mirror of https://github.com/hpcaitech/ColossalAI
aibig-modeldata-parallelismdeep-learningdistributed-computingfoundation-modelsheterogeneous-traininghpcinferencelarge-scalemodel-parallelismpipeline-parallelism
You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
29 lines
1002 B
29 lines
1002 B
7 months ago
|
from ...cuda_extension import _CudaExtension
|
||
|
from ...utils import append_nvcc_threads
|
||
2 years ago
|
|
||
|
|
||
10 months ago
|
class ScaledMaskedSoftmaxCudaExtension(_CudaExtension):
|
||
2 years ago
|
def __init__(self):
|
||
10 months ago
|
super().__init__(name="scaled_masked_softmax_cuda")
|
||
2 years ago
|
|
||
|
def sources_files(self):
|
||
7 months ago
|
ret = [self.csrc_abs_path(fname) for fname in ["kernel/cuda/scaled_masked_softmax_kernel.cu"]] + [
|
||
|
self.pybind_abs_path("softmax/scaled_masked_softmax.cpp")
|
||
10 months ago
|
]
|
||
2 years ago
|
return ret
|
||
|
|
||
|
def cxx_flags(self):
|
||
1 year ago
|
return ["-O3"] + self.version_dependent_macros
|
||
2 years ago
|
|
||
|
def nvcc_flags(self):
|
||
|
extra_cuda_flags = [
|
||
1 year ago
|
"-std=c++14",
|
||
1 year ago
|
"-std=c++17",
|
||
1 year ago
|
"-U__CUDA_NO_HALF_OPERATORS__",
|
||
|
"-U__CUDA_NO_HALF_CONVERSIONS__",
|
||
|
"-U__CUDA_NO_HALF2_OPERATORS__",
|
||
|
"-DTHRUST_IGNORE_CUB_VERSION_CHECK",
|
||
2 years ago
|
]
|
||
7 months ago
|
ret = ["-O3", "--use_fast_math"] + self.version_dependent_macros + extra_cuda_flags + super().nvcc_flags()
|
||
2 years ago
|
return append_nvcc_threads(ret)
|