2023-03-04 12:08:11 +00:00
|
|
|
# This code has been adapted from the DeepSpeed library.
|
|
|
|
# Copyright (c) Microsoft Corporation.
|
|
|
|
|
|
|
|
# Licensed under the MIT License.
|
2023-01-06 12:50:26 +00:00
|
|
|
import importlib
|
2023-01-04 03:38:42 +00:00
|
|
|
import os
|
2023-01-06 12:50:26 +00:00
|
|
|
import time
|
|
|
|
from abc import ABC, abstractmethod
|
2023-01-04 03:38:42 +00:00
|
|
|
from pathlib import Path
|
2023-04-17 03:25:35 +00:00
|
|
|
from typing import List, Optional
|
2023-01-04 03:38:42 +00:00
|
|
|
|
2023-03-03 13:45:05 +00:00
|
|
|
from .utils import check_cuda_availability, check_system_pytorch_cuda_match, print_rank_0
|
2023-02-03 01:47:13 +00:00
|
|
|
|
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
class Builder(ABC):
|
|
|
|
"""
|
|
|
|
Builder is the base class to build extensions for PyTorch.
|
2023-01-04 03:38:42 +00:00
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
Args:
|
|
|
|
name (str): the name of the kernel to be built
|
|
|
|
prebuilt_import_path (str): the path where the extension is installed during pip install
|
2023-01-04 03:38:42 +00:00
|
|
|
"""
|
2023-01-05 05:53:28 +00:00
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
def __init__(self, name: str, prebuilt_import_path: str):
|
|
|
|
self.name = name
|
|
|
|
self.prebuilt_import_path = prebuilt_import_path
|
2023-09-19 06:20:26 +00:00
|
|
|
self.version_dependent_macros = ["-DVERSION_GE_1_1", "-DVERSION_GE_1_3", "-DVERSION_GE_1_5"]
|
2023-01-06 12:50:26 +00:00
|
|
|
|
2023-03-03 13:45:05 +00:00
|
|
|
# we store the op as an attribute to avoid repeated building and loading
|
|
|
|
self.cached_op_module = None
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
assert prebuilt_import_path.startswith(
|
|
|
|
"colossalai._C"
|
|
|
|
), f"The prebuilt_import_path should start with colossalai._C, but got {self.prebuilt_import_path}"
|
2023-01-06 12:50:26 +00:00
|
|
|
|
|
|
|
def relative_to_abs_path(self, code_path: str) -> str:
|
|
|
|
"""
|
|
|
|
This function takes in a path relative to the colossalai root directory and return the absolute path.
|
|
|
|
"""
|
|
|
|
op_builder_module_path = Path(__file__).parent
|
|
|
|
|
|
|
|
# if we install from source
|
|
|
|
# the current file path will be op_builder/builder.py
|
|
|
|
# if we install via pip install colossalai
|
|
|
|
# the current file path will be colossalai/kernel/op_builder/builder.py
|
|
|
|
# this is because that the op_builder inside colossalai is a symlink
|
|
|
|
# this symlink will be replaced with actual files if we install via pypi
|
|
|
|
# thus we cannot tell the colossalai root directory by checking whether the op_builder
|
|
|
|
# is a symlink, we can only tell whether it is inside or outside colossalai
|
2023-09-19 06:20:26 +00:00
|
|
|
if str(op_builder_module_path).endswith("colossalai/kernel/op_builder"):
|
2023-01-06 12:50:26 +00:00
|
|
|
root_path = op_builder_module_path.parent.parent
|
2023-01-04 03:38:42 +00:00
|
|
|
else:
|
2023-09-19 06:20:26 +00:00
|
|
|
root_path = op_builder_module_path.parent.joinpath("colossalai")
|
2023-01-06 12:50:26 +00:00
|
|
|
|
|
|
|
code_abs_path = root_path.joinpath(code_path)
|
|
|
|
return str(code_abs_path)
|
2023-01-04 03:38:42 +00:00
|
|
|
|
|
|
|
def get_cuda_home_include(self):
|
|
|
|
"""
|
|
|
|
return include path inside the cuda home.
|
|
|
|
"""
|
|
|
|
from torch.utils.cpp_extension import CUDA_HOME
|
2023-09-19 06:20:26 +00:00
|
|
|
|
2023-01-04 03:38:42 +00:00
|
|
|
if CUDA_HOME is None:
|
|
|
|
raise RuntimeError("CUDA_HOME is None, please set CUDA_HOME to compile C++/CUDA kernels in ColossalAI.")
|
|
|
|
cuda_include = os.path.join(CUDA_HOME, "include")
|
|
|
|
return cuda_include
|
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
def csrc_abs_path(self, path):
|
2023-09-19 06:20:26 +00:00
|
|
|
return os.path.join(self.relative_to_abs_path("kernel/cuda_native/csrc"), path)
|
2023-01-06 12:50:26 +00:00
|
|
|
|
2023-01-04 03:38:42 +00:00
|
|
|
# functions must be overrided begin
|
2023-01-06 12:50:26 +00:00
|
|
|
@abstractmethod
|
|
|
|
def sources_files(self) -> List[str]:
|
|
|
|
"""
|
|
|
|
This function should return a list of source files for extensions.
|
|
|
|
"""
|
2023-01-04 03:38:42 +00:00
|
|
|
raise NotImplementedError
|
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
@abstractmethod
|
|
|
|
def include_dirs(self) -> List[str]:
|
|
|
|
"""
|
2023-04-18 11:14:59 +00:00
|
|
|
This function should return a list of include files for extensions.
|
2023-01-06 12:50:26 +00:00
|
|
|
"""
|
2023-01-04 03:38:42 +00:00
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
@abstractmethod
|
|
|
|
def cxx_flags(self) -> List[str]:
|
|
|
|
"""
|
|
|
|
This function should return a list of cxx compilation flags for extensions.
|
|
|
|
"""
|
2023-01-04 03:38:42 +00:00
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
@abstractmethod
|
|
|
|
def nvcc_flags(self) -> List[str]:
|
|
|
|
"""
|
|
|
|
This function should return a list of nvcc compilation flags for extensions.
|
|
|
|
"""
|
2023-01-04 03:38:42 +00:00
|
|
|
|
|
|
|
# functions must be overrided over
|
|
|
|
def strip_empty_entries(self, args):
|
2023-09-19 06:20:26 +00:00
|
|
|
"""
|
2023-01-04 03:38:42 +00:00
|
|
|
Drop any empty strings from the list of compile and link flags
|
2023-09-19 06:20:26 +00:00
|
|
|
"""
|
2023-01-04 03:38:42 +00:00
|
|
|
return [x for x in args if len(x) > 0]
|
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
def import_op(self):
|
|
|
|
"""
|
|
|
|
This function will import the op module by its string name.
|
|
|
|
"""
|
|
|
|
return importlib.import_module(self.prebuilt_import_path)
|
|
|
|
|
2023-03-03 13:45:05 +00:00
|
|
|
def check_runtime_build_environment(self):
|
|
|
|
"""
|
|
|
|
Check whether the system environment is ready for extension compilation.
|
|
|
|
"""
|
|
|
|
try:
|
|
|
|
from torch.utils.cpp_extension import CUDA_HOME
|
2023-09-19 06:20:26 +00:00
|
|
|
|
2023-03-03 13:45:05 +00:00
|
|
|
TORCH_AVAILABLE = True
|
|
|
|
except ImportError:
|
|
|
|
TORCH_AVAILABLE = False
|
|
|
|
CUDA_HOME = None
|
|
|
|
|
|
|
|
if not TORCH_AVAILABLE:
|
|
|
|
raise ModuleNotFoundError(
|
2023-09-19 06:20:26 +00:00
|
|
|
"PyTorch is not found. You need to install PyTorch first in order to build CUDA extensions"
|
|
|
|
)
|
2023-03-03 13:45:05 +00:00
|
|
|
|
|
|
|
if CUDA_HOME is None:
|
|
|
|
raise RuntimeError(
|
2023-04-18 11:14:59 +00:00
|
|
|
"CUDA_HOME is not found. You need to export CUDA_HOME environment variable or install CUDA Toolkit first in order to build CUDA extensions"
|
2023-03-03 13:45:05 +00:00
|
|
|
)
|
|
|
|
|
|
|
|
# make sure CUDA is available for compilation during
|
|
|
|
cuda_available = check_cuda_availability()
|
|
|
|
if not cuda_available:
|
2023-04-18 11:14:59 +00:00
|
|
|
raise RuntimeError("CUDA is not available on your system as torch.cuda.is_available() returns False.")
|
2023-03-03 13:45:05 +00:00
|
|
|
|
|
|
|
# make sure system CUDA and pytorch CUDA match, an error will raised inside the function if not
|
|
|
|
check_system_pytorch_cuda_match(CUDA_HOME)
|
|
|
|
|
2023-04-17 03:25:35 +00:00
|
|
|
def load(self, verbose: Optional[bool] = None):
|
2023-01-04 03:38:42 +00:00
|
|
|
"""
|
2023-01-06 12:50:26 +00:00
|
|
|
load the kernel during runtime. If the kernel is not built during pip install, it will build the kernel.
|
|
|
|
If the kernel is built during runtime, it will be stored in `~/.cache/colossalai/torch_extensions/`. If the
|
|
|
|
kernel is built during pip install, it can be accessed through `colossalai._C`.
|
2023-01-04 03:38:42 +00:00
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
Warning: do not load this kernel repeatedly during model execution as it could slow down the training process.
|
2023-01-04 03:38:42 +00:00
|
|
|
|
|
|
|
Args:
|
|
|
|
verbose (bool, optional): show detailed info. Defaults to True.
|
|
|
|
"""
|
2023-04-17 03:25:35 +00:00
|
|
|
if verbose is None:
|
2023-09-19 06:20:26 +00:00
|
|
|
verbose = os.environ.get("CAI_KERNEL_VERBOSE", "0") == "1"
|
2023-03-03 13:45:05 +00:00
|
|
|
# if the kernel has be compiled and cached, we directly use it
|
|
|
|
if self.cached_op_module is not None:
|
|
|
|
return self.cached_op_module
|
2023-01-04 03:38:42 +00:00
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
try:
|
2023-03-03 13:45:05 +00:00
|
|
|
# if the kernel has been pre-built during installation
|
|
|
|
# we just directly import it
|
2023-01-06 12:50:26 +00:00
|
|
|
op_module = self.import_op()
|
|
|
|
if verbose:
|
2023-03-03 13:45:05 +00:00
|
|
|
print_rank_0(
|
2023-09-19 06:20:26 +00:00
|
|
|
f"[extension] OP {self.prebuilt_import_path} has been compiled ahead of time, skip building."
|
|
|
|
)
|
2023-01-06 12:50:26 +00:00
|
|
|
except ImportError:
|
2023-03-03 13:45:05 +00:00
|
|
|
# check environment
|
|
|
|
self.check_runtime_build_environment()
|
|
|
|
|
|
|
|
# time the kernel compilation
|
|
|
|
start_build = time.time()
|
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
# construct the build directory
|
|
|
|
import torch
|
2023-03-03 13:45:05 +00:00
|
|
|
from torch.utils.cpp_extension import load
|
2023-09-19 06:20:26 +00:00
|
|
|
|
|
|
|
torch_version_major = torch.__version__.split(".")[0]
|
|
|
|
torch_version_minor = torch.__version__.split(".")[1]
|
2023-01-06 12:50:26 +00:00
|
|
|
torch_cuda_version = torch.version.cuda
|
2023-09-19 06:20:26 +00:00
|
|
|
home_directory = os.path.expanduser("~")
|
2023-01-06 12:50:26 +00:00
|
|
|
extension_directory = f".cache/colossalai/torch_extensions/torch{torch_version_major}.{torch_version_minor}_cu{torch_cuda_version}"
|
|
|
|
build_directory = os.path.join(home_directory, extension_directory)
|
|
|
|
Path(build_directory).mkdir(parents=True, exist_ok=True)
|
|
|
|
|
|
|
|
if verbose:
|
2023-03-03 13:45:05 +00:00
|
|
|
print_rank_0(f"[extension] Compiling or loading the JIT-built {self.name} kernel during runtime now")
|
2023-01-06 12:50:26 +00:00
|
|
|
|
|
|
|
# load the kernel
|
2023-09-19 06:20:26 +00:00
|
|
|
op_module = load(
|
|
|
|
name=self.name,
|
|
|
|
sources=self.strip_empty_entries(self.sources_files()),
|
|
|
|
extra_include_paths=self.strip_empty_entries(self.include_dirs()),
|
|
|
|
extra_cflags=self.cxx_flags(),
|
|
|
|
extra_cuda_cflags=self.nvcc_flags(),
|
|
|
|
extra_ldflags=[],
|
|
|
|
build_directory=build_directory,
|
|
|
|
verbose=verbose,
|
|
|
|
)
|
2023-01-04 03:38:42 +00:00
|
|
|
|
2023-03-03 13:45:05 +00:00
|
|
|
build_duration = time.time() - start_build
|
|
|
|
|
|
|
|
# log jit compilation time
|
|
|
|
if verbose:
|
|
|
|
print_rank_0(f"[extension] Time to compile or load {self.name} op: {build_duration} seconds")
|
|
|
|
|
|
|
|
# cache the built/loaded kernel
|
|
|
|
self.cached_op_module = op_module
|
2023-01-04 03:38:42 +00:00
|
|
|
|
|
|
|
return op_module
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
def builder(self) -> "CUDAExtension":
|
2023-01-04 03:38:42 +00:00
|
|
|
"""
|
|
|
|
get a CUDAExtension instance used for setup.py
|
|
|
|
"""
|
|
|
|
from torch.utils.cpp_extension import CUDAExtension
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
return CUDAExtension(
|
|
|
|
name=self.prebuilt_import_path,
|
|
|
|
sources=self.strip_empty_entries(self.sources_files()),
|
|
|
|
include_dirs=self.strip_empty_entries(self.include_dirs()),
|
|
|
|
extra_compile_args={
|
|
|
|
"cxx": self.strip_empty_entries(self.cxx_flags()),
|
|
|
|
"nvcc": self.strip_empty_entries(self.nvcc_flags()),
|
|
|
|
},
|
|
|
|
)
|