sg@@ ddlZddlmZddlmZddlZgdZdZGddZGdd Z Gd d Z Gd d Z ejjjejjjejjj dZdj%ej'Z d0dedeejjfdejjfdZejj.j0ejj.j2ejj.j2dZdj%ej'Z d0dedeejj.fdejj.fdZddlmZm Z!de_"de_#dZ$de%fdZ&dZ'de%fdZ(dZ)de%fdZ*de%fdZ+d Z,de%fd!Z-d1d"ed#e%de%fd$Z.d1d"ed#e%de%fd%Z/d1d"ed#e%de%fd&Z0d'Z1de%fd(Z2ejfed)e4* d2d+e%d,e%d-e%d.e%fd/Z5e Z6e Z7y)3N)Union) deprecated)is_builtcuFFTPlanCacheAttrContextPropcuFFTPlanCachecuFFTPlanCacheManager cuBLASModulepreferred_linalg_librarypreferred_blas_librarycufft_plan_cachematmul SDPAParamsenable_cudnn_sdpcudnn_sdp_enabledenable_flash_sdpflash_sdp_enabledenable_mem_efficient_sdpmem_efficient_sdp_enabledmath_sdp_enabledenable_math_sdp"allow_fp16_bf16_reduction_math_sdp$fp16_bf16_reduction_math_sdp_allowedis_flash_attention_availablecan_use_flash_attentioncan_use_efficient_attentioncan_use_cudnn_attention sdp_kernelc6tjjS)z Return whether PyTorch is built with CUDA support. Note that this doesn't necessarily mean CUDA is available; just that if this PyTorch binary were run on a machine with working CUDA drivers and devices, we would be able to use it. )torch_C _has_cudaO/var/www/html/venv/lib/python3.12/site-packages/torch/backends/cuda/__init__.pyrr&s 88  r#ceZdZdZdZdZy)rc ||_||_yN)gettersetter)selfr(r)s r$__init__z&cuFFTPlanCacheAttrContextProp.__init__3s  r#c8|j|jSr')r( device_index)r*objobjtypes r$__get__z%cuFFTPlanCacheAttrContextProp.__get__7s{{3++,,r#ct|jtrt|j|j|j|yr') isinstancer)str RuntimeErrorr-)r*r.vals r$__set__z%cuFFTPlanCacheAttrContextProp.__set__:s3 dkk3 't{{+ + C$$c*r#N)__name__ __module__ __qualname__r+r0r6r"r#r$rr0s-+r#rc|eZdZdZdZeejdZeejejZ dZ y)rz Represent a specific plan cache for a specific `device_index`. The attributes `size` and `max_size`, and method `clear`, can fetch and/ or change properties of the C++ cuFFT plan cache. c||_yr')r-)r*r-s r$r+zcuFFTPlanCache.__init__Hs (r#z.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.c@tj|jSr')r_cufft_clear_plan_cacher-r*s r$clearzcuFFTPlanCache.clearUs,,T->->??r#N) r7r8r9__doc__r+rr_cufft_get_plan_cache_sizesize_cufft_get_plan_cache_max_size_cufft_set_plan_cache_max_sizemax_sizer?r"r#r$rr@sJ) ) (( N D - ,,e.R.RH@r#rc8eZdZdZdZdZdZdZfdZxZ S)ra Represent all cuFFT plan caches, return the cuFFTPlanCache for a given device when indexed. 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. Fc g|_d|_y)NT)caches#_cuFFTPlanCacheManager__initializedr>s r$r+zcuFFTPlanCacheManager.__init__ds !r#ctjjj|}|dks!|tjj k\r-t dtjj d|t |jdk(rG|jjdttjj D|j|S)Nrz/cufft_plan_cache: expected 0 <= device index < z, but got device with index c32K|]}t|ywr')r).0indexs r$ z4cuFFTPlanCacheManager.__getitem__..ps*/u%s) rcuda_utils_get_device_index device_countr4lenrHextendrange)r*devicerMs r$ __getitem__z!cuFFTPlanCacheManager.__getitem__hs !!33F; 19!8!8!::A%**BYBYB[A\]%%*G-  t{{ q KK  389P9P9R3S {{5!!r#cXt|tjj|Sr')getattrrrOcurrent_devicer*names r$ __getattr__z!cuFFTPlanCacheManager.__getattr__us!tEJJ5578$??r#c|jr,t|tjj ||St |||Sr')rIsetattrrrOrZsuper __setattr__)r*r\value __class__s r$raz!cuFFTPlanCacheManager.__setattr__xs?   4 9 9 ;tjjS)z .. warning:: This flag is beta and subject to change. Returns whether flash scaled dot product attention is enabled or not. )rr _get_flash_sdp_enabledr"r#r$rr 88 * * ,,r#enabledcBtjj|y)z| .. warning:: This flag is beta and subject to change. Enables or disables flash scaled dot product attention. N)rr _set_sdp_use_flashrs r$rr  HH(r#c>tjjS)z .. warning:: This flag is beta and subject to change. Returns whether memory efficient scaled dot product attention is enabled or not. )rr _get_mem_efficient_sdp_enabledr"r#r$rr#s 88 2 2 44r#cBtjj|y)z .. warning:: This flag is beta and subject to change. Enables or disables memory efficient scaled dot product attention. N)rr _set_sdp_use_mem_efficientrs r$rr,s  HH''0r#c>tjjS)z .. warning:: This flag is beta and subject to change. Returns whether math scaled dot product attention is enabled or not. )rr _get_math_sdp_enabledr"r#r$rr5s 88 ) ) ++r#cBtjj|y)z{ .. warning:: This flag is beta and subject to change. Enables or disables math scaled dot product attention. N)rr _set_sdp_use_mathrs r$rr>s  HHw'r#cBtjj|y)z .. warning:: This flag is beta and subject to change. Enables or disables fp16/bf16 reduction in math scaled dot product attention. N)rr '_set_math_sdp_allow_fp16_bf16_reductionrs r$rrGs  HH44W=r#c>tjjS)z .. warning:: This flag is beta and subject to change. Returns whether fp16/bf16 reduction in math scaled dot product attention is enabled or not. )rr '_get_math_sdp_allow_fp16_bf16_reductionr"r#r$rrPs 88 ; ; ==r#c>tjjS)a5Check if PyTorch was built with FlashAttention for scaled_dot_product_attention. Returns: True if FlashAttention is built and available; otherwise, False. Note: This function is dependent on a CUDA-enabled build of PyTorch. It will return False in non-CUDA environments. )rr _is_flash_attention_availabler"r#r$rrYs 88 1 1 33r#paramsdebugcBtjj||S)aCheck if FlashAttention can be utilized in scaled_dot_product_attention. Args: params: An instance of SDPAParams containing the tensors for query, key, value, an optional attention mask, dropout rate, and a flag indicating if the attention is causal. debug: Whether to logging.warn debug information as to why FlashAttention could not be run. Defaults to False. Returns: True if FlashAttention can be used with the given parameters; otherwise, False. Note: This function is dependent on a CUDA-enabled build of PyTorch. It will return False in non-CUDA environments. )rr _can_use_flash_attentionrrs r$rrf" 88 , ,VU ;;r#cBtjj||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. )rr _can_use_mem_efficient_attentionrs r$rrzs" 88 4 4VU CCr#cBtjj||S)aCheck if cudnn_attention can be utilized in scaled_dot_product_attention. Args: params: An instance of SDPAParams containing the tensors for query, key, value, an optional attention mask, dropout rate, and a flag indicating if the attention is causal. debug: Whether to logging.warn with information as to why cuDNN attention could not be run. Defaults to False. Returns: True if cuDNN can be used with the given parameters; otherwise, False. Note: This function is dependent on a CUDA-enabled build of PyTorch. It will return False in non-CUDA environments. )rr _can_use_cudnn_attentionrs r$rrrr#c>tjjS)z .. warning:: This flag is beta and subject to change. Returns whether cuDNN scaled dot product attention is enabled or not. )rr _get_cudnn_sdp_enabledr"r#r$rrrr#cBtjj|y)z| .. warning:: This flag is beta and subject to change. Enables or disables cuDNN scaled dot product attention. N)rr _set_sdp_use_cudnnrs r$rrrr#z`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.)category enable_flash enable_mathenable_mem_efficient enable_cudnnc#tKddlm}g}|r|jtj|r|jtj |r|jtj |r|jtj||5} | dddy#wxYw#1swYyxYww)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)torch.nn.attentionrappend SDPBackendFLASH_ATTENTIONEFFICIENT_ATTENTIONMATHCUDNN_ATTENTION)rrrrr backend_listcontexts r$rrs,/LJ667J::;JOO,J667 \ "g M    s0BB8B,B' B8'B))B,,B51B8r')F)TTTT)8 contextlibtypingrtyping_extensionsrr__all__rrrrr r rDefaultCusolverMagmar|joinkeysr}r3r rCublasCublasltrrr torch._Crrrrr8r7rboolrrrrrrrrrrrrrcontextmanager FutureWarningrr r r"r#r$rs(  : + + @@2#4#4L::*xx&&..''00 XX $ $ * * ii 4 4 67:>94 4ehh555 694 XX94zhh##**%%..&&// IIm00238<'2 4ehh333 4'2 XX'2TJ. " -)d)51d1,(T(>>> 4d 4)* r#