Skip to content

GitLab

Commit af801cb4 authored by xyupeng's avatar xyupeng Committed by Frank Lee
Browse files

fix format setup.py (#343)

parent 14820704
Hide whitespace changes
Inline Side-by-side
Showing with 76 additions and 85 deletions
setup.py
View file @ af801cb4
...@@ -34,31 +34,33 @@ def check_cuda_torch_binary_vs_bare_metal(cuda_dir): ...@@ -34,31 +34,33 @@ def check_cuda_torch_binary_vs_bare_metal(cuda_dir):
print(raw_output + "from " + cuda_dir + "/bin\n") print(raw_output + "from " + cuda_dir + "/bin\n")
if bare_metal_major != torch_binary_major: if bare_metal_major != torch_binary_major:
print( print(f'The detected CUDA version ({raw_output}) mismatches the version that was used to compile PyTorch '
f'The detected CUDA version ({raw_output}) mismatches the version that was used to compile PyTorch ({torch.version.cuda}). CUDA extension will not be installed.') f'({torch.version.cuda}). CUDA extension will not be installed.')
return False return False
if bare_metal_minor != torch_binary_minor: if bare_metal_minor != torch_binary_minor:
print("\nWarning: Cuda extensions are being compiled with a version of Cuda that does " print("\nWarning: Cuda extensions are being compiled with a version of Cuda that does "
+ "not match the version used to compile Pytorch binaries. " "not match the version used to compile Pytorch binaries. "
+ "Pytorch binaries were compiled with Cuda {}.\n".format(torch.version.cuda) f"Pytorch binaries were compiled with Cuda {torch.version.cuda}.\n"
+ "In some cases, a minor-version mismatch will not cause later errors: " "In some cases, a minor-version mismatch will not cause later errors: "
+ "https://github.com/NVIDIA/apex/pull/323#discussion_r287021798. ") "https://github.com/NVIDIA/apex/pull/323#discussion_r287021798. ")
return True return True
def check_cuda_availability(cuda_dir): def check_cuda_availability(cuda_dir):
if not torch.cuda.is_available(): if not torch.cuda.is_available():
# https://github.com/NVIDIA/apex/issues/486 # https://github.com/NVIDIA/apex/issues/486
# Extension builds after https://github.com/pytorch/pytorch/pull/23408 attempt to query torch.cuda.get_device_capability(), # Extension builds after https://github.com/pytorch/pytorch/pull/23408 attempt to query
# which will fail if you are compiling in an environment without visible GPUs (e.g. during an nvidia-docker build command). # torch.cuda.get_device_capability(), which will fail if you are compiling in an environment
print('\nWarning: Torch did not find available GPUs on this system.\n', # without visible GPUs (e.g. during an nvidia-docker build command).
'If your intention is to cross-compile, this is not an error.\n' print(
'By default, Colossal-AI will cross-compile for Pascal (compute capabilities 6.0, 6.1, 6.2),\n' '\nWarning: Torch did not find available GPUs on this system.\n',
'Volta (compute capability 7.0), Turing (compute capability 7.5),\n' 'If your intention is to cross-compile, this is not an error.\n'
'and, if the CUDA version is >= 11.0, Ampere (compute capability 8.0).\n' 'By default, Colossal-AI will cross-compile for Pascal (compute capabilities 6.0, 6.1, 6.2),\n'
'If you wish to cross-compile for a single specific architecture,\n' 'Volta (compute capability 7.0), Turing (compute capability 7.5),\n'
'export TORCH_CUDA_ARCH_LIST="compute capability" before running setup.py.\n') 'and, if the CUDA version is >= 11.0, Ampere (compute capability 8.0).\n'
'If you wish to cross-compile for a single specific architecture,\n'
'export TORCH_CUDA_ARCH_LIST="compute capability" before running setup.py.\n')
if os.environ.get("TORCH_CUDA_ARCH_LIST", None) is None: if os.environ.get("TORCH_CUDA_ARCH_LIST", None) is None:
_, bare_metal_major, _ = get_cuda_bare_metal_version(cuda_dir) _, bare_metal_major, _ = get_cuda_bare_metal_version(cuda_dir)
if int(bare_metal_major) == 11: if int(bare_metal_major) == 11:
...@@ -68,8 +70,8 @@ def check_cuda_availability(cuda_dir): ...@@ -68,8 +70,8 @@ def check_cuda_availability(cuda_dir):
return False return False
if cuda_dir is None: if cuda_dir is None:
print( print("nvcc was not found. CUDA extension will not be installed. If you're installing within a container from "
"nvcc was not found. CUDA extension will not be installed. If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, only images whose names contain 'devel' will provide nvcc.") "https://hub.docker.com/r/pytorch/pytorch, only images whose names contain 'devel' will provide nvcc.")
return False return False
return True return True
...@@ -99,15 +101,14 @@ def get_version(): ...@@ -99,15 +101,14 @@ def get_version():
if build_cuda_ext: if build_cuda_ext:
try: try:
import torch import torch
from torch.utils.cpp_extension import (CUDA_HOME, BuildExtension, from torch.utils.cpp_extension import (CUDA_HOME, BuildExtension, CUDAExtension)
CUDAExtension)
print("\n\ntorch.__version__ = {}\n\n".format(torch.__version__)) print("\n\ntorch.__version__ = {}\n\n".format(torch.__version__))
TORCH_MAJOR = int(torch.__version__.split('.')[0]) TORCH_MAJOR = int(torch.__version__.split('.')[0])
TORCH_MINOR = int(torch.__version__.split('.')[1]) TORCH_MINOR = int(torch.__version__.split('.')[1])
if TORCH_MAJOR < 1 or (TORCH_MAJOR == 1 and TORCH_MINOR < 8): if TORCH_MAJOR < 1 or (TORCH_MAJOR == 1 and TORCH_MINOR < 8):
raise RuntimeError("Colossal-AI requires Pytorch 1.8 or newer.\n" raise RuntimeError("Colossal-AI requires Pytorch 1.8 or newer.\n"
+ "The latest stable release can be obtained from https://pytorch.org/") "The latest stable release can be obtained from https://pytorch.org/")
except ImportError: except ImportError:
print('torch is not found. CUDA extension will not be installed') print('torch is not found. CUDA extension will not be installed')
build_cuda_ext = False build_cuda_ext = False
...@@ -115,7 +116,6 @@ if build_cuda_ext: ...@@ -115,7 +116,6 @@ if build_cuda_ext:
if build_cuda_ext: if build_cuda_ext:
build_cuda_ext = check_cuda_availability(CUDA_HOME) and check_cuda_torch_binary_vs_bare_metal(CUDA_HOME) build_cuda_ext = check_cuda_availability(CUDA_HOME) and check_cuda_torch_binary_vs_bare_metal(CUDA_HOME)
if build_cuda_ext: if build_cuda_ext:
# Set up macros for forward/backward compatibility hack around # Set up macros for forward/backward compatibility hack around
# https://github.com/pytorch/pytorch/commit/4404762d7dd955383acee92e6f06b48144a0742e # https://github.com/pytorch/pytorch/commit/4404762d7dd955383acee92e6f06b48144a0742e
...@@ -125,22 +125,20 @@ if build_cuda_ext: ...@@ -125,22 +125,20 @@ if build_cuda_ext:
version_dependent_macros = ['-DVERSION_GE_1_1', '-DVERSION_GE_1_3', '-DVERSION_GE_1_5'] version_dependent_macros = ['-DVERSION_GE_1_1', '-DVERSION_GE_1_3', '-DVERSION_GE_1_5']
def cuda_ext_helper(name, sources, extra_cuda_flags, extra_cxx_flags=[]): def cuda_ext_helper(name, sources, extra_cuda_flags, extra_cxx_flags=[]):
return CUDAExtension(name=name, return CUDAExtension(
sources=[os.path.join('colossalai/kernel/cuda_native/csrc', path) for path in sources], name=name,
include_dirs=[os.path.join( sources=[os.path.join('colossalai/kernel/cuda_native/csrc', path) for path in sources],
this_dir, 'colossalai/kernel/cuda_native/csrc/kernels/include')], include_dirs=[os.path.join(this_dir, 'colossalai/kernel/cuda_native/csrc/kernels/include')],
extra_compile_args={'cxx': ['-O3'] + version_dependent_macros + extra_cxx_flags, extra_compile_args={
'nvcc': append_nvcc_threads(['-O3', 'cxx': ['-O3'] + version_dependent_macros + extra_cxx_flags,
'--use_fast_math'] + version_dependent_macros + extra_cuda_flags)}) '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', ext_modules.append(
'multi_tensor_sgd_kernel.cu', cuda_ext_helper('colossal_C', [
'multi_tensor_scale_kernel.cu', 'colossal_C_frontend.cpp', 'multi_tensor_sgd_kernel.cu', 'multi_tensor_scale_kernel.cu',
'multi_tensor_adam.cu', 'multi_tensor_adam.cu', 'multi_tensor_l2norm_kernel.cu', 'multi_tensor_lamb.cu'
'multi_tensor_l2norm_kernel.cu', ], ['-lineinfo']))
'multi_tensor_lamb.cu'],
['-lineinfo']))
cc_flag = ['-gencode', 'arch=compute_70,code=sm_70'] cc_flag = ['-gencode', 'arch=compute_70,code=sm_70']
_, bare_metal_major, _ = get_cuda_bare_metal_version(CUDA_HOME) _, bare_metal_major, _ = get_cuda_bare_metal_version(CUDA_HOME)
...@@ -148,65 +146,58 @@ if build_cuda_ext: ...@@ -148,65 +146,58 @@ if build_cuda_ext:
cc_flag.append('-gencode') cc_flag.append('-gencode')
cc_flag.append('arch=compute_80,code=sm_80') cc_flag.append('arch=compute_80,code=sm_80')
extra_cuda_flags = ['-U__CUDA_NO_HALF_OPERATORS__', extra_cuda_flags = [
'-U__CUDA_NO_HALF_CONVERSIONS__', '-U__CUDA_NO_HALF_OPERATORS__', '-U__CUDA_NO_HALF_CONVERSIONS__', '--expt-relaxed-constexpr',
'--expt-relaxed-constexpr', '--expt-extended-lambda'
'--expt-extended-lambda'] ]
ext_modules.append(cuda_ext_helper('colossal_scaled_upper_triang_masked_softmax', ext_modules.append(
['scaled_upper_triang_masked_softmax.cpp', cuda_ext_helper('colossal_scaled_upper_triang_masked_softmax',
'scaled_upper_triang_masked_softmax_cuda.cu'], ['scaled_upper_triang_masked_softmax.cpp', 'scaled_upper_triang_masked_softmax_cuda.cu'],
extra_cuda_flags + cc_flag)) extra_cuda_flags + cc_flag))
ext_modules.append(cuda_ext_helper('colossal_scaled_masked_softmax', ext_modules.append(
['scaled_masked_softmax.cpp', 'scaled_masked_softmax_cuda.cu'], cuda_ext_helper('colossal_scaled_masked_softmax',
extra_cuda_flags + cc_flag)) ['scaled_masked_softmax.cpp', 'scaled_masked_softmax_cuda.cu'], extra_cuda_flags + cc_flag))
ext_modules.append(cuda_ext_helper('colossal_moe_cuda', ext_modules.append(
['moe_cuda.cpp', 'moe_cuda_kernel.cu'], cuda_ext_helper('colossal_moe_cuda', ['moe_cuda.cpp', 'moe_cuda_kernel.cu'], extra_cuda_flags + cc_flag))
extra_cuda_flags + cc_flag))
extra_cuda_flags = ['-maxrregcount=50'] extra_cuda_flags = ['-maxrregcount=50']
ext_modules.append(cuda_ext_helper('colossal_layer_norm_cuda', ext_modules.append(
['layer_norm_cuda.cpp', 'layer_norm_cuda_kernel.cu'], cuda_ext_helper('colossal_layer_norm_cuda', ['layer_norm_cuda.cpp', 'layer_norm_cuda_kernel.cu'],
extra_cuda_flags + cc_flag)) extra_cuda_flags + cc_flag))
extra_cuda_flags = ['-std=c++14', extra_cuda_flags = [
'-U__CUDA_NO_HALF_OPERATORS__', '-std=c++14', '-U__CUDA_NO_HALF_OPERATORS__', '-U__CUDA_NO_HALF_CONVERSIONS__', '-U__CUDA_NO_HALF2_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__', '-DTHRUST_IGNORE_CUB_VERSION_CHECK'
'-U__CUDA_NO_HALF2_OPERATORS__', ]
'-DTHRUST_IGNORE_CUB_VERSION_CHECK']
ext_modules.append(
ext_modules.append(cuda_ext_helper('colossal_multihead_attention', cuda_ext_helper('colossal_multihead_attention', [
['multihead_attention_1d.cpp', 'multihead_attention_1d.cpp', 'kernels/cublas_wrappers.cu', 'kernels/transform_kernels.cu',
'kernels/cublas_wrappers.cu', 'kernels/dropout_kernels.cu', 'kernels/normalize_kernels.cu', 'kernels/softmax_kernels.cu',
'kernels/transform_kernels.cu', 'kernels/general_kernels.cu', 'kernels/cuda_util.cu'
'kernels/dropout_kernels.cu', ], extra_cuda_flags + cc_flag))
'kernels/normalize_kernels.cu',
'kernels/softmax_kernels.cu',
'kernels/general_kernels.cu',
'kernels/cuda_util.cu'],
extra_cuda_flags + cc_flag))
extra_cxx_flags = ['-std=c++14', '-lcudart', '-lcublas', '-g', '-Wno-reorder', '-fopenmp', '-march=native'] extra_cxx_flags = ['-std=c++14', '-lcudart', '-lcublas', '-g', '-Wno-reorder', '-fopenmp', '-march=native']
ext_modules.append(cuda_ext_helper('cpu_adam', ext_modules.append(cuda_ext_helper('cpu_adam', ['cpu_adam.cpp'], extra_cuda_flags, extra_cxx_flags))
['cpu_adam.cpp'],
extra_cuda_flags,
extra_cxx_flags))
setup( setup(
name='colossalai', name='colossalai',
version=get_version(), version=get_version(),
packages=find_packages(exclude=('benchmark', packages=find_packages(exclude=(
'docker', 'benchmark',
'tests', 'docker',
'docs', 'tests',
'examples', 'docs',
'tests', 'examples',
'scripts', 'tests',
'requirements', 'scripts',
'*.egg-info',)), 'requirements',
'*.egg-info',
)),
description='An integrated large-scale model training system with efficient parallelization techniques', description='An integrated large-scale model training system with efficient parallelization techniques',
long_description=fetch_readme(), long_description=fetch_readme(),
long_description_content_type='text/markdown', long_description_content_type='text/markdown',
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment