mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2025-09-05 11:02:05 +00:00
refactor kernel (#142)
This commit is contained in:
ver217
GitHub
114
setup.py
114
setup.py
@@ -11,8 +11,7 @@ this_dir = os.path.dirname(os.path.abspath(__file__))
|
||||
|
||||
|
||||
def get_cuda_bare_metal_version(cuda_dir):
|
||||
raw_output = subprocess.check_output(
|
||||
[cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
|
||||
raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
|
||||
output = raw_output.split()
|
||||
release_idx = output.index("release") + 1
|
||||
release = output[release_idx].split(".")
|
||||
@@ -23,8 +22,7 @@ def get_cuda_bare_metal_version(cuda_dir):
|
||||
|
||||
|
||||
def check_cuda_torch_binary_vs_bare_metal(cuda_dir):
|
||||
raw_output, bare_metal_major, bare_metal_minor = get_cuda_bare_metal_version(
|
||||
cuda_dir)
|
||||
raw_output, bare_metal_major, bare_metal_minor = get_cuda_bare_metal_version(cuda_dir)
|
||||
torch_binary_major = torch.version.cuda.split(".")[0]
|
||||
torch_binary_minor = torch.version.cuda.split(".")[1]
|
||||
|
||||
@@ -40,6 +38,13 @@ def check_cuda_torch_binary_vs_bare_metal(cuda_dir):
|
||||
"You can try commenting out this check (at your own risk).")
|
||||
|
||||
|
||||
def append_nvcc_threads(nvcc_extra_args):
|
||||
_, bare_metal_major, bare_metal_minor = get_cuda_bare_metal_version(CUDA_HOME)
|
||||
if int(bare_metal_major) >= 11 and int(bare_metal_minor) >= 2:
|
||||
return nvcc_extra_args + ["--threads", "4"]
|
||||
return nvcc_extra_args
|
||||
|
||||
|
||||
def fetch_requirements(path):
|
||||
with open(path, 'r') as fd:
|
||||
return [r.strip() for r in fd.readlines()]
|
||||
@@ -67,8 +72,8 @@ print("\n\ntorch.__version__ = {}\n\n".format(torch.__version__))
|
||||
TORCH_MAJOR = int(torch.__version__.split('.')[0])
|
||||
TORCH_MINOR = int(torch.__version__.split('.')[1])
|
||||
|
||||
if TORCH_MAJOR == 0 and TORCH_MINOR < 4:
|
||||
raise RuntimeError("Colossal-AI requires Pytorch 0.4 or newer.\n" +
|
||||
if TORCH_MAJOR < 1 or (TORCH_MAJOR == 1 and TORCH_MINOR < 8):
|
||||
raise RuntimeError("Colossal-AI requires Pytorch 1.8 or newer.\n" +
|
||||
"The latest stable release can be obtained from https://pytorch.org/")
|
||||
|
||||
cmdclass = {}
|
||||
@@ -79,22 +84,9 @@ ext_modules = []
|
||||
# and
|
||||
# https://github.com/NVIDIA/apex/issues/456
|
||||
# https://github.com/pytorch/pytorch/commit/eb7b39e02f7d75c26d8a795ea8c7fd911334da7e#diff-4632522f237f1e4e728cb824300403ac
|
||||
version_ge_1_1 = []
|
||||
if (TORCH_MAJOR > 1) or (TORCH_MAJOR == 1 and TORCH_MINOR > 0):
|
||||
version_ge_1_1 = ['-DVERSION_GE_1_1']
|
||||
version_ge_1_3 = []
|
||||
if (TORCH_MAJOR > 1) or (TORCH_MAJOR == 1 and TORCH_MINOR > 2):
|
||||
version_ge_1_3 = ['-DVERSION_GE_1_3']
|
||||
version_ge_1_5 = []
|
||||
if (TORCH_MAJOR > 1) or (TORCH_MAJOR == 1 and TORCH_MINOR > 4):
|
||||
version_ge_1_5 = ['-DVERSION_GE_1_5']
|
||||
version_dependent_macros = version_ge_1_1 + version_ge_1_3 + version_ge_1_5
|
||||
version_dependent_macros = ['-DVERSION_GE_1_1', '-DVERSION_GE_1_3', '-DVERSION_GE_1_5']
|
||||
|
||||
if "--cuda_ext" in sys.argv:
|
||||
if TORCH_MAJOR == 0:
|
||||
raise RuntimeError("--cuda_ext requires Pytorch 1.0 or later, "
|
||||
"found torch.__version__ = {}".format(torch.__version__))
|
||||
|
||||
sys.argv.remove("--cuda_ext")
|
||||
|
||||
if CUDA_HOME is None:
|
||||
@@ -103,19 +95,66 @@ if "--cuda_ext" in sys.argv:
|
||||
else:
|
||||
check_cuda_torch_binary_vs_bare_metal(CUDA_HOME)
|
||||
|
||||
ext_modules.append(
|
||||
CUDAExtension(name='colossal_C',
|
||||
sources=['csrc/colossal_C_frontend.cpp',
|
||||
'csrc/multi_tensor_sgd_kernel.cu',
|
||||
'csrc/multi_tensor_scale_kernel.cu',
|
||||
'csrc/multi_tensor_adam.cu',
|
||||
'csrc/multi_tensor_l2norm_kernel.cu',
|
||||
'csrc/multi_tensor_lamb.cu'],
|
||||
extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
|
||||
'nvcc': ['-lineinfo',
|
||||
'-O3',
|
||||
# '--resource-usage',
|
||||
'--use_fast_math'] + version_dependent_macros}))
|
||||
def cuda_ext_helper(name, sources, extra_cuda_flags):
|
||||
return CUDAExtension(name=name,
|
||||
sources=[os.path.join('colossalai/kernel/cuda_native/csrc', path) for path in sources],
|
||||
include_dirs=[os.path.join(
|
||||
this_dir, 'colossalai/kernel/cuda_native/csrc/kernels/include')],
|
||||
extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
|
||||
'nvcc': append_nvcc_threads(['-O3',
|
||||
'--use_fast_math'] + version_dependent_macros + extra_cuda_flags)})
|
||||
|
||||
ext_modules.append(cuda_ext_helper('colossal_C',
|
||||
['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'],
|
||||
['-lineinfo']))
|
||||
|
||||
cc_flag = ['-gencode', 'arch=compute_70,code=sm_70']
|
||||
_, bare_metal_major, _ = get_cuda_bare_metal_version(CUDA_HOME)
|
||||
if int(bare_metal_major) >= 11:
|
||||
cc_flag.append('-gencode')
|
||||
cc_flag.append('arch=compute_80,code=sm_80')
|
||||
|
||||
extra_cuda_flags = ['-U__CUDA_NO_HALF_OPERATORS__',
|
||||
'-U__CUDA_NO_HALF_CONVERSIONS__',
|
||||
'--expt-relaxed-constexpr',
|
||||
'--expt-extended-lambda']
|
||||
|
||||
ext_modules.append(cuda_ext_helper('colossal_scaled_upper_triang_masked_softmax',
|
||||
['scaled_upper_triang_masked_softmax.cpp',
|
||||
'scaled_upper_triang_masked_softmax_cuda.cu'],
|
||||
extra_cuda_flags + cc_flag))
|
||||
|
||||
ext_modules.append(cuda_ext_helper('colossal_scaled_masked_softmax',
|
||||
['scaled_masked_softmax.cpp', 'scaled_masked_softmax_cuda.cu'],
|
||||
extra_cuda_flags + cc_flag))
|
||||
|
||||
extra_cuda_flags = ['-maxrregcount=50']
|
||||
|
||||
ext_modules.append(cuda_ext_helper('colossal_layer_norm_cuda',
|
||||
['layer_norm_cuda.cpp', 'layer_norm_cuda_kernel.cu'],
|
||||
extra_cuda_flags + cc_flag))
|
||||
|
||||
extra_cuda_flags = ['-std=c++14',
|
||||
'-U__CUDA_NO_HALF_OPERATORS__',
|
||||
'-U__CUDA_NO_HALF_CONVERSIONS__',
|
||||
'-U__CUDA_NO_HALF2_OPERATORS__',
|
||||
'-DTHRUST_IGNORE_CUB_VERSION_CHECK']
|
||||
|
||||
ext_modules.append(cuda_ext_helper('colossal_multihead_attention',
|
||||
['multihead_attention_1d.cpp',
|
||||
'kernels/cublas_wrappers.cu',
|
||||
'kernels/transform_kernels.cu',
|
||||
'kernels/dropout_kernels.cu',
|
||||
'kernels/normalize_kernels.cu',
|
||||
'kernels/softmax_kernels.cu',
|
||||
'kernels/general_kernels.cu',
|
||||
'kernels/cuda_util.cu'],
|
||||
extra_cuda_flags + cc_flag))
|
||||
|
||||
|
||||
install_requires = fetch_requirements('requirements/requirements.txt')
|
||||
@@ -123,14 +162,17 @@ install_requires = fetch_requirements('requirements/requirements.txt')
|
||||
setup(
|
||||
name='colossalai',
|
||||
version='0.0.1-beta',
|
||||
packages=find_packages(exclude=('csrc',
|
||||
packages=find_packages(exclude=('benchmark',
|
||||
'docker',
|
||||
'tests',
|
||||
'docs',
|
||||
'examples',
|
||||
'tests',
|
||||
'scripts',
|
||||
'requirements',
|
||||
'*.egg-info',)),
|
||||
description='An integrated large-scale model training system with efficient parallelization techniques',
|
||||
ext_modules=ext_modules,
|
||||
cmdclass={'build_ext': BuildExtension} if ext_modules else {},
|
||||
package_data={'colossalai': ['kernel/cuda_native/csrc/*']},
|
||||
install_requires=install_requires,
|
||||
)
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user