U yh8@s ddlZddlmZddlmZddlZdddddd d d d d dddddddddddgZddZGdddZGdddZ GdddZ GdddZ ej j jej j jej j jdZdeZd8edeej j fej j d d!d Zej jjej jjej jjd"ZdeZd9edeej jfej jd d#d Zdd$lmZm Z!d%e_"d e_#d&dZ$e%d'd(dZ&d)dZ'e%d'd*dZ(d+dZ)e%d'd,dZ*d:ee%e%d.d/dZ+d;ee%e%d.d0dZ,d1dZ-e%d'd2dZ.ej/ed3e0d4dtdtjd|t|jdkrl|jddt tjD|j|S)Nrz/cufft_plan_cache: expected 0 <= device index < z, but got device with index css|]}t|VqdSr)r).0indexrrr lsz4cuFFTPlanCacheManager.__getitem__..) rcudaZ_utilsZ_get_device_indexZ device_countr(lenr3extendrange)r Zdevicer6rrr __getitem__ds  z!cuFFTPlanCacheManager.__getitem__cCst|tj|Sr)getattrrr8current_devicer namerrr __getattr__qsz!cuFFTPlanCacheManager.__getattr__cs.|jrt|tj||St||SdSr)r4setattrrr8r>super __setattr__r r@value __class__rrrDtsz!cuFFTPlanCacheManager.__setattr__) r+r,r-r0r4r!r<rArD __classcell__rrrGrrUs  c@seZdZddZddZdS)rcCsF|dkrtjS|dkr$tjS|dkr6tjStd|dSNZ allow_tf32Z&allow_fp16_reduced_precision_reductionZ&allow_bf16_reduced_precision_reductionzUnknown attribute )rrZ_get_cublas_allow_tf32Z2_get_cublas_allow_fp16_reduced_precision_reductionZ2_get_cublas_allow_bf16_reduced_precision_reductionAttributeErrorr?rrrrA|s   zcuBLASModule.__getattr__cCsL|dkrtj|S|dkr(tj|S|dkr|tkr,tdtdtjt|n$t|tjjrZtj|ntdtj S)af Override the heuristic PyTorch uses to choose between cuSOLVER and MAGMA for CUDA linear algebra operations. .. 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` N"Unknown input value. Choose from: .Unknown input value type.) r&r'_LinalgBackendsr(_LinalgBackends_strrrZ_set_linalg_preferred_backend_LinalgBackendZ_get_linalg_preferred_backendrMrrrr s,  )ZcublasZcublasltZ hipblasltcCsl|dkr nXt|tr>|tkr,tdtdtjt|n$t|tjjrZtj|ntdtj S)a Override the library PyTorch uses for BLAS operations. Choose between cuBLAS and cuBLASLt. .. warning:: This flag is experimental and subject to change. When PyTorch runs a CUDA BLAS operation it defaults to cuBLAS even if both cuBLAS and cuBLASLt are available. For PyTorch built for ROCm, hipBLAS and hipBLASLt may offer different performance. This flag (a :class:`str`) allows overriding which BLAS library to use. * If `"cublas"` is set then cuBLAS will be used wherever possible. * If `"cublaslt"` is set then cuBLASLt will be used wherever possible. * When no input is given, this function returns the currently preferred library. * User may use the environment variable TORCH_BLAS_PREFER_CUBLASLT=1 to set the preferred library to cuBLASLt 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 library selection is incorrect for your application's inputs. NrOrPrQ) r&r' _BlasBackendsr(_BlasBackends_strrrZ_set_blas_preferred_backend _BlasBackendZ_get_blas_preferred_backendrUrrrr s  ) _SDPAParams _SDPBackendztorch.backends.cudacCs tjS)z .. warning:: This flag is beta and subject to change. Returns whether flash scaled dot product attention is enabled or not. )rrZ_get_flash_sdp_enabledrrrrr sZenabledcCstj|dS)z| .. warning:: This flag is beta and subject to change. Enables or disables flash scaled dot product attention. N)rrZ_set_sdp_use_flashr[rrrrscCs tjS)z .. warning:: This flag is beta and subject to change. Returns whether memory efficient scaled dot product attention is enabled or not. )rrZ_get_mem_efficient_sdp_enabledrrrrrscCstj|dS)z .. warning:: This flag is beta and subject to change. Enables or disables memory efficient scaled dot product attention. N)rrZ_set_sdp_use_mem_efficientr[rrrr'scCs tjS)z .. warning:: This flag is beta and subject to change. Returns whether math scaled dot product attention is enabled or not. )rrZ_get_math_sdp_enabledrrrrr0scCstj|dS)z{ .. warning:: This flag is beta and subject to change. Enables or disables math scaled dot product attention. N)rrZ_set_sdp_use_mathr[rrrr9sF)paramsdebugrNcCstj||S)aCheck if FlashAttention can be utilized in scaled_dot_product_attention. Args: params: An instance of SDPAParams containing the tensors for query, key, value, an optional attention mask, dropout rate, and a flag indicating if the attention is causal. debug: Whether to logging.warn debug information as to why FlashAttention could not be run. Defaults to False. Returns: True if FlashAttention can be used with the given parameters; otherwise, False. Note: This function is dependent on a CUDA-enabled build of PyTorch. It will return False in non-CUDA environments. )rrZ_can_use_flash_attentionr\r]rrrrBscCstj||S)aCheck if efficient_attention can be utilized in scaled_dot_product_attention. Args: params: An instance of SDPAParams containing the tensors for query, key, value, an optional attention mask, dropout rate, and a flag indicating if the attention is causal. debug: Whether to logging.warn with information as to why efficient_attention could not be run. Defaults to False. Returns: True if efficient_attention can be used with the given parameters; otherwise, False. Note: This function is dependent on a CUDA-enabled build of PyTorch. It will return False in non-CUDA environments. )rrZ _can_use_mem_efficient_attentionr^rrrrVscCs tjS)z .. warning:: This flag is beta and subject to change. Returns whether cuDNN scaled dot product attention is enabled or not. )rrZ_get_cudnn_sdp_enabledrrrrrjscCstj|dS)z| .. warning:: This flag is beta and subject to change. Enables or disables cuDNN scaled dot product attention. N)rrZ_set_sdp_use_cudnnr[rrrrssz`torch.backends.cuda.sdp_kernel()` is deprecated. In the future, this context manager will be removed. Please see `torch.nn.attention.sdpa_kernel()` for the new context manager, with updated signature.)categoryT) enable_flash enable_mathenable_mem_efficient enable_cudnnc csvddlm}g}|r |tj|r0|tj|r@|tj|rP|tj||}z |VW5XW5QRXdS)a .. 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. r) sdpa_kernelN)Ztorch.nn.attentionrdappend SDPBackendZFLASH_ATTENTIONZEFFICIENT_ATTENTIONZMATHZCUDNN_ATTENTION)r`rarbrcrdZ backend_listcontextrrrr|s       )N)N)F)F)TTTT)4 contextlibtypingrtyping_extensionsrr__all__rrrrrrrTZDefaultZCusolverZMagmarRjoinkeysrSr'r rXZCublasZCublasltrVrWr Ztorch._CrYr rZrfr,r+rboolrrrrrrrrrcontextmanager FutureWarningrr r rrrrs   & = *