setup.py 15.1 KB
Newer Older
1
2
# Copyright (c) 2023, Tri Dao.

Tri Dao's avatar
Tri Dao committed
3
4
5
import sys
import warnings
import os
6
7
import re
import ast
Tri Dao's avatar
Tri Dao committed
8
from pathlib import Path
Tri Dao's avatar
Tri Dao committed
9
from packaging.version import parse, Version
10
import platform
Tri Dao's avatar
Tri Dao committed
11
12
13
14

from setuptools import setup, find_packages
import subprocess

Pierce Freeman's avatar
Pierce Freeman committed
15
16
import urllib.request
import urllib.error
Tri Dao's avatar
Tri Dao committed
17
18
from wheel.bdist_wheel import bdist_wheel as _bdist_wheel

Tri Dao's avatar
Tri Dao committed
19
import torch
20
21
22
23
24
25
from torch.utils.cpp_extension import (
    BuildExtension,
    CppExtension,
    CUDAExtension,
    CUDA_HOME,
)
Tri Dao's avatar
Tri Dao committed
26
27
28
29
30


with open("README.md", "r", encoding="utf-8") as fh:
    long_description = fh.read()

Tri Dao's avatar
Tri Dao committed
31
32
33
34

# ninja build does not work unless include_dirs are abs path
this_dir = os.path.dirname(os.path.abspath(__file__))

35
PACKAGE_NAME = "flash_attn"
Tri Dao's avatar
Tri Dao committed
36

37
38
39
BASE_WHEEL_URL = (
    "https://github.com/Dao-AILab/flash-attention/releases/download/{tag_name}/{wheel_name}"
)
40
41
42
43
44

# FORCE_BUILD: Force a fresh build locally, instead of attempting to find prebuilt wheels
# SKIP_CUDA_BUILD: Intended to allow CI to use a simple `python setup.py sdist` run to copy over raw files, without any cuda compilation
FORCE_BUILD = os.getenv("FLASH_ATTENTION_FORCE_BUILD", "FALSE") == "TRUE"
SKIP_CUDA_BUILD = os.getenv("FLASH_ATTENTION_SKIP_CUDA_BUILD", "FALSE") == "TRUE"
Tri Dao's avatar
Tri Dao committed
45
46
# For CI, we want the option to build with C++11 ABI since the nvcr images use C++11 ABI
FORCE_CXX11_ABI = os.getenv("FLASH_ATTENTION_FORCE_CXX11_ABI", "FALSE") == "TRUE"
47
48


49
50
def get_platform():
    """
51
    Returns the platform name as used in wheel filenames.
52
    """
53
    if sys.platform.startswith("linux"):
54
        return f'linux_{platform.uname().machine}'
55
56
57
58
59
    elif sys.platform == "darwin":
        mac_version = ".".join(platform.mac_ver()[0].split(".")[:2])
        return f"macosx_{mac_version}_x86_64"
    elif sys.platform == "win32":
        return "win_amd64"
60
    else:
61
        raise ValueError("Unsupported platform: {}".format(sys.platform))
62

Tri Dao's avatar
Tri Dao committed
63
64
65
66
67

def get_cuda_bare_metal_version(cuda_dir):
    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
    output = raw_output.split()
    release_idx = output.index("release") + 1
Tri Dao's avatar
Tri Dao committed
68
    bare_metal_version = parse(output[release_idx].split(",")[0])
Tri Dao's avatar
Tri Dao committed
69

Tri Dao's avatar
Tri Dao committed
70
    return raw_output, bare_metal_version
Tri Dao's avatar
Tri Dao committed
71
72


73
def check_if_cuda_home_none(global_option: str) -> None:
Tri Dao's avatar
Tri Dao committed
74
75
    if CUDA_HOME is not None:
        return
76
77
78
    # warn instead of error because user could be downloading prebuilt wheels, so nvcc won't be necessary
    # in that case.
    warnings.warn(
Tri Dao's avatar
Tri Dao committed
79
80
81
82
83
84
85
        f"{global_option} was requested, but nvcc was not found.  Are you sure your environment has nvcc available?  "
        "If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, "
        "only images whose names contain 'devel' will provide nvcc."
    )


def append_nvcc_threads(nvcc_extra_args):
86
87
    nvcc_threads = os.getenv("NVCC_THREADS") or "4"
    return nvcc_extra_args + ["--threads", nvcc_threads]
Tri Dao's avatar
Tri Dao committed
88
89
90
91
92


cmdclass = {}
ext_modules = []

Tri Dao's avatar
Tri Dao committed
93
94
95
96
# We want this even if SKIP_CUDA_BUILD because when we run python setup.py sdist we want the .hpp
# files included in the source distribution, in case the user compiles from source.
subprocess.run(["git", "submodule", "update", "--init", "csrc/cutlass"])

97
98
99
100
101
102
103
104
105
106
107
108
if not SKIP_CUDA_BUILD:
    print("\n\ntorch.__version__  = {}\n\n".format(torch.__version__))
    TORCH_MAJOR = int(torch.__version__.split(".")[0])
    TORCH_MINOR = int(torch.__version__.split(".")[1])

    # Check, if ATen/CUDAGeneratorImpl.h is found, otherwise use ATen/cuda/CUDAGeneratorImpl.h
    # See https://github.com/pytorch/pytorch/pull/70650
    generator_flag = []
    torch_dir = torch.__path__[0]
    if os.path.exists(os.path.join(torch_dir, "include", "ATen", "CUDAGeneratorImpl.h")):
        generator_flag = ["-DOLD_GENERATOR_PATH"]

109
    check_if_cuda_home_none("flash_attn")
110
111
    # Check, if CUDA11 is installed for compute capability 8.0
    cc_flag = []
112
113
114
    if CUDA_HOME is not None:
        _, bare_metal_version = get_cuda_bare_metal_version(CUDA_HOME)
        if bare_metal_version < Version("11.6"):
115
116
117
118
            raise RuntimeError(
                "FlashAttention is only supported on CUDA 11.6 and above.  "
                "Note: make sure nvcc has a supported version by running nvcc -V."
            )
119
120
    # cc_flag.append("-gencode")
    # cc_flag.append("arch=compute_75,code=sm_75")
Tri Dao's avatar
Tri Dao committed
121
    cc_flag.append("-gencode")
122
    cc_flag.append("arch=compute_80,code=sm_80")
Tri Dao's avatar
Tri Dao committed
123
124
125
126
    if CUDA_HOME is not None:
        if bare_metal_version >= Version("11.8"):
            cc_flag.append("-gencode")
            cc_flag.append("arch=compute_90,code=sm_90")
127

Tri Dao's avatar
Tri Dao committed
128
129
130
131
132
    # HACK: The compiler flag -D_GLIBCXX_USE_CXX11_ABI is set to be the same as
    # torch._C._GLIBCXX_USE_CXX11_ABI
    # https://github.com/pytorch/pytorch/blob/8472c24e3b5b60150096486616d98b7bea01500b/torch/utils/cpp_extension.py#L920
    if FORCE_CXX11_ABI:
        torch._C._GLIBCXX_USE_CXX11_ABI = True
133
134
    ext_modules.append(
        CUDAExtension(
135
            name="flash_attn_2_cuda",
136
            sources=[
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
                "csrc/flash_attn/flash_api.cpp",
                "csrc/flash_attn/src/flash_fwd_hdim32_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim32_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim64_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim64_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim96_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim96_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim128_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim128_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim160_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim160_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim192_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim192_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim224_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim224_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim256_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim256_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim32_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim32_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim64_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim64_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim96_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim96_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim128_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim128_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim160_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim160_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim192_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim192_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim224_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim224_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim256_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim256_bf16_sm80.cu",
Tri Dao's avatar
Tri Dao committed
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
                "csrc/flash_attn/src/flash_fwd_split_hdim32_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim32_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim64_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim64_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim96_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim96_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim128_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim128_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim160_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim160_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim192_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim192_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim224_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim224_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim256_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim256_bf16_sm80.cu",
186
187
188
189
190
191
192
193
194
195
196
197
198
199
            ],
            extra_compile_args={
                "cxx": ["-O3", "-std=c++17"] + generator_flag,
                "nvcc": append_nvcc_threads(
                    [
                        "-O3",
                        "-std=c++17",
                        "-U__CUDA_NO_HALF_OPERATORS__",
                        "-U__CUDA_NO_HALF_CONVERSIONS__",
                        "-U__CUDA_NO_HALF2_OPERATORS__",
                        "-U__CUDA_NO_BFLOAT16_CONVERSIONS__",
                        "--expt-relaxed-constexpr",
                        "--expt-extended-lambda",
                        "--use_fast_math",
Tri Dao's avatar
Tri Dao committed
200
                        # "--ptxas-options=-v",
201
                        # "--ptxas-options=-O2",
202
                        # "-lineinfo",
203
204
205
                        # "-DFLASHATTENTION_DISABLE_BACKWARD",
                        # "-DFLASHATTENTION_DISABLE_DROPOUT",
                        # "-DFLASHATTENTION_DISABLE_ALIBI",
Nicolas Patry's avatar
Nicolas Patry committed
206
                        # "-DFLASHATTENTION_DISABLE_SOFTCAP",
207
208
                        # "-DFLASHATTENTION_DISABLE_UNEVEN_K",
                        # "-DFLASHATTENTION_DISABLE_LOCAL",
209
210
211
212
213
214
                    ]
                    + generator_flag
                    + cc_flag
                ),
            },
            include_dirs=[
215
216
217
                Path(this_dir) / "csrc" / "flash_attn",
                Path(this_dir) / "csrc" / "flash_attn" / "src",
                Path(this_dir) / "csrc" / "cutlass" / "include",
218
219
            ],
        )
Tri Dao's avatar
Tri Dao committed
220
    )
Tri Dao's avatar
Tri Dao committed
221

Tri Dao's avatar
Tri Dao committed
222

223
224
225
226
227
228
229
230
231
232
def get_package_version():
    with open(Path(this_dir) / "flash_attn" / "__init__.py", "r") as f:
        version_match = re.search(r"^__version__\s*=\s*(.*)$", f.read(), re.MULTILINE)
    public_version = ast.literal_eval(version_match.group(1))
    local_version = os.environ.get("FLASH_ATTN_LOCAL_VERSION")
    if local_version:
        return f"{public_version}+{local_version}"
    else:
        return str(public_version)

Tri Dao's avatar
Tri Dao committed
233

234
235
236
237
238
239
def get_wheel_url():
    # Determine the version numbers that will be used to determine the correct wheel
    # We're using the CUDA version used to build torch, not the one currently installed
    # _, cuda_version_raw = get_cuda_bare_metal_version(CUDA_HOME)
    torch_cuda_version = parse(torch.version.cuda)
    torch_version_raw = parse(torch.__version__)
240
241
242
    # For CUDA 11, we only compile for CUDA 11.8, and for CUDA 12 we only compile for CUDA 12.2
    # to save CI time. Minor versions should be compatible.
    torch_cuda_version = parse("11.8") if torch_cuda_version.major == 11 else parse("12.2")
243
244
245
246
247
248
249
250
251
    python_version = f"cp{sys.version_info.major}{sys.version_info.minor}"
    platform_name = get_platform()
    flash_version = get_package_version()
    # cuda_version = f"{cuda_version_raw.major}{cuda_version_raw.minor}"
    cuda_version = f"{torch_cuda_version.major}{torch_cuda_version.minor}"
    torch_version = f"{torch_version_raw.major}.{torch_version_raw.minor}"
    cxx11_abi = str(torch._C._GLIBCXX_USE_CXX11_ABI).upper()

    # Determine wheel URL based on CUDA version, torch version, python version and OS
252
253
    wheel_filename = f"{PACKAGE_NAME}-{flash_version}+cu{cuda_version}torch{torch_version}cxx11abi{cxx11_abi}-{python_version}-{python_version}-{platform_name}.whl"
    wheel_url = BASE_WHEEL_URL.format(tag_name=f"v{flash_version}", wheel_name=wheel_filename)
254
255
256
    return wheel_url, wheel_filename


257
class CachedWheelsCommand(_bdist_wheel):
Tri Dao's avatar
Tri Dao committed
258
259
260
261
262
263
    """
    The CachedWheelsCommand plugs into the default bdist wheel, which is ran by pip when it cannot
    find an existing wheel (which is currently the case for all flash attention installs). We use
    the environment parameters to detect whether there is already a pre-built version of a compatible
    wheel available and short-circuits the standard full build pipeline.
    """
264

Tri Dao's avatar
Tri Dao committed
265
    def run(self):
266
        if FORCE_BUILD:
Pierce Freeman's avatar
Pierce Freeman committed
267
            return super().run()
268

269
        wheel_url, wheel_filename = get_wheel_url()
270
271
272
        print("Guessing wheel URL: ", wheel_url)
        try:
            urllib.request.urlretrieve(wheel_url, wheel_filename)
273
274
275
276
277
278
279
280
281

            # Make the archive
            # Lifted from the root wheel processing command
            # https://github.com/pypa/wheel/blob/cf71108ff9f6ffc36978069acb28824b44ae028e/src/wheel/bdist_wheel.py#LL381C9-L381C85
            if not os.path.exists(self.dist_dir):
                os.makedirs(self.dist_dir)

            impl_tag, abi_tag, plat_tag = self.get_tag()
            archive_basename = f"{self.wheel_dist_name}-{impl_tag}-{abi_tag}-{plat_tag}"
282

283
284
285
            wheel_path = os.path.join(self.dist_dir, archive_basename + ".whl")
            print("Raw wheel path", wheel_path)
            os.rename(wheel_filename, wheel_path)
286
        except (urllib.error.HTTPError, urllib.error.URLError):
287
288
            print("Precompiled wheel not found. Building from source...")
            # If the wheel could not be downloaded, build from source
289
            super().run()
290
291


292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
class NinjaBuildExtension(BuildExtension):
    def __init__(self, *args, **kwargs) -> None:
        # do not override env MAX_JOBS if already exists
        if not os.environ.get("MAX_JOBS"):
            import psutil

            # calculate the maximum allowed NUM_JOBS based on cores
            max_num_jobs_cores = max(1, os.cpu_count() // 2)

            # calculate the maximum allowed NUM_JOBS based on free memory
            free_memory_gb = psutil.virtual_memory().available / (1024 ** 3)  # free memory in GB
            max_num_jobs_memory = int(free_memory_gb / 9)  # each JOB peak memory cost is ~8-9GB when threads = 4

            # pick lower value of jobs based on cores vs memory metric to minimize oom and swap usage during compilation
            max_jobs = max(1, min(max_num_jobs_cores, max_num_jobs_memory))
            os.environ["MAX_JOBS"] = str(max_jobs)

        super().__init__(*args, **kwargs)


Tri Dao's avatar
Tri Dao committed
312
setup(
313
    name=PACKAGE_NAME,
314
    version=get_package_version(),
Tri Dao's avatar
Tri Dao committed
315
    packages=find_packages(
316
317
318
319
320
321
322
323
324
325
        exclude=(
            "build",
            "csrc",
            "include",
            "tests",
            "dist",
            "docs",
            "benchmarks",
            "flash_attn.egg-info",
        )
Tri Dao's avatar
Tri Dao committed
326
327
    ),
    author="Tri Dao",
Tri Dao's avatar
Tri Dao committed
328
    author_email="trid@cs.stanford.edu",
Tri Dao's avatar
Tri Dao committed
329
330
331
    description="Flash Attention: Fast and Memory-Efficient Exact Attention",
    long_description=long_description,
    long_description_content_type="text/markdown",
Tri Dao's avatar
Tri Dao committed
332
    url="https://github.com/Dao-AILab/flash-attention",
Tri Dao's avatar
Tri Dao committed
333
334
    classifiers=[
        "Programming Language :: Python :: 3",
335
        "License :: OSI Approved :: BSD License",
Phil Wang's avatar
Phil Wang committed
336
        "Operating System :: Unix",
Tri Dao's avatar
Tri Dao committed
337
    ],
Tri Dao's avatar
Tri Dao committed
338
    ext_modules=ext_modules,
339
    cmdclass={"bdist_wheel": CachedWheelsCommand, "build_ext": NinjaBuildExtension}
340
341
342
    if ext_modules
    else {
        "bdist_wheel": CachedWheelsCommand,
343
    },
Gustaf's avatar
Gustaf committed
344
345
346
347
348
    python_requires=">=3.7",
    install_requires=[
        "torch",
        "einops",
    ],
349
    setup_requires=[
350
351
352
        "packaging",
        "psutil",
        "ninja",
353
    ],
354
)