i)SSKrSSKrSSKrSSKrSSKJr SSKJrJrJrJ r J r SSK r SSK J r SSKJr SSKJr SSKJr SSKJr SS KJr SS KJr SS KJr \\"S 5"S S\ \555rg)N) resources)AnyDictfinalListOptional) AotiBackend)MoveCondPredicateToCpuPass)ReplaceEdgeOpWithTritonOpPass) experimental)BackendDetails) CompileSpec)conv1d_to_conv2d) SDPBackendzHThis API and all of cuda backend related functionality are experimental.cR\rSrSrSr\S\4Sj5r\S\S\ \4Sj5r \S\ 4Sj5r \S\ 4Sj5r \S\\\44S j5r\S\\\44S j5r\S \\S\\R4S j5r\S \\S\\\R44S j5r\S5rSrg) CudaBackenda  CudaBackend is a backend that compiles a model to run on CUDA devices. It uses the AOTInductor compiler to generate optimized CUDA kernels for the model's operators with libtorch-free. The compiled model can be executed on CUDA devices using the Executorch runtime. returncg)Ncudaclss d/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/executorch/backends/cuda/cuda_backend.pyget_device_nameCudaBackend.get_device_name)s cuda_versioncn^ SUS3m S[S[4U 4SjjnSSKJn U(a0[R R USS 5nU"U5(aU$S H]n[RRU5nU(dM+[R R USS 5nU"U5(dM[Us $ S US 3n[R RU5(aU$[R"S 5nU(aU"U5(aU$S nU"U5(aU$g![a Nf=f)z Find ptxas binary that matches the expected CUDA version. Returns the path to ptxas if found and version matches, None otherwise. z/cuda-/pathrc>[RRU5(dg[RRU5nTU;$)z;Check if ptxas at given path matches expected CUDA version.F)osr!existsrealpath)r!resolvedexpected_version_markers r_validate_ptxas_versionDCudaBackend._find_ptxas_for_version.._validate_ptxas_version5s877>>$''ww''-H*h6 6rr) CUDA_HOMEbinptxas)r* CUDA_PATH CUDA_ROOTz/usr/local/cuda-z /bin/ptxasz/usr/local/cuda/bin/ptxasN) strbooltorch.utils.cpp_extensionr*r#r!join ImportErrorenvirongetr$shutilwhich) rr(r* ptxas_pathenv_var cuda_homeversioned_path ptxas_in_path default_pathr's @r_find_ptxas_for_version#CudaBackend._find_ptxas_for_version-s %+<.": 7# 7$ 7  ;WW\\)UGD *:66%% ?G w/IyWW\\)UGD *:66%% ?,L>D 77>>. ) )! ! W- 4]CC 3 "< 0 0 3   s= 12.9), false otherwise. NF. TTRITON_PTXAS_PATHTORCH_CUDA_ARCH_LISTzTORCH_CUDA_ARCH_LIST is set to z,, skipping automatic architecture detection.r)torchversionrmapintsplitrr>r#r4r5loggingwarningcurrent_deviceget_device_capability Exception)rmajorminorr8device capabilitys r"_setup_cuda_environment_for_fatbin.CudaBackend._setup_cuda_environment_for_fatbin`s7 !==--L#sL$6$6s$;BQ$?@LErzerkeqj{uz(@@N % 2< ./zz~~45A5bjjnnE[6\5]^JKZZ..0F99&AJ4>qM?!JqM?1SBJJ- .  s*E6E$E! " y !(#//,7@@L$'L! NN:C6G;L7<  ") ?> ?)rcr[RRR[R /5$)ah Return SDPA MATH backend context manager for CUDA compilation. This context manager plays as a fallback solution for any remaining PyTorch SDPA operations to use the MATH backend (decomposed SDPA) during AOTInductor compilation. Note: - If SDPA ops are replaced with Triton kernels by ReplaceEdgeOpWithTritonOpPass, this context manager will have no effect on those ops (they are no longer PyTorch SDPA ops). - If SDPA ops are NOT replaced (e.g., when triton_kernel_mode="OFF"), this context manager will force them to use the MATH backend, causing them to be automatically decomposed during compilation. )rInn attention sdpa_kernelrMATHrs r&get_extra_aoti_compile_context_manager2CudaBackend.get_extra_aoti_compile_context_managers& xx!!--z.?@@rrN)__name__ __module__ __qualname____firstlineno____doc__ classmethodr/r staticmethodrr>r0rWrZrrr]rdrrtypingrurr__static_attributes__rrrrrsG  0c0hsm00d%%%NT tCH~   S#X  d;.?DDT2D -D c6::o DDLAArr)rNr#r6r importlibrrrrrrrI%executorch.backends.aoti.aoti_backendr :executorch.backends.cuda.passes.move_cond_predicate_to_cpur 0executorch.backends.cuda.triton.replacement_passr executorch.exir._warningsr 'executorch.exir.backend.backend_detailsr +executorch.exir.backend.compile_spec_schemartorch._inductor.decompositionrtorch.nn.attentionrrrrrrsp 33 =3BC:) NkA+~kAkAr