from. A cwd-relative path returns nothing when
            # the build runs from anywhere but the repository root, and a wheel with no packages
            # in it ships no Python at all.
            where=str(Path(__file__).parent / "src"),
            include=[
                "executorch",
                "executorch.data",
                "executorch.data.bin",
                "executorch.exir",
                "executorch.exir.*",
                "executorch.extension",
                "executorch.extension.flat_tensor",
                "executorch.extension.flat_tensor.*",
                "executorch.extension.pytree",
            ],
            exclude=[
                "*.test",
                "*.test.*",
                "*.tests",
                "*.tests.*",
                "*.__pycache__",
                "*.__pycache__.*",
            ],
        )
    )


_WALK_SKIP_DIRS = frozenset(
    {".git", "pip-out", "cmake-out", "third-party", "third_party", "__pycache__"}
)

_TEST_DIR_NAMES = frozenset({"test", "tests"})


@functools.lru_cache(maxsize=None)
def _top_level_package_dirs() -> FrozenSet[str]:
    """The first path component of every package the wheel ships.

    Derived from the tree rather than listed, so a new top-level directory is covered without an
    edit here. Used to recognize the unprefixed spelling of a first-party import.
    """
    root = Path(__file__).parent / "src" / "executorch"
    if not root.is_dir():
        return frozenset()
    return frozenset(entry.name for entry in root.iterdir() if entry.is_dir())


# Named only by a workflow or by a documented command, so no import reaches them. Listed here
# rather than scanned from .github, which a source distribution does not carry; a test re-derives
# the list so it cannot drift.
_CI_ENTRY_POINTS = (
    "executorch.backends.cuda.batching.test.export_toy_decoder",
    "executorch.backends.mlx.test.run_all_tests",
    "executorch.backends.mlx.test.test_sample",
    "executorch.backends.mlx.test.test_slot_recycling",
    "executorch.backends.samsung.test.utils.run_tests",
    "executorch.backends.test.suite.generate_markdown_summary_json",
    "executorch.examples.models.muse_glimmer.tests.gen_prompt_golden",
    "executorch.examples.models.muse_glimmer.tests.test_mlx_pipeline",
    "executorch.examples.models.muse_glimmer.tests.test_prompt_tokens",
    "executorch.extension.pybindings.test.test_pybindings",
)

# Directories whose test modules are reached without any import statement naming them, so no scan
# of the source can find them: mlx.yml runs each file it discovers under custom_kernel_ops, the
# webgpu scripts import one module per operator, runner.py resolves a suite root out of a dict and
# then walks it, and the llava README documents a `python -m` command. Directories rather than file
# names, so a new test is covered when it is added.
_CI_ENTRY_POINT_DIRS = (
    "executorch.backends.mlx.custom_kernel_ops",
    "executorch.backends.webgpu.test",
    "executorch.backends.test.suite",
    "executorch.examples.models.llava.test",
)


def _is_test_module(dotted: str) -> bool:
    return any(part in _TEST_DIR_NAMES for part in dotted.split("."))


def _module_name(root: Path, path: Path) -> str:
    parts = list(path.relative_to(root).parts)
    if parts[-1] == "__init__.py":
        parts = parts[:-1]
    else:
        parts[-1] = parts[-1][: -len(".py")]
    return ".".join(["executorch"] + parts)


def _first_party_module(name: str) -> Optional[str]:
    """The `executorch.`-prefixed spelling of an import target, or None if it is not ours.

    This repository imports itself two ways. Most code says `executorch.backends.x`, but some
    says `backends.x`, which resolves because pytest puts the repository root on sys.path. Both
    name the same file, so both have to count as a reference or a helper reached only by the
    second spelling is dropped from the wheel while its importers still expect it.
    """
    if name.startswith("executorch."):
        return name
    if name.split(".", 1)[0] in _top_level_package_dirs():
        return f"executorch.{name}"
    return None


def _scan_imports(path: Path, package: str, out: Set[str], dynamic: Set[str]) -> None:
    """Collect into out the executorch modules one file refers to."""
    try:
        tree = ast.parse(path.read_text(encoding="utf-8", errors="replace"))
    except SyntaxError:
        return
    for node in ast.walk(tree):
        out.update(_import_targets(node, package, dynamic))


_GENERATED_DIR_NAMES = frozenset(
    {
        ".venv",
        "venv",
        "build",
        "dist",
        "buck-out",
        ".cache",
        ".hypothesis",
        ".mypy_cache",
        ".pytest_cache",
        ".tox",
        "test-build",
        "arm_test",
        "riscv_test",
    }
)


def _unshipped_directories(root: Path) -> List[Path]:
    """Checkout directories the wheel does not carry, whose imports still have to be followed.

    src/executorch is a subset of the repository, so a file under test/ or tools/ is never
    packaged, yet a module it imports still has to ship.

    Generated directories are left out, because a build tree or an in-tree virtualenv holds an
    INSTALLED copy of this package, and reading it would let the last wheel vote on what the next
    one ships. Listed by name rather than asked of git, because `git check-ignore` needs a working
    repository and answers differently for a pattern with a trailing slash depending on whether the
    directory exists yet, which made the same build behave differently on two platforms.
    """
    repository = Path(__file__).parent
    if not repository.is_dir() or not root.is_dir():
        return []
    shipped = {entry.name for entry in root.iterdir()}
    return [
        entry
        for entry in sorted(repository.iterdir())
        if entry.is_dir()
        and entry.name not in shipped
        and entry.name not in _WALK_SKIP_DIRS
        and entry.name not in _GENERATED_DIR_NAMES
        and entry.name not in ("src", ".github")
    ]


def _import_graph(root: Path) -> Tuple[Set[str], Dict[str, Set[str]], Set[str]]:
    """Every module under root, what each imports, and literal importlib targets.

    The walk covers root, but the SEED covers more: a file elsewhere in the checkout can import
    a module that ships, so its imports are collected too and attributed to a synthetic name.
    Without that, a helper whose only importer lives outside the shipped tree looks unreachable.
    """
    modules: Set[str] = set()
    edges: Dict[str, Set[str]] = {}
    dynamic: Set[str] = set()

    # followlinks, because src/executorch is a tree of symlinks into the repository root.
    for dirpath, dirnames, filenames in os.walk(root, followlinks=True):
        # Vendored trees are skipped by the same test that excludes them from the package list,
        # not only by directory name. A submodule checked out under an ordinary name, FACTO and
        # the tokenizers among them, is otherwise read as first-party, and its imports would keep
        # test modules the wheel has no reason to carry.
        dirnames[:] = [
            d
            for d in dirnames
            if d not in _WALK_SKIP_DIRS
            and not _is_vendored_path(os.path.relpath(os.path.join(dirpath, d), root))
        ]
        for filename in filenames:
            if not filename.endswith(".py"):
                continue
            path = Path(dirpath) / filename
            me = _module_name(root, path)
            modules.add(me)
            package = me if filename == "__init__.py" else me.rsplit(".", 1)[0]
            _scan_imports(path, package, edges.setdefault(me, set()), dynamic)

    # Directories of the checkout that the wheel does not ship, test/ among them. Their files are
    # never packaged, so they are not modules, but what they import must still ship: for example
    # test/end2end/test_end2end.py imports two model helpers out of exir/tests.
    for entry in _unshipped_directories(root):
        for dirpath, dirnames, filenames in os.walk(entry, followlinks=False):
            dirnames[:] = [d for d in dirnames if d not in _WALK_SKIP_DIRS]
            for filename in filenames:
                if not filename.endswith(".py"):
                    continue
                outside = f"<outside>{dirpath}/{filename}"
                _scan_imports(
                    Path(dirpath) / filename,
                    "",
                    edges.setdefault(outside, set()),
                    dynamic,
                )

    return modules, edges, dynamic


def _import_targets(node: ast.AST, package: str, dynamic: Set[str]) -> Set[str]:
    """The executorch modules one AST node refers to."""
    found: Set[str] = set()
    if isinstance(node, ast.Import):
        found.update(
            name for name in (_first_party_module(a.name) for a in node.names) if name
        )
    elif isinstance(node, ast.ImportFrom):
        if node.level:
            if not package:
                # A file outside the shipped tree, so a relative import stays inside that tree
                # and cannot name anything the wheel carries.
                return found
            # A relative import names a real module too, and inside a kept package its target
            # has to ship: stages/__init__.py does `from .export import Export`, so dropping
            # stages.export would break every importer of that package.
            parts = package.split(".")
            if node.level > 1:
                parts = parts[: len(parts) - (node.level - 1)]
            base = ".".join(parts + (node.module.split(".") if node.module else []))
        elif node.module and (prefixed := _first_party_module(node.module)):
            base = prefixed
        else:
            return found
        if base.startswith("executorch"):
            found.add(base)
            # `from pkg import name` may name a submodule rather than an attribute, and there
            # is no way to tell without importing, so both readings are kept.
            found.update(f"{base}.{a.name}" for a in node.names)
    elif (
        isinstance(node, ast.Call)
        and isinstance(node.func, ast.Attribute)
        and node.func.attr == "import_module"
        and node.args
        and isinstance(node.args[0], ast.Constant)
        and isinstance(node.args[0].value, str)
    ):
        target = node.args[0].value
        if target.startswith("."):
            if not package:
                return found
            target = package + target
        resolved = _first_party_module(target)
        if resolved:
            dynamic.add(resolved)
    return found


@functools.lru_cache(maxsize=None)
def _reachable_test_modules() -> FrozenSet[str]:
    """Test modules something can still reach once the wheel is installed.

    A test case that nothing imports is dead weight in the wheel: pytest loads it from a path in
    the checkout, never through the installed name. A shared helper is the opposite, because the
    suites import each other by installed name, so it has to ship or collection breaks.

    Reachable means named by something, anywhere in the checkout, including by a test module
    itself. That looks circular and is not: a test collected from the checkout still resolves
    `from executorch.x.test import helper` through the INSTALLED package, so the helper must be
    in the wheel even though the file importing it is not.
    """
    root = Path(__file__).parent / "src" / "executorch"
    modules, edges, dynamic = _import_graph(root)

    # Every name anything refers to. No transitive walk is needed: this is already the union of
    # every edge target, so following an edge could only rediscover a name that is in here.
    referenced = set(dynamic)
    referenced.update(_CI_ENTRY_POINTS)
    for targets in edges.values():
        referenced.update(targets)

    keep = {name for name in referenced if _is_test_module(name)} & modules
    # Everything under a directory whose tests are run one file at a time by a discovery loop.
    keep |= {
        name
        for name in modules
        if _is_test_module(name)
        and any(
            name == prefix or name.startswith(f"{prefix}.")
            for prefix in _CI_ENTRY_POINT_DIRS
        )
    }
    # Parent packages of anything kept, or the dotted path cannot resolve.
    for name in list(keep):
        parts = name.split(".")
        for end in range(2, len(parts)):
            parent = ".".join(parts[:end])
            if _is_test_module(parent):
                keep.add(parent)
    return frozenset(keep)


_SHADER_TEMPLATE_MARKERS = (
    "parameter_names_with_default_values",
    "shader_variants",
    "generate_variant_forall",
)


@functools.lru_cache(maxsize=None)
def _is_shader_template(path: str) -> bool:
    """Whether a yaml file is a shader codegen input rather than data the wheel needs.

    gen_vulkan_spv.py and gen_wgsl_headers.py expand these into SPIR-V and WGSL headers during
    the cmake build, so the wheel already carries the compiled result. Matched on content rather
    than on a directory list, because the same shape appears under vulkan and webgpu and a path
    list goes stale as soon as a backend adds one. The op and kernel definitions that ARE read
    at run time, edge.yaml among them, carry none of these keys.
    """
    if not path.endswith(".yaml"):
        return False
    full = Path(__file__).parent / path
    try:
        head = full.read_text(encoding="utf-8", errors="replace")
    except OSError:
        return False
    return any(marker in head for marker in _SHADER_TEMPLATE_MARKERS)


def _full_packages() -> List[str]:
    """Every package the full wheel ships.

    Without an explicit list setuptools discovers all of src/executorch, which pulls in the
    Python files and codegen scripts of the vendored third-party checkouts. Those exist to
    build the C++ targets, so once the libraries are built no shipped module imports them.

    Test packages deliberately stay. The suites in this repository import each other through
    the installed name, for example `from executorch.backends.arm.test import common`, so
    dropping them from the wheel stops the suites collecting under a non-editable install.
    """
    return sorted(
        package
        # Anchored on this file rather than the working directory, so the list does not
        # change with where the build or a test was started from.
        for package in find_namespace_packages(
            where=str(Path(__file__).parent / "src"),
            include=["executorch", "executorch.*"],
        )
        # The include patterns above DO match these, since they are ordinary dotted names,
        # which is exactly why they have to be removed here instead.
        if not _is_vendored_path(package.replace(".", "/"))
    )


# The published project names for the CUDA runtime components a CUDA wheel links but
# does not bundle, keyed by CUDA major version. Not derivable from a suffix rule: the
# CUDA 12 wheels carry a "-cu12" suffix while the CUDA 13 ones are published under
# unsuffixed names. A train with no entry here declares nothing rather than guessing a
# name that may not exist.
#
# Only what a shipped library actually loads. Measured on a built wheel, the CUDA
# libraries need the CUDA runtime and nothing else, because cuRAND is used only through
# its device-side header API, which compiles into the object rather than linking a
# library, and the generated model library embeds its kernels rather than compiling them
# at run time, so there is no runtime compiler to satisfy either.
# Bounded to the train the binaries were built against. The shipped libraries record
# NEEDED libcudart.so.<major> and a runtime path into that train's own directory, so a
# resolution to a different major installs a different layout and a different soname and the
# wheel is unimportable. Nothing catches that at install time; it surfaces as an unresolved
# libcudart on first import. The 12 package is already major-specific by name, but the 13 one
# is not, so it needs the specifier to say what the name does not.
_CUDA_RUNTIME_PACKAGES = {
    "12": ("nvidia-cuda-runtime-cu12>=12,<13",),
    "13": ("nvidia-cuda-runtime>=13,<14",),
}

# Where each train installs its libraries under site-packages. CUDA 13 collects them in
# one directory while CUDA 12 gives each component its own, so the search path differs by
# train and cannot be a single literal.
#
# Every declared package needs its directory here, and nothing else belongs. The loader only
# searches what is recorded here, so a missing directory leaves a shipped library unable to find
# a package that is installed, and an extra one implies a dependency the wheel does not have.
_CUDA_LIBRARY_DIRECTORIES = {
    "12": ("nvidia/cuda_runtime/lib",),
    "13": ("nvidia/cu13/lib",),
}

# Arch subdirectory names MKL's exported link interface appends to its prefix.
_MKL_ARCH_DIRECTORIES = ("intel64", "intel64_win", "win-x64")


def _cmake_args() -> List[str]:
    """CMAKE_ARGS split into arguments, tolerating an unbalanced quote.

    shlex is the correct parser for a value that names a shell argument list, but it raises on an
    unbalanced quote, and a path containing an apostrophe is enough to trigger it. Both callers run at
    module scope, so the exception surfaced as a traceback during the build rather than a diagnosable
    error. Falling back to whitespace splitting keeps the build working for the case that caused it.
    """
    raw = os.environ.get("CMAKE_ARGS", "")
    try:
        return shlex.split(raw)
    except ValueError:
        return raw.split()


# Release rows that ship no CUDA. Only the metadata side reads this. The shell classifier is
# deliberately broader, treating every name it does not recognise as CPU, so the two do not
# agree on rows like rocm or xpu and are not meant to. What matters is that a row this side
# calls CPU declares no CUDA runtime, which is what the list is for.
_CPU_ROW_NAMES = ("cpu", "cpu-aarch64")


def _row_is_cpu_only() -> bool:
    """Whether the release row this build belongs to names itself a CPU row.

    The metadata side already reads the row to decide which NVIDIA packages to declare, so the build
    has to read the same input or the two disagree and the wheel ships a delegate it cannot load.
    Absent means unknown rather than CPU, which keeps a plain local build behaving as before.
    """
    raw = (
        (os.environ.get("CU_VERSION") or os.environ.get("DESIRED_CUDA") or "")
        .strip()
        .lower()
    )
    return raw in _CPU_ROW_NAMES


def _cuda_train() -> str:
    """The CUDA major version this wheel is being built for, or "" for a CPU wheel.

    The release row's own field wins when it is set, because a row states the train it
    targets and that is more authoritative than whichever toolkit happens to sit on the
    builder. The wheel build exports CU_VERSION; DESIRED_CUDA is the matrix field name.

    Falling back to the installed toolkit matters for every build that is not a release
    job. The build turns CUDA on by detecting a toolkit, so keying only off the release
    field produced a wheel that carried the CUDA libraries with no dependency declarations
    and no way to find the CUDA runtime.

    Returns "" when the build did not enable CUDA, so a CPU wheel declares nothing even on
    a machine that has a toolkit installed.

    Raises when a release row names a train the installed toolkit does not provide. The
    declared packages and the loader paths both come from this value, so disagreeing with
    the toolkit that compiled the libraries produces a wheel that installs cleanly and then
    cannot load: a cu126 row built against a 13.0 toolkit declares the CUDA 12 runtime for
    binaries that need libcudart.so.13.
    """
    # An explicit OFF first, ahead of the release field. A CPU row on a builder that has a
    # toolkit installed sets both, so reading the row field first would declare a runtime
    # the wheel never loads.
    if not install_utils.is_cmake_option_on(
        _cmake_args(),
        "EXECUTORCH_BUILD_CUDA",
        default=True,
    ):
        return ""

    raw = os.environ.get("CU_VERSION") or os.environ.get("DESIRED_CUDA") or ""
    # A row spelled "cpu" is a CPU row, unless the caller asked for CUDA explicitly. The
    # shortcut exists so a local build named that way with nothing requested does not fall
    # through and raise on the unsupported-train branch below. It must not swallow an explicit
    # request, because that request still reaches CMake, which builds the CUDA libraries: the
    # wheel would then carry them with no dependency declared and no path to the runtime.
    if raw.lower() in _CPU_ROW_NAMES and not install_utils.is_cmake_option_on(
        _cmake_args(), "EXECUTORCH_BUILD_CUDA", default=False
    ):
        return ""
    if raw.lower() in _CPU_ROW_NAMES:
        # Reached only when the caller turned CUDA on for a CPU row. Refused rather than
        # guessed, because the option already reached CMake and built the libraries, so
        # declaring nothing would ship them with no runtime dependency and no way to find
        # it. Named here so the reader is not sent to add "cpu" to a list of CUDA trains.
        raise RuntimeError(
            f"the build names a CPU row while also asking for CUDA "
            f"(EXECUTORCH_BUILD_CUDA=ON in CMAKE_ARGS). Those contradict: the CUDA "
            f"libraries would be built and shipped with no runtime dependency declared. "
            f"Pick one, either drop the option or name a CUDA row instead of {raw!r}."
        )
    # Reduce to digits and match against the same (major, minor) trains the shell classifier
    # uses. Previously this took the first two digits and matched against major only, so a
    # row spelled with an unsupported minor (say cu125) was classified CPU by the shell and
    # CUDA 12 here, and the wheel then declared CUDA runtime packages for a CPU build.
    digits = re.sub(r"[^0-9]", "", raw)
    # The same rows as `trains` below, keeping both numbers rather than reducing to the
    # major, so the guard can tell two rows of one major apart.
    _CUDA_ROW_MINORS = {
        f"{major}{minor}": (major, minor)
        for major, minor in install_utils.SUPPORTED_CUDA_VERSIONS
    }
    trains = {
        f"{major}{minor}": str(major)
        for major, minor in install_utils.SUPPORTED_CUDA_VERSIONS
    }
    requested = trains.get(digits, "")

    # Read the toolkit version directly, without the (major, minor) validator, so the guard
    # below fires on any mismatch rather than only on the three listed pairs.
    detected_version = install_utils._detected_cuda_version()
    detected_major = detected_version[0] if detected_version is not None else None
    detected = (
        str(detected_major)
        if detected_major is not None and str(detected_major) in _CUDA_RUNTIME_PACKAGES
        else ""
    )

    if requested:
        # A row that names a train has to be buildable for that train. Reported here
        # rather than left to produce a mismatched wheel, because nothing downstream
        # compares the two: the metadata comes from the row and the binaries come from
        # the toolkit.
        if detected and detected != requested:
            raise RuntimeError(
                f"this build targets CUDA {requested} (from "
                f"{'CU_VERSION' if os.environ.get('CU_VERSION') else 'DESIRED_CUDA'}="
                f"{raw!r}) but the installed toolkit is CUDA {detected}. The declared "
                "runtime packages and the loader search paths come from the requested "
                "train while the libraries are compiled by the installed one, so the "
                "wheel would install and then fail to load. Install a matching toolkit "
                "or build the row that matches this one."
            )
        # The check above compares majors, which two rows of the same major share. A row
        # names its train down to the minor, so cu130 and cu132 both reduce to 13 and a
        # cu132 row built against a 13.0 toolkit passed. The device code and the version
        # in the wheel's local label both come from the row, so that wheel claims a
        # toolkit it was not compiled by. Compared here rather than folded above so the
        # message can name both numbers.
        if detected_version is not None and digits in _CUDA_ROW_MINORS:
            requested_pair = _CUDA_ROW_MINORS[digits]
            if detected_version != requested_pair:
                requested_text = f"{requested_pair[0]}.{requested_pair[1]}"
                detected_text = f"{detected_version[0]}.{detected_version[1]}"
                raise RuntimeError(
                    f"this build targets CUDA {requested_text} (from "
                    f"{'CU_VERSION' if os.environ.get('CU_VERSION') else 'DESIRED_CUDA'}="
                    f"{raw!r}) but the installed toolkit is CUDA {detected_text}. Those "
                    "share a major version, so the runtime dependency this wheel declares "
                    "is correct while the device code and the version recorded in its "
                    "local label are not. Install a matching toolkit or build the row "
                    "that matches this one."
                )
        return requested

    if raw and not requested:
        # A row named something this packaging does not recognise. Silently reporting the
        # builder's toolkit instead contradicts "the row's field wins" and produced a
        # wheel tagged for one train carrying another.
        supported = ", ".join(
            f"cu{major}{minor}"
            for major, minor in install_utils.SUPPORTED_CUDA_VERSIONS
        )
        raise RuntimeError(
            f"the release row requests CUDA {raw!r}, which is not a train this project "
            f"supports ({supported}). Add it to SUPPORTED_CUDA_VERSIONS in install_utils "
            "and to _CUDA_RUNTIME_PACKAGES and _CUDA_LIBRARY_DIRECTORIES here, or build "
            "a supported row. Falling back to whatever toolkit this builder has would tag "
            "the wheel for one train and fill it with another."
        )

    # Fall back to the installed toolkit, because keying this off a release variable alone produced a wheel
    # that carried the CUDA libraries while declaring no CUDA runtime and recording no way to reach one.
    #
    # Two ways CUDA gets built, and both have to agree with what is declared here. The build gate turns it
    # on when a SUPPORTED train is installed, so a toolkit whose minor is unlisted builds CPU-only and
    # declaring runtime packages for it would make a CPU wheel demand four CUDA wheels. An explicit ON
    # bypasses that gate and reaches CMake directly, where find_package(CUDAToolkit) accepts a toolkit
    # this packaging does not list, so the libraries ship and the runtime has to be declared for them.
    # Asking only whether the train is supported got the first case right and the second wrong.
    explicit_on = install_utils.is_cmake_option_on(
        _cmake_args(),
        "EXECUTORCH_BUILD_CUDA",
        default=False,
    )
    if not install_utils.is_cuda_available() and not explicit_on:
        return ""
    if explicit_on and not detected:
        # The explicit request reaches CMake either way, so returning "" here shipped the
        # CUDA libraries with no runtime declared and no path to one, which is the wheel
        # this whole function exists to prevent. An unlisted minor still resolves to a
        # major and is fine; an unlisted major has nothing to declare.
        installed = (
            f"CUDA {detected_major}"
            if detected_major is not None
            else "no CUDA toolkit"
        )
        raise RuntimeError(
            f"the build asks for CUDA (EXECUTORCH_BUILD_CUDA=ON in CMAKE_ARGS) but found "
            f"{installed}, and this packaging declares a runtime only for CUDA "
            f"{', '.join(sorted(_CUDA_RUNTIME_PACKAGES))}. The libraries would still be "
            "built and shipped with nothing to load them against. Install a toolkit on one "
            "of those majors, or add this one to _CUDA_RUNTIME_PACKAGES and "
            "_CUDA_LIBRARY_DIRECTORIES."
        )
    return detected


def _windows_cuda_toolkit(cmake_configuration_args: List[str]) -> str:
    """The CUDA toolkit root a Windows CUDA build compiles with, or "" when CUDA is off.

    Decided with the same conditions the CUDA gate below uses, because the answer picks the
    generator for the whole configure: a toolkit here with CUDA turned off would still build the
    CPU wheel with Ninja instead of the Visual Studio ClangCL toolset. The compiler is the one
    install_utils reports, so the toolkit that compiles is the train packaging declares.
    """
    arguments = cmake_configuration_args + [item for item in _cmake_args() if item]
    if (
        _is_minimal_build()
        or _row_is_cpu_only()
        or not install_utils.is_cmake_option_on(
            arguments, "EXECUTORCH_BUILD_CUDA", default=True
        )
    ):
        return ""
    explicit = install_utils.is_cmake_option_on(
        arguments, "EXECUTORCH_BUILD_CUDA", default=False
    )
    if not explicit and not install_utils.is_cuda_available():
        return ""
    nvcc = shutil.which(install_utils._selected_nvcc()[0])
    if not nvcc:
        return ""
    return Path(nvcc).resolve().parent.parent.as_posix()


def _cuda_libraries_built(cmake_cache_dir: Optional[str]) -> bool:
    """Whether this build produced the CUDA libraries, read from the CMake cache.

    The build turns CUDA on from the cache, so the cache is the fact that decides what ships. The
    release row's CUDA version is a different question: a build on a toolkit whose train this packaging
    does not recognise still produces the libraries while declaring no train, and gating anything else on
    the train left that wheel carrying libraries with no matching header.

    Falls back to the train when no cache is readable, which is the case for a source distribution where
    nothing was built here anyway.
    """
    cache_path = os.path.join(cmake_cache_dir or "", "CMakeCache.txt")
    if os.path.exists(cache_path):
        return CMakeCache(cache_path=cache_path).is_enabled("EXECUTORCH_BUILD_CUDA")
    return bool(_cuda_train())


def _verify_cuda_runtime_matches_train(cmake_cache_dir: Optional[str]) -> None:
    """Fail the build when the linked CUDA runtime is not the train being declared.

    The declared packages come from the compiler version, while the library that actually gets
    linked comes from find_package(CUDAToolkit). Those are normally the same toolkit, but
    CUDAToolkit_ROOT steers the second and not the first, so they can split inside a single
    find_package call: measured with the compiler at 13.0 and that variable at 12.8,
    CUDAToolkit_VERSION reported 13.0.88 while the binary needed libcudart.so.12. Packaging
    would then declare the CUDA 13 runtime for a wheel that cannot load without CUDA 12.

    Read from the CMake cache rather than from the environment, because the cache records what
    the build resolved rather than what was requested.
    """
    train = _cuda_train()
    if not train:
        return
    cache_path = os.path.join(cmake_cache_dir or "", "CMakeCache.txt")
    if not os.path.exists(cache_path):
        return
    cache = CMakeCache(cache_path=cache_path)
    if not cache.is_enabled("EXECUTORCH_BUILD_CUDA"):
        return
    linked = cache.get("CUDA_cudart_LIBRARY")
    if linked is None or not linked.value:
        return
    # Read the major from the resolved file name rather than from the recorded path, because the
    # conventional way to name a toolkit is the versionless /usr/local/cuda symlink, which carries
    # no version at all. Matching the directory accepted that spelling silently, which is the one
    # the guard's own message tells the user to set. The resolved name ends in the soname the
    # loader will ask for, which is the thing the declared package has to agree with.
    found = re.search(r"libcudart\.so\.(\d+)", os.path.realpath(linked.value))
    if found is None or found.group(1) == train:
        return
    raise RuntimeError(
        f"this build declares the CUDA {train} runtime but linked the CUDA "
        f"{found.group(1)} one from {linked.value!r}, so the wheel would install and then "
        "fail to load. The declared train follows the CUDA compiler while the linked "
        "libraries follow find_package(CUDAToolkit), so point CUDACXX and CUDAToolkit_ROOT "
        "at the same toolkit."
    )


def _cuda_dependencies() -> List[str]:
    """Runtime libraries a CUDA wheel needs but does not bundle.

    Declared rather than vendored, the way the PyTorch CUDA wheels do it, so one copy is
    shared with torch instead of shipping a second one.
    """
    train = _cuda_train()
    # Marked for Linux. These nvidia wheels publish no distribution for the other platforms, so an
    # unmarked requirement would make an install elsewhere fail on a dependency it cannot satisfy.
    # The Windows CUDA wheel needs none of them: its DLLs load only the driver (nvcuda.dll), and the
    # CUDA runtime a model's own library imports comes from the CUDA Toolkit install.
    return [
        f"{name}; platform_system == 'Linux'"
        for name in _CUDA_RUNTIME_PACKAGES.get(train, ())
    ]


# Directories inside the wheel that hold libraries a shipped library links, relative to the package
# root rather than to the linking library, because the wheel ships libraries at more than one depth.
#
# The CUDA libraries are split across two directories and reference each other in both directions:
# the delegate in lib/ links the shims library in backends/cuda/, and the shims library links the
# stream helper back in lib/. So both hops are needed.
#
# Applied to every shipped library rather than mapping each library to the directories it happens to
# need. An unused hop costs nothing at load time, while a missing one produces a wheel that installs
# and then fails to load, and a per-library mapping would have to be revisited every time a library
# moves.
_SIBLING_LIBRARY_DIRECTORIES = ("backends/cuda", "lib", "src/executorch/lib")


def _sibling_library_search_paths(depth: int = 1) -> List[str]:
    """Loader paths that reach another directory inside this same package.

    `depth` is how many directories separate the linking library from the package root, and it has to
    be honoured for the same reason the CUDA hops honour it: the wheel ships libraries at depth one
    (lib/) and depth two (backends/cuda/, extension/pybindings/ and others). Measured with a fixed
    pair sized for one depth, six of twelve hops landed somewhere that does not exist, and the hop
    from lib/ escaped the package entirely into a sibling of it, where an unrelated library with a
    matching SONAME could satisfy the dependency first.
    """
    up = "/".join([".."] * depth)
    token = _loader_relative_token()
    return [f"{token}/{up}/{directory}" for directory in _SIBLING_LIBRARY_DIRECTORIES]


def _loader_relative_token() -> str:
    """The token a runtime search path uses to mean "the directory this file is in".

    ELF spells it $ORIGIN and Mach-O spells it @loader_path. Both are literal text in the
    recorded path, so the wrong one becomes a directory of that name and resolves to
    nothing.
    """
    return "@loader_path" if sys.platform == "darwin" else "$ORIGIN"


def _cuda_runtime_search_paths(depth: int = 1) -> List[str]:
    """Loader paths that reach the CUDA wheels installed beside this one.

    Those wheels install as siblings of this package, so the hop has to climb out of the package first.
    `depth` is how many directories separate the library from the package root, and the wheel ships
    libraries at more than one depth: a hop sized for one of them lands inside this package from the
    other, where nothing is found.
    """
    train = _cuda_train()
    out = "/".join([".."] * (depth + 1))
    return [
        f"{_loader_relative_token()}/{out}/{directory}"
        for directory in _CUDA_LIBRARY_DIRECTORIES.get(train, ())
    ]


def _is_cuda_toolkit_directory(entry: str) -> bool:
    """Whether a runtime search path entry names a library directory inside a CUDA toolkit.

    Matched on the two layouts a toolkit actually installs rather than on the word "cuda" appearing
    somewhere above the directory. Scanning a window of components dropped a torch directory whose build
    root happened to be named after a CUDA version, and torch's directory is the one absolute path a
    shipped library has to keep.

    Position alone cannot separate the two, because a real targets layout puts the cuda-named component at
    the same depth a build root does, so each layout is spelled out instead.
    """
    parts = [part.lower() for part in PurePosixPath(entry).parts]
    if not parts or parts[-1] not in ("lib", "lib64"):
        return False

    def cuda_named(part: str) -> bool:
        return bool(re.fullmatch(r"cuda(?:-\d+(?:\.\d+)*|[-_]?toolkit)?", part))

    # <toolkit>/lib64
    if len(parts) >= 2 and cuda_named(parts[-2]):
        return True
    # <toolkit>/targets/<arch>/lib
    return len(parts) >= 4 and parts[-3] == "targets" and cuda_named(parts[-4])


def _is_unresolved_math_library_directory(entry: str) -> bool:
    """Whether a runtime search path entry is a maths library directory whose prefix resolved empty.

    Torch's exported CMake package creates a caffe2::mkl imported target, and linking torch brings it
    in even though this project never asks for MKL. Its link directories are a hardcoded list of four,
    spelled below MKL_ROOT, which resolves to nothing here, so what the linker records is left
    anchored at the filesystem root: /lib, /lib/intel64, /lib/intel64_win, /lib/win-x64. The bare
    /lib does not survive, because CMake filters its own implicit link directories out of the link
    line.

    An empty prefix is the whole signature, so the entry must be exactly /lib/<arch>. A real
    installation spells the same arch directory below a prefix, as /opt/intel/mkl/lib/intel64, and
    that one is a directory the environment genuinely provides.

    Only the three arch directories are handled. The bare prefix/lib is deliberately left, because
    with a resolving prefix it is an ordinary library directory this project has no business
    dropping, and with an empty one CMake filters it out as an implicit link directory before the
    link line is built, so it never reaches a shipped library.

    Dropping these loses nothing: no shipped library names an MKL or OpenMP runtime in DT_NEEDED, so
    nothing resolves through them, and they sit ahead of the relative hops appended below.
    """
    parts = PurePosixPath(entry).parts
    return (
        len(parts) == 3
        and parts[:2] == ("/", "lib")
        and parts[2] in _MKL_ARCH_DIRECTORIES
    )


def _package_relative_depth(library: Path) -> int:
    """How many directories separate a shipped library from the installed package root.

    Searched from the END of the path. At build time the path is absolute and a source checkout is
    often named after the package too, so taking the first match found the checkout instead of the
    package inside the build output and produced a hop that climbs out of the install directory.
    """
    parts = list(Path(library).parts)
    if "executorch" not in parts:
        return 1
    index = len(parts) - 1 - parts[::-1].index("executorch")
    return max(len(parts) - index - 2, 0)


def _torchao_requirement() -> str:
    """The torchao dependency, pinned to the series install_requirements.py installs.

    Derived from that module rather than written out, so a nightly bump cannot move the
    pin without moving this bound with it. A bump into the next series would otherwise
    silently stop satisfying the lower bound, and installing this package over a
    development checkout would replace the torchao that was just installed.

    Loaded by path, the way install_utils is above, because setuptools executes this
    file without the project directory on sys.path, so a plain import does not resolve.
    """
    path = Path(__file__).parent / "install_requirements.py"
    spec = importlib.util.spec_from_file_location("install_requirements", path)
    if spec is None or spec.loader is None:
        raise ImportError(f"Could not load {path}")
    module = importlib.util.module_from_spec(spec)
    # The module imports install_utils by name at import time, which only resolves
    # because that name is registered below.
    sys.modules.setdefault("install_utils", install_utils)
    spec.loader.exec_module(module)

    version = module.TORCHAO_NIGHTLY_VERSION
    if (
        install_utils.determine_torch_url(module.TORCH_URL_BASE).endswith("/cu134")
        and not module.torchao_from_source()
    ):
        version = module.CU134_TORCHAO_NIGHTLY_VERSION
    major, minor = (int(part) for part in version.split(".")[:2])
    return f"torchao>={version},<{major}.{minor + 1}"


def _base_dependencies() -> List[str]:
    """Runtime dependencies for the full wheel.

    Declared here rather than in pyproject.toml (where `dependencies` is marked
    dynamic) so the minimal build can ship a slimmer set. Keep in sync with the
    project's runtime needs.
    """
    return [
        "expecttest",
        "flatbuffers",
        "hypothesis",
        "kgb",
        "mpmath==1.3.0",
        "numpy>=2.0.0; python_version >= '3.10'",
        "packaging",
        "pandas>=2.2.2; python_version >= '3.10'",
        "parameterized",
        # The Qualcomm backend needs both, on different paths. Any lowering reads the CPU vendor
        # to disable an mkldnn path that crashes on AMD, whether or not an SDK is already set up.
        # Fetching an SDK additionally needs requests, which its downloader imports at module
        # scope. Neither is needed merely to import the backend.
        "py-cpuinfo",
        "requests",
        "pytorch-tokenizers",
        # Shipped code imports torchao at module scope in many places, so a plain install cannot
        # lower a model without it. Among others: the XNNPACK utilities the partitioner uses
        # (backends/xnnpack/utils/utils.py), the Core ML quantizer, and executorch.export itself.
        # The MLX backend needs it too, though indirectly: it registers a torchao operator that
        # only exists once torchao has been imported.
        #
        # The lower bound is the nightly install_requirements.py pins, so that installing this
        # package over a development checkout leaves that pin in place. A bound at the stable
        # release instead would evict it, because a dev release sorts below its own final.
        #
        # The upper bound is what makes naming a pre-release safe. A specifier that names one
        # accepts pre-releases for this requirement, so without the bound pip would resolve a
        # pre-release of the next series wherever one is published, such as an index carrying
        # nightlies. Keep both bounds in the same series.
        _torchao_requirement(),
        "pyyaml",
        "ruamel.yaml",
        "sympy",
        "tabulate",
        # See also third-party/TARGETS for buck's typing-extensions version.
        "typing-extensions>=4.10.0",
        # Keep this version in sync with: ./backends/apple/coreml/scripts/install_requirements.sh
        # Linux is deliberate: the Core ML export flow runs there, so a model can be lowered
        # for Apple hardware from a Linux machine. Linux is narrowed to x86_64 because that is
        # the only Linux architecture coremltools publishes a build for. Everywhere else pip
        # falls back to the source archive and produces a py3-none-any install with none of the
        # compiled extensions, which imports and then cannot write a model, because
        # libmilstoragepython, which stores the weights of an mlprogram, is absent. Measured
        # with coremltools 9.0 on a Linux aarch64 machine. Keep this in sync with the condition
        # in conftest.py; .ci/scripts/tests/test_coreml_markers.py checks that the two agree.
        "coremltools==9.0; (platform_system == 'Darwin' or (platform_system == 'Linux' and platform_machine == 'x86_64')) and python_version < '3.14'",
        # coremltools needs scikit-learn to palettize weights, so it follows coremltools.
        # Without a marker it also installed where coremltools does not, most visibly on
        # Windows, and brought scipy with it.
        #
        # Nothing here imports scikit-learn directly. coremltools does, and not through the
        # _HAS_SKLEARN flag it sets: quantization_utils.py imports sklearn.cluster.KMeans at the
        # point of use, and that is the default clustering path, taken unless the weight is a
        # single column of at least ten thousand float16 values. So the CODEBOOK_WEIGHT_ONLY
        # recipe raises ModuleNotFoundError without this, whatever _HAS_SKLEARN says.
        #
        # The floor is above the version coremltools accepts for its own scikit-learn model
        # converter, which is why that converter reports itself disabled. That is a separate
        # feature nothing here uses, and palettization does not check the version, so the two
        # are unrelated. Measured with coremltools 9.0 and scikit-learn 1.9.0: the converter is
        # off and palettization works.
        "scikit-learn>=1.7.1; (platform_system == 'Darwin' or (platform_system == 'Linux' and platform_machine == 'x86_64')) and python_version < '3.14'",
        "hydra-core>=1.3.0",
        "omegaconf>=2.3.0",
    ]


def _minimal_dependencies() -> List[str]:
    """Runtime dependencies for the minimal (AOT export only) wheel.

    Derived as the subset of _base_dependencies() that executorch.exir needs to
    lower and serialize a .pte, so version pins and markers stay in sync with the
    full set. torch is intentionally absent from both (consumers bring their own).
    mpmath is intentionally dropped too: it is pulled transitively by sympy, whose
    "mpmath<1.4" cap resolves to the same 1.3.0 the full wheel pins. Keep the name
    set below in sync with the `expected` set in .ci/scripts/test_minimal_wheel.sh.
    """
    keep = {
        "flatbuffers",
        "numpy",
        "packaging",
        "pyyaml",
        "ruamel-yaml",
        "sympy",
        "tabulate",
        "typing-extensions",
    }

    def _name(dep: str) -> str:
        # PEP 503 normalized distribution name, e.g. "ruamel.yaml" -> "ruamel-yaml".
        return re.sub(
            r"[-_.]+", "-", re.split(r"[ ;\[<>=!~(]", dep, maxsplit=1)[0]
        ).lower()

    minimal = [dep for dep in _base_dependencies() if _name(dep) in keep]
    # Fail the build loudly if a name in `keep` no longer matches a full-wheel dep
    # (e.g. renamed or removed in _base_dependencies()), instead of silently
    # shipping a minimal wheel that is missing a required dependency.
    unmatched = keep - {_name(dep) for dep in minimal}
    assert not unmatched, f"minimal keep-set names not found in base deps: {unmatched}"
    return minimal


class Version:
    """Static strings that describe the version of the pip package."""

    # Cached values returned by the properties.
    __root_dir_attr: Optional[str] = None
    __string_attr: Optional[str] = None
    __git_hash_attr: Optional[str] = None

    @classmethod
    def _root_dir(cls) -> str:
        """The path to the root of the git repo."""
        if cls.__root_dir_attr is None:
            # This setup.py file lives in the root of the repo.
            cls.__root_dir_attr = str(Path(__file__).parent.resolve())
        return str(cls.__root_dir_attr)

    @classmethod
    def git_hash(cls) -> Optional[str]:
        """The current git hash, if known."""
        if cls.__git_hash_attr is None:
            import subprocess

            try:
                cls.__git_hash_attr = (
                    subprocess.check_output(
                        ["git", "rev-parse", "HEAD"], cwd=cls._root_dir()
                    )
                    .decode("ascii")
                    .strip()
                )
            except subprocess.CalledProcessError:
                cls.__git_hash_attr = ""  # Non-None but empty.
        # A non-None but empty value indicates that we don't know it.
        return cls.__git_hash_attr if cls.__git_hash_attr else None

    @classmethod
    def string(cls) -> str:
        """The version string."""
        if cls.__string_attr is None:
            # If set, BUILD_VERSION should override any local version
            # information. CI will use this to manage, e.g., release vs. nightly
            # versions.
            version = os.getenv("BUILD_VERSION", "").strip()
            if not version:
                # Otherwise, read the version from a local file and add the git
                # commit if available.
                version = (
                    open(os.path.join(cls._root_dir(), "version.txt")).read().strip()
                )
                if cls.git_hash():
                    version += "+" + cls.git_hash()[:7]  # type: ignore[index]
            cls.__string_attr = version
        return cls.__string_attr

    @classmethod
    def write_to_python_file(cls, path: str) -> None:
        """Creates a file similar to PyTorch core's `torch/version.py`."""

        lines = [
            "from typing import Optional",
            '__all__ = ["__version__", "git_version"]',
            f'__version__ = "{cls.string()}"',
            # A string or None.
            f"git_version: Optional[str] = {repr(cls.git_hash())}",
        ]
        with open(path, "w") as fp:
            fp.write("\n".join(lines) + "\n")


# The build type is determined by the DEBUG environment variable. If DEBUG is
# set to a non-empty value, the build type is Debug. Otherwise, the build type
# is Release.
def get_build_type(is_debug=None) -> str:
    debug = int(os.environ.get("DEBUG", 0) or 0) if is_debug is None else is_debug
    return "Debug" if debug else "Release"


def get_dynamic_lib_name(name: str) -> str:
    if _is_windows():
        return f"{name}.dll"
    elif _is_macos():
        return f"lib{name}.dylib"
    else:
        return f"lib{name}.so"


def _dynamic_lib_suffix() -> str:
    """The loadable-library suffix on this platform, including the dot.

    Separate from get_dynamic_lib_name because a file whose prefix is not known
    ahead of time still needs the suffix named: globbing the suffix as well would
    also match an import library, an exports file, or a soname's versioned links.
    """
    if _is_windows():
        return ".dll"
    if _is_macos():
        return ".dylib"
    return ".so"


def get_executable_name(name: str) -> str:
    if _is_windows():
        return name + ".exe"
    else:
        return name


# Visual Studio is a multi-config generator and writes into a per-config subdirectory.
_CFG = "%BUILD_TYPE%/" if _is_windows() else ""


def _windows_import_libraries() -> List["BuiltFile"]:
    """The import library beside each DLL offered to C++ consumers, which they link against."""
    if not _is_windows():
        return []
    entries = [
        ("", "executorch_shared", "executorch", []),
        (
            "extension/threadpool/",
            "executorch_threadpool",
            None,
            ["EXECUTORCH_BUILD_PTHREADPOOL", "EXECUTORCH_BUILD_CPUINFO"],
        ),
        (
            "configurations/",
            "executorch_kernels_optimized",
            None,
            ["EXECUTORCH_BUILD_KERNELS_OPTIMIZED"],
        ),
        (
            "kernels/quantized/",
            "executorch_kernels_quantized",
            None,
            ["EXECUTORCH_BUILD_KERNELS_QUANTIZED"],
        ),
        (
            "backends/xnnpack/",
            "executorch_backend_xnnpack",
            None,
            ["EXECUTORCH_BUILD_XNNPACK"],
        ),
        (
            "backends/cuda/",
            "executorch_backend_cuda",
            None,
            ["EXECUTORCH_BUILD_CUDA"],
        ),
        (
            "extension/cuda/",
            "executorch_extension_cuda",
            None,
            ["EXECUTORCH_BUILD_CUDA"],
        ),
        (
            "backends/openvino/",
            "executorch_backend_openvino",
            None,
            ["EXECUTORCH_BUILD_OPENVINO"],
        ),
    ]
    return [
        BuiltFile(
            src_dir=f"%CMAKE_CACHE_DIR%/{subdir}{_CFG}",
            src_name=f"{built}.lib",
            dst=f"executorch/lib/{shipped or built}.lib",
            dependent_cmake_flags=["EXECUTORCH_BUILD_SHARED", *flags],
        )
        for subdir, built, shipped, flags in entries
    ]


class _BaseExtension(Extension):
    """A base class that maps an abstract source to an abstract destination."""

    # Prefix of the synthetic name a BuiltFile carries. A built file is not a Python
    # module, so its name is deliberately not a module path, and setuptools has to be
    # able to tell it apart from a real extension. See _ExecuTorchDistribution, which
    # keeps these names out of the metadata setuptools derives from ext_modules.
    SYNTHETIC_NAME_PREFIX = "@EXECUTORCH_BuiltFile_"

    def __init__(
        self,
        src: str,
        dst: str,
        name: str,
        dependent_cmake_flags: List[str],
    ):
        # Source path; semantics defined by the subclass.
        self.src: str = src

        # Destination path relative to a namespace defined elsewhere. If this ends
        # in "/", it is treated as a directory. If this is "", it is treated as the
        # root of the namespace.
        # Destination path; semantics defined by the subclass.
        self.dst: str = dst

        # Other parts of setuptools expects .name to exist. For actual extensions
        # this can be the module path, but otherwise it should be somehing unique
        # that doesn't look like a module path.
        self.name: str = name

        self.dependent_cmake_flags = dependent_cmake_flags
        self.cmake_cache: Optional[CMakeCache] = None

        super().__init__(name=self.name, sources=[])

    def _get_build_dir(self, installer: "InstallerBuildExt") -> Path:
        # Share the cmake-out location with CustomBuild.
        build_cmd = installer.get_finalized_command("build")
        if "%CMAKE_CACHE_DIR%" in self.src:
            if not hasattr(build_cmd, "cmake_cache_dir"):
                raise RuntimeError(
                    f"Extension {self.name} has a src {self.src} that contains"
                    " %CMAKE_CACHE_DIR% but CMake does not run in the `build` "
                    "command. Please double check if the command is correct."
                )
            else:
                return Path(build_cmd.cmake_cache_dir)
        else:
            # If the src path doesn't contain %CMAKE_CACHE_DIR% placeholder,
            # try to find it under the current directory.
            return Path(".")

    def is_cmake_artifact_used(self, installer: "InstallerBuildExt") -> bool:
        # The flags name what the build turned on, so read them from the build's cache and
        # not from wherever this entry's source lives. A source tree file resolves to the
        # working directory, which holds no cache, so its flags were silently ignored.
        cmake_cache_dir = getattr(
            installer.get_finalized_command("build"), "cmake_cache_dir", None
        )
        if not cmake_cache_dir:
            # CMake did not run, so there is nothing to test against. An entry that needs
            # the build directory still fails in src_path right after this.
            return True

        cache_path = os.path.join(cmake_cache_dir, "CMakeCache.txt")
        if not os.path.exists(cache_path):
            # If this is not a CMake folder, then assume it's used.
            return True
        elif self.cmake_cache is None:
            self.cmake_cache = CMakeCache(cache_path=cache_path)

        return all(
            self.cmake_cache.is_enabled(flag) for flag in self.dependent_cmake_flags
        )

    def src_path(self, installer: "InstallerBuildExt") -> Path:
        """Returns the path to the source file, resolving globs.

        Args:
            installer: The InstallerBuildExt instance that is installing the
                file.
        """
        build_dir = self._get_build_dir(installer)

        src_path = self.src.replace("%CMAKE_CACHE_DIR%/", "")

        cfg = get_build_type(installer.debug)

        if os.name == "nt":
            # Replace %BUILD_TYPE% with the current build type.
            src_path = src_path.replace("%BUILD_TYPE%", cfg)
        else:
            # Remove %BUILD_TYPE% from the path.
            src_path = src_path.replace("/%BUILD_TYPE%", "")

        # Construct the full source path, resolving globs. If there are no glob
        # pattern characters, this will just ensure that the source file exists.
        srcs = tuple(build_dir.glob(src_path))
        if len(srcs) != 1:
            raise ValueError(
                f"Expecting exactly 1 file matching {self.src} in {build_dir}, "
                f"found {repr(srcs)}. Resolved src pattern: {src_path}."
            )
        return srcs[0]

    def inplace_dir(self, installer: "InstallerBuildExt") -> Path:
        """Returns the path of this file to be installed to, under inplace mode.

        It will be a relative path to the project root directory. For more info
        related to inplace/editable mode, please checkout this doc:
        https://setuptools.pypa.io/en/latest/userguide/development_mode.html
        """
        raise NotImplementedError()


class BuiltFile(_BaseExtension):
    """An extension that installs a single file that was built by cmake.

    This isn't technically a `build_ext` style python extension, but there's no
    dedicated command for installing arbitrary data. It's convenient to use
    this, though, because it lets us manage the files to install as entries in
    `ext_modules`.
    """

    def __init__(
        self,
        src_dir: str,
        src_name: str,
        dst: str,
        dependent_cmake_flags: List[str],
        is_executable: bool = False,
        is_dynamic_lib: bool = False,
    ):
        """Initializes a BuiltFile.

        Args:
            src_dir: The directory of the file to install, relative to the cmake-out
                directory. A placeholder %BUILD_TYPE% will be replaced with the build
                type for multi-config generators (like Visual Studio) where the build
                output is in a subdirectory named after the build type. For single-
                config generators (like Makefile Generators or Ninja), this placeholder
                will be removed.
            src_name: The name of the file to install
            dst: The path to install to, relative to the root of the pip
                package. If dst ends in "/", it is treated as a directory.
                Otherwise it is treated as a filename.
            is_executable: If True, the file is an executable. This is used to
                determine the destination filename for executable.
            is_dynamic_lib: If True, the file is a dynamic library. This is used
                to determine the destination filename for dynamic library.
        """
        if is_executable and is_dynamic_lib:
            raise ValueError("is_executable and is_dynamic_lib cannot be both True.")
        if is_executable:
            src_name = get_executable_name(src_name)
        elif is_dynamic_lib:
            src_name = get_dynamic_lib_name(src_name)
        src = os.path.join(src_dir, src_name)
        # This is not a real extension, so use a unique name that doesn't look
        # like a module path. Some of setuptools's autodiscovery will look for
        # extension names with prefixes that match certain module paths.
        super().__init__(
            src=src,
            dst=dst,
            name=f"{_BaseExtension.SYNTHETIC_NAME_PREFIX}{src}:{dst}",
            dependent_cmake_flags=dependent_cmake_flags,
        )

    def dst_path(self, installer: "InstallerBuildExt") -> Path:
        """Returns the path to the destination file.

        Args:
            installer: The InstallerBuildExt instance that is installing the
                file.
        """
        dst_root = Path(installer.build_lib).resolve()

        if self.dst.endswith("/"):
            # Destination looks like a directory. Use the basename of the source
            # file for its final component.
            return dst_root / Path(self.dst) / self.src_path(installer).name
        else:
            # Destination looks like a file.
            return dst_root / Path(self.dst)

    def inplace_dir(self, installer: "InstallerBuildExt") -> Path:
        """For a `BuiltFile`, we use self.dst as its inplace directory path.
        Need to handle directory vs file.
        """
        # The destination is relative to the installed package, so resolve it against the same
        # package directory an extension uses. Anchoring at the repo root instead only worked for
        # destinations that already had a directory under src/executorch, and silently scattered
        # the rest, which left the CMake package searching a directory nothing was copied into.
        relative = self.dst.removeprefix("executorch/")
        if not relative.endswith("/"):
            relative = os.path.dirname(relative)
        build_py = installer.get_finalized_command("build_py")
        package_dir = os.path.abspath(build_py.get_package_dir("executorch"))
        return Path(package_dir) / relative


class BuiltExtension(_BaseExtension):
    """An extension that installs a python extension that was built by cmake."""

    def __init__(
        self,
        src: str,
        modpath: str,
        dependent_cmake_flags: List[str],
        src_dir: Optional[str] = None,
    ):
        """Initializes a BuiltExtension.

        Args:
            src_dir: The directory of the file to install, relative to the cmake-out
                directory. A placeholder %BUILD_TYPE% will be replaced with the build
                type for multi-config generators (like Visual Studio) where the build
                output is in a subdirectory named after the build type. For single-
                config generators (like Makefile Generators or Ninja), this placeholder
                will be removed.
            src_name: The name of the file to install. If the path ends in `.so`,
            modpath: The dotted path of the python module that maps to the
                extension.
        """
        assert (
            "/" not in modpath
        ), f"modpath must be a dotted python module path: saw '{modpath}'"
        full_src = src
        if src_dir is None and _is_windows():
            src_dir = "%BUILD_TYPE%/"
        if src_dir is not None:
            full_src = os.path.join(src_dir, src)
        self.dependent_cmake_flags = dependent_cmake_flags
        # This is a real extension, so use the modpath as the name.
        super().__init__(
            src=f"%CMAKE_CACHE_DIR%/{full_src}",
            dst=modpath,
            name=modpath,
            dependent_cmake_flags=self.dependent_cmake_flags,
        )

    def src_path(self, installer: "InstallerBuildExt") -> Path:
        """Returns the path to the source file, resolving globs.

        Args:
            installer: The InstallerBuildExt instance that is installing the
                file.
        """
        try:
            return super().src_path(installer)
        except ValueError:
            # Probably couldn't find the file. If the path ends with .so, try
            # looking for a .dylib file instead, in case we're running on macos.
            if self.src.endswith(".so"):
                dylib_src = re.sub(r"\.so$", ".dylib", self.src)
                return BuiltExtension(
                    src=dylib_src,
                    modpath=self.dst,
                    dependent_cmake_flags=self.dependent_cmake_flags,
                ).src_path(installer)
            else:
                raise

    def dst_path(self, installer: "InstallerBuildExt") -> Path:
        """Returns the path to the destination file.

        Args:
            installer: The InstallerBuildExt instance that is installing the
                file.
        """
        # Our destination is a dotted module path. get_ext_fullpath() returns
        # the relative path to the .so/.dylib/etc. file that maps to the module
        # path: that's the file we're creating.
        return Path(installer.get_ext_fullpath(self.dst))

    def inplace_dir(self, installer: "InstallerBuildExt") -> Path:
        """For BuiltExtension, deduce inplace dir path from extension name."""
        build_py = installer.get_finalized_command("build_py")
        modpath = self.name.split(".")
        package = ".".join(modpath[:-1])
        package_dir = os.path.abspath(build_py.get_package_dir(package))

        return Path(package_dir)


class InstallerBuildExt(build_ext):
    """Installs files that were built by cmake."""

    def __init__(self, *args, **kwargs):
        self._ran_build = False
        super().__init__(*args, **kwargs)

    def run(self):
        # Run the build command first in editable mode. Since `build` command
        # will also trigger `build_ext` command, only run this once.
        if self._ran_build:
            return

        if self.editable_mode:
            self._ran_build = True
            self.run_command("build")
        super().run()
        if self.editable_mode:
            # The first substitution runs before any cache exists, so it reads the command line
            # and falls back to off. The preset this build uses turns the tracer on without
            # passing anything, so that fallback published the wrong struct layout for an
            # ordinary editable install. The build has produced a cache by now, so replace the
            # best effort value with what the build actually configured.
            self._resubstitute_tracer_definition()

    def _resubstitute_tracer_definition(self) -> None:
        cache_dir = _tracer_cache_dir(self)
        if cache_dir is None:
            return
        for source, destination in _TRACER_DEFINITION_PATHS:
            if os.path.exists(destination):
                # The first pass consumed the placeholder, so restore the template before
                # substituting again. Without this there is nothing left to replace and the
                # correction silently does nothing.
                shutil.copyfile(source, destination)
                _substitute_tracer_definition(destination, cache_dir)

    def copy_extensions_to_source(self) -> None:
        """For each extension in `ext_modules`, we need to copy the extension
        file from the build directory to the correct location in the local
        directory.

        This should only be triggered when inplace mode (editable mode) is enabled.

        Args:

        Returns:
        """
        for ext in self.extensions:
            if not ext.is_cmake_artifact_used(self):
                continue

            package_dir = ext.inplace_dir(self)

            # Ensure that the destination directory exists.
            self.mkpath(os.fspath(package_dir))

            regular_file = ext.src_path(self)
            inplace_file = os.path.join(
                package_dir, os.path.basename(ext.dst_path(self))
            )

            # Always copy, even if source is older than destination, to ensure
            # that the right extensions for the current Python/platform are
            # used.
            if os.path.exists(regular_file) or not ext.optional:
                self.copy_file(regular_file, inplace_file, level=self.verbose)
                # A copied extension still names its libraries by soname, and the entries that
                # reach them are relative to where it was built. The wheel path repairs that
                # from build_extension, and the editable copy needs the same repair or the
                # import fails on a library the loader cannot find.
                # Not restricted by class: both classes copy a file that records paths, and
                # the shipped libraries are the sibling class, so testing for one of them
                # left every library carrying a build directory and an empty entry that the
                # loader reads as the working directory.
                build_command = self.get_finalized_command("build")
                cache_dir = getattr(build_command, "cmake_cache_dir", None)
                _strip_absolute_runtime_paths(
                    Path(inplace_file), _cuda_libraries_built(cache_dir)
                )

            if ext._needs_stub:
                inplace_stub = self._get_equivalent_stub(ext, inplace_file)
                self._write_stub_file(inplace_stub, ext, compile=True)
                # Always compile stub and remove the original (leave the cache behind)
                # (this behaviour was observed in previous iterations of the code)

    # TODO(dbort): Depend on the "build" command to ensure it runs first

    def build_extension(self, ext: _BaseExtension) -> None:
        if not ext.is_cmake_artifact_used(self):
            return

        src_file: Path = ext.src_path(self)
        dst_file: Path = ext.dst_path(self)

        # Ensure that the destination directory exists.
        if not dst_file.parent.exists():
            self.mkpath(os.fspath(dst_file.parent))

        # Copy the file.
        self.copy_file(os.fspath(src_file), os.fspath(dst_file))

        # Ensure that the destination file is writable, even if the source was
        # not. build_py does this by passing preserve_mode=False to copy_file,
        # but that would clobber the X bit on any executables. TODO(dbort): This
        # probably won't work on Windows.
        if not os.access(src_file, os.W_OK):
            # The owner only. A mode of 0o222 would also grant write to the group
            # and to everyone, and chmod takes an absolute mode so no umask
            # narrows it, which turned a 0o555 build output into 0o777.
            os.chmod(src_file, os.stat(src_file).st_mode | stat.S_IWUSR)

        # The destination too, and before the rewrite below, which opens the file
        # for writing. copy_file preserves mode here on purpose, because this path
        # also copies flatc and preserve_mode=False would drop its executable bit,
        # so a read-only source arrives read-only.
        #
        # This mode is the one the wheel archives, so widening it here ships a
        # world-writable library.
        if not os.access(dst_file, os.W_OK):
            os.chmod(dst_file, os.stat(dst_file).st_mode | stat.S_IWUSR)

        cmake_cache_dir = getattr(
            self.get_finalized_command("build"), "cmake_cache_dir", None
        )
        _strip_absolute_runtime_paths(dst_file, _cuda_libraries_built(cmake_cache_dir))
        # After the rewrite rather than before it, because each tool re-signs the file ad hoc when
        # it finishes, so whichever runs last owns the signature. Running strip last keeps the
        # signature over the bytes that ship. Functionally either order works: measured, the two
        # produce the same size, the same exported symbols, the same rpaths, and both verify.
        #
        # Only the wheel copy is stripped; the editable copy above is a developer's own build
        # output, where the local symbols are what a debugger and a profiler read.
        _strip_local_symbols(dst_file)


def _append_relative_search_paths(entries: List[str], depth: int = 1) -> None:
    """Add the relative hops a shipped library needs, skipping any already present.

    Two kinds, both relative so the wheel works wherever the environment lives:
    the CUDA runtime, which arrives in its own wheel installed beside this one, and a sibling
    ExecuTorch library that the wheel installs in a different directory from the library linking it.

    `depth` sizes the hop out of this package, since the wheel ships libraries at more than one depth.
    """
    for search_path in (
        *_cuda_runtime_search_paths(depth),
        *_sibling_library_search_paths(depth),
    ):
        if search_path not in entries:
            entries.append(search_path)


def _write_runtime_paths(
    library: Path,
    tool: str,
    original: str,
    found: List[str],
    entries: List[str],
    is_mach_o: bool,
) -> None:
    """Record the filtered runtime search path back onto the library.

    ELF holds one value that is replaced outright. Mach-O holds one load command per entry
    with nothing to overwrite, so the difference is applied as deletes and adds.

    Told which format this is rather than reading the suffix, because a macOS Python extension
    is Mach-O while named .so, and deciding here produced patchelf syntax passed to otool.
    """
    if not is_mach_o:
        rewritten = ":".join(entries)
        if rewritten == original:
            return
        subprocess.run(
            [tool, "--set-rpath", rewritten, os.fspath(library)],
            check=True,
        )
        return
    install_name_tool = shutil.which("install_name_tool")
    if install_name_tool is None:
        # The caller checks for this tool before deciding to clean a Mach-O library, so reaching
        # here means the environment changed underneath. Leaving the paths in place would ship a
        # library naming the build machine, so say so rather than continue quietly.
        raise RuntimeError(
            "install_name_tool disappeared while cleaning " + os.fspath(library)
        )
    # Checked rather than best effort. A silent failure here leaves the entry the caller asked to
    # remove, or omits the one it asked to add, and packaging then reports success on a library whose
    # search paths are wrong. The result is a wheel that either carries the build machine's directories
    # or cannot find torch, and neither is visible until something loads it.
    for entry in found:
        if entry not in entries:
            _run_install_name_tool(
                [install_name_tool, "-delete_rpath", entry, os.fspath(library)], library
            )
    for entry in entries:
        if entry not in found:
            _run_install_name_tool(
                [install_name_tool, "-add_rpath", entry, os.fspath(library)], library
            )


def _run_install_name_tool(command: List[str], library: Path) -> None:
    """Run an install_name_tool rewrite, raising with its output if it fails.

    Separate from the call sites so both the delete and the add report the same way, and so a future
    third rewrite cannot reintroduce the silent form.
    """
    result = subprocess.run(command, capture_output=True, check=False)
    if result.returncode != 0:
        raise RuntimeError(
            f"{' '.join(command[1:3])} failed on {library.name}, so its runtime search paths are not "
            f"what this build intended: {(result.stderr or result.stdout).decode(errors='replace').strip()}"
        )


def _strip_local_symbols(library: Path) -> None:
    """Discard the local symbol table from a Mach-O file the wheel ships.

    Only on macOS, and only because its linker keeps these where the GNU one does not. The Linux
    shared libraries ship with no .symtab section, while the macOS ones carried a local symbol for
    every internal function: 448 in the runtime, 4190 in the merged kernels, and 35496 in flatc,
    which alone was 6 MB of the 7 MB this removes.

    `-x` removes local symbols and keeps every external one, so what a consumer can link against
    does not change. Anything beyond that would strip exported symbols and break linking, so the
    flag matters.

    A failure here is not fatal. The file is correct either way, and a wheel that is larger than
    intended is better than a build that stops. That also covers the files this runs on that are
    not Mach-O at all, such as the Metal library and a template: strip declines them, and on BSD
    it declines them with a zero exit, leaving the file byte for byte as it was.
    """
    if not _is_macos():
        return
    strip = shutil.which("strip")
    if strip is None:
        return
    result = subprocess.run(
        [strip, "-x", os.fspath(library)], capture_output=True, check=False
    )
    if result.returncode != 0:
        print(
            f"warning: could not strip local symbols from {library.name}, shipping it as "
            f"built: {(result.stderr or result.stdout).decode(errors='replace').strip()}",
            flush=True,
        )


def _parse_runtime_paths(original: str, is_mach_o: bool) -> List[str]:
    """The runtime search path entries a tool reported, as a list.

    The two formats differ in shape, not just spelling: ELF keeps one colon separated
    string, while Mach-O keeps a separate load command per entry, so one cannot be
    split the way the other is.
    """
    if not is_mach_o:
        return original.split(":")
    found = []
    lines = original.splitlines()
    for index, line in enumerate(lines):
        if "LC_RPATH" not in line:
            continue
        for following in lines[index + 1 : index + 4]:
            stripped = following.strip()
            if stripped.startswith("path "):
                found.append(stripped.split(" (offset", 1)[0][len("path ") :])
                break
    # One entry per distinct path, order preserved. Mach-O can carry the same LC_RPATH in two load
    # commands, and the caller issues one -delete_rpath per element: the second finds nothing left to
    # delete and fails, which now aborts packaging rather than being silently absorbed.
    return list(dict.fromkeys(found))


def _is_usable_runtime_path(
    entry: str,
    safe_to_drop_toolkit_paths: bool,
    has_relative_torch_route: bool,
) -> bool:
    if not entry:
        # The loader reads an empty entry as the process working directory.
        return False
    if not entry.startswith("/"):
        return True
    # Absolute, so decide by what it points at.
    #
    # A directory inside this build cannot exist for a user.
    #
    # A CUDA toolkit directory is dropped for a different reason: the wheel declares the CUDA runtime
    # as a dependency and reaches it through a relative hop, so an absolute toolkit path is both
    # unnecessary and harmful. It sits ahead of the hop, so a user who happens to have a toolkit at
    # that prefix resolves the runtime from there instead of from the declared dependency.
    #
    # Whether that is safe is decided above, because the one case it is not is a CUDA build on an
    # unrecognised train, which has no hop to fall back on.
    #
    # A torch lib directory is dropped when a relative route to torch is already recorded, since
    # that route is what resolves torch on an installed wheel while the absolute one only names a
    # directory from the machine that built it. It is kept when no relative route exists, because
    # then it is the only way this library finds torch.
    #
    # A maths library directory whose prefix resolved empty is dropped for the same reason: nothing in
    # the wheel resolves through it, and it sits ahead of the relative hops appended below.
    #
    # Anything else absolute stays, because it is a dependency the environment provides and the wheel
    # has no relative answer for.
    #
    # The build directories are matched as whole path components rather than as substrings. A bare
    # "/cmake-out" also matches "/home/user/cmake-outputs/torchlibs", which is an unrelated directory
    # a user could really have, and stripping it breaks a dependency the library resolves there.
    # The setuptools staging directory is spelled build/lib.<platform>-<pyver>, for example
    # lib.linux-x86_64-cpython-312, so match the whole shape rather than any part starting "lib.".
    if any(
        part in ("pip-out", "cmake-out")
        or re.fullmatch(r"lib\.[^/]+-(cpython-\d+|\d+(?:\.\d+)*)", part)
        for part in entry.split("/")
    ):
        return False
    # Matched on the layout a CUDA toolkit actually installs, not on the word "cuda" anywhere in the
    # path. A substring test dropped a torch directory that merely sat under a directory named after a
    # CUDA version, which is the one absolute path that has to survive.
    if safe_to_drop_toolkit_paths and _is_cuda_toolkit_directory(entry):
        return False
    if _is_unresolved_math_library_directory(entry):
        return False
    if entry.rstrip("/").endswith("/torch/lib") and has_relative_torch_route:
        return False
    return True


def _strip_absolute_runtime_paths(library: Path, ships_cuda: bool) -> None:
    """Remove unusable runtime search paths from a library the wheel ships.

    These libraries are copied out of the build tree rather than installed, so they
    still carry every directory the linker recorded while resolving their
    dependencies. Two kinds of entry are removed:

    - a directory inside this build, which names the machine that produced the
      wheel and cannot exist for a user
    - an empty entry, which the loader reads as the process working directory

    Other absolute entries are kept. The Python extensions link torch and resolve it
    through the directory the linker recorded, so dropping that would stop them
    importing in an environment where torch is not beside them.

    Best effort for a CPU wheel, because the tool cannot be guaranteed on PATH: the pip
    package does not reliably provide a binary inside a build venv, so failing there would
    break building from source on a machine that simply lacks it. The absolute paths stay
    in place and the wheel still works. It is declared as a build requirement so a release
    build gets it.

    Not best effort for a CUDA wheel. The delegate under backends/cuda/ reaches its
    dependency in lib/ only through the paths written here, so a missing tool produces a
    wheel that installs and then cannot load, and that case stops the build instead.

    What must not happen is both this and its check going quiet together, which is how
    a wheel carrying build-machine directories could ship unnoticed. So the check in
    the release tests treats a missing patchelf as a failure rather than a skip: the
    wheel-build environment has it, and that is where the guarantee belongs.
    """
    # Decided by the platform rather than the suffix, because a Python extension on macOS is
    # Mach-O while being named .so: measured on a shipped macOS wheel, every .dylib came out
    # clean and the extension still carried five build directories, because the suffix test sent
    # it to patchelf, which does not exist there.
    is_mach_o = sys.platform == "darwin"
    if library.suffix not in (".dylib", ".so") and ".so." not in library.name:
        return
    tool = shutil.which("otool") if is_mach_o else shutil.which("patchelf")
    if tool is None or (is_mach_o and shutil.which("install_name_tool") is None):
        needed = "otool and install_name_tool" if is_mach_o else "patchelf"
        if ships_cuda:
            raise RuntimeError(
                f"{needed} was not found on PATH, and this is a CUDA wheel. The relative "
                f"search path that lets the delegate reach the CUDA runtime beside it is "
                f"written here and nowhere else, so the wheel would install and then fail "
                f"to load. Install it and build again."
            )
        log.warn(
            f"{needed} was not found on PATH, so {library.name} keeps the absolute search "
            "paths the linker recorded, including any that name this machine. "
            "test_no_absolute_runtime_paths rejects those, so a wheel built without the tool "
            "fails that check rather than shipping quietly."
        )
        return
    result = subprocess.run(
        (
            [tool, "-l", os.fspath(library)]
            if is_mach_o
            else [tool, "--print-rpath", os.fspath(library)]
        ),
        capture_output=True,
        text=True,
        check=False,
    )
    if result.returncode != 0:
        return
    # A library carrying no runtime search path is not returned on early, because it still needs
    # the hops added below. The linker records nothing when every dependency resolved from a
    # default directory, and whether the builder image keeps the CUDA runtime in one of those is
    # not something this wheel controls. Returning here on that would drop the hops to the CUDA
    # wheels the package declares as dependencies, leaving the delegate with no route to them,
    # and the deciding fact would be a property of the build machine rather than of the wheel.
    # patchelf prints the same empty string for an absent tag and for one holding a single empty
    # entry. Neither is worth keeping, both fall out of the filter below, so the two need no
    # telling apart.
    original = result.stdout.strip()

    # Whether dropping an absolute CUDA toolkit path is safe. It is a cleanup when a relative hop replaces
    # it, and also when this wheel carries no CUDA at all, because then nothing in it loads from that
    # directory and the path only names the build machine. It is a regression only for a CUDA build whose
    # train this packaging does not recognise, which declares no dependency and adds no hop, so dropping
    # the path there would leave the delegate with no route to libcudart.
    safe_to_drop_toolkit_paths = not ships_cuda or bool(
        _cuda_runtime_search_paths(_package_relative_depth(library))
    )

    found = _parse_runtime_paths(original, is_mach_o)
    # Whether the library can still reach torch without the absolute entry.
    has_relative_torch_route = any(
        not entry.startswith("/") and entry.rstrip("/").endswith("/torch/lib")
        for entry in found
    )
    entries = [
        entry
        for entry in found
        if _is_usable_runtime_path(
            entry, safe_to_drop_toolkit_paths, has_relative_torch_route
        )
    ]
    # A CUDA wheel links the CUDA runtime from a separate wheel installed beside this
    # one, so the loader needs a relative hop to reach it. Without this the library
    # resolves the runtime only through the absolute toolkit path the linker recorded,
    # which names the build machine and will not exist for a user who installed from an
    # index. Appended, so a path already present keeps its position.
    _append_relative_search_paths(entries, _package_relative_depth(library))
    _write_runtime_paths(library, tool, original, found, entries, is_mach_o)


class CustomBuildPy(build_py):
    """Copies platform-independent files from the source tree into the output
    package directory.

    Override it so we can copy some files to locations that don't match their
    original relative locations.

    Standard setuptools features like package_data and MANIFEST.in can only
    include or exclude a file in the source tree; they don't have a way to map
    a file to a different relative location under the output package directory.
    """

    def _prune_unstaged_files(self) -> None:
        """Delete .py and .yaml this command staged in an earlier build and no longer wants.

        Restricted to files that exist in the source tree, because build_py is the first of
        build's sub commands and everything a later one stages is still sitting in the build
        directory when this runs. build_ext generates executorch/data/bin/__init__.py, the
        target of the flatc console script, from a template that lives elsewhere, so a walk
        that removed anything absent from build_py's own file list would delete it.
        """
        if self.editable_mode:
            return
        if not self.packages:
            # Nothing to compare against, so every staged file would look unwanted. Refuse
            # rather than empty the build directory.
            return

        wanted = set()
        for package, module, _ in self.find_all_modules():
            parts = package.split(".") if package else []
            wanted.add(os.path.join(self.build_lib, *parts, f"{module}.py"))
        for package in self.packages or ():
            src_dir = self.get_package_dir(package)
            build_dir = os.path.join(*([self.build_lib] + package.split(".")))
            for filename in self.find_data_files(package, src_dir):
                wanted.add(os.path.join(build_dir, os.path.relpath(filename, src_dir)))

        source_root = Path(__file__).parent / "src"
        for dirpath, _, filenames in os.walk(self.build_lib):
            for filename in filenames:
                if not filename.endswith((".py", ".yaml")):
                    continue
                staged = os.path.join(dirpath, filename)
                if staged in wanted:
                    continue
                relative = os.path.relpath(staged, self.build_lib)
                if not (source_root / relative).is_file():
                    # Generated by another command, so build_py must not remove it.
                    continue
                os.remove(staged)

    def find_package_modules(self, package, package_dir):
        modules = super().find_package_modules(package, package_dir)
        if self.editable_mode or not _is_test_module(package):
            # An editable install exposes the whole source tree whatever is listed here, and a
            # package outside a test directory has nothing to drop.
            return modules
        keep = _reachable_test_modules()
        return [
            entry
            for entry in modules
            if entry[1] == "__init__" or f"{package}.{entry[1]}" in keep
        ]

    def find_data_files(self, package, src_dir):
        files = super().find_data_files(package, src_dir)
        if self.editable_mode:
            return files
        root = os.path.dirname(os.path.abspath(__file__))
        return [
            _f for _f in files if not _is_shader_template(os.path.relpath(_f, root))
        ]

    def analyze_manifest(self):
        super().analyze_manifest()
        # Recent versions of setuptools may include bare directory symlinks from version
        # control (e.g. src/executorch/{backends,codegen,data,...} ->
        # ../../<name>) in manifest_files. These exist for editable mode but
        # break regular installs: build_package_data passes them to copy_file,
        # which calls os.path.isfile() and gets False for a symlink-to-directory.
        if not self.editable_mode:
            _root = os.path.dirname(os.path.abspath(__file__))
            for _pkg in list(self.manifest_files):
                self.manifest_files[_pkg] = [
                    _f
                    for _f in self.manifest_files[_pkg]
                    if os.path.isfile(os.path.join(_root, _f))
                    # A directory left out of `packages` is not simply skipped. setuptools
                    # walks up to the nearest listed package and records the file as that
                    # package's data, so a vendored *.yaml still arrives under its parent.
                    # Filter with the same list so the two agree.
                    and not _is_vendored_path(_f)
                    # Shader templates are consumed by the cmake build, not at run time.
                    and not _is_shader_template(_f)
                ]

    def _copy_extra_files(self, src_to_dst, dst_root: str) -> None:
        """Copy the non-Python files the package ships, filling in any placeholders."""
        for src, dst in src_to_dst:
            dst = os.path.join(dst_root, dst)

            # When modifying the filesystem, use the self.* methods defined by
            # Command to benefit from the same logging and dry_run logic as
            # setuptools.

            # Ensure that the destination directory exists.
            self.mkpath(os.path.dirname(dst))
            # Remove any previous copy first, because copy_file skips a
            # destination newer than its source. A second build in the same
            # checkout would otherwise keep a configuration file that was
            # substituted for the previous build's settings.
            if os.path.exists(dst):
                os.remove(dst)
            # Follow the example of the base build_py class by not preserving
            # the mode. This ensures that the output file is read/write even if
            # the input file is read-only.
            self.copy_file(src, dst, preserve_mode=False)
            if os.path.basename(dst) == "executorch-config.cmake":
                tracer_cache_dir = _tracer_cache_dir(self)
                if tracer_cache_dir is None:
                    # An editable install reaches here because setuptools runs build_py before
                    # build_ext, so no cache exists yet. Publishing nothing would leave a developer
                    # with no find_package at all, so write a best effort value from the command
                    # line. The preset this build uses turns the tracer on without passing
                    # anything, so this guess is usually wrong: the pass after the build restores
                    # the template and substitutes what the cache actually says.
                    _substitute_tracer_definition_from_args(dst)
                    _TRACER_DEFINITION_PATHS.append((os.path.abspath(src), dst))
                    continue
                _substitute_tracer_definition(dst, tracer_cache_dir)

    def run(self):
        # Copy python files to the output directory. This set of files is
        # defined by the py_module list and package_data patterns.
        build_py.run(self)

        # A rebuild over a staging directory left by an earlier build keeps whatever that build
        # put there, because build_py only ever copies and never deletes. So a file this build
        # deliberately leaves out is still present from last time, and the wheel packages it.
        # Remove what is no longer wanted rather than only skipping the copy.
        self._prune_unstaged_files()

        # dst_root is the root of the `executorch` module in the output package
        # directory. build_lib is the platform-independent root of the output
        # package, and will look like `pip-out/lib`. It can contain multiple
        # python packages, so be sure to copy the files into the `executorch`
        # package subdirectory.
        if self.editable_mode:
            # In editable mode, the package directory is the original source directory
            dst_root = self.get_package_dir("executorch")
        else:
            dst_root = os.path.join(self.build_lib, "executorch")
        # Create the version file.
        Version.write_to_python_file(os.path.join(dst_root, "version.py"))

        # Manually copy files into the output package directory. These are
        # typically python "resource" files that will live alongside the python
        # code that uses them.
        src_to_dst = [
            # TODO(dbort): See if we can add a custom pyproject.toml section for
            # these, instead of hard-coding them here. See
            # https://setuptools.pypa.io/en/latest/userguide/extension.html
            ("schema/scalar_type.fbs", "exir/_serialize/scalar_type.fbs"),
            ("schema/program.fbs", "exir/_serialize/program.fbs"),
        ]
        if not _is_minimal_build():
            cmake_cache_dir = getattr(
                self.get_finalized_command("build"), "cmake_cache_dir", None
            )
            src_to_dst += [
                (
                    "devtools/bundled_program/schema/bundled_program_schema.fbs",
                    "devtools/bundled_program/serialize/bundled_program_schema.fbs",
                ),
                (
                    "devtools/bundled_program/schema/scalar_type.fbs",
                    "devtools/bundled_program/serialize/scalar_type.fbs",
                ),
                # Install executorch-wheel-config.cmake to pip package.
                (
                    "tools/cmake/executorch-wheel-config.cmake",
                    "share/cmake/executorch-config.cmake",
                ),
                # And again where CMake looks when a consumer points CMAKE_PREFIX_PATH at the
                # package root, which is the ordinary way to use an installed package. CMake
                # searches <prefix>/lib/cmake/<name>, not <prefix>/share/cmake directly, so
                # without this copy the root is not a usable prefix and a consumer needs a
                # path that names this project's layout. The first location stays because the
                # existing contract uses it.
                (
                    "tools/cmake/executorch-wheel-config.cmake",
                    "lib/cmake/executorch/executorch-config.cmake",
                ),
                # Type checkers find a stub only under the module's own name, so
                # the extension's stub ships as _C.pyi, as the Buck build names it.
                (
                    "extension/pybindings/pybindings.pyi",
                    "extension/pybindings/_C.pyi",
                ),
            ]
            # The headers the package installs. Two audiences now: a custom-operator
            # build, which needs the kernel and tensor helpers, and a C++ application
            # using the shipped libraries as an SDK, which needs the documented entry
            # points as well.
            # TODO: Use cmake to gather the headers instead of hard-coding them here.
            # For example:
            # https://discourse.cmake.org/t/installing-headers-the-modern-way-regurgitated-and-revisited/3238/3
            for include_dir in [
                "runtime/core/",
                "runtime/executor/",
                "runtime/kernel/",
                "runtime/backend/",
                "runtime/platform/",
                "extension/kernel_util/",
                "extension/tensor/",
                "extension/threadpool/",
                # Module is how the documentation tells a C++ application to load and
                # run a program. Without it the package ships the libraries to do that
                # and no way to call them, which the C++ consumer check catches.
                "extension/module/",
                # Module's constructors take unique_ptr to the runtime's allocator
                # and loader bases, whose headers already ship. These supply the
                # concrete subclasses a caller has to construct to pass one, such as
                # MallocMemoryAllocator and FileDataLoader.
                "extension/memory_allocator/",
                "extension/data_loader/",
                # The MergedDataMap and FlatTensorDataMap entry points, whose
                # implementations ship inside libexecutorch.so. The .ptd file header
                # is included too, so a caller writing or reading a .ptd by hand has
                # its declarations.
                "extension/named_data_map/merged_data_map.h",
                "extension/flat_tensor/flat_tensor_data_map.h",
                "extension/flat_tensor/serialize/flat_tensor_header.h",
                # ETDump, whose library the package ships as a component. A profiler
                # that cannot be included is a library nobody can call.
                #
                # The whole directory except the filter, which includes a regular
                # expression library the wheel does not carry and whose implementation
                # is not in the shipped library either. Publishing a header that cannot
                # be included is worse than not publishing it, because the failure
                # arrives at compile time in someone else's project.
                "devtools/etdump/etdump_flatcc.h",
                "devtools/etdump/emitter.h",
                "devtools/etdump/utils.h",
                "devtools/etdump/data_sinks/",
            ] + (
                # The CUDA stream helper's public header, the device guard beside it, and the export macros
                # they include. The stream helper's library is shared so the process has one copy of the
                # caller-stream state, and that is a handshake the caller takes part in, so a consumer needs
                # the declarations to take part at all. The device guard's own definitions are compiled into that
                # same library, so a consumer needs this header to reach them. The CUDA allocator is compiled into
                # it too, so another delegate can take the same allocator; its header names CUDA types, so it and
                # the runtime API header it includes need the CUDA toolkit to compile.
                #
                # Only when this wheel carries the CUDA delegate, and decided from the same CMake cache the
                # libraries ship on. Keying it off the release row's CUDA version instead meant a build on
                # an unrecognised toolkit shipped both CUDA libraries and both CMake components with no
                # header, so a consumer got a component it could link and not include.
                [
                    "extension/cuda/caller_stream.h",
                    "extension/cuda/cuda_allocator.h",
                    "extension/cuda/device_guard.h",
                    "extension/cuda/export.h",
                    "extension/cuda/runtime_api.h",
                ]
                if _cuda_libraries_built(cmake_cache_dir)
                else []
            ):
                # A directory entry publishes everything under it, and a file entry publishes
                # just that file. Some directories hold headers a consumer cannot compile
                # against, so those are named individually rather than swept in.
                for src in _headers_to_install(Path(include_dir)):
                    src_to_dst.append(
                        (str(src), os.path.join("include/executorch", str(src)))
                    )
        self._copy_extra_files(src_to_dst, dst_root)

        # Copy CMake-generated Python directories that setuptools missed.
        # Setuptools discovers packages at configuration time, before CMake
        # runs. Directories created by CMake during the build (e.g. by
        # generate.py) are not in the package list and must be copied manually.
        generated_dirs = []
        if not _is_minimal_build():
            generated_dirs.append("backends/mlx/serialization/_generated")
        for rel_dir in generated_dirs:
            src_dir = os.path.join("src/executorch", rel_dir)
            if not os.path.isdir(src_dir):
                continue
            dst_dir = os.path.join(dst_root, rel_dir)
            for dirpath, _dirnames, filenames in os.walk(src_dir):
                for filename in filenames:
                    src_file = os.path.join(dirpath, filename)
                    rel_path = os.path.relpath(src_file, src_dir)
                    dst_file = os.path.join(dst_dir, rel_path)
                    self.mkpath(os.path.dirname(dst_file))
                    self.copy_file(src_file, dst_file, preserve_mode=False)

        if not _is_minimal_build():
            self._write_cmake_version_file(dst_root)

    def _write_cmake_version_file(self, dst_root: str) -> None:
        """Write the CMake package version file, so `find_package(executorch 1.2)` works.

        Generated rather than copied, because the version is only known here:
        version.txt gives the base and BUILD_VERSION overrides it for a nightly. A
        checked-in file would go stale the first time either changed.
        """
        template = os.path.join(
            os.path.dirname(os.path.abspath(__file__)),
            "tools",
            "cmake",
            "executorch-wheel-config-version.cmake.in",
        )
        with open(template) as handle:
            contents = handle.read()
        # Only the numeric release part. A Python version can carry a local segment
        # such as "1.5.0+cpu" or a development suffix, and `find_package(executorch
        # 1.5.0+cpu)` is rejected by CMake as an invalid argument, so a consumer could
        # not name the version this file reports. Strip to the dotted numbers CMake
        # can compare, which is what a consumer asks for in practice.
        # Two variables with different jobs. CMake compares PACKAGE_VERSION, so it has to be the
        # numeric release and nothing else. EXECUTORCH_BUILD_VERSION is documented as the full
        # version, which is what a consumer pinning an exact build compares against, so filling it
        # from the numeric part would make that comparison pass against a different wheel.
        build_version = Version.string()
        cmake_version = re.match(r"\d+(?:\.\d+)*", build_version)
        if not cmake_version:
            # A version file claiming 0 would satisfy every version request, which is worse than
            # not building at all.
            raise RuntimeError(
                f"cannot derive a numeric CMake version from {build_version!r}; the version file "
                "would claim 0 and satisfy every version request"
            )
        contents = contents.replace("@EXECUTORCH_VERSION@", cmake_version.group(0))
        contents = contents.replace("@EXECUTORCH_BUILD_VERSION@", build_version)
        # CMake only reads a version file that sits beside the configuration file it found, so this
        # goes to both locations the configuration is installed to. Writing it to one would leave a
        # version request silently unchecked when the other location was used.
        for destination in (
            os.path.join(dst_root, "share", "cmake", "executorch-config-version.cmake"),
            os.path.join(
                dst_root,
                "lib",
                "cmake",
                "executorch",
                "executorch-config-version.cmake",
            ),
        ):
            self.mkpath(os.path.dirname(destination))
            with open(destination, "w") as handle:
                handle.write(contents)


def _headers_to_install(entry: Path):
    """The headers a copy list entry publishes, skipping any a consumer could not or should not use.

    A directory entry publishes everything under it, and a file entry publishes just that file. A header a
    consumer cannot compile is worse than an absent one, because the failure lands in their project rather
    than here, and the directory entries sweep in a few of those.

    Test directories are skipped as a whole rather than by name. They hold mocks and stubs for this
    project's own tests, nothing the wheel installs includes them, and a consumer linking a mock allocator
    or a stub platform would get behaviour no release intends. Matched on any part starting with "test", so
    a directory named testing_util counts too, which a plain equality check missed.
    """
    candidates = entry.rglob("*.h") if entry.is_dir() else [entry]
    return [
        src
        for src in candidates
        if not _is_unshippable_header(src)
        and not any(part.startswith("test") for part in src.parts[:-1])
    ]


def _is_unshippable_header(src: Path) -> bool:
    """Whether a header is on the list of ones a consumer of the wheel could not compile.

    Compared on the path ending rather than the file name. Two headers here are both called
    tensor_util.h, one a utility that many shipped headers include and one a test helper, so matching the
    bare name either kept the helper or removed the utility everything needs.
    """
    posix = src.as_posix()
    return any(posix.endswith(entry) for entry in _UNSHIPPABLE_HEADERS)


class Buck2EnvironmentFixer(contextlib.AbstractContextManager):
    """Removes HOME from the environment when running as root.

    This script is sometimes run as root in docker containers. buck2 doesn't
    allow running as root unless $HOME is owned by root or is not set.

    TODO(pytorch/test-infra#5091): Remove this once the CI jobs stop running as
    root.
    """

    def __init__(self):
        self.saved_env = {}

    def __enter__(self):
        if os.name != "nt" and os.geteuid() == 0 and "HOME" in os.environ:
            log.info("temporarily unsetting HOME while running as root")
            self.saved_env["HOME"] = os.environ.pop("HOME")
        return self

    def __exit__(self, *args, **kwargs):
        if "HOME" in self.saved_env:
            log.info("restored HOME")
            os.environ["HOME"] = self.saved_env["HOME"]


# TODO(dbort): For editable wheels, may need to update get_source_files(),
# get_outputs(), and get_output_mapping() to satisfy
# https://setuptools.pypa.io/en/latest/userguide/extension.html#setuptools.command.build.SubCommand.get_output_mapping


def _crt_definition(build_type: str) -> str:
    """The define naming the C++ library the Windows DLLs were built with.

    A Debug build links the debug C++ library, and a consumer has to match it, so the
    runtime headers check this against the consumer's own configuration.
    """
    if build_type.lower() == "debug":
        return "ET_PREBUILT_DEBUG_CRT"
    return "ET_PREBUILT_RELEASE_CRT"


def _substitute_tracer_definition(path: str, cmake_cache_dir: str) -> None:
    """Fill in the tracer placeholder in an installed CMake configuration file.

    Read from the cache rather than assumed, because the option is set per platform and a builder can
    override it. A consumer compiling against the wrong setting gets a different object layout for
    the profiling scope classes, which fails silently.
    """
    cache_path = os.path.join(cmake_cache_dir, "CMakeCache.txt")
    if not os.path.exists(cache_path):
        raise RuntimeError(
            f"cannot read {cache_path}, so the tracer definition published to consumers cannot be "
            "derived from what was built"
        )
    enabled = CMakeCache(cache_path=cache_path).is_enabled(
        "EXECUTORCH_ENABLE_EVENT_TRACER"
    )
    build_type = CMakeCache(cache_path=cache_path).get("CMAKE_BUILD_TYPE")
    crt = _crt_definition(build_type.value if build_type else get_build_type())
    with open(path) as handle:
        contents = handle.read()
    with open(path, "w") as handle:
        handle.write(
            contents.replace(
                "@EXECUTORCH_TRACER_DEFINITION@",
                "ET_EVENT_TRACER_ENABLED" if enabled else "",
            ).replace("@EXECUTORCH_CRT_DEFINITION@", crt)
        )


# Where the first substitution wrote, so the post-build pass knows which files to correct.
_TRACER_DEFINITION_PATHS: list = []


def _substitute_tracer_definition_from_args(destination) -> None:
    """Substitute the tracer definition using the arguments that configure the build.

    Used when no CMake cache exists yet, which is the normal case for an editable install. Reading the
    same arguments CMake will read keeps the shipped config consistent with the build it describes.
    """
    # CMake accepts both -DNAME=VALUE and -D NAME=VALUE, and only the first spelling was read,
    # so the second published the tracer-off layout while the libraries carried the tracer-on
    # one. Both spellings are handled here. Absent still means off, because an editable install
    # legitimately reaches this before any cache exists and a consumer must receive a usable
    # definition rather than a raw placeholder.
    tokens = _cmake_args()
    enabled = False
    for index, argument in enumerate(tokens):
        setting = None
        if argument.startswith("-DEXECUTORCH_ENABLE_EVENT_TRACER"):
            setting = argument[2:]
        elif argument == "-D" and index + 1 < len(tokens):
            setting = tokens[index + 1]
        if setting and setting.startswith("EXECUTORCH_ENABLE_EVENT_TRACER"):
            value = setting.split("=", 1)[-1].strip().upper()
            enabled = value in ("ON", "1", "TRUE", "YES")
    text = Path(destination).read_text()
    Path(destination).write_text(
        text.replace(
            "@EXECUTORCH_TRACER_DEFINITION@",
            "ET_EVENT_TRACER_ENABLED" if enabled else "",
        ).replace("@EXECUTORCH_CRT_DEFINITION@", _crt_definition(get_build_type()))
    )


def _tracer_cache_dir(command) -> Optional[str]:
    """The build directory holding CMakeCache.txt, or None when nothing has been built yet.

    None rather than a fallback guess, because the value decides a compile definition that changes
    object layout. A wrong answer is silent: the consumer compiles a different version of the
    profiling scope classes and its traces come back empty with no error.
    """
    build_command = command.get_finalized_command("build")
    cache_dir = getattr(build_command, "cmake_cache_dir", None)
    if not cache_dir:
        # None rather than an error, because nothing has been built yet: an editable install runs
        # build_py before build_ext. The caller then leaves the file alone instead of guessing a
        # setting, so no config claims a tracer state that nothing verified.
        return None
    return os.path.abspath(cache_dir)


class CustomBuild(build):
    def initialize_options(self):
        super().initialize_options()
        # The default build_base directory is called "build", but we have a
        # top-level directory with that name. Setting build_base in setup()
        # doesn't affect this, so override the core build command.
        #
        # See build.initialize_options() in
        # setuptools/_distutils/command/build.py for the default.
        self.build_base = "pip-out"

    def run(self):  # noqa C901
        self.dump_options()
        minimal_build = _is_minimal_build()
        cmake_build_type = get_build_type(self.debug)
        # get_python_lib() typically returns the path to site-packages, where
        # all pip packages in the environment are installed.
        cmake_prefix_path = os.environ.get("CMAKE_PREFIX_PATH", get_python_lib())
        # Put the cmake cache under the temp directory, like
        # "pip-out/temp.<plat>/cmake-out".
        pip_build_dir = os.path.join(
            os.path.dirname(os.path.abspath(__file__)), self.build_temp
        )
        cmake_cache_dir = os.path.join(pip_build_dir, "cmake-out")
        self.mkpath(cmake_cache_dir)

        cmake_configuration_args = [
            f"-DPYTHON_EXECUTABLE={sys.executable}",
            # Let cmake calls like `find_package(Torch)` find cmake config files
            # like `TorchConfig.cmake` that are provided by pip packages.
            f"-DCMAKE_PREFIX_PATH={cmake_prefix_path}",
            f"-DCMAKE_BUILD_TYPE={cmake_build_type}",
        ]

        # Use ClangCL on Windows. A CUDA build uses Ninja with clang-cl instead in a
        # Visual Studio developer environment that has ninja: the CUDA toolkit's Visual
        # Studio integration fails compiler identification under ClangCL on some toolkit
        # and Visual Studio pairs (MSB4023 in its targets file), and switching the toolset
        # to cl.exe instead fails on sources cl.exe cannot compile. The multi-config
        # generator keeps the per-configuration output directories packaging reads. nvcc
        # still compiles device code with cl.exe as its host compiler. Ninja finds the
        # compilers only on PATH, which a developer environment provides and a plain shell
        # does not, so without ninja, clang-cl and cl there the build keeps the Visual
        # Studio generator it always used.
        if _is_windows():
            windows_cuda_home = _windows_cuda_toolkit(cmake_configuration_args)
            if windows_cuda_home and all(
                shutil.which(tool) for tool in ("ninja", "clang-cl", "cl")
            ):
                cmake_configuration_args += [
                    "-GNinja Multi-Config",
                    "-DCMAKE_C_COMPILER=clang-cl",
                    "-DCMAKE_CXX_COMPILER=clang-cl",
                    f"-DCUDAToolkit_ROOT={windows_cuda_home}",
                ]
                # CMake reads CUDACXX only when CMAKE_CUDA_COMPILER is unset, so naming
                # the compiler would drop options a user put there, such as the common
                # -allow-unsupported-compiler for a newer Visual Studio.
                if not os.environ.get("CUDACXX"):
                    cmake_configuration_args += [
                        f"-DCMAKE_CUDA_COMPILER={windows_cuda_home}/bin/nvcc.exe"
                    ]
            else:
                if windows_cuda_home:
                    log.warning(
                        "building the CUDA backend with the Visual Studio generator, "
                        "because ninja, clang-cl and cl are not all on PATH; run from "
                        "a Visual Studio developer prompt with ninja for the Ninja "
                        "build the wheels use"
                    )
                cmake_configuration_args += ["-T ClangCL"]

        # Allow adding extra cmake args through the environment. Used by some
        # tests and demos to expand the set of targets included in the pip
        # package.
        cmake_configuration_args += [item for item in _cmake_args() if item]

        # This is the wheel build, and that option is what gives the shipped kernel
        # libraries the names packaging looks for and what suppresses the
        # versioned soname links a wheel does not carry. Appended after the environment's
        # arguments, and so winning over them, because turning it off here produces a
        # build whose output packaging cannot find rather than a different valid wheel.
        cmake_configuration_args += ["-DEXECUTORCH_BUILD_WHEEL_DO_NOT_USE=ON"]

        if minimal_build:
            cmake_configuration_args += _minimal_cmake_flags()

        # A row that names a CUDA train has already declared the NVIDIA runtime packages in
        # install_requires, which is set before any build runs and therefore cannot consult the
        # build. If no toolkit is reachable here, the wheel would ship no CUDA library while
        # still making pip fetch the CUDA runtime, so stop rather than publish that mismatch.
        #
        # Tested against the same detection that produced the train, not against the stricter
        # (major, minor) validator. The validator rejects a toolkit whose minor is unlisted, and
        # an explicit EXECUTORCH_BUILD_CUDA=ON deliberately builds on such a toolkit, so asking
        # the validator here reported "no toolkit" on a machine whose toolkit had just been used
        # to derive the train, and refused the one case the fallback above exists to support.
        if (
            not minimal_build
            and _cuda_train()
            and install_utils._detected_cuda_major() is None
        ):
            raise RuntimeError(
                "this row names a CUDA train but no CUDA compiler was found, so the wheel "
                "would declare the CUDA runtime and ship no CUDA library. Put nvcc on PATH or "
                "point CUDACXX at it, or build the CPU row instead."
            )

        # Enable the CUDA delegate when a toolkit is present, unless the release row says this is a
        # CPU wheel. Without that second condition a CPU row built on a machine that happens to have
        # a toolkit produced a wheel carrying the delegate while its metadata declared no NVIDIA
        # package and no way to reach one, so the delegate could not load. An explicit option still
        # wins, so a caller can override the row on purpose.
        if (
            not minimal_build
            and install_utils.is_cuda_available()
            and not _row_is_cpu_only()
            and install_utils.is_cmake_option_on(
                cmake_configuration_args, "EXECUTORCH_BUILD_CUDA", default=True
            )
        ):
            cmake_configuration_args += ["-DEXECUTORCH_BUILD_CUDA=ON"]

        # Unlike CUDA, Vulkan also needs its third-party submodules, which
        # aren't in the default checkout, along with glslc. A partial checkout
        # no-ops here rather than failing in CMake.
        vulkan_third_party = os.path.join(
            os.path.dirname(os.path.abspath(__file__)),
            "backends",
            "vulkan",
            "third-party",
        )
        vulkan_submodules_present = all(
            os.path.exists(os.path.join(vulkan_third_party, *parts))
            for parts in (
                ("volk", "volk.c"),
                ("Vulkan-Headers", "include", "vulkan", "vulkan.h"),
                ("VulkanMemoryAllocator", "include", "vk_mem_alloc.h"),
            )
        )
        if (
            not minimal_build
            and vulkan_submodules_present
            and install_utils.is_vulkan_available()
            and install_utils.is_cmake_option_on(
                cmake_configuration_args, "EXECUTORCH_BUILD_VULKAN", default=True
            )
        ):
            cmake_configuration_args += ["-DEXECUTORCH_BUILD_VULKAN=ON"]

        # Check if QNN SDK is available (via QNN_SDK_ROOT env var), and if so,
        # enable building the Qualcomm backend by default.
        qnn_sdk_root = os.environ.get("QNN_SDK_ROOT", "").strip()
        if (
            not minimal_build
            and qnn_sdk_root
            and install_utils.is_cmake_option_on(
                cmake_configuration_args, "EXECUTORCH_BUILD_QNN", default=True
            )
        ):
            cmake_configuration_args += [
                "-DEXECUTORCH_BUILD_QNN=ON",
                f"-DQNN_SDK_ROOT={qnn_sdk_root}",
            ]

        # Enable the OpenVINO backend on Linux and Windows. The backend loads
        # the OpenVINO runtime at run time, so it has no build-time SDK
        # dependency.
        if (
            not minimal_build
            and sys.platform in ("linux", "win32")
            and install_utils.is_cmake_option_on(
                cmake_configuration_args,
                "EXECUTORCH_BUILD_OPENVINO",
                default=True,
            )
        ):
            cmake_configuration_args += ["-DEXECUTORCH_BUILD_OPENVINO=ON"]

        with Buck2EnvironmentFixer():
            # Generate the cmake cache from scratch to ensure that the cache state
            # is predictable.
            if os.path.exists(cmake_cache_dir):
                log.info(f"clearing {cmake_cache_dir}")
                shutil.rmtree(cmake_cache_dir)

            subprocess.run(
                [
                    "cmake",
                    *cmake_configuration_args,
                    "--preset",
                    "pybind",
                    "-B",
                    cmake_cache_dir,
                ],
                check=True,
            )

        cmake_cache = CMakeCache(
            cache_path=os.path.join(cmake_cache_dir, "CMakeCache.txt")
        )
        # Checked here rather than after the build, because the cache already records which
        # toolkit was resolved and reporting a mismatch is cheaper than compiling first.
        _verify_cuda_runtime_matches_train(cmake_cache_dir)
        cmake_build_args = [
            # Default build parallelism based on number of cores, but allow
            # overriding through the environment.
            "-j{parallelism}".format(
                parallelism=os.environ.get(
                    "CMAKE_BUILD_PARALLEL_LEVEL", os.cpu_count() - 1
                )
            ),
            # CMAKE_BUILD_TYPE variable specifies the build type (configuration) for
            # single-configuration generators (e.g., Makefile Generators or Ninja).
            # For multi-config generators (like Visual Studio), CMAKE_BUILD_TYPE
            # isn’t directly applicable.
            # During the build step, --config specifies the configuration to build
            # for multi-config generators.
            f"--config={cmake_build_type}",
        ]

        # Allow adding extra build args through the environment. Used by some
        # tests and demos to expand the set of targets included in the pip
        # package.
        cmake_build_args += [
            item
            for item in re.split(r"\s+", os.environ.get("CMAKE_BUILD_ARGS", ""))
            if item
        ]

        if minimal_build:
            # The minimal wheel only needs flatc. Every other target is gated off
            # by _minimal_cmake_flags(), so skip the entire non-minimal target
            # list explicitly rather than relying on each flag being OFF.
            cmake_build_args += ["--target", "flatbuffers_ep"]
        else:
            if cmake_cache.is_enabled("EXECUTORCH_BUILD_PYBIND"):
                cmake_build_args += ["--target", "portable_lib"]
                cmake_build_args += ["--target", "data_loader"]
                cmake_build_args += ["--target", "selective_build"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_EXTENSION_LLM_RUNNER"):
                cmake_build_args += ["--target", "_llm_runner"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_VULKAN"):
                cmake_build_args += ["--target", "vulkan_backend"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_CMSIS_NN_PYBINDS"):
                cmake_build_args += ["--target", "cmsis_nn"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_CUDA"):
                # The stream helper is not named here: the delegate links it PUBLIC, so
                # asking for the delegate builds it. It used to be named under an
                # EXECUTORCH_BUILD_SHARED condition that its packaging entry does not
                # carry, which read as if a non-shared CUDA build shipped a file nothing
                # had asked to build.
                cmake_build_args += ["--target", "aoti_cuda_backend"]
                cmake_build_args += ["--target", "aoti_common_shims_slim"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_EXTENSION_MODULE"):
                cmake_build_args += ["--target", "extension_module"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_EXTENSION_TRAINING"):
                cmake_build_args += ["--target", "_training_lib"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_COREML"):
                cmake_build_args += ["--target", "executorchcoreml"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_MLX"):
                cmake_build_args += ["--target", "mlxdelegate"]

            # The libraries a C++ consumer links out of the wheel. Most of them used to
            # be reached as dependencies of the Python extension, which meant a shared
            # build with the bindings off asked for none of them and packaging then
            # looked for files no target had been told to produce. Each condition here
            # is the one its packaging entry carries, so the two lists cannot drift.
            if cmake_cache.is_enabled("EXECUTORCH_BUILD_SHARED"):
                cmake_build_args += ["--target", "executorch_shared"]
                cmake_build_args += ["--target", "etdump"]
                if cmake_cache.is_enabled("EXECUTORCH_COREML_DELEGATE_LIBRARY_BUILT"):
                    cmake_build_args += ["--target", "coremldelegate"]
                if cmake_cache.is_enabled(
                    "EXECUTORCH_BUILD_PTHREADPOOL"
                ) and cmake_cache.is_enabled("EXECUTORCH_BUILD_CPUINFO"):
                    cmake_build_args += ["--target", "extension_threadpool"]
                if cmake_cache.is_enabled("EXECUTORCH_BUILD_KERNELS_OPTIMIZED"):
                    cmake_build_args += ["--target", "optimized_native_cpu_ops_lib"]
                if cmake_cache.is_enabled("EXECUTORCH_BUILD_KERNELS_QUANTIZED"):
                    cmake_build_args += ["--target", "executorch_quantized_ops"]
                if cmake_cache.is_enabled("EXECUTORCH_BUILD_KERNELS_TORCHAO"):
                    cmake_build_args += ["--target", "executorch_kernels_torchao"]
                if cmake_cache.is_enabled("EXECUTORCH_BUILD_XNNPACK"):
                    cmake_build_args += ["--target", "xnnpack_backend"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_KERNELS_LLM_AOT"):
                cmake_build_args += ["--target", "custom_ops_aot_lib"]
                cmake_build_args += ["--target", "quantized_ops_aot_lib"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_QNN"):
                cmake_build_args += ["--target", "qnn_executorch_backend"]
                cmake_build_args += ["--target", "PyQnnManagerAdaptor"]

            if cmake_cache.is_enabled("EXECUTORCH_BUILD_OPENVINO"):
                cmake_build_args += ["--target", "openvino_backend"]

        # Set PYTHONPATH to the location of the pip package.
        os.environ["PYTHONPATH"] = (
            site.getsitepackages()[0] + ";" + os.environ.get("PYTHONPATH", "")
        )
        # Build the system.
        self.spawn(["cmake", "--build", cmake_cache_dir, *cmake_build_args])
        # Share the cmake-out location with _BaseExtension.
        self.cmake_cache_dir = cmake_cache_dir
        # Finally, run the underlying subcommands like build_py, build_ext.
        build.run(self)


class _ExecuTorchDistribution(Distribution):
    """A Distribution that hides the synthetic BuiltFile names from metadata.

    A prebuilt file that the wheel merely copies is declared as an Extension, because
    a non-empty ext_modules is how setuptools decides a wheel is platform specific.
    Those entries are not Python modules, so they carry a synthetic name rather than a
    module path. setuptools does not know that: it derives top_level.txt from every
    ext_modules entry's name, so each recipe string was landing in the published
    metadata as a top level import name.
    """

    def iter_distribution_names(self):
        for name in super().iter_distribution_names():
            if name.startswith(_BaseExtension.SYNTHETIC_NAME_PREFIX):
                continue
            yield name


setup_kwargs = {}
if _is_minimal_build():
    setup_kwargs["packages"] = _minimal_packages()
    setup_kwargs["install_requires"] = _minimal_dependencies()
else:
    setup_kwargs["packages"] = _full_packages()
    # A CUDA wheel links the CUDA runtime but does not bundle it, so the wheels that
    # carry it are declared here. A CPU wheel adds nothing.
    setup_kwargs["install_requires"] = _base_dependencies() + _cuda_dependencies()


setup(
    version=Version.string(),
    distclass=_ExecuTorchDistribution,
    cmdclass={
        "build": CustomBuild,
        "build_ext": InstallerBuildExt,
        "build_py": CustomBuildPy,
    },
    # Note that setuptools uses the presence of ext_modules as the main signal
    # that a wheel is platform-specific. If we install any platform-specific
    # files, this list must be non-empty. Therefore, we should always install
    # platform-specific files using InstallerBuildExt.
    ext_modules=[
        BuiltFile(
            src_dir="%CMAKE_CACHE_DIR%/third-party/flatc_ep/bin/",
            src_name="flatc",
            dst="executorch/data/bin/",
            is_executable=True,
            dependent_cmake_flags=[],
        ),
        BuiltFile(
            src_dir="tools/wheel",
            src_name="pip_data_bin_init.py.in",
            dst="executorch/data/bin/__init__.py",
            dependent_cmake_flags=[],
        ),
        *(
            []
            if _is_minimal_build()
            else [
                # Install the shared runtime the Python extension links, rather
                # than having the extension contain its own copy. Named without a
                # version, so a consumer's find_library(executorch) resolves it: that
                # matches libexecutorch.so and not libexecutorch.so.1. A version is
                # only useful where something upgrades the library independently of
                # what links it, which never happens inside a wheel.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/" + _CFG,
                    src_name=get_dynamic_lib_name("executorch"),
                    dst="executorch/lib/" + get_dynamic_lib_name("executorch"),
                    dependent_cmake_flags=["EXECUTORCH_BUILD_SHARED"],
                ),
                # Install the profiler next to it, as its own library rather than
                # code fused into the Python extension, so a process has one copy of
                # it however many consumers load.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/devtools/etdump/" + _CFG,
                    src_name=get_dynamic_lib_name("executorch_etdump"),
                    dst="executorch/lib/" + get_dynamic_lib_name("executorch_etdump"),
                    # Not gated on EXECUTORCH_BUILD_DEVTOOLS. The shared build adds
                    # the devtools subdirectory itself, so the library exists
                    # whenever the shared build does. The Python extension carries a
                    # hard dependency on it, so requiring the option here left a
                    # wheel whose extension could not load at all.
                    dependent_cmake_flags=["EXECUTORCH_BUILD_SHARED"],
                ),
                # Install the shared thread pool next to it. It is a separate
                # library so that a process has one pool rather than one per
                # component that uses it.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/extension/threadpool/" + _CFG,
                    src_name=get_dynamic_lib_name("executorch_threadpool"),
                    dst="executorch/lib/"
                    + get_dynamic_lib_name("executorch_threadpool"),
                    # The target only exists when both of its dependencies are
                    # enabled, so packaging has to require them too or a shared
                    # build with either turned off looks for a file that was
                    # never built.
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_PTHREADPOOL",
                        "EXECUTORCH_BUILD_CPUINFO",
                    ],
                ),
                # Install the merged CPU kernels beside them, so the operators are
                # registered once per process rather than once per component.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/configurations/" + _CFG,
                    src_name=get_dynamic_lib_name("executorch_kernels_optimized"),
                    dst="executorch/lib/"
                    + get_dynamic_lib_name("executorch_kernels_optimized"),
                    # The target is only created when the optimized kernels are
                    # enabled, so packaging has to require that too rather than
                    # looking for a file a shared build may never have produced.
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_KERNELS_OPTIMIZED",
                    ],
                ),
                # For build systems that read pkg-config rather than CMake packages.
                # Not on Windows, where CMakeLists.txt does not generate it.
                *(
                    []
                    if _is_windows()
                    else [
                        BuiltFile(
                            src_dir="%CMAKE_CACHE_DIR%/",
                            src_name="executorch-wheel.pc",
                            dst="executorch/lib/pkgconfig/executorch.pc",
                            dependent_cmake_flags=["EXECUTORCH_BUILD_SHARED"],
                        ),
                    ]
                ),
                # The CUDA delegate and the process-wide CUDA stream helper, for a
                # wheel built from a CUDA index. Only present when the build asks for
                # CUDA, so packaging requires that rather than looking for files a
                # CPU-only build never produced.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/backends/cuda/" + _CFG,
                    src_name=get_dynamic_lib_name("executorch_backend_cuda"),
                    dst="executorch/lib/"
                    + get_dynamic_lib_name("executorch_backend_cuda"),
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_CUDA",
                    ],
                ),
                # The stream helper the delegate and the shim layer both record as a
                # dependency. Globbed rather than named, because the file name depends on
                # the build: a shared build renames it to libexecutorch_extension_cuda.so
                # to match the other shipped components, and any other build leaves it as
                # libextension_cuda.so. The shim ships whenever CUDA is on, so naming only
                # the shared spelling left the non-shared build shipping a shim whose
                # DT_NEEDED resolved to nothing. Two names means is_dynamic_lib cannot be
                # used, since it builds one name and prepends a prefix the shared spelling
                # does not have, so the prefix is globbed and the suffix is named. The
                # build type is in the directory the way the sibling entries have it.
                # Naming the suffix matters: this entry accepts exactly one file, and a
                # bare wildcard also matches what a build leaves beside the library, an
                # import library and an exports file on MSVC, or a soname's versioned
                # links, and packaging then fails on a layout that is perfectly valid.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/extension/cuda/%BUILD_TYPE%/",
                    src_name="*extension_cuda" + _dynamic_lib_suffix(),
                    dst="executorch/lib/",
                    dependent_cmake_flags=["EXECUTORCH_BUILD_CUDA"],
                ),
                # The quantized kernels, as their own library rather than code
                # fused into the AOT-only extension beside the Python bindings.
                # A C++ application running a quantized model could not link
                # them before.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/kernels/quantized/" + _CFG,
                    src_name=get_dynamic_lib_name("executorch_kernels_quantized"),
                    dst="executorch/lib/"
                    + get_dynamic_lib_name("executorch_kernels_quantized"),
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_KERNELS_QUANTIZED",
                    ],
                ),
                # The TorchAO kernels, so a C++ application running a model that
                # uses them can link them from the wheel. The Apple framework build
                # already ships these; this is the wheel's equivalent. Written to
                # the cache root because the wrapper target is declared there.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/",
                    src_name=get_dynamic_lib_name("executorch_kernels_torchao"),
                    dst="executorch/lib/"
                    + get_dynamic_lib_name("executorch_kernels_torchao"),
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_KERNELS_TORCHAO",
                    ],
                ),
                # The OpenVINO delegate, so a C++ application can link it from the
                # wheel. Only the adapter ships here: the OpenVINO runtime itself is
                # loaded at run time and comes from the openvino extra.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/backends/openvino/%BUILD_TYPE%/",
                    src_name="*executorch_backend_openvino" + _dynamic_lib_suffix(),
                    dst="executorch/lib/",
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_OPENVINO",
                    ],
                ),
                # Install the XNNPACK delegate beside them, so a process has one
                # copy of it instead of one per component that uses it.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/backends/xnnpack/" + _CFG,
                    src_name=get_dynamic_lib_name("executorch_backend_xnnpack"),
                    dst="executorch/lib/"
                    + get_dynamic_lib_name("executorch_backend_xnnpack"),
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_XNNPACK",
                    ],
                ),
                # Install the MLX delegate beside them, so a C++ consumer can link
                # it out of the wheel rather than only reaching it from Python.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/backends/mlx/",
                    src_name=get_dynamic_lib_name("executorch_backend_mlx"),
                    dst="executorch/lib/"
                    + get_dynamic_lib_name("executorch_backend_mlx"),
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_MLX",
                    ],
                ),
                # Install the Core ML delegate beside them, so a C++ consumer can
                # link it out of the wheel rather than only reaching it from Python.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/backends/apple/coreml/",
                    src_name=get_dynamic_lib_name("executorch_backend_coreml"),
                    dst="executorch/lib/"
                    + get_dynamic_lib_name("executorch_backend_coreml"),
                    dependent_cmake_flags=[
                        # Not EXECUTORCH_BUILD_COREML: that is also on for the Python
                        # extension on non-Apple platforms, where this shared library
                        # is not built. This flag is set only when it actually is.
                        "EXECUTORCH_COREML_DELEGATE_LIBRARY_BUILT",
                    ],
                ),
                *_windows_import_libraries(),
                # Install the prebuilt pybindings extension wrapper for the runtime,
                # portable kernels, and a selection of backends. This lets users
                # load and execute .pte files from python.
                BuiltExtension(
                    src="_C.cp*" if _is_windows() else "_C.*",
                    modpath="executorch.extension.pybindings._C",
                    dependent_cmake_flags=["EXECUTORCH_BUILD_PYBIND"],
                ),
                # Install the data_loader pybindings extension which provides the
                # PyDataLoader type for external pybinding extensions.
                BuiltExtension(
                    src="data_loader.cp*" if _is_windows() else "data_loader.*",
                    modpath="executorch.extension.pybindings.data_loader",
                    dependent_cmake_flags=["EXECUTORCH_BUILD_PYBIND"],
                ),
                # MLX metallib (Metal GPU kernels) must be colocated with the image
                # that carries MLX code, because MLX resolves its own address with
                # dladdr() and looks for mlx.metallib in that directory at runtime.
                # Which image that is depends on how the delegate was built: a shared
                # delegate carries MLX itself, while a static one is absorbed into the
                # Python extension. The build decides which of these two entries
                # applies, so exactly one copy ships.
                # After submodule migration, the path is backends/mlx/mlx/...
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/backends/mlx/mlx/mlx/backend/metal/kernels/",
                    src_name="mlx.metallib",
                    dst="executorch/lib/",
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_MLX",
                    ],
                ),
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/backends/mlx/mlx/mlx/backend/metal/kernels/",
                    src_name="mlx.metallib",
                    dst="executorch/extension/pybindings/",
                    dependent_cmake_flags=["EXECUTORCH_MLX_METALLIB_IN_PYBINDINGS"],
                ),
                BuiltExtension(
                    src="extension/training/_training_lib.*",  # @lint-ignore https://github.com/pytorch/executorch/blob/cb3eba0d7f630bc8cec0a9cc1df8ae2f17af3f7a/scripts/lint_xrefs.sh
                    modpath="executorch.extension.training.pybindings._training_lib",
                    dependent_cmake_flags=["EXECUTORCH_BUILD_EXTENSION_TRAINING"],
                ),
                BuiltExtension(
                    src_dir="%CMAKE_CACHE_DIR%/codegen/tools/%BUILD_TYPE%/",
                    src="selective_build.cp*" if _is_windows() else "selective_build.*",
                    modpath="executorch.codegen.tools.selective_build",
                    dependent_cmake_flags=["EXECUTORCH_BUILD_PYBIND"],
                ),
                BuiltExtension(
                    src="cmsis_nn.cp*" if _is_windows() else "cmsis_nn.*",
                    src_dir="backends/cortex_m/cmsis_nn-build/%BUILD_TYPE%/",
                    modpath="executorch.backends.cortex_m.library._cmsis_nn.cmsis_nn",
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_CMSIS_NN_PYBINDS",
                    ],
                ),
                BuiltExtension(
                    src="extension/llm/runner/_llm_runner.*",  # @lint-ignore https://github.com/pytorch/executorch/blob/cb3eba0d7f630bc8cec0a9cc1df8ae2f17af3f7a/scripts/lint_xrefs.sh
                    modpath="executorch.extension.llm.runner._llm_runner",
                    dependent_cmake_flags=["EXECUTORCH_BUILD_EXTENSION_LLM_RUNNER"],
                ),
                BuiltExtension(
                    src="executorchcoreml.*",
                    src_dir="backends/apple/coreml",
                    modpath="executorch.backends.apple.coreml.executorchcoreml",
                    dependent_cmake_flags=["EXECUTORCH_BUILD_COREML"],
                ),
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/extension/llm/custom_ops/%BUILD_TYPE%/",
                    src_name="custom_ops_aot_lib",
                    dst="executorch/extension/llm/custom_ops/",
                    is_dynamic_lib=True,
                    dependent_cmake_flags=["EXECUTORCH_BUILD_KERNELS_LLM_AOT"],
                ),
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/kernels/quantized/%BUILD_TYPE%/",
                    src_name="quantized_ops_aot_lib",
                    dst="executorch/kernels/quantized/",
                    is_dynamic_lib=True,
                    dependent_cmake_flags=["EXECUTORCH_BUILD_KERNELS_LLM_AOT"],
                ),
                # A Windows import library for the delegate's shim DLL. Lowering for
                # Windows is a cross compile, so a Linux CUDA build needs it too.
                BuiltFile(
                    src_dir="backends/cuda/runtime/",
                    src_name="aoti_cuda_shims.lib",
                    dst="executorch/data/lib/",
                    dependent_cmake_flags=["EXECUTORCH_BUILD_CUDA"],
                ),
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/backends/cuda/%BUILD_TYPE%/",
                    src_name="aoti_cuda_shims",
                    dst="executorch/backends/cuda/",
                    is_dynamic_lib=True,
                    dependent_cmake_flags=["EXECUTORCH_BUILD_CUDA"],
                ),
                # The stream helper this library needs ships in lib/ from here on,
                # alongside the other components a C++ consumer links.
                # The Qualcomm delegate, beside the others so a C++ consumer links it
                # the same way. Only the adapter ships here: the Qualcomm runtime is
                # resolved by name at run time and comes from the vendor SDK, which a
                # user installs separately. The shared build renames the output to
                # match its siblings, so this looks for that name.
                BuiltFile(
                    src_dir="%CMAKE_CACHE_DIR%/backends/qualcomm/%BUILD_TYPE%/",
                    src_name="*executorch_backend_qnn" + _dynamic_lib_suffix(),
                    dst="executorch/lib/",
                    dependent_cmake_flags=[
                        "EXECUTORCH_BUILD_SHARED",
                        "EXECUTORCH_BUILD_QNN",
                    ],
                ),
                BuiltExtension(
                    src_dir="backends/qualcomm/%BUILD_TYPE%/",
                    src=(
                        "PyQnnManagerAdaptor*.pyd"
                        if _is_windows()
                        else "PyQnnManagerAdaptor.*"
                    ),
                    modpath="executorch.backends.qualcomm.python.PyQnnManagerAdaptor",
                    dependent_cmake_flags=["EXECUTORCH_BUILD_QNN"],
                ),
            ]
        ),
    ],
    **setup_kwargs,
)