setup.py 18.7 KB
Newer Older
1
import contextlib
2
3
4
import io
import os
import re
5
import subprocess
6
import warnings
7
8
from pathlib import Path
from typing import List, Set
9

10
from packaging.version import parse, Version
Woosuk Kwon's avatar
Woosuk Kwon committed
11
import setuptools
Woosuk Kwon's avatar
Woosuk Kwon committed
12
import torch
13
import torch.utils.cpp_extension as torch_cpp_ext
14
from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CUDA_HOME, ROCM_HOME
15

zhuwenwen's avatar
zhuwenwen committed
16
17
18
19
from typing import Optional, Union
import subprocess
from pathlib import Path

20
ROOT_DIR = os.path.dirname(__file__)
21

22
23
24
25
26
# If you are developing the C++ backend of vLLM, consider building vLLM with
# `python setup.py develop` since it will give you incremental builds.
# The downside is that this method is deprecated, see
# https://github.com/pypa/setuptools/issues/917

27
28
MAIN_CUDA_VERSION = "12.1"

29
# Supported NVIDIA GPU architectures.
30
NVIDIA_SUPPORTED_ARCHS = {"7.0", "7.5", "8.0", "8.6", "8.9", "9.0"}
zhuwenwen's avatar
zhuwenwen committed
31
ROCM_SUPPORTED_ARCHS = {"gfx908", "gfx90a", "gfx906", "gfx926", "gfx942", "gfx1100"}
32
33
34
35
36
37
38
# SUPPORTED_ARCHS = NVIDIA_SUPPORTED_ARCHS.union(ROCM_SUPPORTED_ARCHS)


def _is_hip() -> bool:
    return torch.version.hip is not None


39
40
41
42
def _is_neuron() -> bool:
    torch_neuronx_installed = True
    try:
        subprocess.run(["neuron-ls"], capture_output=True, check=True)
43
    except FileNotFoundError:
44
45
46
47
        torch_neuronx_installed = False
    return torch_neuronx_installed


48
def _is_cuda() -> bool:
49
    return (torch.version.cuda is not None) and not _is_neuron()
50

51

52
# Compiler flags.
53
CXX_FLAGS = ["-g", "-O2", "-std=c++17"]
54
# TODO(woosuk): Should we use -O3?
zhuwenwen's avatar
zhuwenwen committed
55
NVCC_FLAGS = ["-O2", "-std=c++17","--gpu-max-threads-per-block=1024"]
Woosuk Kwon's avatar
Woosuk Kwon committed
56

57
58
59
60
61
62
if _is_hip():
    if ROCM_HOME is None:
        raise RuntimeError(
            "Cannot find ROCM_HOME. ROCm must be available to build the package."
        )
    NVCC_FLAGS += ["-DUSE_ROCM"]
63
64
    NVCC_FLAGS += ["-U__HIP_NO_HALF_CONVERSIONS__"]
    NVCC_FLAGS += ["-U__HIP_NO_HALF_OPERATORS__"]
65
66
67
68
69

if _is_cuda() and CUDA_HOME is None:
    raise RuntimeError(
        "Cannot find CUDA_HOME. CUDA must be available to build the package.")

70
71
72
ABI = 1 if torch._C._GLIBCXX_USE_CXX11_ABI else 0
CXX_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
NVCC_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
73

74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94

def get_hipcc_rocm_version():
    # Run the hipcc --version command
    result = subprocess.run(['hipcc', '--version'],
                            stdout=subprocess.PIPE,
                            stderr=subprocess.STDOUT,
                            text=True)

    # Check if the command was executed successfully
    if result.returncode != 0:
        print("Error running 'hipcc --version'")
        return None

    # Extract the version using a regular expression
    match = re.search(r'HIP version: (\S+)', result.stdout)
    if match:
        # Return the version string
        return match.group(1)
    else:
        print("Could not find HIP version in the output")
        return None
Woosuk Kwon's avatar
Woosuk Kwon committed
95

96

97
98
99
100
101
def glob(pattern: str):
    root = Path(__name__).parent
    return [str(p) for p in root.glob(pattern)]


102
103
104
def get_neuronxcc_version():
    import sysconfig
    site_dir = sysconfig.get_paths()["purelib"]
105
106
    version_file = os.path.join(site_dir, "neuronxcc", "version",
                                "__init__.py")
107
108
109
110
111
112
113
114
115
116
117
118
119
120

    # Check if the command was executed successfully
    with open(version_file, "rt") as fp:
        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:
        raise RuntimeError("Could not find HIP version in the output")


121
122
123
124
125
126
127
128
129
130
131
132
133
def get_nvcc_cuda_version(cuda_dir: str) -> Version:
    """Get the CUDA version from nvcc.

    Adapted from https://github.com/NVIDIA/apex/blob/8b7a1ff183741dd8f9b87e7bafd04cfde99cea28/setup.py
    """
    nvcc_output = subprocess.check_output([cuda_dir + "/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


134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
def get_pytorch_rocm_arch() -> Set[str]:
    """Get the cross section of Pytorch,and vllm supported gfx arches

    ROCM can get the supported gfx architectures in one of two ways
    Either through the PYTORCH_ROCM_ARCH env var, or output from
    rocm_agent_enumerator.

    In either case we can generate a list of supported arch's and
    cross reference with VLLM's own ROCM_SUPPORTED_ARCHs.
    """
    env_arch_list = os.environ.get("PYTORCH_ROCM_ARCH", None)

    # If we don't have PYTORCH_ROCM_ARCH specified pull the list from rocm_agent_enumerator
    if env_arch_list is None:
        command = "rocm_agent_enumerator"
        env_arch_list = subprocess.check_output([command]).decode('utf-8')\
                        .strip().replace("\n", ";")
        arch_source_str = "rocm_agent_enumerator"
    else:
        arch_source_str = "PYTORCH_ROCM_ARCH env variable"

    # List are separated by ; or space.
    pytorch_rocm_arch = set(env_arch_list.replace(" ", ";").split(";"))

    # Filter out the invalid architectures and print a warning.
    arch_list = pytorch_rocm_arch.intersection(ROCM_SUPPORTED_ARCHS)

    # If none of the specified architectures are valid, raise an error.
    if not arch_list:
        raise RuntimeError(
            f"None of the ROCM architectures in {arch_source_str} "
            f"({env_arch_list}) is supported. "
            f"Supported ROCM architectures are: {ROCM_SUPPORTED_ARCHS}.")
    invalid_arch_list = pytorch_rocm_arch - ROCM_SUPPORTED_ARCHS
    if invalid_arch_list:
        warnings.warn(
            f"Unsupported ROCM architectures ({invalid_arch_list}) are "
            f"excluded from the {arch_source_str} output "
            f"({env_arch_list}). Supported ROCM architectures are: "
            f"{ROCM_SUPPORTED_ARCHS}.",
            stacklevel=2)
    return arch_list


178
179
180
181
182
183
184
def get_torch_arch_list() -> Set[str]:
    # TORCH_CUDA_ARCH_LIST can have one or more architectures,
    # e.g. "8.0" or "7.5,8.0,8.6+PTX". Here, the "8.6+PTX" option asks the
    # compiler to additionally include PTX code that can be runtime-compiled
    # and executed on the 8.6 or newer architectures. While the PTX code will
    # not give the best performance on the newer architectures, it provides
    # forward compatibility.
185
186
    env_arch_list = os.environ.get("TORCH_CUDA_ARCH_LIST", None)
    if env_arch_list is None:
187
188
189
        return set()

    # List are separated by ; or space.
190
191
192
193
194
    torch_arch_list = set(env_arch_list.replace(" ", ";").split(";"))
    if not torch_arch_list:
        return set()

    # Filter out the invalid architectures and print a warning.
195
196
197
    valid_archs = NVIDIA_SUPPORTED_ARCHS.union(
        {s + "+PTX"
         for s in NVIDIA_SUPPORTED_ARCHS})
198
199
200
201
    arch_list = torch_arch_list.intersection(valid_archs)
    # If none of the specified architectures are valid, raise an error.
    if not arch_list:
        raise RuntimeError(
202
            "None of the CUDA architectures in `TORCH_CUDA_ARCH_LIST` env "
203
            f"variable ({env_arch_list}) is supported. "
204
            f"Supported CUDA architectures are: {valid_archs}.")
205
206
207
    invalid_arch_list = torch_arch_list - valid_archs
    if invalid_arch_list:
        warnings.warn(
208
            f"Unsupported CUDA architectures ({invalid_arch_list}) are "
209
            "excluded from the `TORCH_CUDA_ARCH_LIST` env variable "
210
            f"({env_arch_list}). Supported CUDA architectures are: "
211
212
            f"{valid_archs}.",
            stacklevel=2)
213
    return arch_list
214
215


216
217
218
219
220
221
222
if _is_hip():
    rocm_arches = get_pytorch_rocm_arch()
    NVCC_FLAGS += ["--offload-arch=" + arch for arch in rocm_arches]
else:
    # First, check the TORCH_CUDA_ARCH_LIST environment variable.
    compute_capabilities = get_torch_arch_list()

223
if _is_cuda() and not compute_capabilities:
224
225
226
227
228
229
230
231
232
    # If TORCH_CUDA_ARCH_LIST is not defined or empty, target all available
    # GPUs on the current machine.
    device_count = torch.cuda.device_count()
    for i in range(device_count):
        major, minor = torch.cuda.get_device_capability(i)
        if major < 7:
            raise RuntimeError(
                "GPUs with compute capability below 7.0 are not supported.")
        compute_capabilities.add(f"{major}.{minor}")
233

234
235
ext_modules = []

236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
if _is_cuda():
    nvcc_cuda_version = get_nvcc_cuda_version(CUDA_HOME)
    if not compute_capabilities:
        # If no GPU is specified nor available, add all supported architectures
        # based on the NVCC CUDA version.
        compute_capabilities = NVIDIA_SUPPORTED_ARCHS.copy()
        if nvcc_cuda_version < Version("11.1"):
            compute_capabilities.remove("8.6")
        if nvcc_cuda_version < Version("11.8"):
            compute_capabilities.remove("8.9")
            compute_capabilities.remove("9.0")
    # Validate the NVCC CUDA version.
    if nvcc_cuda_version < Version("11.0"):
        raise RuntimeError(
            "CUDA 11.0 or higher is required to build the package.")
    if (nvcc_cuda_version < Version("11.1")
            and any(cc.startswith("8.6") for cc in compute_capabilities)):
        raise RuntimeError(
            "CUDA 11.1 or higher is required for compute capability 8.6.")
255
    if nvcc_cuda_version < Version("11.8"):
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
        if any(cc.startswith("8.9") for cc in compute_capabilities):
            # CUDA 11.8 is required to generate the code targeting compute capability 8.9.
            # However, GPUs with compute capability 8.9 can also run the code generated by
            # the previous versions of CUDA 11 and targeting compute capability 8.0.
            # Therefore, if CUDA 11.8 is not available, we target compute capability 8.0
            # instead of 8.9.
            warnings.warn(
                "CUDA 11.8 or higher is required for compute capability 8.9. "
                "Targeting compute capability 8.0 instead.",
                stacklevel=2)
            compute_capabilities = set(cc for cc in compute_capabilities
                                       if not cc.startswith("8.9"))
            compute_capabilities.add("8.0+PTX")
        if any(cc.startswith("9.0") for cc in compute_capabilities):
            raise RuntimeError(
                "CUDA 11.8 or higher is required for compute capability 9.0.")

273
274
    NVCC_FLAGS_PUNICA = NVCC_FLAGS.copy()

275
276
277
278
279
280
281
282
    # Add target compute capabilities to NVCC flags.
    for capability in compute_capabilities:
        num = capability[0] + capability[2]
        NVCC_FLAGS += ["-gencode", f"arch=compute_{num},code=sm_{num}"]
        if capability.endswith("+PTX"):
            NVCC_FLAGS += [
                "-gencode", f"arch=compute_{num},code=compute_{num}"
            ]
283
284
285
286
287
288
289
290
        if int(capability[0]) >= 8:
            NVCC_FLAGS_PUNICA += [
                "-gencode", f"arch=compute_{num},code=sm_{num}"
            ]
            if capability.endswith("+PTX"):
                NVCC_FLAGS_PUNICA += [
                    "-gencode", f"arch=compute_{num},code=compute_{num}"
                ]
291
292
293
294
295
296
297

    # Use NVCC threads to parallelize the build.
    if nvcc_cuda_version >= Version("11.2"):
        nvcc_threads = int(os.getenv("NVCC_THREADS", 8))
        num_threads = min(os.cpu_count(), nvcc_threads)
        NVCC_FLAGS += ["--threads", str(num_threads)]

298
299
300
    if nvcc_cuda_version >= Version("11.8"):
        NVCC_FLAGS += ["-DENABLE_FP8_E5M2"]

301
302
303
304
305
306
307
308
309
310
311
312
    # changes for punica kernels
    NVCC_FLAGS += torch_cpp_ext.COMMON_NVCC_FLAGS
    REMOVE_NVCC_FLAGS = [
        '-D__CUDA_NO_HALF_OPERATORS__',
        '-D__CUDA_NO_HALF_CONVERSIONS__',
        '-D__CUDA_NO_BFLOAT16_CONVERSIONS__',
        '-D__CUDA_NO_HALF2_OPERATORS__',
    ]
    for flag in REMOVE_NVCC_FLAGS:
        with contextlib.suppress(ValueError):
            torch_cpp_ext.COMMON_NVCC_FLAGS.remove(flag)

313
    install_punica = bool(int(os.getenv("VLLM_INSTALL_PUNICA_KERNELS", "0")))
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
    device_count = torch.cuda.device_count()
    for i in range(device_count):
        major, minor = torch.cuda.get_device_capability(i)
        if major < 8:
            install_punica = False
            break
    if install_punica:
        ext_modules.append(
            CUDAExtension(
                name="vllm._punica_C",
                sources=["csrc/punica/punica_ops.cc"] +
                glob("csrc/punica/bgmv/*.cu"),
                extra_compile_args={
                    "cxx": CXX_FLAGS,
                    "nvcc": NVCC_FLAGS_PUNICA,
                },
            ))
331
332
elif _is_neuron():
    neuronxcc_version = get_neuronxcc_version()
333

334
335
336
337
338
339
340
vllm_extension_sources = [
    "csrc/cache_kernels.cu",
    "csrc/attention/attention_kernels.cu",
    "csrc/pos_encoding_kernels.cu",
    "csrc/activation_kernels.cu",
    "csrc/layernorm_kernels.cu",
    "csrc/quantization/squeezellm/quant_cuda_kernel.cu",
kliuae's avatar
kliuae committed
341
    "csrc/quantization/gptq/q_gemm.cu",
342
    "csrc/cuda_utils_kernels.cu",
343
    "csrc/moe_align_block_size_kernels.cu",
344
345
346
347
348
    "csrc/pybind.cpp",
]

if _is_cuda():
    vllm_extension_sources.append("csrc/quantization/awq/gemm_kernels.cu")
349
    vllm_extension_sources.append("csrc/custom_all_reduce.cu")
Woosuk Kwon's avatar
Woosuk Kwon committed
350

351
352
353
354
355
356
357
358
359
360
361
    # Add MoE kernels.
    ext_modules.append(
        CUDAExtension(
            name="vllm._moe_C",
            sources=glob("csrc/moe/*.cu") + glob("csrc/moe/*.cpp"),
            extra_compile_args={
                "cxx": CXX_FLAGS,
                "nvcc": NVCC_FLAGS,
            },
        ))

362
363
364
365
366
367
368
369
if not _is_neuron():
    vllm_extension = CUDAExtension(
        name="vllm._C",
        sources=vllm_extension_sources,
        extra_compile_args={
            "cxx": CXX_FLAGS,
            "nvcc": NVCC_FLAGS,
        },
370
        libraries=["cuda"] if _is_cuda() else [],
371
372
    )
    ext_modules.append(vllm_extension)
373

374

375
376
377
378
def get_path(*filepath) -> str:
    return os.path.join(ROOT_DIR, *filepath)


379
def find_version(filepath: str) -> str:
380
381
382
383
384
    """Extract version information from the given filepath.

    Adapted from https://github.com/ray-project/ray/blob/0b190ee1160eeca9796bc091e07eaebf4c85b511/python/setup.py
    """
    with open(filepath) as fp:
Woosuk Kwon's avatar
Woosuk Kwon committed
385
386
        version_match = re.search(r"^__version__ = ['\"]([^'\"]*)['\"]",
                                  fp.read(), re.M)
387
388
389
390
391
        if version_match:
            return version_match.group(1)
        raise RuntimeError("Unable to find version string.")


zhuwenwen's avatar
zhuwenwen committed
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
432
def get_abi():
    try:
        command = "echo '#include <string>' | gcc -x c++ -E -dM - | fgrep _GLIBCXX_USE_CXX11_ABI" 
        result = subprocess.run(command, shell=True, capture_output=True, text=True) 
        output = result.stdout.strip() 
        abi = "abi" + output.split(" ")[-1]
        return abi
    except Exception:
        return 'abiUnknown'


def get_sha(root: Union[str, Path]) -> str:
    try:
        return subprocess.check_output(['git', 'rev-parse', 'HEAD'], cwd=root).decode('ascii').strip()
    except Exception:
        return 'Unknown'

def get_version_add(sha: Optional[str] = None) -> str:
    vllm_root = os.path.dirname(os.path.abspath(__file__))
    add_version_path = os.path.join(os.path.join(vllm_root, "vllm"), "version.py")
    if sha != 'Unknown':
        if sha is None:
            sha = get_sha(vllm_root)
        version = 'git' + sha[:7]

    # abi version
    version += "." + get_abi()

    # dtk version
    if os.getenv("ROCM_PATH"):
        rocm_path = os.getenv('ROCM_PATH', "")
        rocm_version_path = os.path.join(rocm_path, '.info', "rocm_version")
        with open(rocm_version_path, 'r',encoding='utf-8') as file:
            lines = file.readlines()
        rocm_version=lines[0][:-2].replace(".", "")
        version += ".dtk" + rocm_version

    # torch version
    version += ".torch" + torch.__version__[:3]

    with open(add_version_path, encoding="utf-8",mode="w") as file:
zhuwenwen's avatar
zhuwenwen committed
433
434
        file.write("__version__='0.3.2'\n")
        file.write("__dcu_version__='0.3.2+{}'\n".format(version))
zhuwenwen's avatar
zhuwenwen committed
435
436
437
438
439
440
441
442
443
444
445
    file.close()
    
    
def get_version():
    get_version_add()
    version_file = 'vllm/version.py'
    with open(version_file, encoding='utf-8') as f:
        exec(compile(f.read(), version_file, 'exec'))
    return locals()['__dcu_version__']


446
447
def get_vllm_version() -> str:
    version = find_version(get_path("vllm", "__init__.py"))
448
449
450
451
452
453

    if _is_hip():
        # Get the HIP version
        hipcc_version = get_hipcc_rocm_version()
        if hipcc_version != MAIN_CUDA_VERSION:
            rocm_version_str = hipcc_version.replace(".", "")[:3]
zhuwenwen's avatar
zhuwenwen committed
454
455
        #     version += f"+rocm{rocm_version_str}"
        version = get_version()
456
457
458
459
460
461
    elif _is_neuron():
        # Get the Neuron version
        neuron_version = str(neuronxcc_version)
        if neuron_version != MAIN_CUDA_VERSION:
            neuron_version_str = neuron_version.replace(".", "")[:3]
            version += f"+neuron{neuron_version_str}"
462
463
464
465
466
467
    else:
        cuda_version = str(nvcc_cuda_version)
        if cuda_version != MAIN_CUDA_VERSION:
            cuda_version_str = cuda_version.replace(".", "")[:3]
            version += f"+cu{cuda_version_str}"

468
469
470
    return version


471
def read_readme() -> str:
Stephen Krider's avatar
Stephen Krider committed
472
473
474
475
476
477
    """Read the README file if present."""
    p = get_path("README.md")
    if os.path.isfile(p):
        return io.open(get_path("README.md"), "r", encoding="utf-8").read()
    else:
        return ""
478
479


480
481
def get_requirements() -> List[str]:
    """Get Python package dependencies from requirements.txt."""
482
483
484
    if _is_hip():
        with open(get_path("requirements-rocm.txt")) as f:
            requirements = f.read().strip().split("\n")
485
486
487
    elif _is_neuron():
        with open(get_path("requirements-neuron.txt")) as f:
            requirements = f.read().strip().split("\n")
488
489
490
    else:
        with open(get_path("requirements.txt")) as f:
            requirements = f.read().strip().split("\n")
491
492
493
    return requirements


Simon Mo's avatar
Simon Mo committed
494
495
496
497
498
package_data = {"vllm": ["py.typed"]}
if os.environ.get("VLLM_USE_PRECOMPILED"):
    ext_modules = []
    package_data["vllm"].append("*.so")

Woosuk Kwon's avatar
Woosuk Kwon committed
499
setuptools.setup(
Woosuk Kwon's avatar
Woosuk Kwon committed
500
    name="vllm",
501
    version=get_vllm_version(),
Woosuk Kwon's avatar
Woosuk Kwon committed
502
    author="vLLM Team",
503
    license="Apache 2.0",
Woosuk Kwon's avatar
Woosuk Kwon committed
504
505
    description=("A high-throughput and memory-efficient inference and "
                 "serving engine for LLMs"),
506
507
    long_description=read_readme(),
    long_description_content_type="text/markdown",
508
    url="https://github.com/vllm-project/vllm",
509
    project_urls={
510
511
        "Homepage": "https://github.com/vllm-project/vllm",
        "Documentation": "https://vllm.readthedocs.io/en/latest/",
512
513
514
515
516
    },
    classifiers=[
        "Programming Language :: Python :: 3.8",
        "Programming Language :: Python :: 3.9",
        "Programming Language :: Python :: 3.10",
Woosuk Kwon's avatar
Woosuk Kwon committed
517
        "Programming Language :: Python :: 3.11",
518
519
520
        "License :: OSI Approved :: Apache Software License",
        "Topic :: Scientific/Engineering :: Artificial Intelligence",
    ],
Woosuk Kwon's avatar
Woosuk Kwon committed
521
522
    packages=setuptools.find_packages(exclude=("benchmarks", "csrc", "docs",
                                               "examples", "tests")),
523
524
    python_requires=">=3.8",
    install_requires=get_requirements(),
Woosuk Kwon's avatar
Woosuk Kwon committed
525
    ext_modules=ext_modules,
526
    cmdclass={"build_ext": BuildExtension} if not _is_neuron() else {},
Simon Mo's avatar
Simon Mo committed
527
    package_data=package_data,
Woosuk Kwon's avatar
Woosuk Kwon committed
528
)