utils.py 10.3 KB
Newer Older
1
# Copyright (c) 2022-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
4
5
6
#
# See LICENSE for license information.

"""Installation script."""

7
8
import functools
import glob
9
import importlib
10
11
12
13
14
import os
import re
import shutil
import subprocess
import sys
15
import platform
16
from pathlib import Path
17
from importlib.metadata import version as get_version
18
from subprocess import CalledProcessError
19
from typing import List, Optional, Tuple, Union
20
21


22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
# Needs to stay consistent with .pre-commit-config.yaml config.
def min_python_version() -> Tuple[int]:
    """Minimum supported Python version."""
    return (3, 10, 0)


def min_python_version_str() -> str:
    """String representing minimum supported Python version."""
    return ".".join(map(str, min_python_version()))


if sys.version_info < min_python_version():
    raise RuntimeError(
        f"Transformer Engine requires Python {min_python_version_str()} or newer, "
        f"but found Python {platform.python_version()}."
    )


40
@functools.lru_cache(maxsize=None)
41
42
def debug_build_enabled() -> bool:
    """Whether to build with a debug configuration"""
43
    return bool(int(os.getenv("NVTE_BUILD_DEBUG", "0")))
44
45


46
47
48
49
50
51
52
53
@functools.lru_cache(maxsize=None)
def get_max_jobs_for_parallel_build() -> int:
    """Number of parallel jobs for Nina build"""

    # Default: maximum parallel jobs
    num_jobs = 0

    # Check environment variable
54
55
    if os.getenv("NVTE_BUILD_MAX_JOBS"):
        num_jobs = int(os.getenv("NVTE_BUILD_MAX_JOBS"))
56
57
58
59
60
61
62
63
64
65
66
67
    elif os.getenv("MAX_JOBS"):
        num_jobs = int(os.getenv("MAX_JOBS"))

    # Check command-line arguments
    for arg in sys.argv.copy():
        if arg.startswith("--parallel="):
            num_jobs = int(arg.replace("--parallel=", ""))
            sys.argv.remove(arg)

    return num_jobs


Phuong Nguyen's avatar
Phuong Nguyen committed
68
def all_files_in_dir(path, name_extension=None):
69
70
71
    all_files = []
    for dirname, _, names in os.walk(path):
        for name in names:
72
            if name_extension is not None and not name.endswith(f".{name_extension}"):
Phuong Nguyen's avatar
Phuong Nguyen committed
73
                continue
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
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
            all_files.append(Path(dirname, name))
    return all_files


def remove_dups(_list: List):
    return list(set(_list))


def found_cmake() -> bool:
    """ "Check if valid CMake is available

    CMake 3.18 or newer is required.

    """

    # Check if CMake is available
    try:
        _cmake_bin = cmake_bin()
    except FileNotFoundError:
        return False

    # Query CMake for version info
    output = subprocess.run(
        [_cmake_bin, "--version"],
        capture_output=True,
        check=True,
        universal_newlines=True,
    )
    match = re.search(r"version\s*([\d.]+)", output.stdout)
    version = match.group(1).split(".")
    version = tuple(int(v) for v in version)
    return version >= (3, 18)


def cmake_bin() -> Path:
    """Get CMake executable

    Throws FileNotFoundError if not found.

    """

    # Search in CMake Python package
    _cmake_bin: Optional[Path] = None
    try:
        from cmake import CMAKE_BIN_DIR
    except ImportError:
        pass
    else:
        _cmake_bin = Path(CMAKE_BIN_DIR).resolve() / "cmake"
        if not _cmake_bin.is_file():
            _cmake_bin = None

    # Search in path
    if _cmake_bin is None:
        _cmake_bin = shutil.which("cmake")
        if _cmake_bin is not None:
            _cmake_bin = Path(_cmake_bin).resolve()

    # Return executable if found
    if _cmake_bin is None:
        raise FileNotFoundError("Could not find CMake executable")
    return _cmake_bin


def found_ninja() -> bool:
    """ "Check if Ninja is available"""
    return shutil.which("ninja") is not None


def found_pybind11() -> bool:
    """ "Check if pybind11 is available"""

    # Check if Python package is installed
    try:
        import pybind11
    except ImportError:
        pass
    else:
        return True

    # Check if CMake can find pybind11
    if not found_cmake():
        return False
    try:
        subprocess.run(
            [
                "cmake",
                "--find-package",
                "-DMODE=EXIST",
                "-DNAME=pybind11",
                "-DCOMPILER_ID=CXX",
                "-DLANGUAGE=CXX",
            ],
            stdout=subprocess.DEVNULL,
            stderr=subprocess.DEVNULL,
            check=True,
        )
    except (CalledProcessError, OSError):
        pass
    else:
        return True
    return False


178
@functools.lru_cache(maxsize=None)
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
def cuda_toolkit_include_path() -> Tuple[str, str]:
    """Returns root path for cuda toolkit includes.

    return `None` if CUDA is not found."""
    # Try finding CUDA
    cuda_home: Optional[Path] = None
    if cuda_home is None and os.getenv("CUDA_HOME"):
        # Check in CUDA_HOME
        cuda_home = Path(os.getenv("CUDA_HOME")) / "include"
    if cuda_home is None:
        # Check in NVCC
        nvcc_bin = shutil.which("nvcc")
        if nvcc_bin is not None:
            cuda_home = Path(nvcc_bin.rstrip("/bin/nvcc")) / "include"
    if cuda_home is None:
        # Last-ditch guess in /usr/local/cuda
        if Path("/usr/local/cuda").is_dir():
            cuda_home = Path("/usr/local/cuda") / "include"
    return cuda_home


@functools.lru_cache(maxsize=None)
def nvcc_path() -> Tuple[str, str]:
    """Returns the NVCC binary path.
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223

    Throws FileNotFoundError if NVCC is not found."""
    # Try finding NVCC
    nvcc_bin: Optional[Path] = None
    if nvcc_bin is None and os.getenv("CUDA_HOME"):
        # Check in CUDA_HOME
        cuda_home = Path(os.getenv("CUDA_HOME"))
        nvcc_bin = cuda_home / "bin" / "nvcc"
    if nvcc_bin is None:
        # Check if nvcc is in path
        nvcc_bin = shutil.which("nvcc")
        if nvcc_bin is not None:
            cuda_home = Path(nvcc_bin.rstrip("/bin/nvcc"))
            nvcc_bin = Path(nvcc_bin)
    if nvcc_bin is None:
        # Last-ditch guess in /usr/local/cuda
        cuda_home = Path("/usr/local/cuda")
        nvcc_bin = cuda_home / "bin" / "nvcc"
    if not nvcc_bin.is_file():
        raise FileNotFoundError(f"Could not find NVCC at {nvcc_bin}")

224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
    return nvcc_bin


@functools.lru_cache(maxsize=None)
def get_cuda_include_dirs() -> Tuple[str, str]:
    """Returns the CUDA header directory."""

    # If cuda is installed via toolkit, all necessary headers
    # are bundled inside the top level cuda directory.
    if cuda_toolkit_include_path() is not None:
        return [cuda_toolkit_include_path()]

    # Use pip wheels to include all headers.
    try:
        import nvidia
    except ModuleNotFoundError as e:
        raise RuntimeError("CUDA not found.")

    cuda_root = Path(nvidia.__file__).parent
    return [
244
245
246
        subdir / "include"
        for subdir in cuda_root.iterdir()
        if subdir.is_dir() and (subdir / "include").is_dir()
247
    ]
248
249


250
251
@functools.lru_cache(maxsize=None)
def cuda_archs() -> str:
252
253
254
    archs = os.getenv("NVTE_CUDA_ARCHS")
    if archs is None:
        version = cuda_version()
255
        if version >= (13, 0):
256
            archs = "75;80;89;90;100;120"
257
        elif version >= (12, 8):
258
            archs = "70;80;89;90;100;120"
259
        else:
260
261
            archs = "70;80;89;90"
    return archs
262
263


264
def cuda_version() -> Tuple[int, ...]:
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
    """CUDA Toolkit version as a (major, minor) tuple.

    Try to get cuda version by locating the nvcc executable and running nvcc --version. If
    nvcc is not found, look for the cuda runtime package pip `nvidia-cuda-runtime-cu12`
    and check pip version.
    """

    try:
        nvcc_bin = nvcc_path()
    except FileNotFoundError as e:
        pass
    else:
        output = subprocess.run(
            [nvcc_bin, "-V"],
            capture_output=True,
            check=True,
            universal_newlines=True,
        )
        match = re.search(r"release\s*([\d.]+)", output.stdout)
        version = match.group(1).split(".")
        return tuple(int(v) for v in version)

    try:
288
        version_str = get_version("nvidia-cuda-runtime-cu12")
289
290
291
292
        version_tuple = tuple(int(part) for part in version_str.split(".") if part.isdigit())
        return version_tuple
    except importlib.metadata.PackageNotFoundError:
        raise RuntimeError("Could neither find NVCC executable nor CUDA runtime Python package.")
293
294
295
296
297


def get_frameworks() -> List[str]:
    """DL frameworks to build support for"""
    _frameworks: List[str] = []
298
    supported_frameworks = ["pytorch", "jax"]
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334

    # Check environment variable
    if os.getenv("NVTE_FRAMEWORK"):
        _frameworks.extend(os.getenv("NVTE_FRAMEWORK").split(","))

    # Check command-line arguments
    for arg in sys.argv.copy():
        if arg.startswith("--framework="):
            _frameworks.extend(arg.replace("--framework=", "").split(","))
            sys.argv.remove(arg)

    # Detect installed frameworks if not explicitly specified
    if not _frameworks:
        try:
            import torch
        except ImportError:
            pass
        else:
            _frameworks.append("pytorch")
        try:
            import jax
        except ImportError:
            pass
        else:
            _frameworks.append("jax")

    # Special framework names
    if "all" in _frameworks:
        _frameworks = supported_frameworks.copy()
    if "none" in _frameworks:
        _frameworks = []

    # Check that frameworks are valid
    _frameworks = [framework.lower() for framework in _frameworks]
    for framework in _frameworks:
        if framework not in supported_frameworks:
335
            raise ValueError(f"Transformer Engine does not support framework={framework}")
336
337
338
339

    return _frameworks


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
def copy_common_headers(
    src_dir: Union[Path, str],
    dst_dir: Union[Path, str],
) -> None:
    """Copy headers from core library

    src_dir should be the transformer_engine directory within the root
    Transformer Engine repository. All .h and .cuh files within
    transformer_engine/common are copied into dst_dir. Relative paths
    are preserved.

    """

    # Find common header files in src dir
    headers = glob.glob(
        os.path.join(str(src_dir), "common", "**", "*.h"),
        recursive=True,
    )
    headers.extend(
        glob.glob(
            os.path.join(str(src_dir), "common", "**", "*.cuh"),
            recursive=True,
        )
    )
    headers = [Path(path) for path in headers]

    # Copy common header files to dst dir
    src_dir = Path(src_dir)
    dst_dir = Path(dst_dir)
    for path in headers:
        new_path = dst_dir / path.relative_to(src_dir)
        new_path.parent.mkdir(exist_ok=True, parents=True)
        shutil.copy(path, new_path)