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.
30 lines
1.1 KiB
30 lines
1.1 KiB
7 months ago
|
from ...cuda_extension import _CudaExtension
|
||
|
from ...utils import get_cuda_cc_flag
|
||
10 months ago
|
|
||
|
|
||
|
class FusedOptimizerCudaExtension(_CudaExtension):
|
||
|
def __init__(self):
|
||
|
super().__init__(name="fused_optim_cuda")
|
||
|
|
||
|
def sources_files(self):
|
||
|
ret = [
|
||
|
self.csrc_abs_path(fname)
|
||
|
for fname in [
|
||
7 months ago
|
"kernel/cuda/multi_tensor_sgd_kernel.cu",
|
||
|
"kernel/cuda/multi_tensor_scale_kernel.cu",
|
||
|
"kernel/cuda/multi_tensor_adam_kernel.cu",
|
||
|
"kernel/cuda/multi_tensor_l2norm_kernel.cu",
|
||
|
"kernel/cuda/multi_tensor_lamb_kernel.cu",
|
||
10 months ago
|
]
|
||
7 months ago
|
] + [self.pybind_abs_path("optimizer/optimizer.cpp")]
|
||
10 months ago
|
return ret
|
||
|
|
||
|
def cxx_flags(self):
|
||
|
version_dependent_macros = ["-DVERSION_GE_1_1", "-DVERSION_GE_1_3", "-DVERSION_GE_1_5"]
|
||
|
return ["-O3"] + version_dependent_macros
|
||
|
|
||
|
def nvcc_flags(self):
|
||
|
extra_cuda_flags = ["-lineinfo"]
|
||
|
extra_cuda_flags.extend(get_cuda_cc_flag())
|
||
7 months ago
|
return ["-O3", "--use_fast_math"] + extra_cuda_flags + super().nvcc_flags()
|