Source code for torch
"""
The torch package contains data structures for multi-dimensional
tensors and defines mathematical operations over these tensors.
Additionally, it provides many utilities for efficient serialization of
Tensors and arbitrary types, and other useful utilities.
It has a CUDA counterpart, that enables you to run your tensor computations
on an NVIDIA GPU with compute capability >= 3.0.
"""
# mypy: allow-untyped-defs
import builtins
import ctypes
import glob
import importlib
import inspect
import math
import os
import platform
import sys
import textwrap
import threading
from typing import (
Any as _Any,
Callable as _Callable,
Dict as _Dict,
get_origin as _get_origin,
Optional as _Optional,
overload as _overload,
Set as _Set,
Tuple as _Tuple,
Type as _Type,
TYPE_CHECKING,
TypeVar as _TypeVar,
Union as _Union,
)
from typing_extensions import ParamSpec as _ParamSpec
if TYPE_CHECKING:
from .types import IntLikeType
# multipy/deploy is setting this import before importing torch, this is the most
# reliable way we have to detect if we're running within deploy.
# https://github.com/pytorch/multipy/blob/d60f34ad38c371e441fe7ffdb77a3c3dda5a5d19/multipy/runtime/interpreter/interpreter_impl.cpp#L134-L137
def _running_with_deploy() -> builtins.bool:
return sys.modules.get("torch._meta_registrations", None) is object
from torch._utils import (
_functionalize_sync as _sync,
_import_dotted_name,
classproperty,
)
from torch._utils_internal import (
get_file_path,
prepare_multiprocessing_environment,
USE_GLOBAL_DEPS,
USE_RTLD_GLOBAL_WITH_LIBTORCH,
)
# TODO(torch_deploy) figure out how to freeze version.py in fbcode build
if _running_with_deploy():
__version__ = "torch-deploy-1.8"
# TODO: Remove this ugly hack when deploy typing extensions are updated to 4.10+
if not TYPE_CHECKING:
import typing_extensions
_TypeIs = typing_extensions.TypeGuard
typing_extensions.TypeIs = _TypeIs
else:
from typing_extensions import TypeIs as _TypeIs
from torch.torch_version import __version__ as __version__
__all__ = [
"BoolStorage",
"BoolTensor",
"ByteStorage",
"ByteTensor",
"CharStorage",
"CharTensor",
"DoubleStorage",
"DoubleTensor",
"FloatStorage",
"FloatTensor",
"GradScaler",
"IntStorage",
"IntTensor",
"LongStorage",
"LongTensor",
"ShortStorage",
"ShortTensor",
"SymBool",
"SymFloat",
"SymInt",
"Tensor",
"TypedStorage",
"UntypedStorage",
"are_deterministic_algorithms_enabled",
"autocast",
"chunk",
"compile",
"cond",
"enable_grad",
"export",
"get_default_device",
"get_deterministic_debug_mode",
"get_device_module",
"get_float32_matmul_precision",
"get_rng_state",
"inference_mode",
"initial_seed",
"is_deterministic_algorithms_warn_only_enabled",
"is_storage",
"is_tensor",
"is_warn_always_enabled",
"load",
"lobpcg",
"manual_seed",
"matmul",
"no_grad",
"rand",
"randn",
"save",
"seed",
"set_default_device",
"set_default_tensor_type",
"set_deterministic_debug_mode",
"set_float32_matmul_precision",
"set_printoptions",
"set_rng_state",
"set_warn_always",
"split",
"stack",
"sym_float",
"sym_fresh_size",
"sym_int",
"sym_ite",
"sym_max",
"sym_min",
"sym_not",
"sym_sum",
"typename",
"unravel_index",
"use_deterministic_algorithms",
"vmap",
]
# Please keep this list sorted
assert __all__ == sorted(__all__)
################################################################################
# Load the extension module
################################################################################
if sys.platform == "win32":
def _load_dll_libraries() -> None:
import sysconfig
from torch.version import cuda as cuda_version
pfiles_path = os.getenv("ProgramFiles", r"C:\Program Files")
py_dll_path = os.path.join(sys.exec_prefix, "Library", "bin")
th_dll_path = os.path.join(os.path.dirname(__file__), "lib")
usebase_path = os.path.join(
sysconfig.get_config_var("userbase"), "Library", "bin"
)
# When users create a virtualenv that inherits the base environment,
# we will need to add the corresponding library directory into
# DLL search directories. Otherwise, it will rely on `PATH` which
# is dependent on user settings.
if sys.exec_prefix != sys.base_exec_prefix:
base_py_dll_path = os.path.join(sys.base_exec_prefix, "Library", "bin")
else:
base_py_dll_path = ""
dll_paths = [
p
for p in (th_dll_path, py_dll_path, base_py_dll_path, usebase_path)
if os.path.exists(p)
]
if not builtins.any(
os.path.exists(os.path.join(p, "nvToolsExt64_1.dll")) for p in dll_paths
):
nvtoolsext_dll_path = os.path.join(
os.getenv(
"NVTOOLSEXT_PATH",
os.path.join(pfiles_path, "NVIDIA Corporation", "NvToolsExt"),
),
"bin",
"x64",
)
else:
nvtoolsext_dll_path = ""
if cuda_version and builtins.all(
not glob.glob(os.path.join(p, "cudart64*.dll")) for p in dll_paths
):
cuda_version_1 = cuda_version.replace(".", "_")
cuda_path_var = "CUDA_PATH_V" + cuda_version_1
default_path = os.path.join(
pfiles_path, "NVIDIA GPU Computing Toolkit", "CUDA", f"v{cuda_version}"
)
cuda_path = os.path.join(os.getenv(cuda_path_var, default_path), "bin")
else:
cuda_path = ""
dll_paths.extend(
p for p in (nvtoolsext_dll_path, cuda_path) if os.path.exists(p)
)
kernel32 = ctypes.WinDLL("kernel32.dll", use_last_error=True)
with_load_library_flags = hasattr(kernel32, "AddDllDirectory")
prev_error_mode = kernel32.SetErrorMode(0x0001)
kernel32.LoadLibraryW.restype = ctypes.c_void_p
if with_load_library_flags:
kernel32.LoadLibraryExW.restype = ctypes.c_void_p
for dll_path in dll_paths:
os.add_dll_directory(dll_path)
try:
ctypes.CDLL("vcruntime140.dll")
ctypes.CDLL("msvcp140.dll")
if platform.machine() != "ARM64":
ctypes.CDLL("vcruntime140_1.dll")
except OSError:
print(
textwrap.dedent(
"""
Microsoft Visual C++ Redistributable is not installed, this may lead to the DLL load failure.
It can be downloaded at https://aka.ms/vs/16/release/vc_redist.x64.exe
"""
).strip()
)
dlls = glob.glob(os.path.join(th_dll_path, "*.dll"))
path_patched = False
for dll in dlls:
is_loaded = False
if with_load_library_flags:
res = kernel32.LoadLibraryExW(dll, None, 0x00001100)
last_error = ctypes.get_last_error()
if res is None and last_error != 126:
err = ctypes.WinError(last_error)
err.strerror += (
f' Error loading "{dll}" or one of its dependencies.'
)
raise err
elif res is not None:
is_loaded = True
if not is_loaded:
if not path_patched:
os.environ["PATH"] = ";".join(dll_paths + [os.environ["PATH"]])
path_patched = True
res = kernel32.LoadLibraryW(dll)
if res is None:
err = ctypes.WinError(ctypes.get_last_error())
err.strerror += (
f' Error loading "{dll}" or one of its dependencies.'
)
raise err
kernel32.SetErrorMode(prev_error_mode)
_load_dll_libraries()
del _load_dll_libraries
def _preload_cuda_deps(lib_folder: str, lib_name: str) -> None:
"""Preloads cuda deps if they could not be found otherwise."""
# Should only be called on Linux if default path resolution have failed
assert platform.system() == "Linux", "Should only be called on Linux"
lib_path = None
for path in sys.path:
nvidia_path = os.path.join(path, "nvidia")
if not os.path.exists(nvidia_path):
continue
candidate_lib_paths = glob.glob(
os.path.join(nvidia_path, lib_folder, "lib", lib_name)
)
if candidate_lib_paths and not lib_path:
lib_path = candidate_lib_paths[0]
if lib_path:
break
if not lib_path:
raise ValueError(f"{lib_name} not found in the system path {sys.path}")
ctypes.CDLL(lib_path)
# See Note [Global dependencies]
def _load_global_deps() -> None:
if _running_with_deploy() or platform.system() == "Windows":
return
# Determine the file extension based on the platform
lib_ext = ".dylib" if platform.system() == "Darwin" else ".so"
lib_name = f"libtorch_global_deps{lib_ext}"
here = os.path.abspath(__file__)
global_deps_lib_path = os.path.join(os.path.dirname(here), "lib", lib_name)
try:
ctypes.CDLL(global_deps_lib_path, mode=ctypes.RTLD_GLOBAL)
except OSError as err:
# Can only happen for wheel with cuda libs as PYPI deps
# As PyTorch is not purelib, but nvidia-*-cu12 is
cuda_libs: _Dict[str, str] = {
"cublas": "libcublas.so.*[0-9]",
"cudnn": "libcudnn.so.*[0-9]",
"cuda_nvrtc": "libnvrtc.so.*[0-9]",
"cuda_runtime": "libcudart.so.*[0-9]",
"cuda_cupti": "libcupti.so.*[0-9]",
"cufft": "libcufft.so.*[0-9]",
"curand": "libcurand.so.*[0-9]",
"nvjitlink": "libnvJitLink.so.*[0-9]",
"cusparse": "libcusparse.so.*[0-9]",
"cusolver": "libcusolver.so.*[0-9]",
"nccl": "libnccl.so.*[0-9]",
"nvtx": "libnvToolsExt.so.*[0-9]",
}
is_cuda_lib_err = [
lib for lib in cuda_libs.values() if lib.split(".")[0] in err.args[0]
]
if not is_cuda_lib_err:
raise err
for lib_folder, lib_name in cuda_libs.items():
_preload_cuda_deps(lib_folder, lib_name)
ctypes.CDLL(global_deps_lib_path, mode=ctypes.RTLD_GLOBAL)
if (USE_RTLD_GLOBAL_WITH_LIBTORCH or os.getenv("TORCH_USE_RTLD_GLOBAL")) and (
_running_with_deploy() or platform.system() != "Windows"
):
# Do it the hard way. You might want to load libtorch with RTLD_GLOBAL in a
# few circumstances:
#
# 1. You're in a build environment (e.g., fbcode) where
# libtorch_global_deps is not available, but you still need
# to get mkl to link in with RTLD_GLOBAL or it will just
# not work.
#
# 2. You're trying to run PyTorch under UBSAN and you need
# to ensure that only one copy of libtorch is loaded, so
# vptr checks work properly
#
# If you're using this setting, you must verify that all the libraries
# you load consistently use the same libstdc++, or you may have
# mysterious segfaults.
#
old_flags = sys.getdlopenflags()
sys.setdlopenflags(os.RTLD_GLOBAL | os.RTLD_LAZY)
from torch._C import * # noqa: F403
sys.setdlopenflags(old_flags)
del old_flags
else:
# Easy way. You want this most of the time, because it will prevent
# C++ symbols from libtorch clobbering C++ symbols from other
# libraries, leading to mysterious segfaults.
#
# If building in an environment where libtorch_global_deps isn't available
# like parts of fbsource, but where RTLD_GLOBAL causes segfaults, you will
# want USE_RTLD_GLOBAL_WITH_LIBTORCH = False and USE_GLOBAL_DEPS = False
#
# See Note [Global dependencies]
if USE_GLOBAL_DEPS:
_load_global_deps()
from torch._C import * # noqa: F403
[docs]class SymInt:
"""
Like an int (including magic methods), but redirects all operations on the
wrapped node. This is used in particular to symbolically record operations
in the symbolic shape workflow.
"""
def __init__(self, node):
# This field MUST be named node; C++ binding code assumes that this
# class has a field named node that stores SymNode
self.node = node
def __bool__(self):
return builtins.bool(self != 0)
def __int__(self):
return self.node.int_()
def __index__(self):
return self.node.int_()
# Magic methods installed by torch.fx.experimental.sym_node
def __round__(self, ndigits=None):
return self
def __truediv__(self, other):
if isinstance(other, (builtins.float, SymFloat)):
return sym_float(self).__float_truediv__(other)
if not isinstance(other, (builtins.int, SymInt)):
return NotImplemented
return self.__int_truediv__(other)
def __rtruediv__(self, other):
if isinstance(other, (builtins.float, SymFloat)):
return sym_float(self).__rfloat_truediv__(other)
if not isinstance(other, (builtins.int, SymInt)):
return NotImplemented
return self.__rint_truediv__(other)
def __floordiv__(self, other):
if isinstance(other, (builtins.float, SymFloat)):
return sym_float(math.floor(sym_float(self) / other))
if not isinstance(other, (builtins.int, SymInt)):
return NotImplemented
return self.__int_floordiv__(other)
def __rfloordiv__(self, other):
if isinstance(other, (builtins.float, SymFloat)):
return sym_float(math.floor(other / sym_float(self)))
if not isinstance(other, (builtins.int, SymInt)):
return NotImplemented
return self.__rint_floordiv__(other)
# nb: complex is impossible to handle correctly lol, with
# negative base and integral float need to diverge semantics and
# just always return complex. Neener neener pretend this problem
# doesn't exist
def __pow__(self, other):
if isinstance(other, (builtins.float, SymFloat)):
return sym_float(self).__pow__(other)
if not isinstance(other, (builtins.int, SymInt)):
return NotImplemented
# Guards! This guard is necessary because we need to know it to
# determine the output type of this operation
if other >= 0:
return self.__pow_by_natural__(other)
else:
# Mercifully, when the exponent is negative, Python just promotes
# to doubles and does a float pow:
#
# if (Py_SIZE(b) < 0 && c == NULL) {
# /* if exponent is negative and there's no modulus:
# return a float. This works because we know
# that this calls float_pow() which converts its
# arguments to double. */
# Py_DECREF(a);
# Py_DECREF(b);
# return PyFloat_Type.tp_as_number->nb_power(v, w, x);
# }
return sym_float(self).__pow__(sym_float(other))
def __rpow__(self, other):
if isinstance(other, (builtins.float, SymFloat)):
return sym_float(self).__rpow__(other)
if not isinstance(other, (builtins.int, SymInt)):
return NotImplemented
if self >= 0: # self is exponent
return self.__rpow_by_natural__(other)
else:
return sym_float(self).__rpow__(sym_float(other))
def __eq__(self, other: object) -> builtins.bool:
raise TypeError("type stub not overridden")
def __lt__(self, other) -> builtins.bool:
raise TypeError("type stub not overridden")
def __gt__(self, other) -> builtins.bool:
raise TypeError("type stub not overridden")
def __le__(self, other) -> builtins.bool:
raise TypeError("type stub not overridden")
def __ge__(self, other) -> builtins.bool:
raise TypeError("type stub not overridden")
def __add__(self, other) -> "SymInt":
raise TypeError("type stub not overridden")
def __radd__(self, other) -> "SymInt":
raise TypeError("type stub not overridden")
def __rmul__(self, other) -> "SymInt":
raise TypeError("type stub not overridden")
def __mod__(self, other: "IntLikeType") -> "SymInt":
raise TypeError("type stub not overridden")
def __mul__(self, other) -> "SymInt":
raise TypeError("type stub not overridden")
def __pow_by_natural__(self, other) -> "SymInt":
raise TypeError("type stub not overridden")
def __rpow_by_natural__(self, other) -> "SymInt":
raise TypeError("type stub not overridden")
def __int_truediv__(self, other) -> "SymFloat":
raise TypeError("type stub not overridden")
def __rint_truediv__(self, other) -> "SymFloat":
raise TypeError("type stub not overridden")
def __int_floordiv__(self, other) -> "SymFloat":
raise TypeError("type stub not overridden")
def __rint_floordiv__(self, other) -> "SymFloat":
raise TypeError("type stub not overridden")
def __sym_max__(self, other):
raise TypeError("type stub not overridden")
def __sym_min__(self, other):
raise TypeError("type stub not overridden")
def __sym_float__(self):
raise TypeError("type stub not overridden")
def __neg__(self):
raise TypeError("type stub not overridden")
def __sub__(self, other: "IntLikeType") -> "SymInt":
raise TypeError("type stub not overridden")
def __rsub__(self, other: "IntLikeType") -> "SymInt":
raise TypeError("type stub not overridden")
def __and__(self, other) -> "SymInt":
raise TypeError("type stub not overridden")
def __or__(self, other) -> "SymInt":
raise TypeError("type stub not overridden")
def __repr__(self):
return self.node._graph_repr()
def _sympy_(self):
return self.node.expr
def __hash__(self) -> builtins.int:
if self.node.is_nested_int():
return hash(self.node.nested_int())
else:
# We could support constant SymInts as well, but not doing it for now
raise TypeError("unhashable type: non-nested SymInt")
# TODO: Force specialization
# This can't be done because the TypeError here is load bearing
# for einops
# https://github.com/arogozhnikov/einops/blob/6181e1e95dc58c00a3143c1726da1c6ee0463164/einops/einops.py#L237
# return hash(builtins.int(self))
[docs] def as_integer_ratio(self) -> _Tuple["SymInt", builtins.int]:
"""Represent this int as an exact integer ratio"""
return self, 1
def bit_length(self) -> builtins.int:
# TODO: A more relaxed guard is possible here, where you guard to
# allow all integer quantities which would result in the same bit
# length. We can also just make a dedicated Sympy function for
# computing this quantity and represent it symbolically.
return builtins.int(self).bit_length()
def conjugate(self) -> "SymInt":
return self
[docs]class SymFloat:
"""
Like an float (including magic methods), but redirects all operations on the
wrapped node. This is used in particular to symbolically record operations
in the symbolic shape workflow.
"""
def __init__(self, node):
# This field MUST be named node; C++ binding code assumes that this
# class has a field named node that stores SymNode
self.node = node
def __truediv__(self, other):
if not isinstance(other, (builtins.int, builtins.float, SymInt, SymFloat)):
return NotImplemented
return self.__float_truediv__(sym_float(other))
def __rtruediv__(self, other):
if not isinstance(other, (builtins.int, builtins.float, SymInt, SymFloat)):
return NotImplemented
return self.__rfloat_truediv__(sym_float(other))
def __floordiv__(self, other):
if not isinstance(other, (builtins.int, builtins.float, SymInt, SymFloat)):
return NotImplemented
return sym_float(math.floor(self / sym_float(other)))
def __rfloordiv__(self, other):
if not isinstance(other, (builtins.int, builtins.float, SymInt, SymFloat)):
return NotImplemented
return sym_float(math.floor(sym_float(other) / self))
def __bool__(self):
return self.node.bool_()
def __float__(self):
return self.node.guard_float("", 0)
# Symbolic power does NOT work with negative base, this is to avoid
# potential complex outputs
def __pow__(self, other):
if not isinstance(other, (builtins.int, builtins.float, SymInt, SymFloat)):
return NotImplemented
torch._check(self >= 0)
return self.__float_pow__(other)
def __rpow__(self, other):
if not isinstance(other, (builtins.int, builtins.float, SymInt, SymFloat)):
return NotImplemented
torch._check(other >= 0)
return self.__rfloat_pow__(other)
# Magic methods installed by torch.fx.experimental.sym_node
def __eq__(self, other: object) -> builtins.bool:
raise TypeError("type stub not overridden")
def __lt__(self, other) -> builtins.bool:
raise TypeError("type stub not overridden")
def __gt__(self, other) -> builtins.bool:
raise TypeError("type stub not overridden")
def __le__(self, other) -> builtins.bool:
raise TypeError("type stub not overridden")
def __ge__(self, other) -> builtins.bool:
raise TypeError("type stub not overridden")
def __float_pow__(self, other) -> "SymFloat":
raise TypeError("type stub not overridden")
def __rfloat_pow__(self, other) -> "SymFloat":
raise TypeError("type stub not overridden")
def __float_truediv__(self, other) -> "SymFloat":
raise TypeError("type stub not overridden")
def __rfloat_truediv__(self, other) -> "SymFloat":
raise TypeError("type stub not overridden")
def __trunc__(self):
raise TypeError("type stub not overridden")
def __sym_max__(self, other):
raise TypeError("type stub not overridden")
def __sym_min__(self, other):
raise TypeError("type stub not overridden")
def __sym_int__(self):
raise TypeError("type stub not overridden")
[docs] def is_integer(self):
"""Return True if the float is an integer."""
raise TypeError("type stub not overridden")
[docs] def as_integer_ratio(self) -> _Tuple[builtins.int, builtins.int]:
"""Represent this float as an exact integer ratio"""
return builtins.float(self).as_integer_ratio()
def __repr__(self):
return self.node._graph_repr()
def _sympy_(self):
return self.node.expr
def __hash__(self):
return hash(builtins.float(self))
[docs] def conjugate(self) -> "SymFloat":
"""Returns the complex conjugate of the float."""
return self
[docs] def hex(self) -> str:
"""Returns the hexadecimal representation of the float."""
return self.node.guard_float("", 0).hex()
[docs]class SymBool:
"""
Like an bool (including magic methods), but redirects all operations on the
wrapped node. This is used in particular to symbolically record operations
in the symbolic shape workflow.
Unlike regular bools, regular boolean operators will force extra guards instead
of symbolically evaluate. Use the bitwise operators instead to handle this.
"""
def __init__(self, node):
# This field MUST be named node; C++ binding code assumes that this
# class has a field named node that stores SymNode
self.node = node
def __bool__(self):
return self.node.bool_()
def __int__(self):
return builtins.int(self.node.bool_())
# Magic methods installed by torch.fx.experimental.sym_node
def __and__(self, other) -> "SymBool":
raise TypeError("type stub not overridden")
def __or__(self, other) -> "SymBool":
raise TypeError("type stub not overridden")
# We very carefully define __sym_not__, and not a number of other
# plausible alternatives:
#
# - We do not override __not__ because this is not a real magic
# method; you cannot override the meaning of the not builtin in
# Python. We use the name 'sym_not' to clarify that in user code you
# cannot use the builtin not or operator.not_ or operator.__not__ and
# hit this magic method; you must use our custom sym_not operator.
#
# - We do not override the __invert__ method because SymBool is
# meant to be usable in situations where bool is expected. However,
# bitwise negation ~a does the wrong thing with booleans (because
# bool is a subclass of int, so ~1 = -2 which is not falseish.)
# This would be a giant footgun, so we get around it by defining
# our own operator. Note that bitwise and/or do the right thing,
# so we reuse the conventional operators there for readability.
#
def __sym_not__(self) -> "SymBool":
raise TypeError("type stub not overridden")
def __sym_ite__(self, then_val, else_val):
raise TypeError("type stub not overridden")
def __eq__(self, other) -> builtins.bool:
raise TypeError("type stub not overridden")
def __repr__(self):
return self.node._graph_repr()
def _sympy_(self):
return self.node.expr
def __hash__(self):
if self.node.is_constant():
return hash(self.node.bool_())
else:
# Force specialization
return hash(builtins.bool(self))
[docs]def sym_not(a):
r"""SymInt-aware utility for logical negation.
Args:
a (SymBool or bool): Object to negate
"""
import sympy
if overrides.has_torch_function_unary(a):
return overrides.handle_torch_function(sym_not, (a,), a)
if hasattr(a, "__sym_not__"):
return a.__sym_not__()
if isinstance(a, sympy.Basic):
return ~a # type: ignore[operator]
return not a
[docs]def sym_float(a):
r"""SymInt-aware utility for float casting.
Args:
a (SymInt, SymFloat, or object): Object to cast
"""
if overrides.has_torch_function_unary(a):
return overrides.handle_torch_function(sym_float, (a,), a)
if isinstance(a, SymFloat):
return a
elif hasattr(a, "__sym_float__"):
return a.__sym_float__()
return builtins.float(a) # type: ignore[operator]
[docs]def sym_int(a):
r"""SymInt-aware utility for int casting.
Args:
a (SymInt, SymFloat, or object): Object to cast
"""
if overrides.has_torch_function_unary(a):
return overrides.handle_torch_function(sym_int, (a,), a)
if isinstance(a, SymInt):
return a
elif isinstance(a, SymFloat):
return math.trunc(a)
return builtins.int(a) # type: ignore[operator]
[docs]def sym_max(a, b):
"""
SymInt-aware utility for max which avoids branching on a < b.
Unlike builtins.max(), this only works for int/float, and it always
promotes to float if any argument is float (unlike builtins.max, which
will faithfully preserve the type of the input argument).
"""
if overrides.has_torch_function((a, b)):
return overrides.handle_torch_function(sym_max, (a, b), a, b)
if isinstance(a, (SymInt, SymFloat)):
return a.__sym_max__(b)
elif isinstance(b, (SymInt, SymFloat)):
# Due to promotion semantics, this is operator is commutative:
# max(1, 1.0) === max(1.0, 1) === 1.0
return b.__sym_max__(a)
# TODO: Probably can make bool work too, just lazy
all_types, float_types = __all_and_float_types()
assert isinstance(a, all_types), type(a)
assert isinstance(b, all_types), type(b)
if isinstance(a, float_types) or isinstance(b, float_types):
return builtins.float(builtins.max(a, b))
else:
return builtins.max(a, b)
def __all_and_float_types() -> _Tuple[_Tuple[_Type, ...], _Tuple[_Type, ...]]:
try:
import numpy as np
all_types: _Tuple[_Type, ...] = (
np.integer,
np.floating,
builtins.int,
builtins.float,
)
float_types: _Tuple[_Type, ...] = (np.floating, builtins.float)
except ModuleNotFoundError:
all_types = (builtins.int, builtins.float)
float_types = (builtins.float,)
return all_types, float_types
[docs]def sym_min(a, b):
"""SymInt-aware utility for min()."""
if overrides.has_torch_function((a, b)):
return overrides.handle_torch_function(sym_min, (a, b), a, b)
if isinstance(a, (SymInt, SymFloat)):
return a.__sym_min__(b)
elif isinstance(b, (SymInt, SymFloat)):
return b.__sym_min__(a)
all_types, float_types = __all_and_float_types()
assert isinstance(a, all_types), type(a)
assert isinstance(b, all_types), type(b)
if isinstance(a, float_types) or isinstance(b, float_types):
return builtins.float(builtins.min(a, b))
else:
return builtins.min(a, b)
[docs]def sym_sum(args):
"""
N-ary add which is faster to compute for long lists than iterated binary
addition. Only does something special for integers.
"""
if overrides.has_torch_function(args):
return overrides.handle_torch_function(sym_sum, args, args)
found = None
for a in args:
if not isinstance(a, (SymInt, builtins.int)):
return builtins.sum(args)
if isinstance(a, SymInt):
found = a.node
if found is None:
return builtins.sum(args)
from torch.fx.experimental.sym_node import to_node, wrap_node
return wrap_node(found.sym_sum(tuple(to_node(found, a) for a in args)))
# Drop in replacement for math.sqrt, math.sin, math.cos etc
def _get_sym_math_fn(name):
def fn(a):
if overrides.has_torch_function_unary(a):
return overrides.handle_torch_function(fn, (a,), a)
if isinstance(a, SymInt):
a = torch.sym_float(a)
if hasattr(a, f"__sym_{name}__"):
return getattr(a, f"__sym_{name}__")()
return getattr(math, name)(a)
return fn
__fn, __name, __sym_name = None, "", ""
for __name in (
"sqrt",
"cos",
"cosh",
"sin",
"sinh",
"tan",
"tanh",
"asin",
"acos",
"atan",
"log2",
):
__sym_name = f"_sym_{__name}"
__fn = _get_sym_math_fn(__name)
__fn.__qualname__ = __fn.__name__ = __sym_name
globals()[__sym_name] = __fn
del __fn, __name, __sym_name, _get_sym_math_fn
# Adding temporary shortcut
sym_sqrt = globals()["_sym_sqrt"]
__all__.append("sym_sqrt")
[docs]def sym_ite(b, t, f):
if overrides.has_torch_function((b, t, f)):
return overrides.handle_torch_function(sym_ite, (b, t, f), b, t, f)
assert isinstance(b, (SymBool, builtins.bool)) and type(t) == type(f)
if isinstance(b, SymBool):
return b.__sym_ite__(t, f)
return t if b else f
# Create a fresh unbacked int, from an (possibly unbacked int) expression.
# Check to see if we can load C extensions, and if not provide some guidance
# on what the problem might be.
try:
# _initExtension is chosen (arbitrarily) as a sentinel.
from torch._C import _initExtension
except ImportError:
import torch._C as _C_for_compiled_check
# The __file__ check only works for Python 3.7 and above.
if _C_for_compiled_check.__file__ is None:
raise ImportError(
textwrap.dedent(
"""
Failed to load PyTorch C extensions:
It appears that PyTorch has loaded the `torch/_C` folder
of the PyTorch repository rather than the C extensions which
are expected in the `torch._C` namespace. This can occur when
using the `install` workflow. e.g.
$ python setup.py install && python -c "import torch"
This error can generally be solved using the `develop` workflow
$ python setup.py develop && python -c "import torch" # This should succeed
or by running Python from a different directory.
"""
).strip()
) from None
raise # If __file__ is not None the cause is unknown, so just re-raise.
# The torch._C submodule is already loaded via `from torch._C import *` above
# Make an explicit reference to the _C submodule to appease linters
from torch import _C as _C
__name, __obj = "", None
for __name in dir(_C):
if __name[0] != "_" and not __name.endswith("Base"):
__all__.append(__name)
__obj = getattr(_C, __name)
if callable(__obj) or inspect.isclass(__obj):
if __obj.__module__ != __name__: # "torch"
# TODO: fix their module from C++ side
if __name not in {
"DisableTorchFunctionSubclass",
"DisableTorchFunction",
"Generator",
}:
__obj.__module__ = __name__ # "torch"
elif __name == "TensorBase":
# issue 109438 / pr 109940. Prevent TensorBase from being copied into torch.
delattr(sys.modules[__name__], __name)
del __name, __obj
if not TYPE_CHECKING:
# issue 38137 and python issue 43367. Submodules of a C extension are
# non-standard, and attributes of those submodules cannot be pickled since
# pickle expect to be able to import them as "from _C.sub import attr"
# which fails with "_C is not a package
def _import_extension_to_sys_modules(module, memo=None):
if memo is None:
memo = set()
if module in memo:
return
memo.add(module)
module_name = module.__name__
for name in dir(module):
member = getattr(module, name)
member_name = getattr(member, "__name__", "")
if inspect.ismodule(member) and member_name.startswith(module_name):
sys.modules.setdefault(member_name, member)
# Recurse for submodules (e.g., `_C._dynamo.eval_frame`)
_import_extension_to_sys_modules(member, memo)
_import_extension_to_sys_modules(_C)
del _import_extension_to_sys_modules
################################################################################
# Define basic utilities
################################################################################
def typename(obj: _Any, /) -> str:
"""
String representation of the type of an object.
This function returns a fully qualified string representation of an object's type.
Args:
obj (object): The object whose type to represent
Returns:
str: the type of the object `o`
Example:
>>> x = torch.tensor([1, 2, 3])
>>> torch.typename(x)
'torch.LongTensor'
>>> torch.typename(torch.nn.Parameter)
'torch.nn.parameter.Parameter'
"""
if isinstance(obj, torch.Tensor):
return obj.type()
module = getattr(obj, "__module__", "") or ""
qualname = ""
if hasattr(obj, "__qualname__"):
qualname = obj.__qualname__
elif hasattr(obj, "__name__"):
qualname = obj.__name__
else:
module = obj.__class__.__module__ or ""
qualname = obj.__class__.__qualname__
if module in {"", "builtins"}:
return qualname
return f"{module}.{qualname}"
[docs]def is_tensor(obj: _Any, /) -> _TypeIs["torch.Tensor"]:
r"""Returns True if `obj` is a PyTorch tensor.
Note that this function is simply doing ``isinstance(obj, Tensor)``.
Using that ``isinstance`` check is better for typechecking with mypy,
and more explicit - so it's recommended to use that instead of
``is_tensor``.
Args:
obj (object): Object to test
Example::
>>> x = torch.tensor([1, 2, 3])
>>> torch.is_tensor(x)
True
"""
return isinstance(obj, torch.Tensor)
[docs]def is_storage(obj: _Any, /) -> _TypeIs[_Union["TypedStorage", "UntypedStorage"]]:
r"""Returns True if `obj` is a PyTorch storage object.
Args:
obj (Object): Object to test
"""
return type(obj) in _storage_classes
_GLOBAL_DEVICE_CONTEXT = threading.local()
[docs]def get_default_device() -> "torch.device":
r"""Gets the default ``torch.Tensor`` to be allocated on ``device``"""
global _GLOBAL_DEVICE_CONTEXT
if hasattr(_GLOBAL_DEVICE_CONTEXT, "device_context"):
device = _GLOBAL_DEVICE_CONTEXT.device_context.device
if device.index is not None:
return device
else:
# TODO: Call like get_device_index() method corresponding to
# each device type
return torch.tensor([]).device
else:
return torch.device("cpu")
[docs]def set_default_device(
device: _Optional[_Union["torch.device", str, builtins.int]],
) -> None:
"""Sets the default ``torch.Tensor`` to be allocated on ``device``. This
does not affect factory function calls which are called with an explicit
``device`` argument. Factory calls will be performed as if they
were passed ``device`` as an argument.
To only temporarily change the default device instead of setting it
globally, use ``with torch.device(device):`` instead.
The default device is initially ``cpu``. If you set the default tensor
device to another device (e.g., ``cuda``) without a device index, tensors
will be allocated on whatever the current device for the device type,
even after :func:`torch.cuda.set_device` is called.
.. warning::
This function imposes a slight performance cost on every Python
call to the torch API (not just factory functions). If this
is causing problems for you, please comment on
https://github.com/pytorch/pytorch/issues/92701
.. note::
This doesn't affect functions that create tensors that share the same memory as the input, like:
:func:`torch.from_numpy` and :func:`torch.frombuffer`
Args:
device (device or string): the device to set as default
Example::
>>> # xdoctest: +SKIP("requires cuda, changes global state")
>>> torch.get_default_device()
device(type='cpu')
>>> torch.set_default_device('cuda') # current device is 0
>>> torch.get_default_device()
device(type='cuda', index=0)
>>> torch.set_default_device('cuda')
>>> torch.cuda.set_device('cuda:1') # current device is 1
>>> torch.get_default_device()
device(type='cuda', index=1)
>>> torch.set_default_device('cuda:1')
>>> torch.get_default_device()
device(type='cuda', index=1)
"""
global _GLOBAL_DEVICE_CONTEXT
if hasattr(_GLOBAL_DEVICE_CONTEXT, "device_context"):
device_context = _GLOBAL_DEVICE_CONTEXT.device_context
if device_context is not None:
device_context.__exit__(None, None, None)
if device is None:
device_context = None
else:
from torch.utils._device import DeviceContext
device_context = DeviceContext(device)
device_context.__enter__()
_GLOBAL_DEVICE_CONTEXT.device_context = device_context
[docs]def set_default_tensor_type(t: _Union[_Type["torch.Tensor"], str], /) -> None:
r"""
.. warning::
This function is deprecated as of PyTorch 2.1, please use :func:`torch.set_default_dtype()` and
:func:`torch.set_default_device()` as alternatives.
Sets the default ``torch.Tensor`` type to floating point tensor type
``t``. This type will also be used as default floating point type for
type inference in :func:`torch.tensor`.
The default floating point tensor type is initially ``torch.FloatTensor``.
Args:
t (type or string): the floating point tensor type or its name
Example::
>>> # xdoctest: +SKIP("Other tests may have changed the default type. Can we reset it?")
>>> torch.tensor([1.2, 3]).dtype # initial default for floating point is torch.float32
torch.float32
>>> torch.set_default_tensor_type(torch.DoubleTensor)
>>> torch.tensor([1.2, 3]).dtype # a new floating point tensor
torch.float64
"""
if isinstance(t, str):
t = _import_dotted_name(t)
_C._set_default_tensor_type(t)
[docs]def set_default_dtype(d: "torch.dtype", /) -> None:
r"""
Sets the default floating point dtype to :attr:`d`. Supports floating point dtype
as inputs. Other dtypes will cause torch to raise an exception.
When PyTorch is initialized its default floating point dtype is torch.float32,
and the intent of set_default_dtype(torch.float64) is to facilitate NumPy-like
type inference. The default floating point dtype is used to:
1. Implicitly determine the default complex dtype. When the default floating type is float16,
the default complex dtype is complex32. For float32, the default complex dtype is complex64.
For float64, it is complex128. For bfloat16, an exception will be raised because
there is no corresponding complex type for bfloat16.
2. Infer the dtype for tensors constructed using Python floats or complex Python
numbers. See examples below.
3. Determine the result of type promotion between bool and integer tensors and
Python floats and complex Python numbers.
Args:
d (:class:`torch.dtype`): the floating point dtype to make the default.
Example:
>>> # xdoctest: +SKIP("Other tests may have changed the default type. Can we reset it?")
>>> # initial default for floating point is torch.float32
>>> # Python floats are interpreted as float32
>>> torch.tensor([1.2, 3]).dtype
torch.float32
>>> # initial default for floating point is torch.complex64
>>> # Complex Python numbers are interpreted as complex64
>>> torch.tensor([1.2, 3j]).dtype
torch.complex64
>>> torch.set_default_dtype(torch.float64)
>>> # Python floats are now interpreted as float64
>>> torch.tensor([1.2, 3]).dtype # a new floating point tensor
torch.float64
>>> # Complex Python numbers are now interpreted as complex128
>>> torch.tensor([1.2, 3j]).dtype # a new complex tensor
torch.complex128
>>> torch.set_default_dtype(torch.float16)
>>> # Python floats are now interpreted as float16
>>> torch.tensor([1.2, 3]).dtype # a new floating point tensor
torch.float16
>>> # Complex Python numbers are now interpreted as complex128
>>> torch.tensor([1.2, 3j]).dtype # a new complex tensor
torch.complex32
"""
_C._set_default_dtype(d)
[docs]def use_deterministic_algorithms(
mode: builtins.bool,
*,
warn_only: builtins.bool = False,
) -> None:
r"""Sets whether PyTorch operations must use "deterministic"
algorithms. That is, algorithms which, given the same input, and when
run on the same software and hardware, always produce the same output.
When enabled, operations will use deterministic algorithms when available,
and if only nondeterministic algorithms are available they will throw a
:class:`RuntimeError` when called.
.. note:: This setting alone is not always enough to make an application
reproducible. Refer to :ref:`reproducibility` for more information.
.. note:: :func:`torch.set_deterministic_debug_mode` offers an alternative
interface for this feature.
The following normally-nondeterministic operations will act
deterministically when ``mode=True``:
* :class:`torch.nn.Conv1d` when called on CUDA tensor
* :class:`torch.nn.Conv2d` when called on CUDA tensor
* :class:`torch.nn.Conv3d` when called on CUDA tensor
* :class:`torch.nn.ConvTranspose1d` when called on CUDA tensor
* :class:`torch.nn.ConvTranspose2d` when called on CUDA tensor
* :class:`torch.nn.ConvTranspose3d` when called on CUDA tensor
* :class:`torch.nn.ReplicationPad2d` when attempting to differentiate a CUDA tensor
* :func:`torch.bmm` when called on sparse-dense CUDA tensors
* :func:`torch.Tensor.__getitem__` when attempting to differentiate a CPU tensor
and the index is a list of tensors
* :func:`torch.Tensor.index_put` with ``accumulate=False``
* :func:`torch.Tensor.index_put` with ``accumulate=True`` when called on a CPU
tensor
* :func:`torch.Tensor.put_` with ``accumulate=True`` when called on a CPU
tensor
* :func:`torch.Tensor.scatter_add_` when called on a CUDA tensor
* :func:`torch.gather` when called on a CUDA tensor that requires grad
* :func:`torch.index_add` when called on CUDA tensor
* :func:`torch.index_select` when attempting to differentiate a CUDA tensor
* :func:`torch.repeat_interleave` when attempting to differentiate a CUDA tensor
* :func:`torch.Tensor.index_copy` when called on a CPU or CUDA tensor
* :func:`torch.Tensor.scatter` when `src` type is Tensor and called on CUDA tensor
* :func:`torch.Tensor.scatter_reduce` when ``reduce='sum'`` or ``reduce='mean'`` and called on CUDA tensor
The following normally-nondeterministic operations will throw a
:class:`RuntimeError` when ``mode=True``:
* :class:`torch.nn.AvgPool3d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.AdaptiveAvgPool2d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.AdaptiveAvgPool3d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.MaxPool3d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.AdaptiveMaxPool2d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.FractionalMaxPool2d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.FractionalMaxPool3d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.MaxUnpool1d`
* :class:`torch.nn.MaxUnpool2d`
* :class:`torch.nn.MaxUnpool3d`
* :func:`torch.nn.functional.interpolate` when attempting to differentiate a CUDA tensor
and one of the following modes is used:
- ``linear``
- ``bilinear``
- ``bicubic``
- ``trilinear``
* :class:`torch.nn.ReflectionPad1d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.ReflectionPad2d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.ReflectionPad3d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.ReplicationPad1d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.ReplicationPad3d` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.NLLLoss` when called on a CUDA tensor
* :class:`torch.nn.CTCLoss` when attempting to differentiate a CUDA tensor
* :class:`torch.nn.EmbeddingBag` when attempting to differentiate a CUDA tensor when
``mode='max'``
* :func:`torch.Tensor.put_` when ``accumulate=False``
* :func:`torch.Tensor.put_` when ``accumulate=True`` and called on a CUDA tensor
* :func:`torch.histc` when called on a CUDA tensor
* :func:`torch.bincount` when called on a CUDA tensor and ``weights``
tensor is given
* :func:`torch.kthvalue` with called on a CUDA tensor
* :func:`torch.median` with indices output when called on a CUDA tensor
* :func:`torch.nn.functional.grid_sample` when attempting to differentiate a CUDA tensor
* :func:`torch.cumsum` when called on a CUDA tensor when dtype is floating point or complex
* :func:`torch.Tensor.scatter_reduce` when ``reduce='prod'`` and called on CUDA tensor
* :func:`torch.Tensor.resize_` when called with a quantized tensor
In addition, several operations fill uninitialized memory when this setting
is turned on and when
:attr:`torch.utils.deterministic.fill_uninitialized_memory` is turned on.
See the documentation for that attribute for more information.
A handful of CUDA operations are nondeterministic if the CUDA version is
10.2 or greater, unless the environment variable ``CUBLAS_WORKSPACE_CONFIG=:4096:8``
or ``CUBLAS_WORKSPACE_CONFIG=:16:8`` is set. See the CUDA documentation for more
details: `<https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility>`_
If one of these environment variable configurations is not set, a :class:`RuntimeError`
will be raised from these operations when called with CUDA tensors:
* :func:`torch.mm`
* :func:`torch.mv`
* :func:`torch.bmm`
Note that deterministic operations tend to have worse performance than
nondeterministic operations.
.. note::
This flag does not detect or prevent nondeterministic behavior caused
by calling an inplace operation on a tensor with an internal memory
overlap or by giving such a tensor as the :attr:`out` argument for an
operation. In these cases, multiple writes of different data may target
a single memory location, and the order of writes is not guaranteed.
Args:
mode (:class:`bool`): If True, makes potentially nondeterministic
operations switch to a deterministic algorithm or throw a runtime
error. If False, allows nondeterministic operations.
Keyword args:
warn_only (:class:`bool`, optional): If True, operations that do not
have a deterministic implementation will throw a warning instead of
an error. Default: ``False``
Example::
>>> # xdoctest: +SKIP
>>> torch.use_deterministic_algorithms(True)
# Forward mode nondeterministic error
>>> torch.randn(10, device='cuda').kthvalue(1)
...
RuntimeError: kthvalue CUDA does not have a deterministic implementation...
# Backward mode nondeterministic error
>>> torch.nn.AvgPool3d(1)(torch.randn(3, 4, 5, 6, requires_grad=True).cuda()).sum().backward()
...
RuntimeError: avg_pool3d_backward_cuda does not have a deterministic implementation...
"""
_C._set_deterministic_algorithms(mode, warn_only=warn_only)
[docs]def are_deterministic_algorithms_enabled() -> builtins.bool:
r"""Returns True if the global deterministic flag is turned on. Refer to
:func:`torch.use_deterministic_algorithms` documentation for more details.
"""
return _C._get_deterministic_algorithms()
[docs]def is_deterministic_algorithms_warn_only_enabled() -> builtins.bool:
r"""Returns True if the global deterministic flag is set to warn only.
Refer to :func:`torch.use_deterministic_algorithms` documentation for more
details.
"""
return _C._get_deterministic_algorithms_warn_only()
[docs]def set_deterministic_debug_mode(debug_mode: _Union[builtins.int, str]) -> None:
r"""Sets the debug mode for deterministic operations.
.. note:: This is an alternative interface for
:func:`torch.use_deterministic_algorithms`. Refer to that function's
documentation for details about affected operations.
Args:
debug_mode(str or int): If "default" or 0, don't error or warn on
nondeterministic operations. If "warn" or 1, warn on
nondeterministic operations. If "error" or 2, error on
nondeterministic operations.
"""
# NOTE: builtins.int is used here because int in this scope resolves
# to torch.int
if not isinstance(debug_mode, (builtins.int, str)):
raise TypeError(f"debug_mode must be str or int, but got {type(debug_mode)}")
if isinstance(debug_mode, str):
if debug_mode == "default":
debug_mode = 0
elif debug_mode == "warn":
debug_mode = 1
elif debug_mode == "error":
debug_mode = 2
else:
raise RuntimeError(
"invalid value of debug_mode, expected one of `default`, "
f"`warn`, `error`, but got {debug_mode}"
)
if debug_mode == 0:
_C._set_deterministic_algorithms(False)
elif debug_mode == 1:
_C._set_deterministic_algorithms(True, warn_only=True)
elif debug_mode == 2:
_C._set_deterministic_algorithms(True)
else:
raise RuntimeError(
"invalid value of debug_mode, expected 0, 1, or 2, " f"but got {debug_mode}"
)
[docs]def get_deterministic_debug_mode() -> builtins.int:
r"""Returns the current value of the debug mode for deterministic
operations. Refer to :func:`torch.set_deterministic_debug_mode`
documentation for more details.
"""
if _C._get_deterministic_algorithms():
if _C._get_deterministic_algorithms_warn_only():
return 1
else:
return 2
else:
return 0
[docs]def get_float32_matmul_precision() -> str:
r"""Returns the current value of float32 matrix multiplication precision. Refer to
:func:`torch.set_float32_matmul_precision` documentation for more details.
"""
return _C._get_float32_matmul_precision()
[docs]def set_float32_matmul_precision(precision: str) -> None:
r"""Sets the internal precision of float32 matrix multiplications.
Running float32 matrix multiplications in lower precision may significantly increase
performance, and in some programs the loss of precision has a negligible impact.
Supports three settings:
* "highest", float32 matrix multiplications use the float32 datatype (24 mantissa
bits with 23 bits explicitly stored) for internal computations.
* "high", float32 matrix multiplications either use the TensorFloat32 datatype (10
mantissa bits explicitly stored) or treat each float32 number as the sum of two bfloat16 numbers
(approximately 16 mantissa bits with 14 bits explicitly stored), if the appropriate fast matrix multiplication
algorithms are available. Otherwise float32 matrix multiplications are computed
as if the precision is "highest". See below for more information on the bfloat16
approach.
* "medium", float32 matrix multiplications use the bfloat16 datatype (8 mantissa
bits with 7 bits explicitly stored) for internal computations, if a fast matrix multiplication algorithm
using that datatype internally is available. Otherwise float32
matrix multiplications are computed as if the precision is "high".
When using "high" precision, float32 multiplications may use a bfloat16-based algorithm
that is more complicated than simply truncating to some smaller number mantissa bits
(e.g. 10 for TensorFloat32, 7 for bfloat16 explicitly stored). Refer to [Henry2019]_ for a complete
description of this algorithm. To briefly explain here, the first step is to realize
that we can perfectly encode a single float32 number as the sum of three bfloat16
numbers (because float32 has 23 mantissa bits while bfloat16 has 7 explicitly stored, and both have the
same number of exponent bits). This means that the product of two float32 numbers can
be exactly given by the sum of nine products of bfloat16 numbers. We can then trade
accuracy for speed by dropping some of these products. The "high" precision algorithm
specifically keeps only the three most significant products, which conveniently excludes
all of the products involving the last 8 mantissa bits of either input. This means that
we can represent our inputs as the sum of two bfloat16 numbers rather than three.
Because bfloat16 fused-multiply-add (FMA) instructions are typically >10x faster than
float32 ones, it's faster to do three multiplications and 2 additions with bfloat16
precision than it is to do a single multiplication with float32 precision.
.. [Henry2019] http://arxiv.org/abs/1904.06376
.. note::
This does not change the output dtype of float32 matrix multiplications,
it controls how the internal computation of the matrix multiplication is performed.
.. note::
This does not change the precision of convolution operations. Other flags,
like `torch.backends.cudnn.allow_tf32`, may control the precision of convolution
operations.
.. note::
This flag currently only affects one native device type: CUDA.
If "high" or "medium" are set then the TensorFloat32 datatype will be used
when computing float32 matrix multiplications, equivalent to setting
`torch.backends.cuda.matmul.allow_tf32 = True`. When "highest" (the default)
is set then the float32 datatype is used for internal computations, equivalent
to setting `torch.backends.cuda.matmul.allow_tf32 = False`.
Args:
precision(str): can be set to "highest" (default), "high", or "medium" (see above).
"""
_C._set_float32_matmul_precision(precision)
[docs]def set_warn_always(b: builtins.bool, /) -> None:
r"""When this flag is False (default) then some PyTorch warnings may only
appear once per process. This helps avoid excessive warning information.
Setting it to True causes these warnings to always appear, which may be
helpful when debugging.
Args:
b (:class:`bool`): If True, force warnings to always be emitted
If False, set to the default behaviour
"""
_C._set_warnAlways(b)
[docs]def is_warn_always_enabled() -> builtins.bool:
r"""Returns True if the global warn_always flag is turned on. Refer to
:func:`torch.set_warn_always` documentation for more details.
"""
return _C._get_warnAlways()
################################################################################
# Define error checking functions
################################################################################
# These error checking functions must be kept consistent with their C++
# equivalents. Their C++ equivalents are mentioned where applicable.
def _check_with(
error_type,
cond: _Union[builtins.bool, SymBool],
message: _Callable[[], str],
): # noqa: F811
if not isinstance(cond, (builtins.bool, SymBool)):
raise TypeError(f"cond must be a bool, but got {type(cond)}")
from torch.fx.experimental.symbolic_shapes import expect_true
if expect_true(cond):
return
# error_type must be a subclass of Exception and not subclass of Warning
assert issubclass(error_type, Exception) and not issubclass(error_type, Warning)
if message is None:
message_evaluated = (
"Expected cond to be True, but got False. (Could this error "
"message be improved? If so, please report an enhancement request "
"to PyTorch.)"
)
else:
if not callable(message):
raise TypeError("message must be a callable")
message_evaluated = str(message())
raise error_type(message_evaluated)
def _check(cond, message=None): # noqa: F811
r"""Throws error containing an optional message if the specified condition
is False.
Error type: ``RuntimeError``
C++ equivalent: ``TORCH_CHECK``
Args:
cond (:class:`bool`): If False, throw error
message (Callable, optional): Callable that returns either a string or
an object that has a ``__str__()`` method to be used as the error
message. Default: ``None``
"""
_check_with(RuntimeError, cond, message)
def _check_is_size(i, message=None):
"""Checks that a given integer is a valid size (i.e., is non-negative).
You should use this over _check(i >= 0) because we can use the semantic
information (that i is a size) to make some further inferences in case
i is an unbacked SymInt.
NB: Do NOT use this in contexts where a -1 size would be valid (indicating
to infer the size from context, or if you should wrap-around or truncate).
Only use this if the only valid value is an honest to goodness size.
"""
# This is responsible for the expect_true
_check(i >= 0, message)
from torch.fx.experimental.symbolic_shapes import _advise_is_size
_advise_is_size(i)
def _check_index(cond, message=None): # noqa: F811
r"""Throws error containing an optional message if the specified condition
is False.
Error type: ``IndexError``
C++ equivalent: ``TORCH_CHECK_INDEX``
Args:
cond (:class:`bool`): If False, throw error
message (Callable, optional): Callable that returns either a string or
an object that has a ``__str__()`` method to be used as the error
message. Default: ``None``
"""
_check_with(IndexError, cond, message)
def _check_value(cond, message=None): # noqa: F811
r"""Throws error containing an optional message if the specified condition
is False.
Error type: ``ValueError``
C++ equivalent: ``TORCH_CHECK_VALUE``
Args:
cond (:class:`bool`): If False, throw error
message (Callable, optional): Callable that returns either a string or
an object that has a ``__str__()`` method to be used as the error
message. Default: ``None``
"""
_check_with(ValueError, cond, message)
def _check_type(cond, message=None): # noqa: F811
r"""Throws error containing an optional message if the specified condition
is False.
Error type: ``TypeError``
C++ equivalent: ``TORCH_CHECK_TYPE``
Args:
cond (:class:`bool`): If False, throw error
message (Callable, optional): Callable that returns either a string or
an object that has a ``__str__()`` method to be used as the error
message. Default: ``None``
"""
_check_with(TypeError, cond, message)
def _check_not_implemented(cond, message=None): # noqa: F811
r"""Throws error containing an optional message if the specified condition
is False.
Error type: ``NotImplementedError``
C++ equivalent: ``TORCH_CHECK_NOT_IMPLEMENTED``
Args:
cond (:class:`bool`): If False, throw error
message (Callable, optional): Callable that returns either a string or
an object that has a ``__str__()`` method to be used as the error
message. Default: ``None``
"""
_check_with(NotImplementedError, cond, message)
def _check_tensor_all_with(error_type, cond, message=None): # noqa: F811
if not is_tensor(cond):
raise TypeError(f"cond must be a tensor, but got {type(cond)}")
if not cond.dtype == torch.bool:
raise TypeError(f"cond tensor must have dtype torch.bool, but got {cond.dtype}")
_check_with(error_type, cond._is_all_true().item(), message) # type: ignore[arg-type]
# C++ equivalent: `TORCH_CHECK_TENSOR_ALL`
def _check_tensor_all(cond, message=None): # noqa: F811
r"""Throws error containing an optional message if the specified condition
is False.
Error type: ``RuntimeError``
C++ equivalent: ``TORCH_CHECK_TENSOR_ALL``
Args:
cond (:class:`torch.Tensor`): Tensor of dtype ``torch.bool``. If any
element is ``False``, throw error
message (Callable, optional): Callable that returns either a string or
an object that has a ``__str__()`` method to be used as the error
message. Default: ``None``
"""
_check_tensor_all_with(RuntimeError, cond, message)
################################################################################
# Define numeric constants
################################################################################
# For Python Array API (https://data-apis.org/array-api/latest/API_specification/constants.html) and
# NumPy consistency (https://numpy.org/devdocs/reference/constants.html)
from math import e, inf, nan, pi
newaxis: None = None
__all__.extend(["e", "pi", "nan", "inf", "newaxis"])
################################################################################
# Define Storage and Tensor classes
################################################################################
from torch._tensor import Tensor # usort: skip
# needs to be after torch.Tensor is defined to avoid circular dependencies
from torch import storage as storage # usort: skip
from torch.storage import (
_LegacyStorage,
_StorageBase,
_warn_typed_storage_removal,
TypedStorage,
UntypedStorage,
)
# NOTE: New <type>Storage classes should never be added. When adding a new
# dtype, use torch.storage.TypedStorage directly.
[docs]class ComplexDoubleStorage(_LegacyStorage):
@classproperty
def _dtype(self):
return torch.cdouble
[docs]class ComplexFloatStorage(_LegacyStorage):
@classproperty
def _dtype(self):
return torch.cfloat
_storage_classes: _Set[_Type[_Union[TypedStorage, UntypedStorage]]] = {
UntypedStorage,
DoubleStorage,
FloatStorage,
LongStorage,
IntStorage,
ShortStorage,
CharStorage,
ByteStorage,
HalfStorage,
BoolStorage,
QUInt8Storage,
QInt8Storage,
QInt32Storage,
BFloat16Storage,
ComplexFloatStorage,
ComplexDoubleStorage,
QUInt4x2Storage,
QUInt2x4Storage,
TypedStorage,
}
# The _tensor_classes set is initialized by the call to initialize_python_bindings.
_tensor_classes: _Set[_Type["torch.Tensor"]] = set()
# If you edit these imports, please update torch/__init__.py.in as well
from torch import amp as amp, random as random, serialization as serialization
from torch._tensor_str import set_printoptions
from torch.amp import autocast, GradScaler
from torch.random import get_rng_state, initial_seed, manual_seed, seed, set_rng_state
from torch.serialization import load, save
################################################################################
# Initialize extension
################################################################################
# Shared memory manager needs to know the exact location of manager executable
def _manager_path():
if _running_with_deploy() or platform.system() == "Windows":
return b""
path = get_file_path("torch", "bin", "torch_shm_manager")
prepare_multiprocessing_environment(get_file_path("torch"))
if not os.path.exists(path):
raise RuntimeError("Unable to find torch_shm_manager at " + path)
return path.encode("utf-8")
_C._initExtension(_manager_path())
del _manager_path
# Appease the type checker: it can't deal with direct setting of globals().
# Note that we will see "too many" functions when reexporting this way; there
# is not a good way to fix this problem. Perhaps, try to redesign VariableFunctions
# so that this import is good enough
if TYPE_CHECKING:
# Some type signatures pulled in from _VariableFunctions here clash with
# signatures already imported. For now these clashes are ignored; see
# PR #43339 for details.
from torch._C._VariableFunctions import * # type: ignore[assignment, misc] # noqa: F403
# Fixup segment_reduce visibility
_segment_reduce = segment_reduce
del segment_reduce # noqa: F821
# Ops not to be exposed in `torch` namespace,
# mostly helper ops.
PRIVATE_OPS = ("unique_dim",)
__name, __obj = "", None
for __name in dir(_C._VariableFunctions):
if __name.startswith("__") or __name in PRIVATE_OPS:
continue
__obj = getattr(_C._VariableFunctions, __name)
__obj.__module__ = __name__ # "torch"
# Hide some APIs that should not be public
if __name == "segment_reduce":
# TODO: Once the undocumented FC window is passed, remove the line bellow
globals()[__name] = __obj
__name = "_" + __name
globals()[__name] = __obj
if not __name.startswith("_"):
__all__.append(__name)
del __name, __obj
################################################################################
# Add torch.dtype instances to the public API
################################################################################
import torch
__all__.extend(
name for name in dir(torch) if isinstance(getattr(torch, name), torch.dtype)
)
################################################################################
# Import TorchDynamo's lazy APIs to avoid circular dependenices
################################################################################
# needs to be before from torch.functional import * to avoid circular dependencies
from torch._compile import _disable_dynamo # usort: skip
################################################################################
# Import interface functions defined in Python
################################################################################
# needs to be after the above ATen bindings so we can overwrite from Python side
from torch import _VF as _VF, functional as functional # usort: skip
from torch.functional import * # usort: skip # noqa: F403
################################################################################
# Remove unnecessary members
################################################################################
del _StorageBase
del _LegacyStorage
################################################################################
# Define _assert
################################################################################
# needs to be before the submodule imports to avoid circular dependencies
[docs]def _assert(condition, message):
r"""A wrapper around Python's assert which is symbolically traceable."""
if type(condition) is not torch.Tensor and overrides.has_torch_function(
(condition,)
):
return overrides.handle_torch_function(
_assert, (condition,), condition, message
)
assert condition, message
################################################################################
# Import most common subpackages
################################################################################
# Use the redundant form so that type checkers know that these are a part of
# the public API. The "regular" import lines are there solely for the runtime
# side effect of adding to the imported module's members for other users.
# needs to be before import torch.nn as nn to avoid circular dependencies
from torch.autograd import ( # usort: skip
enable_grad as enable_grad,
inference_mode as inference_mode,
no_grad as no_grad,
set_grad_enabled as set_grad_enabled,
)
from torch import (
__config__ as __config__,
__future__ as __future__,
_awaits as _awaits,
accelerator as accelerator,
autograd as autograd,
backends as backends,
cpu as cpu,
cuda as cuda,
distributed as distributed,
distributions as distributions,
fft as fft,
futures as futures,
hub as hub,
jit as jit,
linalg as linalg,
mps as mps,
mtia as mtia,
multiprocessing as multiprocessing,
nested as nested,
nn as nn,
optim as optim,
overrides as overrides,
profiler as profiler,
sparse as sparse,
special as special,
testing as testing,
types as types,
utils as utils,
xpu as xpu,
)
from torch.signal import windows as windows
# Quantized, sparse, AO, etc. should be last to get imported, as nothing
# is expected to depend on them.
from torch import ao as ao # usort: skip
# nn.quant* depends on ao -- so should be after those.
import torch.nn.intrinsic
import torch.nn.qat
import torch.nn.quantizable
import torch.nn.quantized
_C._init_names(list(_storage_classes))
# attach docstrings to torch and tensor functions
from torch import _size_docs, _storage_docs, _tensor_docs, _torch_docs
del _torch_docs, _tensor_docs, _storage_docs, _size_docs
[docs]def compiled_with_cxx11_abi() -> builtins.bool:
r"""Returns whether PyTorch was built with _GLIBCXX_USE_CXX11_ABI=1"""
return _C._GLIBCXX_USE_CXX11_ABI
from torch import _library as _library, _ops as _ops
# Import the ops and classes "namespace"
from torch._ops import ops as ops # usort: skip
from torch._classes import classes as classes # usort: skip
sys.modules.setdefault(f"{__name__}.ops", ops)
sys.modules.setdefault(f"{__name__}.classes", classes)
# quantization depends on torch.fx and torch.ops
# Import quantization
from torch import quantization as quantization # usort: skip
# Import the quasi random sampler
from torch import quasirandom as quasirandom # usort: skip
# If you are seeing this, it means that this call site was not checked if
# the memory format could be preserved, and it was switched to old default
# behaviour of contiguous
legacy_contiguous_format = contiguous_format # defined by _C._initExtension()
# Register fork handler to initialize OpenMP in child processes (see gh-28389)
from torch.multiprocessing._atfork import register_after_fork
register_after_fork(torch.get_num_threads)
del register_after_fork
# Import tools that require fully imported torch (for applying
# torch.jit.script as a decorator, for instance):
from torch._lobpcg import lobpcg as lobpcg
# These were previously defined in native_functions.yaml and appeared on the
# `torch` namespace, but we moved them to c10 dispatch to facilitate custom
# class usage. We add these lines here to preserve backward compatibility.
quantized_lstm = ops.aten.quantized_lstm
quantized_gru = ops.aten.quantized_gru
# Import experimental masked operations support. See
# [RFC-0016](https://github.com/pytorch/rfcs/pull/27) for more
# information.
from torch import masked as masked
# Import removed ops with error message about removal
from torch._linalg_utils import ( # type: ignore[misc]
_symeig as symeig,
eig,
lstsq,
matrix_rank,
solve,
)
from torch.utils.dlpack import from_dlpack, to_dlpack
class _TorchCompileInductorWrapper:
compiler_name = "inductor"
def __init__(self, mode, options, dynamic):
self.config: _Dict[str, _Any] = {}
self.dynamic = dynamic
self.apply_mode(mode)
self.apply_options(options)
if self.config.get("triton.cudagraphs", False):
os.environ["DISABLE_CUPTI_LAZY_REINIT"] = "1"
# FIXME: CUDA Graph does not work well with CUPTI teardown.
# 1) crashes on 1st lazy CUPTI re-init after teardown (CUDA 11)
# 2) crashes on 2nd non-lazy CUPTI re-init after teardown (CUDA 12)
# Workaround: turn off CUPTI teardown when using CUDA Graphs.
os.environ["TEARDOWN_CUPTI"] = "0"
def __eq__(self, other):
return (
isinstance(other, _TorchCompileInductorWrapper)
and self.config == other.config
and self.dynamic == other.dynamic
)
def apply_mode(self, mode: _Optional[str]):
if mode is None or mode == "default":
pass
elif mode in {"reduce-overhead", "max-autotune", "max-autotune-no-cudagraphs"}:
from torch._inductor import list_mode_options
self.apply_options(list_mode_options(mode, self.dynamic))
else:
raise RuntimeError(
f"Unrecognized mode={mode}, should be one of: default, reduce-overhead, max-autotune, max-autotune-no-cudagraphs"
)
def apply_options(self, options: _Optional[_Dict[str, _Any]]):
from torch._inductor.compiler_bisector import CompilerBisector
if bisect_changes := CompilerBisector.get_config_change("inductor"):
options = {} if options is None else options
options = (
{**bisect_changes} if options is None else {**options, **bisect_changes} # type: ignore[dict-item]
)
if not options:
return
from torch._inductor import config
current_config: _Dict[str, _Any] = config.get_config_copy()
for key, val in options.items():
attr_name = key.replace("-", "_")
if attr_name not in current_config:
raise RuntimeError(
f"Unexpected optimization option {key}, known options are {list(current_config.keys())}"
)
attr_type = config.get_type(attr_name) # type: ignore[attr-defined]
# Subscriptable generic types don't support isinstance so skip the type
# check. There doesn't seem to be a good way of checking membership without
# 3rd party libraries.
if _get_origin(attr_type) is None:
if not isinstance(val, attr_type):
val_type_str = type(val).__name__
expected_type_str = type(current_config[attr_name]).__name__
raise RuntimeError(
f"Unexpected type of attr {key}, got {val_type_str} should be {expected_type_str}"
)
self.config[attr_name] = val
def __call__(self, model_, inputs_):
from torch._inductor.compile_fx import compile_fx
return compile_fx(model_, inputs_, config_patches=self.config)
def get_compiler_config(self):
from torch._inductor.compile_fx import get_patched_config_dict
return get_patched_config_dict(config_patches=self.config)
def reset(self):
from torch._inductor import config
if "triton.cudagraphs" in self.config or config.triton.cudagraphs:
if self.config.get("triton.cudagraphs", True):
from torch._inductor.cudagraph_trees import reset_cudagraph_trees
reset_cudagraph_trees()
class _TorchCompileWrapper:
def __init__(self, backend, mode, options, dynamic):
from torch._dynamo.backends.registry import lookup_backend
if isinstance(backend, str):
self.compiler_name = backend
elif hasattr(backend, "__name__"):
self.compiler_name = backend.__name__
else:
self.compiler_name = str(backend)
self.dynamic = dynamic
self.compiler_fn = lookup_backend(backend)
self.kwargs = {}
# only pass the args if they non-empty
if mode and mode != "default":
self.kwargs["mode"] = mode
if options:
self.kwargs["options"] = options
def __eq__(self, other):
return (
isinstance(other, _TorchCompileWrapper)
and self.compiler_fn == other.compiler_fn
and self.kwargs == other.kwargs
and self.dynamic == other.dynamic
)
def __call__(self, model_, inputs_):
return self.compiler_fn(model_, inputs_, **self.kwargs)
def reset(self):
if hasattr(self.compiler_fn, "reset"):
self.compiler_fn.reset()
_InputT = _ParamSpec("_InputT")
_RetT = _TypeVar("_RetT")
@_overload
def compile(
model: _Callable[_InputT, _RetT],
*,
fullgraph: builtins.bool = False,
dynamic: _Optional[builtins.bool] = None,
backend: _Union[str, _Callable] = "inductor",
mode: _Union[str, None] = None,
options: _Optional[_Dict[str, _Union[str, builtins.int, builtins.bool]]] = None,
disable: builtins.bool = False,
) -> _Callable[_InputT, _RetT]: ...
@_overload
def compile(
model: None = None,
*,
fullgraph: builtins.bool = False,
dynamic: _Optional[builtins.bool] = None,
backend: _Union[str, _Callable] = "inductor",
mode: _Union[str, None] = None,
options: _Optional[_Dict[str, _Union[str, builtins.int, builtins.bool]]] = None,
disable: builtins.bool = False,
) -> _Callable[[_Callable[_InputT, _RetT]], _Callable[_InputT, _RetT]]: ...
[docs]def compile(
model: _Optional[_Callable] = None,
*,
fullgraph: builtins.bool = False,
dynamic: _Optional[builtins.bool] = None,
backend: _Union[str, _Callable] = "inductor",
mode: _Union[str, None] = None,
options: _Optional[_Dict[str, _Union[str, builtins.int, builtins.bool]]] = None,
disable: builtins.bool = False,
) -> _Union[
_Callable[[_Callable[_InputT, _RetT]], _Callable[_InputT, _RetT]],
_Callable[_InputT, _RetT],
]:
"""
Optimizes given model/function using TorchDynamo and specified backend.
If you are compiling an :class:`torch.nn.Module`, you can also use :meth:`torch.nn.Module.compile`
to compile the module inplace without changing its structure.
Concretely, for every frame executed within the compiled region, we will attempt
to compile it and cache the compiled result on the code object for future
use. A single frame may be compiled multiple times if previous compiled
results are not applicable for subsequent calls (this is called a "guard
failure), you can use TORCH_LOGS=guards to debug these situations.
Multiple compiled results can be associated with a frame up to
``torch._dynamo.config.cache_size_limit``, which defaults to 8; at which
point we will fall back to eager. Note that compile caches are per
*code object*, not frame; if you dynamically create multiple copies of a
function, they will all share the same code cache.
Args:
model (Callable): Module/function to optimize
fullgraph (bool): If False (default), torch.compile attempts to discover compileable regions
in the function that it will optimize. If True, then we require that the entire function be
capturable into a single graph. If this is not possible (that is, if there are graph breaks),
then this will raise an error.
dynamic (bool or None): Use dynamic shape tracing. When this is True, we will up-front attempt
to generate a kernel that is as dynamic as possible to avoid recompilations when
sizes change. This may not always work as some operations/optimizations will
force specialization; use TORCH_LOGS=dynamic to debug overspecialization.
When this is False, we will NEVER generate dynamic kernels, we will always specialize.
By default (None), we automatically detect if dynamism has occurred and compile a more
dynamic kernel upon recompile.
backend (str or Callable): backend to be used
- "inductor" is the default backend, which is a good balance between performance and overhead
- Non experimental in-tree backends can be seen with `torch._dynamo.list_backends()`
- Experimental or debug in-tree backends can be seen with `torch._dynamo.list_backends(None)`
- To register an out-of-tree custom backend:
https://pytorch.org/docs/main/torch.compiler_custom_backends.html#registering-custom-backends
mode (str): Can be either "default", "reduce-overhead", "max-autotune" or "max-autotune-no-cudagraphs"
- "default" is the default mode, which is a good balance between performance and overhead
- "reduce-overhead" is a mode that reduces the overhead of python with CUDA graphs,
useful for small batches. Reduction of overhead can come at the cost of more memory
usage, as we will cache the workspace memory required for the invocation so that we
do not have to reallocate it on subsequent runs. Reduction of overhead is not guaranteed
to work; today, we only reduce overhead for CUDA only graphs which do not mutate inputs.
There are other circumstances where CUDA graphs are not applicable; use TORCH_LOG=perf_hints
to debug.
- "max-autotune" is a mode that leverages Triton or template based matrix multiplications
on supported devices and Triton based convolutions on GPU.
It enables CUDA graphs by default on GPU.
- "max-autotune-no-cudagraphs" is a mode similar to "max-autotune" but without CUDA graphs
- To see the exact configs that each mode sets you can call `torch._inductor.list_mode_options()`
options (dict): A dictionary of options to pass to the backend. Some notable ones to try out are
- `epilogue_fusion` which fuses pointwise ops into templates. Requires `max_autotune` to also be set
- `max_autotune` which will profile to pick the best matmul configuration
- `fallback_random` which is useful when debugging accuracy issues
- `shape_padding` which pads matrix shapes to better align loads on GPUs especially for tensor cores
- `triton.cudagraphs` which will reduce the overhead of python with CUDA graphs
- `trace.enabled` which is the most useful debugging flag to turn on
- `trace.graph_diagram` which will show you a picture of your graph after fusion
- For inductor you can see the full list of configs that it supports by calling `torch._inductor.list_options()`
disable (bool): Turn torch.compile() into a no-op for testing
Example::
@torch.compile(options={"triton.cudagraphs": True}, fullgraph=True)
def foo(x):
return torch.sin(x) + torch.cos(x)
"""
_C._log_api_usage_once("torch.compile")
if sys.version_info >= (3, 14):
raise RuntimeError("Dynamo is not supported on Python 3.14+")
elif sys.version_info >= (3, 13) and not sys._is_gil_enabled():
raise RuntimeError("Dynamo is not supported on Python with GIL disabled")
# Decorator mode
if model is None:
def fn(model: _Callable[_InputT, _RetT]) -> _Callable[_InputT, _RetT]:
if model is None:
raise RuntimeError("Model can't be None")
return compile(
model,
fullgraph=fullgraph,
dynamic=dynamic,
backend=backend,
mode=mode,
options=options,
disable=disable,
)
return fn
if mode is not None and options is not None:
raise RuntimeError(
"Either mode or options can be specified, but both can't be specified at the same time."
)
if mode is None and options is None:
mode = "default"
from torch._inductor.compiler_bisector import CompilerBisector
if bisect_backend := CompilerBisector.get_backend():
backend = bisect_backend
if backend == "inductor":
backend = _TorchCompileInductorWrapper(mode, options, dynamic)
else:
backend = _TorchCompileWrapper(backend, mode, options, dynamic)
return torch._dynamo.optimize(
backend=backend,
nopython=fullgraph,
dynamic=dynamic,
disable=disable,
)(model) # type: ignore[return-value]
def _register_device_module(device_type, module):
r"""Register an external runtime module of the specific :attr:`device_type`
supported by torch.
After the :attr:`module` is registered correctly, the user can refer
the external runtime module as part of torch with attribute torch.xxx.
"""
# Make sure the device_type represent a supported device type for torch.
device_type = torch.device(device_type).type
m = sys.modules[__name__]
if hasattr(m, device_type):
raise RuntimeError(
f"The runtime module of '{device_type}' has already "
f"been registered with '{getattr(m, device_type)}'"
)
setattr(m, device_type, module)
torch_module_name = ".".join([__name__, device_type])
sys.modules[torch_module_name] = module
from torch import (
export as export,
func as func,
library as library,
return_types as return_types,
)
from torch._higher_order_ops import cond as cond, while_loop as while_loop
from torch.func import vmap as vmap
if not TYPE_CHECKING:
from torch import _meta_registrations
# Enable CUDA Sanitizer
if "TORCH_CUDA_SANITIZER" in os.environ:
import torch.cuda._sanitizer as csan
csan.enable_cuda_sanitizer()
# Populate magic methods on SymInt and SymFloat
import torch.fx.experimental.sym_node
from torch import fx as fx
# Register MPS specific decomps
torch.backends.mps._init()
if not _running_with_deploy():
from torch import compiler as compiler
class _TritonLibrary:
lib = torch.library.Library("triton", "DEF")
ops_table: _Dict[_Tuple[str, str], _Callable] = {}
@classmethod
def registerOp(cls, op_key, full_schema, op_impl, dispatch_key):
if (op_key, dispatch_key) not in cls.ops_table:
cls.lib.define(full_schema)
cls.lib.impl("triton::" + op_key, op_impl, dispatch_key)
cls.ops_table[(op_key, dispatch_key)] = op_impl
return cls.ops_table[(op_key, dispatch_key)]
# Deprecated attributes
_deprecated_attrs = {
"has_mps": torch.backends.mps.is_built,
"has_cuda": torch.backends.cuda.is_built,
"has_cudnn": torch.backends.cudnn.is_available,
"has_mkldnn": torch.backends.mkldnn.is_available,
}
if TYPE_CHECKING:
# Import the following modules during type checking to enable code intelligence features,
# such as auto-completion in tools like pylance, even when these modules are not explicitly
# imported in user code.
from torch import (
_dynamo as _dynamo,
_inductor as _inductor,
_subclasses as _subclasses,
onnx as onnx,
)
else:
_lazy_modules = {
"_dynamo",
"_inductor",
"_export",
# ONNX must be imported after _dynamo, _ops, _subclasses, fx, func and jit
"onnx",
}
def __getattr__(name):
# Deprecated attrs
replacement = _deprecated_attrs.get(name)
if replacement is not None:
import warnings
warnings.warn(
f"'{name}' is deprecated, please use '{replacement.__module__}.{replacement.__name__}()'",
stacklevel=2,
)
return replacement()
# Lazy modules
if name in _lazy_modules:
return importlib.import_module(f".{name}", __name__)
raise AttributeError(f"module '{__name__}' has no attribute '{name}'")
[docs]def get_device_module(device: _Optional[_Union[torch.device, str]] = None):
"""
Returns the module associated with a given device(e.g., torch.device('cuda'), "mtia:0", "xpu", ...).
If no device is given, return the module for the current accelerator or CPU if none is present.
"""
if isinstance(device, torch.device):
device_module_name = device.type
elif isinstance(device, str):
device_module_name = torch.device(device).type
elif device is None:
# Using default accelerator type. If no accelerator is available, it automatically returns CPU device.
device_module_name = torch._C._get_accelerator().type
else:
raise RuntimeError(
f"Invalid value of device '{device}', expect torch.device, str, or None"
)
device_module = getattr(torch, device_module_name, None)
if device_module is None:
raise RuntimeError(
f"Device '{device_module_name}' does not have a corresponding module registered as 'torch.{device_module_name}'."
)
return device_module
def _constrain_as_size(
symbol,
min: _Optional[builtins.int] = None,
max: _Optional[builtins.int] = None,
):
"""
This indicates that a given int is size-like, and can be used in any context where a size is expected.
You will typically use this when reading out integers from Tensors, e.g., max.item() or lengths.tolist()
which then need to be used as tensor constructors. Providing these assertions to PyTorch can help resolve
GuardOnDataDependentSymNode errors upon export, since we cannot guard on unbacked SymInts.
This function has unusual semantics in some circumstances in framework
code, we will treat this int as >= 2 (when we do a size-oblivious guard).
This makes it easier to use the unbacked int in size contexts,
as we will often attempt to guard on a size being zero/one
(e.g., when computing the contiguity of a tensor, or testing if
broadcasting can occur), which will not work on unbacked SymInts.
However, if we conservatively assume that the size is not zero/one, we will
end up with a graph that will still work even if the size is zero/one.
For more details, see https://docs.google.com/document/d/1HSuTTVvYH1pTew89Rtpeu84Ht3nQEFTYhAX3Ypa_xJs/edit
```
"""
torch.sym_constrain_range_for_size(symbol, min=min, max=max)
from torch import _logging
_logging._init_logs()
def _import_device_backends():
"""
Leverage the Python plugin mechanism to load out-of-the-tree device extensions.
See this RFC: https://github.com/pytorch/pytorch/issues/122468
"""
from importlib.metadata import entry_points
group_name = "torch.backends"
if sys.version_info < (3, 10):
backend_extensions = entry_points().get(group_name, ())
else:
backend_extensions = entry_points(group=group_name)
for backend_extension in backend_extensions:
try:
# Load the extension
entrypoint = backend_extension.load()
# Call the entrypoint
entrypoint()
except Exception as err:
raise RuntimeError(
f"Failed to load the backend extension: {backend_extension.name}. "
f"You can disable extension auto-loading with TORCH_DEVICE_BACKEND_AUTOLOAD=0."
) from err
def _is_device_backend_autoload_enabled() -> builtins.bool:
"""
Whether autoloading out-of-the-tree device extensions is enabled.
The switch depends on the value of the environment variable
`TORCH_DEVICE_BACKEND_AUTOLOAD`.
Returns:
bool: Whether to enable autoloading the extensions. Enabled by default.
Examples:
>>> torch._is_device_backend_autoload_enabled()
True
"""
# enabled by default
return os.getenv("TORCH_DEVICE_BACKEND_AUTOLOAD", "1") == "1"
if _is_device_backend_autoload_enabled():
_import_device_backends()
def _as_tensor_fullprec(t):
"""
Like torch.as_tensor, but when given Python data types it will keep
them in full precision. Used for calling convention for Dynamo.
"""
ty = type(t)
if ty is builtins.float:
return torch.as_tensor(t, dtype=torch.float64)
elif ty is builtins.int:
return torch.as_tensor(t, dtype=torch.int64)
else:
return torch.as_tensor(t)
