"tests/vscode:/vscode.git/clone" did not exist on "1ad3a636b17c4730bce4517a78b4056f0626fe50"
moe.py 1.04 KB
Newer Older
1
2
import os

3
4
from .builder import Builder
from .utils import append_nvcc_threads, get_cuda_cc_flag
5
6
7
8


class MOEBuilder(Builder):

9
10
11
    NAME = "moe"
    PREBUILT_IMPORT_PATH = "colossalai._C.moe"

12
    def __init__(self):
13
        super().__init__(name=MOEBuilder.NAME, prebuilt_import_path=MOEBuilder.PREBUILT_IMPORT_PATH)
14
15

    def include_dirs(self):
16
17
18
19
20
        ret = [
            self.csrc_abs_path("kernels/include"),
            self.get_cuda_home_include()
        ]
        return ret
21
22

    def sources_files(self):
23
24
        ret = [self.csrc_abs_path(fname) for fname in ['moe_cuda.cpp', 'moe_cuda_kernel.cu']]
        return ret
25
26

    def cxx_flags(self):
27
        return ['-O3'] + self.version_dependent_macros
28
29
30
31
32
33
34
35

    def nvcc_flags(self):
        extra_cuda_flags = [
            '-U__CUDA_NO_HALF_OPERATORS__', '-U__CUDA_NO_HALF_CONVERSIONS__', '--expt-relaxed-constexpr',
            '--expt-extended-lambda'
        ]
        extra_cuda_flags.extend(get_cuda_cc_flag())
        ret = ['-O3', '--use_fast_math'] + extra_cuda_flags
36
        return append_nvcc_threads(ret)