[docs]defis_built():r"""Returns whether PyTorch is built with CUDA support. Note that this doesn't necessarily mean CUDA is available; just that if this PyTorch binary were run a machine with working CUDA drivers and devices, we would be able to use it."""returntorch._C._has_cuda
classcuFFTPlanCacheAttrContextProp:# Like regular ContextProp, but uses the `.device_index` attribute from the# calling object as the first argument to the getter and setter.def__init__(self,getter,setter):self.getter=getterself.setter=setterdef__get__(self,obj,objtype):returnself.getter(obj.device_index)def__set__(self,obj,val):ifisinstance(self.setter,str):raiseRuntimeError(self.setter)self.setter(obj.device_index,val)classcuFFTPlanCache:r""" Represents a specific plan cache for a specific `device_index`. The attributes `size` and `max_size`, and method `clear`, can fetch and/ or change properties of the C++ cuFFT plan cache. """def__init__(self,device_index):self.device_index=device_indexsize=cuFFTPlanCacheAttrContextProp(torch._cufft_get_plan_cache_size,".size is a read-only property showing the number of plans currently in the ""cache. To change the cache capacity, set cufft_plan_cache.max_size.",)max_size=cuFFTPlanCacheAttrContextProp(torch._cufft_get_plan_cache_max_size,torch._cufft_set_plan_cache_max_size)defclear(self):returntorch._cufft_clear_plan_cache(self.device_index)classcuFFTPlanCacheManager:r""" Represents all cuFFT plan caches. When indexed with a device object/index, this object returns the `cuFFTPlanCache` corresponding to that device. Finally, this object, when used directly as a `cuFFTPlanCache` object (e.g., setting the `.max_size`) attribute, the current device's cuFFT plan cache is used. """__initialized=Falsedef__init__(self):self.caches=[]self.__initialized=Truedef__getitem__(self,device):index=torch.cuda._utils._get_device_index(device)ifindex<0orindex>=torch.cuda.device_count():raiseRuntimeError(f"cufft_plan_cache: expected 0 <= device index < {torch.cuda.device_count()}, but got "f"device with index {index}")iflen(self.caches)==0:self.caches.extend(cuFFTPlanCache(index)forindexinrange(torch.cuda.device_count()))returnself.caches[index]def__getattr__(self,name):returngetattr(self[torch.cuda.current_device()],name)def__setattr__(self,name,value):ifself.__initialized:returnsetattr(self[torch.cuda.current_device()],name,value)else:returnsuper().__setattr__(name,value)classcuBLASModule:def__getattr__(self,name):ifname=="allow_tf32":returntorch._C._get_cublas_allow_tf32()elifname=="allow_fp16_reduced_precision_reduction":returntorch._C._get_cublas_allow_fp16_reduced_precision_reduction()elifname=="allow_bf16_reduced_precision_reduction":returntorch._C._get_cublas_allow_bf16_reduced_precision_reduction()raiseAssertionError("Unknown attribute "+name)def__setattr__(self,name,value):ifname=="allow_tf32":returntorch._C._set_cublas_allow_tf32(value)elifname=="allow_fp16_reduced_precision_reduction":returntorch._C._set_cublas_allow_fp16_reduced_precision_reduction(value)elifname=="allow_bf16_reduced_precision_reduction":returntorch._C._set_cublas_allow_bf16_reduced_precision_reduction(value)raiseAssertionError("Unknown attribute "+name)_LinalgBackends={"default":torch._C._LinalgBackend.Default,"cusolver":torch._C._LinalgBackend.Cusolver,"magma":torch._C._LinalgBackend.Magma,}_LinalgBackends_str=", ".join(_LinalgBackends.keys())
[docs]defpreferred_linalg_library(backend:Union[None,str,torch._C._LinalgBackend]=None)->torch._C._LinalgBackend:r""" .. warning:: This flag is experimental and subject to change. When PyTorch runs a CUDA linear algebra operation it often uses the cuSOLVER or MAGMA libraries, and if both are available it decides which to use with a heuristic. This flag (a :class:`str`) allows overriding those heuristics. * If `"cusolver"` is set then cuSOLVER will be used wherever possible. * If `"magma"` is set then MAGMA will be used wherever possible. * If `"default"` (the default) is set then heuristics will be used to pick between cuSOLVER and MAGMA if both are available. * When no input is given, this function returns the currently preferred library. * User may use the environment variable TORCH_LINALG_PREFER_CUSOLVER=1 to set the preferred library to cuSOLVER globally. This flag only sets the initial value of the preferred library and the preferred library may still be overridden by this function call later in your script. Note: When a library is preferred other libraries may still be used if the preferred library doesn't implement the operation(s) called. This flag may achieve better performance if PyTorch's heuristic library selection is incorrect for your application's inputs. Currently supported linalg operators: * :func:`torch.linalg.inv` * :func:`torch.linalg.inv_ex` * :func:`torch.linalg.cholesky` * :func:`torch.linalg.cholesky_ex` * :func:`torch.cholesky_solve` * :func:`torch.cholesky_inverse` * :func:`torch.linalg.lu_factor` * :func:`torch.linalg.lu` * :func:`torch.linalg.lu_solve` * :func:`torch.linalg.qr` * :func:`torch.linalg.eigh` * :func:`torch.linalg.eighvals` * :func:`torch.linalg.svd` * :func:`torch.linalg.svdvals` """ifbackendisNone:passelifisinstance(backend,str):ifbackendnotin_LinalgBackends:raiseRuntimeError("Unknown input value. "f"Choose from: {_LinalgBackends_str}.")torch._C._set_linalg_preferred_backend(_LinalgBackends[backend])elifisinstance(backend,torch._C._LinalgBackend):torch._C._set_linalg_preferred_backend(backend)else:raiseRuntimeError("Unknown input value type.")returntorch._C._get_linalg_preferred_backend()
[docs]classSDPBackend(IntEnum):r"""Enum class for the scaled dot product attention backends. .. warning:: This class is in beta and subject to change. This class needs to stay aligned with the enum defined in: pytorch/aten/src/ATen/native/transformers/sdp_utils_cpp.h """ERROR=-1MATH=0FLASH_ATTENTION=1EFFICIENT_ATTENTION=2
[docs]defflash_sdp_enabled():r""" .. warning:: This flag is beta and subject to change. Returns whether flash scaled dot product attention is enabled or not. """returntorch._C._get_flash_sdp_enabled()
[docs]defenable_flash_sdp(enabled:bool):r""" .. warning:: This flag is beta and subject to change. Enables or disables flash scaled dot product attention. """torch._C._set_sdp_use_flash(enabled)
[docs]defmem_efficient_sdp_enabled():r""" .. warning:: This flag is beta and subject to change. Returns whether memory efficient scaled dot product attention is enabled or not. """returntorch._C._get_mem_efficient_sdp_enabled()
[docs]defenable_mem_efficient_sdp(enabled:bool):r""" .. warning:: This flag is beta and subject to change. Enables or disables memory efficient scaled dot product attention. """torch._C._set_sdp_use_mem_efficient(enabled)
[docs]defmath_sdp_enabled():r""" .. warning:: This flag is beta and subject to change. Returns whether math scaled dot product attention is enabled or not. """returntorch._C._get_math_sdp_enabled()
[docs]defenable_math_sdp(enabled:bool):r""" .. warning:: This flag is beta and subject to change. Enables or disables math scaled dot product attention. """torch._C._set_sdp_use_math(enabled)
[docs]@contextlib.contextmanagerdefsdp_kernel(enable_flash:bool=True,enable_math:bool=True,enable_mem_efficient:bool=True,):r""" .. warning:: This flag is beta and subject to change. This context manager can be used to temporarily enable or disable any of the three backends for scaled dot product attention. Upon exiting the context manager, the previous state of the flags will be restored. """previous_flash:bool=flash_sdp_enabled()previous_mem_efficient:bool=mem_efficient_sdp_enabled()previous_math:bool=math_sdp_enabled()try:enable_flash_sdp(enable_flash)enable_mem_efficient_sdp(enable_mem_efficient)enable_math_sdp(enable_math)yield{}finally:enable_flash_sdp(previous_flash)enable_mem_efficient_sdp(previous_mem_efficient)enable_math_sdp(previous_math)
To analyze traffic and optimize your experience, we serve cookies on this site. By clicking or navigating, you agree to allow our usage of cookies. As the current maintainers of this site, Facebook’s Cookies Policy applies. Learn more, including about available controls: Cookies Policy.