cpu_adam.py 1.26 KB
Newer Older
1
2
3
4
5
6
from .builder import Builder
from .utils import append_nvcc_threads


class CPUAdamBuilder(Builder):
    NAME = "cpu_adam"
7
    PREBUILT_IMPORT_PATH = "colossalai._C.cpu_adam"
8
9

    def __init__(self):
10
        super().__init__(name=CPUAdamBuilder.NAME, prebuilt_import_path=CPUAdamBuilder.PREBUILT_IMPORT_PATH)
11
        self.version_dependent_macros = ["-DVERSION_GE_1_1", "-DVERSION_GE_1_3", "-DVERSION_GE_1_5"]
12
13
14
15

    # necessary 4 functions
    def sources_files(self):
        ret = [
16
            self.csrc_abs_path("cpu_adam.cpp"),
17
        ]
18
        return ret
19
20

    def include_dirs(self):
21
        return [self.csrc_abs_path("includes"), self.get_cuda_home_include()]
22
23

    def cxx_flags(self):
24
25
        extra_cxx_flags = ["-std=c++14", "-lcudart", "-lcublas", "-g", "-Wno-reorder", "-fopenmp", "-march=native"]
        return ["-O3"] + self.version_dependent_macros + extra_cxx_flags
26
27
28

    def nvcc_flags(self):
        extra_cuda_flags = [
29
30
31
32
33
            "-std=c++14",
            "-U__CUDA_NO_HALF_OPERATORS__",
            "-U__CUDA_NO_HALF_CONVERSIONS__",
            "-U__CUDA_NO_HALF2_OPERATORS__",
            "-DTHRUST_IGNORE_CUB_VERSION_CHECK",
34
        ]
35
        ret = ["-O3", "--use_fast_math"] + self.version_dependent_macros + extra_cuda_flags
36
        return append_nvcc_threads(ret)