github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-11-20 18:00:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

[setup] support pre-build and jit-build of cuda kernels (#2374)

Browse Source 
* [setup] support pre-build and jit-build of cuda kernels

* polish code

* polish code

* polish code

* polish code

* polish code

* polish code
This commit is contained in:
Frank Lee
2023-01-06 20:50:26 +08:00
committed by GitHub
parent 12c8bf38d7
commit 40d376c566
36 changed files with 414 additions and 390 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
op_builder/fused_optim.py
View File

@@ -1,20 +1,19 @@
import os
from .builder import Builder, get_cuda_cc_flag
from .builder import Builder
from .utils import get_cuda_cc_flag
class FusedOptimBuilder(Builder):
NAME = 'fused_optim'
BASE_DIR = "cuda_native/csrc"
NAME = "fused_optim"
PREBUILT_IMPORT_PATH = "colossalai._C.fused_optim"
def __init__(self):
self.name = FusedOptimBuilder.NAME
super().__init__()
self.version_dependent_macros = ['-DVERSION_GE_1_1', '-DVERSION_GE_1_3', '-DVERSION_GE_1_5']
super().__init__(name=FusedOptimBuilder.NAME, prebuilt_import_path=FusedOptimBuilder.PREBUILT_IMPORT_PATH)
def sources_files(self):
ret = [
self.colossalai_src_path(os.path.join(FusedOptimBuilder.BASE_DIR, fname)) for fname in [
self.csrc_abs_path(fname) for fname in [
'colossal_C_frontend.cpp', 'multi_tensor_sgd_kernel.cu', 'multi_tensor_scale_kernel.cu',
'multi_tensor_adam.cu', 'multi_tensor_l2norm_kernel.cu', 'multi_tensor_lamb.cu'
]
@@ -22,12 +21,12 @@ class FusedOptimBuilder(Builder):
return ret
def include_dirs(self):
ret = [os.path.join(FusedOptimBuilder.BASE_DIR, "includes"), self.get_cuda_home_include()]
return [self.colossalai_src_path(path) for path in ret]
ret = [self.csrc_abs_path('kernels/include'), self.get_cuda_home_include()]
return ret
def cxx_flags(self):
extra_cxx_flags = []
return ['-O3'] + self.version_dependent_macros + extra_cxx_flags
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']