cpu_adam.py 1.27 KB
Newer Older
1
2
3
4
5
6
7
8
import os

from .builder import Builder
from .utils import append_nvcc_threads


class CPUAdamBuilder(Builder):
    NAME = "cpu_adam"
9
    PREBUILT_IMPORT_PATH = "colossalai._C.cpu_adam"
10
11

    def __init__(self):
12
        super().__init__(name=CPUAdamBuilder.NAME, prebuilt_import_path=CPUAdamBuilder.PREBUILT_IMPORT_PATH)
13
14
15
16
17
        self.version_dependent_macros = ['-DVERSION_GE_1_1', '-DVERSION_GE_1_3', '-DVERSION_GE_1_5']

    # necessary 4 functions
    def sources_files(self):
        ret = [
18
            self.csrc_abs_path('cpu_adam.cpp'),
19
        ]
20
        return ret
21
22
23

    def include_dirs(self):
        return [
24
            self.csrc_abs_path("includes"),
25
26
27
28
29
30
31
32
33
34
35
36
            self.get_cuda_home_include()
        ]

    def cxx_flags(self):
        extra_cxx_flags = ['-std=c++14', '-lcudart', '-lcublas', '-g', '-Wno-reorder', '-fopenmp', '-march=native']
        return ['-O3'] + self.version_dependent_macros + extra_cxx_flags

    def nvcc_flags(self):
        extra_cuda_flags = [
            '-std=c++14', '-U__CUDA_NO_HALF_OPERATORS__', '-U__CUDA_NO_HALF_CONVERSIONS__',
            '-U__CUDA_NO_HALF2_OPERATORS__', '-DTHRUST_IGNORE_CUB_VERSION_CHECK'
        ]
37
38
        ret = ['-O3', '--use_fast_math'] + self.version_dependent_macros + extra_cuda_flags
        return append_nvcc_threads(ret)