setup.py 24.7 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import ctypes
4
import importlib.util
5
import json
6
import logging
7
8
import os
import re
9
import subprocess
bnellnm's avatar
bnellnm committed
10
import sys
11
from pathlib import Path
12
from shutil import which
13

Woosuk Kwon's avatar
Woosuk Kwon committed
14
import torch
15
from packaging.version import Version, parse
16
from setuptools import Extension, setup
17
from setuptools.command.build_ext import build_ext
18
from setuptools_scm import get_version
19
from torch.utils.cpp_extension import CUDA_HOME, ROCM_HOME
20

21
22
23
24
25
26
27
28
29

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


30
ROOT_DIR = Path(__file__).parent
31
logger = logging.getLogger(__name__)
32
33
34
35
36
37

# 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'))

VLLM_TARGET_DEVICE = envs.VLLM_TARGET_DEVICE
38

39
if sys.platform.startswith("darwin") and VLLM_TARGET_DEVICE != "cpu":
40
    logger.warning(
41
42
43
44
45
46
        "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."
47
48
49
        "Building on %s, "
        "so vLLM may not be able to run correctly", sys.platform)
    VLLM_TARGET_DEVICE = "empty"
50
elif (sys.platform.startswith("linux") and torch.version.cuda is None
51
52
53
      and os.getenv("VLLM_TARGET_DEVICE") is None
      and torch.version.hip is None):
    # if cuda or hip is not available and VLLM_TARGET_DEVICE is not set,
54
55
    # fallback to cpu
    VLLM_TARGET_DEVICE = "cpu"
56

57
MAIN_CUDA_VERSION = "12.4"
58

bnellnm's avatar
bnellnm committed
59
60
61
62
63
64
65
66
67
68
69
70
71

def is_sccache_available() -> bool:
    return which("sccache") is not None


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


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


72
73
74
75
76
77
78
79
80
81
82
83
def is_url_available(url: str) -> bool:
    from urllib.request import urlopen

    status = None
    try:
        with urlopen(url) as f:
            status = f.status
    except Exception:
        return False
    return status == 200


bnellnm's avatar
bnellnm committed
84
85
86
class CMakeExtension(Extension):

    def __init__(self, name: str, cmake_lists_dir: str = '.', **kwa) -> None:
87
        super().__init__(name, sources=[], py_limited_api=True, **kwa)
bnellnm's avatar
bnellnm committed
88
89
90
91
92
        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.
93
    did_config: dict[str, bool] = {}
bnellnm's avatar
bnellnm committed
94
95
96
97
98

    #
    # Determine number of compilation jobs and optionally nvcc compile threads.
    #
    def compute_num_jobs(self):
99
100
        # `num_jobs` is either the value of the MAX_JOBS environment variable
        # (if defined) or the number of CPUs available.
101
        num_jobs = envs.MAX_JOBS
102
103
        if num_jobs is not None:
            num_jobs = int(num_jobs)
104
            logger.info("Using MAX_JOBS=%d as the number of jobs.", num_jobs)
105
106
107
108
109
110
111
        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
112

113
        nvcc_threads = None
114
115
116
117
118
        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.
119
            nvcc_threads = envs.NVCC_THREADS
120
121
            if nvcc_threads is not None:
                nvcc_threads = int(nvcc_threads)
122
123
124
                logger.info(
                    "Using NVCC_THREADS=%d as the number of nvcc threads.",
                    nvcc_threads)
125
126
127
            else:
                nvcc_threads = 1
            num_jobs = max(1, num_jobs // nvcc_threads)
bnellnm's avatar
bnellnm committed
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144

        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"
145
        cfg = envs.CMAKE_BUILD_TYPE or default_cfg
bnellnm's avatar
bnellnm committed
146
147
148

        cmake_args = [
            '-DCMAKE_BUILD_TYPE={}'.format(cfg),
149
            '-DVLLM_TARGET_DEVICE={}'.format(VLLM_TARGET_DEVICE),
bnellnm's avatar
bnellnm committed
150
151
        ]

152
        verbose = envs.VERBOSE
bnellnm's avatar
bnellnm committed
153
154
155
156
157
        if verbose:
            cmake_args += ['-DCMAKE_VERBOSE_MAKEFILE=ON']

        if is_sccache_available():
            cmake_args += [
158
                '-DCMAKE_C_COMPILER_LAUNCHER=sccache',
bnellnm's avatar
bnellnm committed
159
160
                '-DCMAKE_CXX_COMPILER_LAUNCHER=sccache',
                '-DCMAKE_CUDA_COMPILER_LAUNCHER=sccache',
161
                '-DCMAKE_HIP_COMPILER_LAUNCHER=sccache',
bnellnm's avatar
bnellnm committed
162
163
164
            ]
        elif is_ccache_available():
            cmake_args += [
165
                '-DCMAKE_C_COMPILER_LAUNCHER=ccache',
bnellnm's avatar
bnellnm committed
166
167
                '-DCMAKE_CXX_COMPILER_LAUNCHER=ccache',
                '-DCMAKE_CUDA_COMPILER_LAUNCHER=ccache',
168
                '-DCMAKE_HIP_COMPILER_LAUNCHER=ccache',
bnellnm's avatar
bnellnm committed
169
170
171
172
173
174
            ]

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

175
176
177
178
        # 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))]

179
180
181
182
183
184
185
186
        # 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)]

bnellnm's avatar
bnellnm committed
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
        #
        # 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 = []
204
205
206
        # Make sure we use the nvcc from CUDA_HOME
        if _is_cuda():
            cmake_args += [f'-DCMAKE_CUDA_COMPILER={CUDA_HOME}/bin/nvcc']
bnellnm's avatar
bnellnm committed
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
        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)

222
        targets = []
223
224
225
226

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

bnellnm's avatar
bnellnm committed
227
228
229
        # Build all the extensions
        for ext in self.extensions:
            self.configure(ext)
230
            targets.append(target_name(ext.name))
bnellnm's avatar
bnellnm committed
231

232
        num_jobs, _ = self.compute_num_jobs()
bnellnm's avatar
bnellnm committed
233

234
235
236
237
238
239
        build_args = [
            "--build",
            ".",
            f"-j={num_jobs}",
            *[f"--target={name}" for name in targets],
        ]
bnellnm's avatar
bnellnm committed
240

241
        subprocess.check_call(["cmake", *build_args], cwd=self.build_temp)
242

243
244
245
246
247
248
249
250
251
252
253
        # 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.
254
255
256
            # We assume only the final component of extension prefix is added by
            # CMake, this is currently true for current extensions but may not
            # always be the case.
257
            prefix = outdir
258
            if '.' in ext.name:
259
260
261
262
263
264
265
266
267
                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)

268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
    def run(self):
        # First, run the standard build_ext command to compile the extensions
        super().run()

        # 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"))
        for file in files:
            dst_file = os.path.join("vllm/vllm_flash_attn",
                                    os.path.basename(file))
            print(f"Copying {file} to {dst_file}")
            self.copy_file(file, dst_file)

283

284
285
286
class repackage_wheel(build_ext):
    """Extracts libraries and other files from an existing wheel."""

287
    def get_base_commit_in_main_branch(self) -> str:
288
289
290
        # Force to use the nightly wheel. This is mainly used for CI testing.
        if envs.VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL:
            return "nightly"
291
292

        try:
293
294
295
296
297
298
299
            # Get the latest commit hash of the upstream main branch.
            resp_json = subprocess.check_output([
                "curl", "-s",
                "https://api.github.com/repos/vllm-project/vllm/commits/main"
            ]).decode("utf-8")
            upstream_main_commit = json.loads(resp_json)["sha"]

300
301
302
303
304
305
306
307
308
309
310
311
312
            # 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"
                ])
313
314
315

            # Then get the commit hash of the current branch that is the same as
            # the upstream main commit.
316
317
318
            current_branch = subprocess.check_output(
                ["git", "branch", "--show-current"]).decode("utf-8").strip()

319
320
321
            base_commit = subprocess.check_output([
                "git", "merge-base", f"{upstream_main_commit}", current_branch
            ]).decode("utf-8").strip()
322
            return base_commit
323
324
        except ValueError as err:
            raise ValueError(err) from None
325
326
327
328
329
330
        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"
331

332
    def run(self) -> None:
333
334
335
        assert _is_cuda(
        ), "VLLM_USE_PRECOMPILED is only supported for CUDA builds"

336
337
338
339
        wheel_location = os.getenv("VLLM_PRECOMPILED_WHEEL_LOCATION", None)
        if wheel_location is None:
            base_commit = self.get_base_commit_in_main_branch()
            wheel_location = f"https://wheels.vllm.ai/{base_commit}/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl"
340
341
342
343
            # Fallback to nightly wheel if latest commit wheel is unavailable,
            # in this rare case, the nightly release CI hasn't finished on main.
            if not is_url_available(wheel_location):
                wheel_location = "https://wheels.vllm.ai/nightly/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl"
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
        import zipfile

        if os.path.isfile(wheel_location):
            wheel_path = wheel_location
            print(f"Using existing wheel={wheel_path}")
        else:
            # Download the wheel from a given URL, assume
            # the filename is the last part of the URL
            wheel_filename = wheel_location.split("/")[-1]

            import tempfile

            # create a temporary directory to store the wheel
            temp_dir = tempfile.mkdtemp(prefix="vllm-wheels")
            wheel_path = os.path.join(temp_dir, wheel_filename)

            print(f"Downloading wheel from {wheel_location} to {wheel_path}")

            from urllib.request import urlretrieve

            try:
                urlretrieve(wheel_location, filename=wheel_path)
            except Exception as e:
                from setuptools.errors import SetupError

                raise SetupError(
                    f"Failed to get vLLM wheel from {wheel_location}") from e

        with zipfile.ZipFile(wheel_path) as wheel:
            files_to_copy = [
                "vllm/_C.abi3.so",
                "vllm/_moe_C.abi3.so",
377
                "vllm/_flashmla_C.abi3.so",
378
379
                "vllm/vllm_flash_attn/_vllm_fa2_C.abi3.so",
                "vllm/vllm_flash_attn/_vllm_fa3_C.abi3.so",
380
                "vllm/vllm_flash_attn/flash_attn_interface.py",
381
                "vllm/cumem_allocator.abi3.so",
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
                # "vllm/_version.py", # not available in nightly wheels yet
            ]
            file_members = filter(lambda x: x.filename in files_to_copy,
                                  wheel.filelist)

            for file in file_members:
                print(f"Extracting and including {file.filename} "
                      "from existing wheel")
                package_name = os.path.dirname(file.filename).replace("/", ".")
                file_name = os.path.basename(file.filename)

                if package_name not in package_data:
                    package_data[package_name] = []

                wheel.extract(file)
                if file_name.endswith(".py"):
                    # python files shouldn't be added to package_data
                    continue

                package_data[package_name].append(file_name)


404
def _is_hpu() -> bool:
405
406
407
408
409
410
411
    # if VLLM_TARGET_DEVICE env var was set explicitly, skip HPU autodetection
    if os.getenv("VLLM_TARGET_DEVICE", None) == VLLM_TARGET_DEVICE:
        return VLLM_TARGET_DEVICE == "hpu"

    # if VLLM_TARGET_DEVICE was not set explicitly, check if hl-smi succeeds,
    # and if it doesn't, check if habanalabs driver is loaded
    is_hpu_available = False
412
    try:
413
414
        out = subprocess.run(["hl-smi"], capture_output=True, check=True)
        is_hpu_available = out.returncode == 0
415
    except (FileNotFoundError, PermissionError, subprocess.CalledProcessError):
416
        if sys.platform.startswith("linux"):
417
418
419
420
421
422
            try:
                output = subprocess.check_output(
                    'lsmod | grep habanalabs | wc -l', shell=True)
                is_hpu_available = int(output) > 0
            except (ValueError, FileNotFoundError, PermissionError,
                    subprocess.CalledProcessError):
423
424
                pass
    return is_hpu_available
425
426


427
428
429
430
def _no_device() -> bool:
    return VLLM_TARGET_DEVICE == "empty"


431
def _is_cuda() -> bool:
432
433
    has_cuda = torch.version.cuda is not None
    return (VLLM_TARGET_DEVICE == "cuda" and has_cuda
434
            and not (_is_neuron() or _is_tpu() or _is_hpu()))
435
436


437
def _is_hip() -> bool:
438
439
    return (VLLM_TARGET_DEVICE == "cuda"
            or VLLM_TARGET_DEVICE == "rocm") and torch.version.hip is not None
440
441


442
def _is_neuron() -> bool:
443
    return VLLM_TARGET_DEVICE == "neuron"
444
445


446
447
448
449
def _is_tpu() -> bool:
    return VLLM_TARGET_DEVICE == "tpu"


450
451
452
453
def _is_cpu() -> bool:
    return VLLM_TARGET_DEVICE == "cpu"


454
455
456
457
def _is_xpu() -> bool:
    return VLLM_TARGET_DEVICE == "xpu"


458
459
460
461
def _build_custom_ops() -> bool:
    return _is_cuda() or _is_hip() or _is_cpu()


462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
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()
481

482
483
        if (get_rocm_core_version(ctypes.byref(major), ctypes.byref(minor),
                                  ctypes.byref(patch)) == 0):
484
            return f"{major.value}.{minor.value}.{patch.value}"
485
        return None
486
    except Exception:
487
        return None
Woosuk Kwon's avatar
Woosuk Kwon committed
488

489

490
491
492
def get_neuronxcc_version():
    import sysconfig
    site_dir = sysconfig.get_paths()["purelib"]
493
494
    version_file = os.path.join(site_dir, "neuronxcc", "version",
                                "__init__.py")
495
496

    # Check if the command was executed successfully
497
    with open(version_file) as fp:
498
499
500
501
502
503
504
505
        content = fp.read()

    # Extract the version using a regular expression
    match = re.search(r"__version__ = '(\S+)'", content)
    if match:
        # Return the version string
        return match.group(1)
    else:
506
        raise RuntimeError("Could not find Neuron version in the output")
507
508


bnellnm's avatar
bnellnm committed
509
def get_nvcc_cuda_version() -> Version:
510
511
512
513
    """Get the CUDA version from nvcc.

    Adapted from https://github.com/NVIDIA/apex/blob/8b7a1ff183741dd8f9b87e7bafd04cfde99cea28/setup.py
    """
514
    assert CUDA_HOME is not None, "CUDA_HOME is not set"
bnellnm's avatar
bnellnm committed
515
    nvcc_output = subprocess.check_output([CUDA_HOME + "/bin/nvcc", "-V"],
516
517
518
519
520
521
522
                                          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


523
524
525
526
527
528
529
530
def get_gaudi_sw_version():
    """
    Returns the driver version.
    """
    # Enable console printing for `hl-smi` check
    output = subprocess.run("hl-smi",
                            shell=True,
                            text=True,
531
                            capture_output=True,
532
533
534
535
536
537
538
                            env={"ENABLE_CONSOLE": "true"})
    if output.returncode == 0 and output.stdout:
        return output.stdout.split("\n")[2].replace(
            " ", "").split(":")[1][:-1].split("-")[0]
    return "0.0.0"  # when hl-smi is not available


539
def get_vllm_version() -> str:
540
    version = get_version(write_to="vllm/_version.py")
541
    sep = "+" if "+" not in version else "."  # dev versions might contain +
542

543
    if _no_device():
544
        if envs.VLLM_TARGET_DEVICE == "empty":
545
            version += f"{sep}empty"
546
    elif _is_cuda():
547
        if envs.VLLM_USE_PRECOMPILED:
548
            version += f"{sep}precompiled"
549
550
551
552
553
554
555
        else:
            cuda_version = str(get_nvcc_cuda_version())
            if cuda_version != 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}"
556
    elif _is_hip():
557
558
559
560
        # Get the Rocm Version
        rocm_version = get_rocm_version() or torch.version.hip
        if rocm_version and rocm_version != MAIN_CUDA_VERSION:
            version += f"{sep}rocm{rocm_version.replace('.', '')[:3]}"
561
562
    elif _is_neuron():
        # Get the Neuron version
bnellnm's avatar
bnellnm committed
563
        neuron_version = str(get_neuronxcc_version())
564
565
        if neuron_version != MAIN_CUDA_VERSION:
            neuron_version_str = neuron_version.replace(".", "")[:3]
566
            version += f"{sep}neuron{neuron_version_str}"
567
568
569
570
571
572
    elif _is_hpu():
        # Get the Intel Gaudi Software Suite version
        gaudi_sw_version = str(get_gaudi_sw_version())
        if gaudi_sw_version != MAIN_CUDA_VERSION:
            gaudi_sw_version = gaudi_sw_version.replace(".", "")[:3]
            version += f"{sep}gaudi{gaudi_sw_version}"
573
    elif _is_tpu():
574
        version += f"{sep}tpu"
575
    elif _is_cpu():
576
577
        if envs.VLLM_TARGET_DEVICE == "cpu":
            version += f"{sep}cpu"
578
    elif _is_xpu():
579
        version += f"{sep}xpu"
580
    else:
581
        raise RuntimeError("Unknown runtime environment")
582

583
584
585
    return version


586
def get_requirements() -> list[str]:
587
    """Get Python package dependencies from requirements.txt."""
588
    requirements_dir = ROOT_DIR / "requirements"
589

590
    def _read_requirements(filename: str) -> list[str]:
591
        with open(requirements_dir / filename) as f:
592
            requirements = f.read().strip().split("\n")
593
594
595
596
        resolved_requirements = []
        for line in requirements:
            if line.startswith("-r "):
                resolved_requirements += _read_requirements(line.split()[1])
597
598
            elif not line.startswith("--") and not line.startswith(
                    "#") and line.strip() != "":
599
600
601
                resolved_requirements.append(line)
        return resolved_requirements

602
    if _no_device():
603
        requirements = _read_requirements("common.txt")
604
    elif _is_cuda():
605
        requirements = _read_requirements("cuda.txt")
606
        cuda_major, cuda_minor = torch.version.cuda.split(".")
607
608
        modified_requirements = []
        for req in requirements:
609
610
            if ("vllm-flash-attn" in req and cuda_major != "12"):
                # vllm-flash-attn is built only for CUDA 12.x.
611
612
613
                # Skip for other versions.
                continue
            modified_requirements.append(req)
614
        requirements = modified_requirements
615
    elif _is_hip():
616
        requirements = _read_requirements("rocm.txt")
617
    elif _is_neuron():
618
        requirements = _read_requirements("neuron.txt")
619
    elif _is_hpu():
620
        requirements = _read_requirements("hpu.txt")
621
    elif _is_tpu():
622
        requirements = _read_requirements("tpu.txt")
623
    elif _is_cpu():
624
        requirements = _read_requirements("cpu.txt")
625
    elif _is_xpu():
626
        requirements = _read_requirements("xpu.txt")
627
628
    else:
        raise ValueError(
629
            "Unsupported platform, please use CUDA, ROCm, Neuron, HPU, "
630
            "or CPU.")
631
632
633
    return requirements


bnellnm's avatar
bnellnm committed
634
635
ext_modules = []

636
if _is_cuda() or _is_hip():
bnellnm's avatar
bnellnm committed
637
638
    ext_modules.append(CMakeExtension(name="vllm._moe_C"))

639
640
641
if _is_hip():
    ext_modules.append(CMakeExtension(name="vllm._rocm_C"))

642
if _is_cuda():
643
    ext_modules.append(CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa2_C"))
644
645
    if envs.VLLM_USE_PRECOMPILED or get_nvcc_cuda_version() >= Version("12.3"):
        # FA3 requires CUDA 12.3 or later
646
647
        ext_modules.append(
            CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa3_C"))
648
649
650
651
        # 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))
652
    ext_modules.append(CMakeExtension(name="vllm.cumem_allocator"))
653

654
if _build_custom_ops():
bnellnm's avatar
bnellnm committed
655
656
    ext_modules.append(CMakeExtension(name="vllm._C"))

657
package_data = {
658
659
660
661
662
    "vllm": [
        "py.typed",
        "model_executor/layers/fused_moe/configs/*.json",
        "model_executor/layers/quantization/utils/configs/*.json",
    ]
663
}
Simon Mo's avatar
Simon Mo committed
664

665
666
667
if _no_device():
    ext_modules = []

668
669
670
671
672
673
674
675
if not ext_modules:
    cmdclass = {}
else:
    cmdclass = {
        "build_ext":
        repackage_wheel if envs.VLLM_USE_PRECOMPILED else cmake_build_ext
    }

bnellnm's avatar
bnellnm committed
676
setup(
677
    # static metadata should rather go in pyproject.toml
678
    version=get_vllm_version(),
Woosuk Kwon's avatar
Woosuk Kwon committed
679
    ext_modules=ext_modules,
680
    install_requires=get_requirements(),
681
    extras_require={
682
        "tensorizer": ["tensorizer>=2.9.0"],
683
        "fastsafetensors": ["fastsafetensors >= 0.1.10"],
684
        "runai": ["runai-model-streamer", "runai-model-streamer-s3", "boto3"],
685
        "audio": ["librosa", "soundfile"],  # Required for audio processing
686
        "video": []  # Kept for backwards compatibility
687
    },
688
    cmdclass=cmdclass,
Simon Mo's avatar
Simon Mo committed
689
    package_data=package_data,
Woosuk Kwon's avatar
Woosuk Kwon committed
690
)