setup.py 29.4 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import ctypes
5
import importlib.util
6
import json
7
import logging
8
import os
9
import re
10
import shutil
11
import subprocess
bnellnm's avatar
bnellnm committed
12
import sys
13
import sysconfig
14
from pathlib import Path
15
from shutil import which
16

Woosuk Kwon's avatar
Woosuk Kwon committed
17
import torch
18
from packaging.version import Version, parse
19
from setuptools import Extension, setup
20
from setuptools.command.build_ext import build_ext
21
from setuptools_scm import get_version
22
from torch.utils.cpp_extension import CUDA_HOME, ROCM_HOME
23

24
25
26
27
28
29
30
31
32

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


33
ROOT_DIR = Path(__file__).parent
34
logger = logging.getLogger(__name__)
35
36
37

# cannot import envs directly because it depends on vllm,
#  which is not installed yet
38
envs = load_module_from_path("envs", os.path.join(ROOT_DIR, "vllm", "envs.py"))
39
40

VLLM_TARGET_DEVICE = envs.VLLM_TARGET_DEVICE
41

42
if sys.platform.startswith("darwin") and VLLM_TARGET_DEVICE != "cpu":
43
    logger.warning("VLLM_TARGET_DEVICE automatically set to `cpu` due to macOS")
44
    VLLM_TARGET_DEVICE = "cpu"
45
elif not (sys.platform.startswith("linux") or sys.platform.startswith("darwin")):
46
47
    logger.warning(
        "vLLM only supports Linux platform (including WSL) and MacOS."
48
        "Building on %s, "
49
50
51
        "so vLLM may not be able to run correctly",
        sys.platform,
    )
52
    VLLM_TARGET_DEVICE = "empty"
53
54
55
56
57
58
elif (
    sys.platform.startswith("linux")
    and torch.version.cuda is None
    and os.getenv("VLLM_TARGET_DEVICE") is None
    and torch.version.hip is None
):
59
    # if cuda or hip is not available and VLLM_TARGET_DEVICE is not set,
60
61
    # fallback to cpu
    VLLM_TARGET_DEVICE = "cpu"
62

bnellnm's avatar
bnellnm committed
63
64

def is_sccache_available() -> bool:
65
66
67
    return which("sccache") is not None and not bool(
        int(os.getenv("VLLM_DISABLE_SCCACHE", "0"))
    )
bnellnm's avatar
bnellnm committed
68
69
70
71
72
73
74
75
76
77


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


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


78
79
80
81
def is_freethreaded():
    return bool(sysconfig.get_config_var("Py_GIL_DISABLED"))


bnellnm's avatar
bnellnm committed
82
class CMakeExtension(Extension):
83
    def __init__(self, name: str, cmake_lists_dir: str = ".", **kwa) -> None:
84
        super().__init__(name, sources=[], py_limited_api=not is_freethreaded(), **kwa)
bnellnm's avatar
bnellnm committed
85
86
87
88
89
        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.
90
    did_config: dict[str, bool] = {}
bnellnm's avatar
bnellnm committed
91
92
93
94
95

    #
    # Determine number of compilation jobs and optionally nvcc compile threads.
    #
    def compute_num_jobs(self):
96
97
        # `num_jobs` is either the value of the MAX_JOBS environment variable
        # (if defined) or the number of CPUs available.
98
        num_jobs = envs.MAX_JOBS
99
100
        if num_jobs is not None:
            num_jobs = int(num_jobs)
101
            logger.info("Using MAX_JOBS=%d as the number of jobs.", num_jobs)
102
103
104
105
106
107
108
        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()
bnellnm's avatar
bnellnm committed
109

110
        nvcc_threads = None
111
112
113
114
115
        if _is_cuda() and get_nvcc_cuda_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.
116
            nvcc_threads = envs.NVCC_THREADS
117
118
            if nvcc_threads is not None:
                nvcc_threads = int(nvcc_threads)
119
                logger.info(
120
121
                    "Using NVCC_THREADS=%d as the number of nvcc threads.", nvcc_threads
                )
122
123
124
            else:
                nvcc_threads = 1
            num_jobs = max(1, num_jobs // nvcc_threads)
bnellnm's avatar
bnellnm committed
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141

        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"
142
        cfg = envs.CMAKE_BUILD_TYPE or default_cfg
bnellnm's avatar
bnellnm committed
143
144

        cmake_args = [
145
146
            "-DCMAKE_BUILD_TYPE={}".format(cfg),
            "-DVLLM_TARGET_DEVICE={}".format(VLLM_TARGET_DEVICE),
bnellnm's avatar
bnellnm committed
147
148
        ]

149
        verbose = envs.VERBOSE
bnellnm's avatar
bnellnm committed
150
        if verbose:
151
            cmake_args += ["-DCMAKE_VERBOSE_MAKEFILE=ON"]
bnellnm's avatar
bnellnm committed
152
153
154

        if is_sccache_available():
            cmake_args += [
155
156
157
158
                "-DCMAKE_C_COMPILER_LAUNCHER=sccache",
                "-DCMAKE_CXX_COMPILER_LAUNCHER=sccache",
                "-DCMAKE_CUDA_COMPILER_LAUNCHER=sccache",
                "-DCMAKE_HIP_COMPILER_LAUNCHER=sccache",
bnellnm's avatar
bnellnm committed
159
160
161
            ]
        elif is_ccache_available():
            cmake_args += [
162
163
164
165
                "-DCMAKE_C_COMPILER_LAUNCHER=ccache",
                "-DCMAKE_CXX_COMPILER_LAUNCHER=ccache",
                "-DCMAKE_CUDA_COMPILER_LAUNCHER=ccache",
                "-DCMAKE_HIP_COMPILER_LAUNCHER=ccache",
bnellnm's avatar
bnellnm committed
166
167
168
169
            ]

        # Pass the python executable to cmake so it can find an exact
        # match.
170
        cmake_args += ["-DVLLM_PYTHON_EXECUTABLE={}".format(sys.executable)]
bnellnm's avatar
bnellnm committed
171

172
173
        # Pass the python path to cmake so it can reuse the build dependencies
        # on subsequent calls to python.
174
        cmake_args += ["-DVLLM_PYTHON_PATH={}".format(":".join(sys.path))]
175

176
177
178
179
180
181
        # 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)
182
        cmake_args += ["-DFETCHCONTENT_BASE_DIR={}".format(fc_base_dir)]
183

bnellnm's avatar
bnellnm committed
184
185
186
187
188
189
        #
        # Setup parallelism and build tool
        #
        num_jobs, nvcc_threads = self.compute_num_jobs()

        if nvcc_threads:
190
            cmake_args += ["-DNVCC_THREADS={}".format(nvcc_threads)]
bnellnm's avatar
bnellnm committed
191
192

        if is_ninja_available():
193
            build_tool = ["-G", "Ninja"]
bnellnm's avatar
bnellnm committed
194
            cmake_args += [
195
196
                "-DCMAKE_JOB_POOL_COMPILE:STRING=compile",
                "-DCMAKE_JOB_POOLS:STRING=compile={}".format(num_jobs),
bnellnm's avatar
bnellnm committed
197
198
199
200
            ]
        else:
            # Default build tool to whatever cmake picks.
            build_tool = []
201
202
        # Make sure we use the nvcc from CUDA_HOME
        if _is_cuda():
203
            cmake_args += [f"-DCMAKE_CUDA_COMPILER={CUDA_HOME}/bin/nvcc"]
204
205
        elif _is_hip():
            cmake_args += [f"-DROCM_PATH={ROCM_HOME}"]
206
207
208
209
210

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

bnellnm's avatar
bnellnm committed
211
        subprocess.check_call(
212
213
214
            ["cmake", ext.cmake_lists_dir, *build_tool, *cmake_args],
            cwd=self.build_temp,
        )
bnellnm's avatar
bnellnm committed
215
216
217
218

    def build_extensions(self) -> None:
        # Ensure that CMake is present and working
        try:
219
            subprocess.check_output(["cmake", "--version"])
bnellnm's avatar
bnellnm committed
220
        except OSError as e:
221
            raise RuntimeError("Cannot find CMake executable") from e
bnellnm's avatar
bnellnm committed
222
223
224
225
226

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

227
        targets = []
228
229
230
231

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

bnellnm's avatar
bnellnm committed
232
233
234
        # Build all the extensions
        for ext in self.extensions:
            self.configure(ext)
235
            targets.append(target_name(ext.name))
bnellnm's avatar
bnellnm committed
236

237
        num_jobs, _ = self.compute_num_jobs()
bnellnm's avatar
bnellnm committed
238

239
240
241
242
243
244
        build_args = [
            "--build",
            ".",
            f"-j={num_jobs}",
            *[f"--target={name}" for name in targets],
        ]
bnellnm's avatar
bnellnm committed
245

246
        subprocess.check_call(["cmake", *build_args], cwd=self.build_temp)
247

248
249
250
251
252
253
254
255
256
257
258
259
        # 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
260
            for _ in range(ext.name.count(".")):
261
262
263
264
                prefix = prefix.parent

            # prefix here should actually be the same for all components
            install_args = [
265
266
267
268
269
270
271
                "cmake",
                "--install",
                ".",
                "--prefix",
                prefix,
                "--component",
                target_name(ext.name),
272
273
274
            ]
            subprocess.check_call(install_args, cwd=self.build_temp)

275
276
277
278
    def run(self):
        # First, run the standard build_ext command to compile the extensions
        super().run()

279
        # copy vllm/vllm_flash_attn/**/*.py from self.build_lib to current
280
281
        # directory so that they can be included in the editable build
        import glob
282
283
284
285
286

        files = glob.glob(
            os.path.join(self.build_lib, "vllm", "vllm_flash_attn", "**", "*.py"),
            recursive=True,
        )
287
        for file in files:
288
289
290
            dst_file = os.path.join(
                "vllm/vllm_flash_attn", file.split("vllm/vllm_flash_attn/")[-1]
            )
291
            print(f"Copying {file} to {dst_file}")
292
            os.makedirs(os.path.dirname(dst_file), exist_ok=True)
293
294
            self.copy_file(file, dst_file)

295
296
297
298
299
300
301
302
303
304
305
306
307
308
        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,
            )

309

310
311
312
class precompiled_build_ext(build_ext):
    """Disables extension building when using precompiled binaries."""

313
    def run(self) -> None:
314
        return
315

316
317
318
319
320
321
    def build_extensions(self) -> None:
        print("Skipping build_ext: using precompiled extensions.")
        return


class precompiled_wheel_utils:
322
323
    """Extracts libraries and other files from an existing wheel."""

324
    @staticmethod
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
    def fetch_metadata_for_variant(
        commit: str, variant: str | None
    ) -> tuple[list[dict], str]:
        """
        Fetches metadata for a specific variant of the precompiled wheel.
        """
        variant_dir = f"{variant}/" if variant is not None else ""
        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}")
        from urllib.request import urlopen

        with urlopen(meta_url) as resp:
            # urlopen raises HTTPError on unexpected status code
            wheels = json.loads(resp.read().decode("utf-8"))
        return wheels, repo_url

    @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 from nightly repo (current main commit via
           VLLM_PRECOMPILED_WHEEL_VARIANT)
        3. the variant corresponding to VLLM_MAIN_CUDA_VERSION from nightly repo
        4. the default variant from nightly repo (current main commit)
        """
        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:
            import platform

            arch = platform.machine()
            # try to fetch the wheel metadata from the nightly wheel repo
            main_variant = "cu" + envs.VLLM_MAIN_CUDA_VERSION.replace(".", "")
            variant = os.getenv("VLLM_PRECOMPILED_WHEEL_VARIANT", main_variant)
            commit = os.getenv(
                "VLLM_PRECOMPILED_WHEEL_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
    ) -> dict:
432
433
434
435
436
437
        import tempfile
        import zipfile

        temp_dir = None
        try:
            if not os.path.isfile(wheel_url_or_path):
438
439
                # use provided filename first, then derive from URL
                wheel_filename = download_filename or wheel_url_or_path.split("/")[-1]
440
441
                temp_dir = tempfile.mkdtemp(prefix="vllm-wheels")
                wheel_path = os.path.join(temp_dir, wheel_filename)
442
                print(f"Downloading wheel from {wheel_url_or_path} to {wheel_path}")
443
                from urllib.request import urlretrieve
444

445
446
447
448
449
450
451
452
453
454
455
456
                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:
                files_to_copy = [
                    "vllm/_C.abi3.so",
                    "vllm/_moe_C.abi3.so",
                    "vllm/_flashmla_C.abi3.so",
457
458
                    "vllm/_flashmla_extension_C.abi3.so",
                    "vllm/_sparse_flashmla_C.abi3.so",
459
460
461
462
463
                    "vllm/vllm_flash_attn/_vllm_fa2_C.abi3.so",
                    "vllm/vllm_flash_attn/_vllm_fa3_C.abi3.so",
                    "vllm/cumem_allocator.abi3.so",
                ]

464
                flash_attn_regex = re.compile(
465
466
                    r"vllm/vllm_flash_attn/(?:[^/.][^/]*/)*(?!\.)[^/]*\.py"
                )
467
468
469
                triton_kernels_regex = re.compile(
                    r"vllm/third_party/triton_kernels/(?:[^/.][^/]*/)*(?!\.)[^/]*\.py"
                )
470
                file_members = list(
471
472
                    filter(lambda x: x.filename in files_to_copy, wheel.filelist)
                )
473
                file_members += list(
474
475
476
477
478
479
                    filter(lambda x: flash_attn_regex.match(x.filename), wheel.filelist)
                )
                file_members += list(
                    filter(
                        lambda x: triton_kernels_regex.match(x.filename), wheel.filelist
                    )
480
                )
481
482
483
484
485

                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)
486
487
488
489
                    with (
                        wheel.open(file.filename) as src,
                        open(target_path, "wb") as dst,
                    ):
490
491
492
493
                        shutil.copyfileobj(src, dst)

                    pkg = os.path.dirname(file.filename).replace("/", ".")
                    package_data_patch.setdefault(pkg, []).append(
494
495
                        os.path.basename(file.filename)
                    )
496
497
498
499
500
501
502
503
504

            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:
505
506
507
        # Force to use the nightly wheel. This is mainly used for CI testing.
        if envs.VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL:
            return "nightly"
508
509

        try:
510
            # Get the latest commit hash of the upstream main branch.
511
512
513
514
515
516
517
            resp_json = subprocess.check_output(
                [
                    "curl",
                    "-s",
                    "https://api.github.com/repos/vllm-project/vllm/commits/main",
                ]
            ).decode("utf-8")
518
519
            upstream_main_commit = json.loads(resp_json)["sha"]

520
521
522
523
            # In Docker build context, .git may be immutable or missing.
            if envs.VLLM_DOCKER_BUILD_CONTEXT:
                return upstream_main_commit

524
525
526
            # Check if the upstream_main_commit exists in the local repo
            try:
                subprocess.check_output(
527
528
                    ["git", "cat-file", "-e", f"{upstream_main_commit}"]
                )
529
530
531
532
533
            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.
534
535
536
                subprocess.check_call(
                    ["git", "fetch", "https://github.com/vllm-project/vllm", "main"]
                )
537
538
539

            # Then get the commit hash of the current branch that is the same as
            # the upstream main commit.
540
541
542
543
544
            current_branch = (
                subprocess.check_output(["git", "branch", "--show-current"])
                .decode("utf-8")
                .strip()
            )
545

546
547
548
549
550
551
552
            base_commit = (
                subprocess.check_output(
                    ["git", "merge-base", f"{upstream_main_commit}", current_branch]
                )
                .decode("utf-8")
                .strip()
            )
553
            return base_commit
554
555
        except ValueError as err:
            raise ValueError(err) from None
556
557
558
559
        except Exception as err:
            logger.warning(
                "Failed to get the base commit in the main branch. "
                "Using the nightly wheel. The libraries in this "
560
561
562
                "wheel may not be compatible with your dev branch: %s",
                err,
            )
563
            return "nightly"
564
565


566
567
568
569
def _no_device() -> bool:
    return VLLM_TARGET_DEVICE == "empty"


570
def _is_cuda() -> bool:
571
    has_cuda = torch.version.cuda is not None
572
    return VLLM_TARGET_DEVICE == "cuda" and has_cuda and not _is_tpu()
573
574


575
def _is_hip() -> bool:
576
577
578
    return (
        VLLM_TARGET_DEVICE == "cuda" or VLLM_TARGET_DEVICE == "rocm"
    ) and torch.version.hip is not None
579
580


581
582
583
584
def _is_tpu() -> bool:
    return VLLM_TARGET_DEVICE == "tpu"


585
586
587
588
def _is_cpu() -> bool:
    return VLLM_TARGET_DEVICE == "cpu"


589
590
591
592
def _is_xpu() -> bool:
    return VLLM_TARGET_DEVICE == "xpu"


593
594
595
596
def _build_custom_ops() -> bool:
    return _is_cuda() or _is_hip() or _is_cpu()


597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
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:
        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()
616

617
618
619
620
621
622
        if (
            get_rocm_core_version(
                ctypes.byref(major), ctypes.byref(minor), ctypes.byref(patch)
            )
            == 0
        ):
623
            return f"{major.value}.{minor.value}.{patch.value}"
624
        return None
625
    except Exception:
626
        return None
Woosuk Kwon's avatar
Woosuk Kwon committed
627

628

bnellnm's avatar
bnellnm committed
629
def get_nvcc_cuda_version() -> Version:
630
631
632
633
    """Get the CUDA version from nvcc.

    Adapted from https://github.com/NVIDIA/apex/blob/8b7a1ff183741dd8f9b87e7bafd04cfde99cea28/setup.py
    """
634
    assert CUDA_HOME is not None, "CUDA_HOME is not set"
635
636
637
    nvcc_output = subprocess.check_output(
        [CUDA_HOME + "/bin/nvcc", "-V"], universal_newlines=True
    )
638
639
640
641
642
643
    output = nvcc_output.split()
    release_idx = output.index("release") + 1
    nvcc_cuda_version = parse(output[release_idx].split(",")[0])
    return nvcc_cuda_version


644
def get_vllm_version() -> str:
645
646
647
    # 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"):
648
649
650
        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")
651

652
    version = get_version(write_to="vllm/_version.py")
653
    sep = "+" if "+" not in version else "."  # dev versions might contain +
654

655
    if _no_device():
656
        if envs.VLLM_TARGET_DEVICE == "empty":
657
            version += f"{sep}empty"
658
    elif _is_cuda():
659
        if envs.VLLM_USE_PRECOMPILED and not envs.VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX:
660
            version += f"{sep}precompiled"
661
662
        else:
            cuda_version = str(get_nvcc_cuda_version())
663
            if cuda_version != envs.VLLM_MAIN_CUDA_VERSION:
664
665
666
667
                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}"
668
    elif _is_hip():
669
670
        # Get the Rocm Version
        rocm_version = get_rocm_version() or torch.version.hip
671
        if rocm_version and rocm_version != envs.VLLM_MAIN_CUDA_VERSION:
672
            version += f"{sep}rocm{rocm_version.replace('.', '')[:3]}"
673
    elif _is_tpu():
674
        version += f"{sep}tpu"
675
    elif _is_cpu():
676
677
        if envs.VLLM_TARGET_DEVICE == "cpu":
            version += f"{sep}cpu"
678
    elif _is_xpu():
679
        version += f"{sep}xpu"
680
    else:
681
        raise RuntimeError("Unknown runtime environment")
682

683
684
685
    return version


686
def get_requirements() -> list[str]:
687
    """Get Python package dependencies from requirements.txt."""
688
    requirements_dir = ROOT_DIR / "requirements"
689

690
    def _read_requirements(filename: str) -> list[str]:
691
        with open(requirements_dir / filename) as f:
692
            requirements = f.read().strip().split("\n")
693
694
695
696
        resolved_requirements = []
        for line in requirements:
            if line.startswith("-r "):
                resolved_requirements += _read_requirements(line.split()[1])
697
698
699
700
701
            elif (
                not line.startswith("--")
                and not line.startswith("#")
                and line.strip() != ""
            ):
702
703
704
                resolved_requirements.append(line)
        return resolved_requirements

705
    if _no_device():
706
        requirements = _read_requirements("common.txt")
707
    elif _is_cuda():
708
        requirements = _read_requirements("cuda.txt")
709
        cuda_major, cuda_minor = torch.version.cuda.split(".")
710
711
        modified_requirements = []
        for req in requirements:
712
            if "vllm-flash-attn" in req and cuda_major != "12":
713
                # vllm-flash-attn is built only for CUDA 12.x.
714
715
716
                # Skip for other versions.
                continue
            modified_requirements.append(req)
717
        requirements = modified_requirements
718
    elif _is_hip():
719
        requirements = _read_requirements("rocm.txt")
720
    elif _is_tpu():
721
        requirements = _read_requirements("tpu.txt")
722
    elif _is_cpu():
723
        requirements = _read_requirements("cpu.txt")
724
    elif _is_xpu():
725
        requirements = _read_requirements("xpu.txt")
726
    else:
727
        raise ValueError("Unsupported platform, please use CUDA, ROCm, or CPU.")
728
729
730
    return requirements


bnellnm's avatar
bnellnm committed
731
732
ext_modules = []

733
if _is_cuda() or _is_hip():
bnellnm's avatar
bnellnm committed
734
    ext_modules.append(CMakeExtension(name="vllm._moe_C"))
735
    ext_modules.append(CMakeExtension(name="vllm.cumem_allocator"))
736
737
738
    # 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))
bnellnm's avatar
bnellnm committed
739

740
741
742
if _is_hip():
    ext_modules.append(CMakeExtension(name="vllm._rocm_C"))

743
if _is_cuda():
744
    ext_modules.append(CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa2_C"))
745
746
    if envs.VLLM_USE_PRECOMPILED or get_nvcc_cuda_version() >= Version("12.3"):
        # FA3 requires CUDA 12.3 or later
747
        ext_modules.append(CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa3_C"))
748
749
        # Optional since this doesn't get built (produce an .so file) when
        # not targeting a hopper system
750
        ext_modules.append(CMakeExtension(name="vllm._flashmla_C", optional=True))
751
        ext_modules.append(
752
753
            CMakeExtension(name="vllm._flashmla_extension_C", optional=True)
        )
754

755
if _build_custom_ops():
bnellnm's avatar
bnellnm committed
756
757
    ext_modules.append(CMakeExtension(name="vllm._C"))

758
package_data = {
759
760
761
762
763
    "vllm": [
        "py.typed",
        "model_executor/layers/fused_moe/configs/*.json",
        "model_executor/layers/quantization/utils/configs/*.json",
    ]
764
}
Simon Mo's avatar
Simon Mo committed
765

766

767
768
# If using precompiled, extract and patch package_data (in advance of setup)
if envs.VLLM_USE_PRECOMPILED:
769
770
771
772
    wheel_url, download_filename = precompiled_wheel_utils.determine_wheel_url()
    patch = precompiled_wheel_utils.extract_precompiled_and_patch_package(
        wheel_url, download_filename
    )
773
774
775
    for pkg, files in patch.items():
        package_data.setdefault(pkg, []).extend(files)

776
777
778
if _no_device():
    ext_modules = []

779
if not ext_modules:
780
781
    cmdclass = {}
else:
782
    cmdclass = {
783
784
785
        "build_ext": precompiled_build_ext
        if envs.VLLM_USE_PRECOMPILED
        else cmake_build_ext
786
    }
787

bnellnm's avatar
bnellnm committed
788
setup(
789
    # static metadata should rather go in pyproject.toml
790
    version=get_vllm_version(),
Woosuk Kwon's avatar
Woosuk Kwon committed
791
    ext_modules=ext_modules,
792
    install_requires=get_requirements(),
793
    extras_require={
794
        "bench": ["pandas", "matplotlib", "seaborn", "datasets"],
795
        "tensorizer": ["tensorizer==2.10.1"],
796
        "fastsafetensors": ["fastsafetensors >= 0.1.10"],
797
        "runai": ["runai-model-streamer[s3,gcs] >= 0.15.0"],
798
799
800
801
802
        "audio": [
            "librosa",
            "soundfile",
            "mistral_common[audio]",
        ],  # Required for audio processing
803
        "video": [],  # Kept for backwards compatibility
804
        "flashinfer": [],  # Kept for backwards compatibility
805
806
        # Optional deps for AMD FP4 quantization support
        "petit-kernel": ["petit-kernel"],
807
    },
808
    cmdclass=cmdclass,
Simon Mo's avatar
Simon Mo committed
809
    package_data=package_data,
Woosuk Kwon's avatar
Woosuk Kwon committed
810
)