# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import ctypes
import importlib.util
import json
import logging
import os
import re
import shutil
import subprocess
import sys
import sysconfig
from pathlib import Path
from shutil import which

import torch
from packaging.version import Version, parse
from setuptools import Extension, setup
from setuptools.command.build_ext import build_ext
from setuptools_rust.build import build_rust
from setuptools_scm import get_version
from torch.utils.cpp_extension import CUDA_HOME, ROCM_HOME


def load_module_from_path(module_name, path):
    spec = importlib.util.spec_from_file_location(module_name, path)
    module = importlib.util.module_from_spec(spec)
    sys.modules[module_name] = module
    spec.loader.exec_module(module)
    return module


ROOT_DIR = Path(__file__).parent
logger = logging.getLogger(__name__)

PRECOMPILED_RUST_FRONTEND_PATH = ROOT_DIR / "vllm" / "vllm-rs"
# setuptools-rust installs PyO3 artifacts as `<module>.<ext-suffix>`, where the
# suffix ends with `.so` on Linux and macOS alike (e.g. `_rust_foo.abi3.so`).
PRECOMPILED_RUST_EXTENSION_MEMBER_REGEX = re.compile(r"vllm/_rust_[^/]*\.so$")

# cannot import envs directly because it depends on vllm,
#  which is not installed yet
envs = load_module_from_path("envs", os.path.join(ROOT_DIR, "vllm", "envs.py"))
rust_build = load_module_from_path(
    "rust_build", os.path.join(ROOT_DIR, "tools", "build_rust.py")
)

VLLM_TARGET_DEVICE = envs.VLLM_TARGET_DEVICE
USE_PRECOMPILED_EXTENSIONS = envs.VLLM_USE_PRECOMPILED
# VLLM_USE_PRECOMPILED implies precompiled rust frontend too.
USE_PRECOMPILED_RUST_FRONTEND = (
    envs.VLLM_USE_PRECOMPILED or envs.VLLM_USE_PRECOMPILED_RUST
)


def should_require_rust_frontend() -> bool:
    value = os.getenv("VLLM_REQUIRE_RUST_FRONTEND", "")
    return value.lower() not in ("", "0", "false", "no")


def get_precompiled_rust_extension_paths() -> list[Path]:
    return sorted((ROOT_DIR / "vllm").glob("_rust_*.so"))


def get_missing_precompiled_rust_extension_modules() -> list[str]:
    present = {
        path.name.split(".", 1)[0] for path in get_precompiled_rust_extension_paths()
    }
    return [
        module_name
        for module_name in rust_build.rust_py_extension_module_names()
        if module_name not in present
    ]


def has_precompiled_rust_extensions() -> bool:
    return not get_missing_precompiled_rust_extension_modules()


def is_metadata_only_build() -> bool:
    return bool({"egg_info", "dist_info"}.intersection(sys.argv[1:]))


if sys.platform.startswith("darwin") and VLLM_TARGET_DEVICE != "cpu":
    logger.warning("VLLM_TARGET_DEVICE automatically set to `cpu` due to macOS")
    VLLM_TARGET_DEVICE = "cpu"
elif not (sys.platform.startswith("linux") or sys.platform.startswith("darwin")):
    logger.warning(
        "vLLM only supports Linux platform (including WSL) and MacOS."
        "Building on %s, "
        "so vLLM may not be able to run correctly",
        sys.platform,
    )
    VLLM_TARGET_DEVICE = "empty"
elif sys.platform.startswith("linux") and os.getenv("VLLM_TARGET_DEVICE") is None:
    if torch.version.hip is not None:
        VLLM_TARGET_DEVICE = "rocm"
        logger.info("Auto-detected ROCm")
    elif torch.version.xpu is not None:
        VLLM_TARGET_DEVICE = "xpu"
        logger.info("Auto-detected XPU")
    elif torch.version.cuda is not None:
        VLLM_TARGET_DEVICE = "cuda"
        logger.info("Auto-detected CUDA")
    else:
        VLLM_TARGET_DEVICE = "cpu"


def is_sccache_available() -> bool:
    return which("sccache") is not None and not bool(
        int(os.getenv("VLLM_DISABLE_SCCACHE", "0"))
    )


def is_ccache_available() -> bool:
    return which("ccache") is not None


def is_ninja_available() -> bool:
    return which("ninja") is not None


def is_freethreaded():
    return bool(sysconfig.get_config_var("Py_GIL_DISABLED"))


def should_bundle_tcmalloc() -> bool:
    import platform

    return (
        VLLM_TARGET_DEVICE == "cpu"
        and sys.platform.startswith("linux")
        and platform.machine() in ("aarch64", "x86_64", "s390x")
    )


def find_tcmalloc() -> Path | None:
    try:
        # get all shared libs the dynamic loader knows about
        output = subprocess.check_output(
            ["ldconfig", "-p"],
            text=True,
            stderr=subprocess.DEVNULL,
        )
    except Exception:
        return None

    # search for libtcmalloc and libtcmalloc_minimal
    for library_pattern in (
        r"\blibtcmalloc_minimal\.so\.(\d+)\b",
        r"\blibtcmalloc\.so\.(\d+)\b",
    ):
        candidates: list[tuple[int, Path]] = []
        for line in output.splitlines():
            match = re.search(library_pattern, line)
            if match is None or "=>" not in line:
                continue
            candidate = Path(line.split("=>")[1].strip())
            if candidate.exists():
                candidates.append((int(match.group(1)), candidate))

        if candidates:
            # if multiple candidates are found, pick the one with the highest
            # version number
            return max(candidates, key=lambda item: item[0])[1]

    return None


def bundle_tcmalloc(build_lib: str) -> None:
    tcmalloc_library = find_tcmalloc()
    if tcmalloc_library is None:
        logger.warning(
            "Failed to locate tcmalloc. For best performance, "
            "please install tcmalloc (e.g. `sudo apt-get "
            "install -y --no-install-recommends libtcmalloc-minimal4`)"
        )
        return

    bundle_dir = os.path.join(build_lib, "vllm", "libs")
    os.makedirs(bundle_dir, exist_ok=True)
    bundle_path = os.path.join(bundle_dir, tcmalloc_library.name)
    shutil.copy2(tcmalloc_library, bundle_path)
    logger.info("Bundled tcmalloc into wheel: %s", bundle_path)


class CMakeExtension(Extension):
    def __init__(self, name: str, cmake_lists_dir: str = ".", **kwa) -> None:
        super().__init__(name, sources=[], py_limited_api=not is_freethreaded(), **kwa)
        self.cmake_lists_dir = os.path.abspath(cmake_lists_dir)


class cmake_build_ext(build_ext):
    # A dict of extension directories that have been configured.
    did_config: dict[str, bool] = {}

    #
    # Determine number of compilation jobs and optionally nvcc compile threads.
    #
    def compute_num_jobs(self):
        # `num_jobs` is either the value of the MAX_JOBS environment variable
        # (if defined) or the number of CPUs available.
        num_jobs = envs.MAX_JOBS
        if num_jobs is not None:
            num_jobs = int(num_jobs)
            logger.info("Using MAX_JOBS=%d as the number of jobs.", num_jobs)
        else:
            try:
                # os.sched_getaffinity() isn't universally available, so fall
                #  back to os.cpu_count() if we get an error here.
                num_jobs = len(os.sched_getaffinity(0))
            except AttributeError:
                num_jobs = os.cpu_count()

        nvcc_threads = None
        if _is_cuda() and CUDA_HOME is not None:
            try:
                nvcc_version = get_nvcc_cuda_version()
                if nvcc_version >= Version("11.2"):
                    # `nvcc_threads` is either the value of the NVCC_THREADS
                    # environment variable (if defined) or 1.
                    # when it is set, we reduce `num_jobs` to avoid
                    # overloading the system.
                    nvcc_threads = envs.NVCC_THREADS
                    if nvcc_threads is not None:
                        nvcc_threads = int(nvcc_threads)
                        logger.info(
                            "Using NVCC_THREADS=%d as the number of nvcc threads.",
                            nvcc_threads,
                        )
                    else:
                        nvcc_threads = 1
                    num_jobs = max(1, num_jobs // nvcc_threads)
            except Exception as e:
                logger.warning("Failed to get NVCC version: %s", e)

        return num_jobs, nvcc_threads

    #
    # Perform cmake configuration for a single extension.
    #
    def configure(self, ext: CMakeExtension) -> None:
        # If we've already configured using the CMakeLists.txt for
        # this extension, exit early.
        if ext.cmake_lists_dir in cmake_build_ext.did_config:
            return

        cmake_build_ext.did_config[ext.cmake_lists_dir] = True

        # Select the build type.
        # Note: optimization level + debug info are set by the build type
        default_cfg = "Debug" if self.debug else "RelWithDebInfo"
        cfg = envs.CMAKE_BUILD_TYPE or default_cfg

        cmake_args = [
            "-DCMAKE_BUILD_TYPE={}".format(cfg),
            "-DVLLM_TARGET_DEVICE={}".format(VLLM_TARGET_DEVICE),
        ]

        verbose = envs.VERBOSE
        if verbose:
            cmake_args += ["-DCMAKE_VERBOSE_MAKEFILE=ON"]

        if is_sccache_available():
            cmake_args += [
                "-DCMAKE_C_COMPILER_LAUNCHER=sccache",
                "-DCMAKE_CXX_COMPILER_LAUNCHER=sccache",
                "-DCMAKE_CUDA_COMPILER_LAUNCHER=sccache",
                "-DCMAKE_HIP_COMPILER_LAUNCHER=sccache",
            ]
        elif is_ccache_available():
            cmake_args += [
                "-DCMAKE_C_COMPILER_LAUNCHER=ccache",
                "-DCMAKE_CXX_COMPILER_LAUNCHER=ccache",
                "-DCMAKE_CUDA_COMPILER_LAUNCHER=ccache",
                "-DCMAKE_HIP_COMPILER_LAUNCHER=ccache",
            ]

        # Pass the python executable to cmake so it can find an exact
        # match.
        cmake_args += ["-DVLLM_PYTHON_EXECUTABLE={}".format(sys.executable)]

        # Pass the python path to cmake so it can reuse the build dependencies
        # on subsequent calls to python.
        cmake_args += ["-DVLLM_PYTHON_PATH={}".format(":".join(sys.path))]

        # Override the base directory for FetchContent downloads to $ROOT/.deps
        # This allows sharing dependencies between profiles,
        # and plays more nicely with sccache.
        # To override this, set the FETCHCONTENT_BASE_DIR environment variable.
        fc_base_dir = os.path.join(ROOT_DIR, ".deps")
        fc_base_dir = os.environ.get("FETCHCONTENT_BASE_DIR", fc_base_dir)
        cmake_args += ["-DFETCHCONTENT_BASE_DIR={}".format(fc_base_dir)]

        #
        # Setup parallelism and build tool
        #
        num_jobs, nvcc_threads = self.compute_num_jobs()

        if nvcc_threads:
            cmake_args += ["-DNVCC_THREADS={}".format(nvcc_threads)]

        if is_ninja_available():
            build_tool = ["-G", "Ninja"]
            cmake_args += [
                "-DCMAKE_JOB_POOL_COMPILE:STRING=compile",
                "-DCMAKE_JOB_POOLS:STRING=compile={}".format(num_jobs),
            ]
        else:
            # Default build tool to whatever cmake picks.
            build_tool = []
        # Make sure we use the nvcc from CUDA_HOME
        if _is_cuda() and CUDA_HOME is not None:
            cmake_args += [f"-DCMAKE_CUDA_COMPILER={CUDA_HOME}/bin/nvcc"]
        elif _is_hip() and ROCM_HOME is not None:
            cmake_args += [f"-DROCM_PATH={ROCM_HOME}"]

        other_cmake_args = os.environ.get("CMAKE_ARGS")
        if other_cmake_args:
            cmake_args += other_cmake_args.split()

        subprocess.check_call(
            ["cmake", ext.cmake_lists_dir, *build_tool, *cmake_args],
            cwd=self.build_temp,
        )

    def build_extensions(self) -> None:
        # Ensure that CMake is present and working
        try:
            subprocess.check_output(["cmake", "--version"])
        except OSError as e:
            raise RuntimeError("Cannot find CMake executable") from e

        # Create build directory if it does not exist.
        if not os.path.exists(self.build_temp):
            os.makedirs(self.build_temp)

        targets = []

        def target_name(s: str) -> str:
            return s.removeprefix("vllm.").removeprefix("vllm_flash_attn.")

        # Build all the extensions
        for ext in self.extensions:
            self.configure(ext)
            targets.append(target_name(ext.name))

        num_jobs, _ = self.compute_num_jobs()

        build_args = [
            "--build",
            ".",
            f"-j={num_jobs}",
            *[f"--target={name}" for name in targets],
        ]

        subprocess.check_call(["cmake", *build_args], cwd=self.build_temp)

        # Install the libraries
        for ext in self.extensions:
            # Install the extension into the proper location
            outdir = Path(self.get_ext_fullpath(ext.name)).parent.absolute()

            # Skip if the install directory is the same as the build directory
            if outdir == self.build_temp:
                continue

            # CMake appends the extension prefix to the install path,
            # and outdir already contains that prefix, so we need to remove it.
            prefix = outdir
            for _ in range(ext.name.count(".")):
                prefix = prefix.parent

            # prefix here should actually be the same for all components
            install_args = [
                "cmake",
                "--install",
                ".",
                "--prefix",
                prefix,
                "--component",
                target_name(ext.name),
            ]
            subprocess.check_call(install_args, cwd=self.build_temp)

    def run(self):
        # First, run the standard build_ext command to compile the extensions
        super().run()

        # bundle tcmalloc into CPU wheels for best OOB perf
        if should_bundle_tcmalloc():
            bundle_tcmalloc(self.build_lib)

        # copy vllm/vllm_flash_attn/**/*.py from self.build_lib to current
        # directory so that they can be included in the editable build
        import glob

        files = glob.glob(
            os.path.join(self.build_lib, "vllm", "vllm_flash_attn", "**", "*.py"),
            recursive=True,
        )
        for file in files:
            dst_file = os.path.join(
                "vllm/vllm_flash_attn", file.split("vllm/vllm_flash_attn/")[-1]
            )
            print(f"Copying {file} to {dst_file}")
            os.makedirs(os.path.dirname(dst_file), exist_ok=True)
            self.copy_file(file, dst_file)

        if _is_cuda() or _is_hip():
            # copy vllm/third_party/triton_kernels/**/*.py from self.build_lib
            # to current directory so that they can be included in the editable
            # build
            print(
                f"Copying {self.build_lib}/vllm/third_party/triton_kernels "
                "to vllm/third_party/triton_kernels"
            )
            shutil.copytree(
                f"{self.build_lib}/vllm/third_party/triton_kernels",
                "vllm/third_party/triton_kernels",
                dirs_exist_ok=True,
            )

        if _is_cuda():
            # copy vendored deep_gemm package from build_lib to source tree
            # for editable installs
            deep_gemm_build = os.path.join(
                self.build_lib, "vllm", "third_party", "deep_gemm"
            )
            if os.path.exists(deep_gemm_build):
                print(f"Copying {deep_gemm_build} to vllm/third_party/deep_gemm")
                shutil.copytree(
                    deep_gemm_build,
                    "vllm/third_party/deep_gemm",
                    dirs_exist_ok=True,
                )

            # copy vendored fmha_sm100 package from build_lib to source tree
            # for editable installs
            fmha_sm100_build = os.path.join(
                self.build_lib, "vllm", "third_party", "fmha_sm100"
            )
            if os.path.exists(fmha_sm100_build):
                print(f"Copying {fmha_sm100_build} to vllm/third_party/fmha_sm100")
                shutil.copytree(
                    fmha_sm100_build,
                    "vllm/third_party/fmha_sm100",
                    dirs_exist_ok=True,
                )

            tml_fa4_build = os.path.join(
                self.build_lib, "vllm", "third_party", "tml_fa4"
            )
            if os.path.exists(tml_fa4_build):
                print(f"Copying {tml_fa4_build} to vllm/third_party/tml_fa4")
                shutil.copytree(
                    tml_fa4_build,
                    "vllm/third_party/tml_fa4",
                    dirs_exist_ok=True,
                )


class precompiled_build_ext(build_ext):
    """Disables extension building when using precompiled binaries."""

    def run(self) -> None:
        return

    def build_extensions(self) -> None:
        print("Skipping build_ext: using precompiled extensions.")
        return


class precompiled_build_rust(build_rust):
    """Skips local Rust builds when all precompiled Rust artifacts are present."""

    def run(self) -> None:
        missing = []
        if not PRECOMPILED_RUST_FRONTEND_PATH.exists():
            missing.append(str(PRECOMPILED_RUST_FRONTEND_PATH))
        missing_rust_extensions = get_missing_precompiled_rust_extension_modules()
        if missing_rust_extensions:
            missing.extend(
                str(ROOT_DIR / "vllm" / f"{module_name}*.so")
                for module_name in missing_rust_extensions
            )

        if not missing:
            logger.info(
                "Skipping local Rust build: using precompiled %s and %s",
                PRECOMPILED_RUST_FRONTEND_PATH,
                get_precompiled_rust_extension_paths(),
            )
            return

        logger.warning(
            "Precompiled wheel did not provide all Rust artifacts (%s); "
            "falling back to local Rust build.",
            ", ".join(missing),
        )
        super().run()


class precompiled_wheel_utils:
    """Extracts libraries and other files from an existing wheel."""

    @staticmethod
    def fetch_metadata_for_variant(
        commit: str,
        variant: str | None,
        *,
        rocm: bool = False,
    ) -> tuple[list[dict], str]:
        """
        Fetches metadata for a specific variant of the precompiled wheel.

        For non-ROCm, fetches vllm metadata.

        For ROCm, discovers all first-level packages and combines their
        metadata into a single list.
        """
        import json
        from html.parser import HTMLParser
        from urllib.request import urlopen

        variant_dir = f"{variant}/" if variant is not None else ""

        if not rocm:
            # Keep original behavior
            repo_url = f"https://wheels.vllm.ai/{commit}/{variant_dir}vllm/"
            meta_url = repo_url + "metadata.json"
            print(f"Trying to fetch nightly build metadata from {meta_url}")
            with urlopen(meta_url) as resp:
                wheels = json.loads(resp.read().decode("utf-8"))

            return wheels, repo_url

        # ROCm: discover all packages under the variant directory.
        repo_url = f"https://wheels.vllm.ai/rocm/{commit}/{variant_dir}"

        class LinkParser(HTMLParser):
            def __init__(self):
                super().__init__()
                self.links: list[str] = []

            def handle_starttag(self, tag, attrs):
                if tag == "a":
                    href = dict(attrs).get("href")
                    if href:
                        self.links.append(href)

        with urlopen(repo_url) as resp:
            parser = LinkParser()
            parser.feed(resp.read().decode("utf-8"))

        packages = [
            href.rstrip("/")
            for href in parser.links
            if href.endswith("/")
            and href not in ("../", "./", "/")
            and "/" not in href.rstrip("/")
        ]

        wheels: list[dict] = []

        for package in packages:
            meta_url = f"{repo_url}{package}/metadata.json"
            print(f"Trying to fetch nightly build metadata from {meta_url}")
            with urlopen(meta_url) as resp:
                package_wheels = json.loads(resp.read().decode("utf-8"))
            wheels.extend(package_wheels)
        return wheels, repo_url

    @staticmethod
    def is_rocm_system() -> bool:
        """Detect ROCm without relying on torch (for build environment)."""
        if os.getenv("ROCM_PATH"):
            return True
        if os.path.isdir("/opt/rocm"):
            return True
        if which("rocminfo") is not None:
            return True
        try:
            import torch

            return torch.version.hip is not None
        except ImportError:
            return False

    @staticmethod
    def detect_system_cuda_variant() -> str:
        """Auto-detect CUDA variant from torch, nvidia-smi, or env default."""

        # Map CUDA major version to hosted wheel variants on wheels.vllm.ai
        supported = {12: "cu129", 13: "cu130"}

        # Respect explicitly set VLLM_MAIN_CUDA_VERSION
        if envs.is_set("VLLM_MAIN_CUDA_VERSION"):
            v = envs.VLLM_MAIN_CUDA_VERSION
            print(f"Using VLLM_MAIN_CUDA_VERSION={v}")
            return "cu" + v.replace(".", "")[:3]

        # Try torch.version.cuda
        cuda_version = None
        try:
            import torch

            cuda_version = torch.version.cuda
        except Exception:
            pass

        # Try nvidia-smi
        if not cuda_version:
            try:
                out = subprocess.run(
                    ["nvidia-smi"], capture_output=True, text=True, timeout=10
                )
                if m := re.search(r"CUDA Version:\s*(\d+\.\d+)", out.stdout):
                    cuda_version = m.group(1)
            except Exception:
                pass

        # Fall back to default
        if not cuda_version:
            cuda_version = envs.VLLM_MAIN_CUDA_VERSION

        # Map to supported variant
        major = int(cuda_version.split(".")[0])
        variant = supported.get(major, supported[max(supported)])
        print(f"Detected CUDA {cuda_version}, using variant {variant}")
        return variant

    @staticmethod
    def rocm_version_to_variant(rocm_version: str) -> str:
        """Convert a ROCm version string to a wheel variant, e.g. 7.2.3 -> rocm723."""
        return "rocm" + rocm_version.replace(".", "")

    @staticmethod
    def detect_system_rocm_variant() -> str | None:
        """Auto-detect the ROCm wheel variant from the installed ROCm stack."""
        rocm_version = get_rocm_version()
        if not rocm_version:
            try:
                import torch

                rocm_version = torch.version.hip
            except Exception:
                pass
        if not rocm_version:
            return None
        variant = precompiled_wheel_utils.rocm_version_to_variant(rocm_version)
        print(f"Detected ROCm {rocm_version}, using variant {variant}")
        return variant

    @staticmethod
    def fetch_available_rocm_variants(commit: str) -> list[str]:
        """List ROCm wheel variants published for a commit on wheels.vllm.ai."""
        from urllib.request import urlopen

        index_url = f"https://wheels.vllm.ai/rocm/{commit}/"
        print(f"Fetching available ROCm variants from {index_url}")
        try:
            with urlopen(index_url) as resp:
                html = resp.read().decode("utf-8")
        except Exception as e:
            logger.warning(
                "Failed to fetch ROCm variant index for commit %s: %s", commit, e
            )
            return []
        variants = sorted(set(re.findall(r"rocm\d+", html)))
        print(f"Available ROCm variants for commit {commit}: {variants}")
        return variants

    @staticmethod
    def resolve_rocm_wheel_variant(
        commit: str, variant_override: str | None
    ) -> str | None:
        """Resolve a ROCm wheel variant for a commit from wheels.vllm.ai."""
        env_variant = precompiled_wheel_utils.detect_system_rocm_variant()
        available = precompiled_wheel_utils.fetch_available_rocm_variants(commit)

        if variant_override is not None:
            if env_variant and variant_override != env_variant:
                logger.warning(
                    "VLLM_PRECOMPILED_WHEEL_VARIANT=%s does not match the "
                    "detected environment ROCm variant %s; refusing to use a "
                    "different ROCm patch wheel",
                    variant_override,
                    env_variant,
                )
                return None
            if available and variant_override not in available:
                logger.warning(
                    "Requested ROCm variant %s is not available for commit %s "
                    "(available: %s)",
                    variant_override,
                    commit,
                    available,
                )
                return None
            return variant_override

        if env_variant is None:
            if available:
                print(
                    "Could not detect ROCm version from the environment; "
                    f"available variants for commit {commit}: {available}"
                )
            return None

        if env_variant in available:
            return env_variant

        logger.warning(
            "Environment ROCm variant %s is not available for commit %s "
            "(available: %s). Precompiled wheels may not be compatible.",
            env_variant,
            commit,
            available,
        )
        return None

    @staticmethod
    def warn_if_rocm_torch_version_mismatch(
        wheels: list[dict], repo_url: str, arch: str
    ) -> None:
        """Warn if installed torch differs from the custom ROCm build on
        wheels.vllm.ai and suggest the correct install command."""
        try:
            installed = torch.__version__
        except Exception:
            return

        def _wheel_version(pkg: str) -> str | None:
            for w in wheels:
                if w.get("package_name") == pkg and arch in w.get("platform_tag", ""):
                    v = w["version"]
                    if w.get("variant") and "+" not in v:
                        v = f"{v}+{w['variant']}"
                    return v
            return None

        expected = _wheel_version("torch")
        if expected is None or installed == expected:
            return

        pkgs = f"torch=={expected}"
        triton_ver = _wheel_version("triton")
        if triton_ver:
            pkgs += f" triton=={triton_ver}"

        logger.warning(
            "Installed PyTorch %s does not match the custom build %s "
            "shipped with vLLM ROCm wheels. The ABI may differ from "
            "official releases. If you hit extension load errors, "
            "reinstall from the vLLM index:\n"
            "  pip install %s --extra-index-url %s\n"
            "  uv pip install %s --extra-index-url %s "
            "--index-strategy unsafe-best-match",
            installed,
            expected,
            pkgs,
            repo_url,
            pkgs,
            repo_url,
        )

    @staticmethod
    def find_local_rocm_wheel() -> str | None:
        """Search for a local vllm wheel in common locations."""
        import glob

        for pattern in ["/vllm-workspace/dist/vllm-*.whl", "./dist/vllm-*.whl"]:
            wheels = glob.glob(pattern)
            if wheels:
                return sorted(wheels)[-1]
        return None

    @staticmethod
    def fetch_wheel_from_pypi_index(index_url: str, package: str = "vllm") -> str:
        """Fetch the latest wheel URL from a PyPI-style simple index."""
        import platform
        from html.parser import HTMLParser
        from urllib.parse import urljoin
        from urllib.request import urlopen

        arch = platform.machine()

        class WheelLinkParser(HTMLParser):
            def __init__(self):
                super().__init__()
                self.wheels = []

            def handle_starttag(self, tag, attrs):
                if tag == "a":
                    for name, value in attrs:
                        if name == "href" and value.endswith(".whl"):
                            self.wheels.append(value)

        simple_url = f"{index_url.rstrip('/')}/{package}/"
        print(f"Fetching wheel list from {simple_url}")
        with urlopen(simple_url) as resp:
            html = resp.read().decode("utf-8")

        parser = WheelLinkParser()
        parser.feed(html)

        for wheel in reversed(parser.wheels):
            if arch in wheel:
                if wheel.startswith("http"):
                    return wheel
                return urljoin(simple_url, wheel)

        raise ValueError(f"No compatible wheel found for {arch} at {simple_url}")

    @staticmethod
    def determine_wheel_url_rocm() -> tuple[str, str | None]:
        """Determine the precompiled wheel for ROCm."""
        # Search for local wheel first
        local_wheel = precompiled_wheel_utils.find_local_rocm_wheel()
        if local_wheel is not None:
            print(f"Found local ROCm wheel: {local_wheel}")
            return local_wheel, None

        import platform

        arch = platform.machine()
        commit = os.getenv("VLLM_PRECOMPILED_WHEEL_COMMIT", "").lower()
        if not commit or len(commit) != 40:
            print(
                f"VLLM_PRECOMPILED_WHEEL_COMMIT not valid: {commit}"
                ", trying to fetch base commit in main branch"
            )
            commit = precompiled_wheel_utils.get_base_commit_in_main_branch()
        variant = precompiled_wheel_utils.resolve_rocm_wheel_variant(
            commit, os.getenv("VLLM_PRECOMPILED_WHEEL_VARIANT", None)
        )
        print(f"Using precompiled ROCm wheel commit {commit} with variant {variant}")
        wheels, repo_url = None, None
        if variant is not None:
            try:
                wheels, repo_url = precompiled_wheel_utils.fetch_metadata_for_variant(
                    commit, variant, rocm=True
                )
                precompiled_wheel_utils.warn_if_rocm_torch_version_mismatch(
                    wheels, repo_url, arch
                )
            except Exception as e:
                logger.warning(
                    "Failed to fetch ROCm wheel metadata for variant %s: %s",
                    variant,
                    e,
                )
        if wheels is not None and repo_url is not None:
            from urllib.parse import urljoin

            for wheel in wheels:
                if wheel.get("package_name") == "vllm" and arch in wheel.get(
                    "platform_tag", ""
                ):
                    print(f"Found precompiled wheel metadata: {wheel}")
                    if "path" not in wheel:
                        raise ValueError(f"Wheel metadata missing path: {wheel}")
                    wheel_url = urljoin(f"{repo_url}vllm/", wheel["path"])
                    download_filename = wheel.get("filename")
                    print(f"Using precompiled wheel URL: {wheel_url}")
                    return wheel_url, download_filename
            logger.warning(
                "No precompiled vllm wheel found for architecture %s "
                "from repo %s. All available wheels: %s",
                arch,
                repo_url,
                wheels,
            )

        # Fall back to AMD's PyPI index
        index_url = os.getenv(
            "VLLM_ROCM_WHEEL_INDEX", "https://pypi.amd.com/vllm-rocm/simple"
        )
        print(f"Fetching ROCm precompiled wheel from {index_url}")
        wheel_url = precompiled_wheel_utils.fetch_wheel_from_pypi_index(index_url)
        download_filename = wheel_url.split("/")[-1].split("#")[0]
        print(f"Using ROCm precompiled wheel: {wheel_url}")
        return wheel_url, download_filename

    @staticmethod
    def determine_wheel_url() -> tuple[str, str | None]:
        """
        Try to determine the precompiled wheel URL or path to use.
        The order of preference is:
        1. user-specified wheel location (can be either local or remote, via
           VLLM_PRECOMPILED_WHEEL_LOCATION)
        2. user-specified variant (VLLM_PRECOMPILED_WHEEL_VARIANT) from nightly repo
           or auto-detected CUDA variant based on system (torch, nvidia-smi)
        3. the default variant from nightly repo

        If downloading from the nightly repo, the commit can be specified via
        VLLM_PRECOMPILED_WHEEL_COMMIT; otherwise, the head commit in the main branch
        is used.
        """
        wheel_location = os.getenv("VLLM_PRECOMPILED_WHEEL_LOCATION", None)
        if wheel_location is not None:
            print(f"Using user-specified precompiled wheel location: {wheel_location}")
            return wheel_location, None
        else:
            # ROCm: resolve wheels from wheels.vllm.ai with environment-matched variant
            if precompiled_wheel_utils.is_rocm_system():
                return precompiled_wheel_utils.determine_wheel_url_rocm()

            import platform

            arch = platform.machine()
            # try to fetch the wheel metadata from the nightly wheel repo,
            # detecting CUDA variant from system if not specified
            variant = os.getenv("VLLM_PRECOMPILED_WHEEL_VARIANT", None)
            if variant is None:
                variant = precompiled_wheel_utils.detect_system_cuda_variant()
            commit = os.getenv("VLLM_PRECOMPILED_WHEEL_COMMIT", "").lower()
            if not commit or len(commit) != 40:
                print(
                    f"VLLM_PRECOMPILED_WHEEL_COMMIT not valid: {commit}"
                    ", trying to fetch base commit in main branch"
                )
                commit = precompiled_wheel_utils.get_base_commit_in_main_branch()
            print(f"Using precompiled wheel commit {commit} with variant {variant}")
            try_default = False
            wheels, repo_url, download_filename = None, None, None
            try:
                wheels, repo_url = precompiled_wheel_utils.fetch_metadata_for_variant(
                    commit, variant
                )
            except Exception as e:
                logger.warning(
                    "Failed to fetch precompiled wheel metadata for variant %s: %s",
                    variant,
                    e,
                )
                try_default = True  # try outside handler to keep the stacktrace simple
            if try_default:
                print("Trying the default variant from remote")
                wheels, repo_url = precompiled_wheel_utils.fetch_metadata_for_variant(
                    commit, None
                )
                # if this also fails, then we have nothing more to try / cache
            assert wheels is not None and repo_url is not None, (
                "Failed to fetch precompiled wheel metadata"
            )
            # The metadata.json has the following format:
            # see .buildkite/scripts/generate-nightly-index.py for details
            """[{
    "package_name": "vllm",
    "version": "0.11.2.dev278+gdbc3d9991",
    "build_tag": null,
    "python_tag": "cp38",
    "abi_tag": "abi3",
    "platform_tag": "manylinux1_x86_64",
    "variant": null,
    "filename": "vllm-0.11.2.dev278+gdbc3d9991-cp38-abi3-manylinux1_x86_64.whl",
    "path": "../vllm-0.11.2.dev278%2Bgdbc3d9991-cp38-abi3-manylinux1_x86_64.whl"
    },
    ...]"""
            from urllib.parse import urljoin

            for wheel in wheels:
                # TODO: maybe check more compatibility later? (python_tag, abi_tag, etc)
                if wheel.get("package_name") == "vllm" and arch in wheel.get(
                    "platform_tag", ""
                ):
                    print(f"Found precompiled wheel metadata: {wheel}")
                    if "path" not in wheel:
                        raise ValueError(f"Wheel metadata missing path: {wheel}")
                    wheel_url = urljoin(repo_url, wheel["path"])
                    download_filename = wheel.get("filename")
                    print(f"Using precompiled wheel URL: {wheel_url}")
                    break
            else:
                raise ValueError(
                    f"No precompiled vllm wheel found for architecture {arch} "
                    f"from repo {repo_url}. All available wheels: {wheels}"
                )

        return wheel_url, download_filename

    @staticmethod
    def extract_precompiled_and_patch_package(
        wheel_url_or_path: str,
        download_filename: str | None,
        *,
        extract_extensions: bool,
        extract_rust_frontend: bool,
    ) -> dict:
        import tempfile
        import zipfile

        temp_dir = None
        try:
            if not os.path.isfile(wheel_url_or_path):
                # use provided filename first, then derive from URL
                wheel_filename = download_filename or wheel_url_or_path.split("/")[-1]
                temp_dir = tempfile.mkdtemp(prefix="vllm-wheels")
                wheel_path = os.path.join(temp_dir, wheel_filename)
                print(f"Downloading wheel from {wheel_url_or_path} to {wheel_path}")
                from urllib.request import urlretrieve

                urlretrieve(wheel_url_or_path, filename=wheel_path)
            else:
                wheel_path = wheel_url_or_path
                print(f"Using existing wheel at {wheel_path}")

            package_data_patch = {}

            with zipfile.ZipFile(wheel_path) as wheel:
                exact_members = set()
                if extract_extensions:
                    exact_members.update(
                        {
                            "vllm/_C.abi3.so",
                            "vllm/_C_stable_libtorch.abi3.so",
                            "vllm/_moe_C_stable_libtorch.abi3.so",
                            "vllm/_qutlass_C.abi3.so",
                            "vllm/_flashmla_C.abi3.so",
                            "vllm/_flashmla_extension_C.abi3.so",
                            "vllm/_flashkda_C.abi3.so",
                            "vllm/_sparse_flashmla_C.abi3.so",
                            "vllm/vllm_flash_attn/_vllm_fa2_C.abi3.so",
                            "vllm/vllm_flash_attn/_vllm_fa3_C.abi3.so",
                            "vllm/cumem_allocator.abi3.so",
                            "vllm/spinloop.abi3.so",
                            "vllm/fs_io_C.abi3.so",
                            # ROCm-specific libraries
                            "vllm/_rocm_C.abi3.so",
                        }
                    )
                if extract_rust_frontend:
                    exact_members.add("vllm/vllm-rs")

                flash_attn_regex = re.compile(
                    r"vllm/vllm_flash_attn/(?:[^/.][^/]*/)*(?!\.)[^/]*\.py"
                )
                # __init__.py and flash_attn_interface.py are source-controlled
                # in vllm and should not be overwritten (matches cmake exclusions)
                flash_attn_files_to_skip = {
                    "vllm/vllm_flash_attn/__init__.py",
                    "vllm/vllm_flash_attn/flash_attn_interface.py",
                }
                triton_kernels_regex = re.compile(
                    r"vllm/third_party/triton_kernels/(?:[^/.][^/]*/)*(?!\.)[^/]*\.py"
                )
                flashmla_regex = re.compile(
                    r"vllm/third_party/flashmla/(?:[^/.][^/]*/)*(?!\.)[^/]*\.py"
                )
                # DeepGEMM: extract all files (.py, .so, .cuh, .h, .hpp, etc.)
                deep_gemm_regex = re.compile(r"vllm/third_party/deep_gemm/.*")
                fmha_sm100_regex = re.compile(r"vllm/third_party/fmha_sm100/.*")
                tml_fa4_regex = re.compile(r"vllm/third_party/tml_fa4/.*")
                file_members = []
                for member in wheel.filelist:
                    if member.filename in exact_members:
                        file_members.append(member)
                        continue
                    if (
                        extract_rust_frontend
                        and PRECOMPILED_RUST_EXTENSION_MEMBER_REGEX.match(
                            member.filename
                        )
                    ):
                        file_members.append(member)
                        continue

                    if not extract_extensions:
                        continue

                    if (
                        (
                            flash_attn_regex.match(member.filename)
                            and member.filename not in flash_attn_files_to_skip
                        )
                        or triton_kernels_regex.match(member.filename)
                        or flashmla_regex.match(member.filename)
                        or deep_gemm_regex.match(member.filename)
                        or tml_fa4_regex.match(member.filename)
                        or fmha_sm100_regex.match(member.filename)
                    ):
                        file_members.append(member)

                for file in file_members:
                    print(f"[extract] {file.filename}")
                    target_path = os.path.join(".", file.filename)
                    os.makedirs(os.path.dirname(target_path), exist_ok=True)
                    with (
                        wheel.open(file.filename) as src,
                        open(target_path, "wb") as dst,
                    ):
                        shutil.copyfileobj(src, dst)
                    mode = file.external_attr >> 16
                    if mode:
                        os.chmod(target_path, mode)

                    pkg = os.path.dirname(file.filename).replace("/", ".")
                    package_data_patch.setdefault(pkg, []).append(
                        os.path.basename(file.filename)
                    )

            return package_data_patch
        finally:
            if temp_dir is not None:
                print(f"Removing temporary directory {temp_dir}")
                shutil.rmtree(temp_dir)

    @staticmethod
    def get_base_commit_in_main_branch() -> str:
        try:
            # Get the latest commit hash of the upstream main branch.
            curl_cmd = [
                "curl",
                "-s",
                "https://api.github.com/repos/vllm-project/vllm/commits/main",
            ]
            github_token = os.getenv("GH_TOKEN", os.getenv("GITHUB_TOKEN"))
            if github_token:
                curl_cmd += [
                    "-H",
                    f"Authorization: token {github_token}",
                ]
            resp_json = subprocess.check_output(curl_cmd).decode("utf-8")
            upstream_main_commit = json.loads(resp_json)["sha"]
            print(f"Upstream main branch latest commit: {upstream_main_commit}")

            # In Docker build context, .git may be immutable or missing.
            if envs.VLLM_DOCKER_BUILD_CONTEXT:
                return upstream_main_commit

            # Check if the upstream_main_commit exists in the local repo
            try:
                subprocess.check_output(
                    ["git", "cat-file", "-e", f"{upstream_main_commit}"]
                )
            except subprocess.CalledProcessError:
                # If not present, fetch it from the remote repository.
                # Note that this does not update any local branches,
                # but ensures that this commit ref and its history are
                # available in our local repo.
                subprocess.check_call(
                    ["git", "fetch", "https://github.com/vllm-project/vllm", "main"]
                )

            # Then get the commit hash of the current branch that is the same as
            # the upstream main commit.
            current_branch = (
                subprocess.check_output(["git", "branch", "--show-current"])
                .decode("utf-8")
                .strip()
            )

            base_commit = (
                subprocess.check_output(
                    ["git", "merge-base", f"{upstream_main_commit}", current_branch]
                )
                .decode("utf-8")
                .strip()
            )
            return base_commit
        except ValueError as err:
            raise ValueError(err) from None
        except Exception as err:
            logger.warning(
                "Failed to get the base commit in the main branch. "
                "Using the nightly wheel. The libraries in this "
                "wheel may not be compatible with your dev branch: %s",
                err,
            )
            return "nightly"


def _no_device() -> bool:
    return VLLM_TARGET_DEVICE == "empty"


def _is_cuda() -> bool:
    has_cuda = torch.version.cuda is not None
    return VLLM_TARGET_DEVICE == "cuda" and has_cuda and not _is_tpu()


def _is_hip() -> bool:
    return (
        VLLM_TARGET_DEVICE == "cuda" or VLLM_TARGET_DEVICE == "rocm"
    ) and torch.version.hip is not None


def _is_tpu() -> bool:
    return VLLM_TARGET_DEVICE == "tpu"


def _is_cpu() -> bool:
    return VLLM_TARGET_DEVICE == "cpu"


def _is_xpu() -> bool:
    return VLLM_TARGET_DEVICE == "xpu"


def _build_custom_ops() -> bool:
    return _is_cuda() or _is_hip()


def get_rocm_version():
    # Get the Rocm version from the ROCM_HOME/bin/librocm-core.so
    # see https://github.com/ROCm/rocm-core/blob/d11f5c20d500f729c393680a01fa902ebf92094b/rocm_version.cpp#L21
    try:
        if ROCM_HOME is None:
            return None
        librocm_core_file = Path(ROCM_HOME) / "lib" / "librocm-core.so"
        if not librocm_core_file.is_file():
            return None
        librocm_core = ctypes.CDLL(librocm_core_file)
        VerErrors = ctypes.c_uint32
        get_rocm_core_version = librocm_core.getROCmVersion
        get_rocm_core_version.restype = VerErrors
        get_rocm_core_version.argtypes = [
            ctypes.POINTER(ctypes.c_uint32),
            ctypes.POINTER(ctypes.c_uint32),
            ctypes.POINTER(ctypes.c_uint32),
        ]
        major = ctypes.c_uint32()
        minor = ctypes.c_uint32()
        patch = ctypes.c_uint32()

        if (
            get_rocm_core_version(
                ctypes.byref(major), ctypes.byref(minor), ctypes.byref(patch)
            )
            == 0
        ):
            return f"{major.value}.{minor.value}.{patch.value}"
        return None
    except Exception:
        return None


def get_nvcc_cuda_version() -> Version:
    """Get the CUDA version from nvcc.

    Adapted from https://github.com/NVIDIA/apex/blob/8b7a1ff183741dd8f9b87e7bafd04cfde99cea28/setup.py
    """
    assert CUDA_HOME is not None, "CUDA_HOME is not set"
    nvcc_output = subprocess.check_output(
        [CUDA_HOME + "/bin/nvcc", "-V"], universal_newlines=True
    )
    output = nvcc_output.split()
    release_idx = output.index("release") + 1
    nvcc_cuda_version = parse(output[release_idx].split(",")[0])
    return nvcc_cuda_version


def get_vllm_version() -> str:
    # Allow overriding the version. This is useful to build platform-specific
    # wheels (e.g. CPU, TPU) without modifying the source.
    if env_version := os.getenv("VLLM_VERSION_OVERRIDE"):
        print(f"Overriding VLLM version with {env_version} from VLLM_VERSION_OVERRIDE")
        os.environ["SETUPTOOLS_SCM_PRETEND_VERSION"] = env_version
        return get_version(write_to="vllm/_version.py")

    version = get_version(write_to="vllm/_version.py")
    sep = "+" if "+" not in version else "."  # dev versions might contain +

    if _no_device():
        if envs.VLLM_TARGET_DEVICE == "empty":
            version += f"{sep}empty"
    elif _is_cuda():
        if USE_PRECOMPILED_EXTENSIONS and not envs.VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX:
            version += f"{sep}precompiled"
        else:
            cuda_version = str(get_nvcc_cuda_version())
            if cuda_version != envs.VLLM_MAIN_CUDA_VERSION:
                cuda_version_str = cuda_version.replace(".", "")[:3]
                # skip this for source tarball, required for pypi
                if "sdist" not in sys.argv:
                    version += f"{sep}cu{cuda_version_str}"
    elif _is_hip():
        # Get the Rocm Version
        rocm_version = get_rocm_version() or torch.version.hip
        if rocm_version and rocm_version != envs.VLLM_MAIN_CUDA_VERSION:
            version += f"{sep}rocm{rocm_version.replace('.', '')[:3]}"
    elif _is_tpu():
        version += f"{sep}tpu"
    elif _is_cpu():
        # Check the local VLLM_TARGET_DEVICE (may be set by auto-detect above),
        # not envs.VLLM_TARGET_DEVICE, so CPU-only hosts still get `+cpu`.
        if VLLM_TARGET_DEVICE == "cpu":
            version += f"{sep}cpu"
    elif _is_xpu():
        version += f"{sep}xpu"
    else:
        raise RuntimeError("Unknown runtime environment")

    return version


def get_requirements() -> list[str]:
    """Get Python package dependencies from requirements.txt."""
    requirements_dir = ROOT_DIR / "requirements"

    def _read_requirements(filename: str) -> list[str]:
        with open(requirements_dir / filename) as f:
            requirements = f.read().strip().split("\n")
        resolved_requirements = []
        for line in requirements:
            if line.startswith("-r "):
                resolved_requirements += _read_requirements(line.split()[1])
            elif (
                not line.startswith("--")
                and not line.startswith("#")
                and line.strip() != ""
            ):
                resolved_requirements.append(line)
        return resolved_requirements

    if _no_device():
        requirements = _read_requirements("common.txt")
    elif _is_cuda():
        requirements = _read_requirements("cuda.txt")
        cuda_major, cuda_minor = torch.version.cuda.split(".")
        modified_requirements = []
        for req in requirements:
            if "vllm-flash-attn" in req and cuda_major != "12":
                # vllm-flash-attn is built only for CUDA 12.x.
                # Skip for other versions.
                continue
            if "flashinfer-cubin" in req:
                # Not on PyPI since 0.6.14 (only https://flashinfer.ai/whl), so
                # it cannot be a wheel dependency; flashinfer falls back to
                # fetching cubins at runtime when the package is absent.
                continue
            if "nvidia-cutlass-dsl[cu13]" in req and cuda_major == "12":
                # [cu13] extra is the default; strip it on CUDA 12 builds.
                req = req.replace("nvidia-cutlass-dsl[cu13]", "nvidia-cutlass-dsl")
            if "humming-kernels[cu13]" in req and cuda_major == "12":
                req = req.replace("humming-kernels[cu13]", "humming-kernels[cu12]")
            modified_requirements.append(req)
        requirements = modified_requirements
    elif _is_hip():
        requirements = _read_requirements("rocm.txt")
    elif _is_tpu():
        requirements = _read_requirements("tpu.txt")
    elif _is_cpu():
        requirements = _read_requirements("cpu.txt")
    elif _is_xpu():
        requirements = _read_requirements("xpu.txt")
    else:
        raise ValueError("Unsupported platform, please use CUDA, ROCm, or CPU.")
    return requirements


ext_modules = []

if _is_cuda() or _is_hip():
    ext_modules.append(CMakeExtension(name="vllm.cumem_allocator"))
    # Optional since this doesn't get built (produce an .so file). This is just
    # copying the relevant .py files from the source repository.
    ext_modules.append(CMakeExtension(name="vllm.triton_kernels", optional=True))

if not _is_xpu() and sys.version_info >= (3, 11):
    ext_modules.append(CMakeExtension(name="vllm.spinloop"))
    ext_modules.append(CMakeExtension(name="vllm.fs_io_C"))

if _is_hip():
    ext_modules.append(CMakeExtension(name="vllm._rocm_C"))

if _is_cuda():
    ext_modules.append(CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa2_C"))
    if USE_PRECOMPILED_EXTENSIONS or (
        CUDA_HOME and get_nvcc_cuda_version() >= Version("12.3")
    ):
        # FA3 requires CUDA 12.3 or later
        ext_modules.append(CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa3_C"))
    # FA4 CuteDSL - Python-only component for FA4's cute DSL support
    # Optional since this doesn't produce a .so file, just copies Python files
    ext_modules.append(
        CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa4_cutedsl_C", optional=True)
    )
    if USE_PRECOMPILED_EXTENSIONS or (
        CUDA_HOME and get_nvcc_cuda_version() >= Version("12.9")
    ):
        # FlashMLA requires CUDA 12.9 or later
        # Optional since this doesn't get built (produce an .so file) when
        # not targeting a hopper system
        ext_modules.append(CMakeExtension(name="vllm._flashmla_C", optional=True))
        ext_modules.append(
            CMakeExtension(name="vllm._flashmla_extension_C", optional=True)
        )
    if USE_PRECOMPILED_EXTENSIONS or (
        CUDA_HOME and get_nvcc_cuda_version() >= Version("12.0")
    ):
        ext_modules.append(CMakeExtension(name="vllm._flashkda_C", optional=True))
    if envs.VLLM_USE_PRECOMPILED or (
        CUDA_HOME and get_nvcc_cuda_version() >= Version("12.3")
    ):
        # DeepGEMM requires CUDA 12.3+ (SM90/SM100)
        # Optional since it won't build on unsupported architectures
        ext_modules.append(CMakeExtension(name="vllm._deep_gemm_C", optional=True))
        ext_modules.append(CMakeExtension(name="vllm._qutlass_C", optional=True))
    # fmha_sm100 is a Python/CuTe-DSL package installed into vllm.third_party.
    ext_modules.append(CMakeExtension(name="vllm.fmha_sm100", optional=True))
    # tml-fa4 is copied into an isolated vllm.third_party package.
    ext_modules.append(CMakeExtension(name="vllm.tml_fa4", optional=True))

if _is_cpu():
    import platform

    if platform.machine() in ("x86_64", "AMD64"):
        ext_modules.append(CMakeExtension(name="vllm._C"))
        ext_modules.append(CMakeExtension(name="vllm._C_AVX512"))
        ext_modules.append(CMakeExtension(name="vllm._C_AVX2"))
    else:
        ext_modules.append(CMakeExtension(name="vllm._C"))

if _build_custom_ops():
    if _is_hip():
        ext_modules.append(CMakeExtension(name="vllm._C"))
    if _is_cuda() or _is_hip():
        ext_modules.append(CMakeExtension(name="vllm._C_stable_libtorch"))
        ext_modules.append(CMakeExtension(name="vllm._moe_C_stable_libtorch"))

package_data = {
    "vllm": [
        "py.typed",
        "libs/*.so*",
        "model_executor/layers/fused_moe/configs/*.json",
        "model_executor/layers/quantization/utils/configs/*.json",
        "entrypoints/serve/instrumentator/static/*.js",
        "entrypoints/serve/instrumentator/static/*.css",
        "distributed/kv_transfer/kv_connector/v1/hf3fs/utils/*.cpp",
        "third_party/flash_linear_attention/LICENSE",
        # DeepGEMM JIT include headers (vendored via cmake)
        "third_party/deep_gemm/include/**/*.cuh",
        "third_party/deep_gemm/include/**/*.h",
        "third_party/deep_gemm/include/**/*.hpp",
        # fmha_sm100 sparse CuTe-DSL helper kernels (vendored via cmake)
        "third_party/fmha_sm100/csrc/**/*.cu",
        "third_party/fmha_sm100/csrc/**/*.h",
        "third_party/fmha_sm100/csrc/**/*.jinja",
        "third_party/fmha_sm100/csrc/**/*.cu.jinja",
        "third_party/fmha_sm100/cute/**/*.cu",
        "third_party/fmha_sm100/cutlass/include/**/*.h",
        "third_party/fmha_sm100/cutlass/include/**/*.hpp",
        "third_party/fmha_sm100/cutlass/tools/util/include/**/*.h",
        "third_party/fmha_sm100/cutlass/tools/util/include/**/*.hpp",
    ]
}


def add_vllm_package_data(filename: str) -> None:
    vllm_files = package_data.setdefault("vllm", [])
    if filename not in vllm_files:
        vllm_files.append(filename)


# PEP 517 invokes setup.py for metadata before invoking the actual build.
if USE_PRECOMPILED_RUST_FRONTEND and not is_metadata_only_build():
    wheel_url, download_filename = precompiled_wheel_utils.determine_wheel_url()
    patch = precompiled_wheel_utils.extract_precompiled_and_patch_package(
        wheel_url,
        download_filename,
        extract_extensions=USE_PRECOMPILED_EXTENSIONS,
        extract_rust_frontend=True,
    )
    for pkg, files in patch.items():
        package_data.setdefault(pkg, []).extend(files)

# If the rust frontend binary is already present in the source tree (e.g.,
# pre-built in a separate Docker build stage), ship it as-is.
if PRECOMPILED_RUST_FRONTEND_PATH.exists():
    add_vllm_package_data("vllm-rs")
for rust_extension_path in get_precompiled_rust_extension_paths():
    add_vllm_package_data(rust_extension_path.name)

if _no_device():
    ext_modules = []

if not ext_modules:
    cmdclass = {}
else:
    cmdclass = {
        "build_ext": precompiled_build_ext
        if USE_PRECOMPILED_EXTENSIONS
        else cmake_build_ext,
    }
if (
    USE_PRECOMPILED_RUST_FRONTEND
    or PRECOMPILED_RUST_FRONTEND_PATH.exists()
    or has_precompiled_rust_extensions()
):
    cmdclass["build_rust"] = precompiled_build_rust

# Resolve the Python version first because get_vllm_version() may set
# SETUPTOOLS_SCM_PRETEND_VERSION, which the Rust version should inherit.
vllm_version = get_vllm_version()
rust_build.prepare_build_environment()

# Rust artifacts, built via setuptools-rust and installed into the package
# directory alongside the Python modules.
rust_extensions = rust_build.rust_extensions(
    optional=not should_require_rust_frontend()
)

setup(
    # static metadata should rather go in pyproject.toml
    version=vllm_version,
    ext_modules=ext_modules,
    rust_extensions=rust_extensions,
    install_requires=get_requirements(),
    extras_require={
        # AMD Zen CPU optimizations via zentorch
        "zen": ["zentorch==2.13.0.0"],
        "bench": ["pandas", "matplotlib", "seaborn", "datasets", "scipy", "plotly"],
        "tensorizer": ["tensorizer==2.10.1"],
        "fastsafetensors": ["fastsafetensors >= 0.3.3"],
        "instanttensor": ["instanttensor >= 0.1.9"],
        "runai": ["runai-model-streamer[s3,gcs,azure] >= 0.15.7"],
        "audio": [
            "av",
            "scipy",
            "soundfile",
            "soxr",
            "mistral_common[audio]",
        ],  # Required for audio processing
        "video": [],  # Kept for backwards compatibility
        # NVIDIA DeepStream (NVDEC) GPU video-decode backend. Linux x86-64
        # only; also needs system GStreamer + libv4l (see docs).
        "deepstream": ["nvidia-deepstream-videodecode-cu13>=9.0.2"],
        "flashinfer": [],  # Kept for backwards compatibility
        "b12x": ["b12x==1.2.4"],
        # Optional deps for Helion kernel development
        # NOTE: When updating helion version, also update CI files:
        #   - .buildkite/test_areas/kernels.yaml
        #   - .buildkite/test-amd.yaml
        "helion": ["helion==1.4.0"],
        # Optional deps for gRPC server (vllm serve --grpc)
        "grpc": ["smg-grpc-servicer[vllm] >= 0.5.2"],
        # Optional deps for OpenTelemetry tracing
        "otel": [
            "opentelemetry-sdk>=1.26.0",
            "opentelemetry-api>=1.26.0",
            "opentelemetry-exporter-otlp>=1.26.0",
            "opentelemetry-semantic-conventions-ai>=0.4.1",
        ],
        # extra quantization plugin
        "extra-quant": ["vllm-gguf-plugin>=0.0.2"],
    },
    cmdclass=cmdclass,
    package_data=package_data,
)