3jR>SSKrSSKrSSKJr SSKrSSKJr SrSSK J r S\R4SjrS\R4S jrS\R4S jrS \SS4S jrS \S\4S jrS\R4SjrS\R4SjrS\R4SjrS\\4SjrS$S\S\S\S-S\S-S\S-S\S\\\44Sjjr"SS5r"SS5rS%S\\-S\\S-S\\ \S4-4Sjjr!S&S \S!\S"\S\4S#jjr g!\ a SrSrNf=f)'N)Any)runtimeTF)_get_device_indexreturnc SSKn[R"[UR S5S55nURUl URUlUR UlUR$UlUR(UlU$![ [ 4ad [RS:Xa7[R"S[RRSS35nN[R"S5nNf=f)Nramdhip64win32 amdhip64_.dllzlibamdhip64.so)rocm_sdkctypesCDLLstrfind_libraries ImportError IndexErrorsysplatformtorchversionhiphipGetErrorStringcuGetErrorStringhipModuleLoadDatacuModuleLoadDatahipModuleGetFunctioncuModuleGetFunctionhipModuleLaunchKernelcuLaunchKernelhipFuncSetAttributecuFuncSetAttribute)r libs K/home/wildlama/miniconda3/lib/python3.13/site-packages/torch/cuda/_utils.py_get_hip_runtime_libraryr$s 0kk#h55jA!DEF00C00C!66C22C 44C J  $0 <<7 "++ %--*;*;A*>)?tDEC++./C 0s5BAD)DDc[RS:Xa[R"S5$[R"S5$)Nr z nvcuda.dllz libcuda.so.1)rrr rr#_get_cuda_libraryr(-s, ||w{{<(({{>**r'ch[RR(a [5$[ 5$N)rrrr$r(r&r'r#_get_gpu_runtime_libraryr+5s! }}')) ""r'resultcUS:Xag[R"5n[5nURU[R"U55 UR bUR R 5OSn[SU35e)NrUnknown CUDA error CUDA error: )r c_char_pr+rbyrefvaluedecode RuntimeError)r,err_strlibcuda error_messages r# _check_cudar8=sn {ooG&(G VV\\'%:;")--"; AU m_5 66r'cZ[(d [S5eUtpU[RR:waO[R "U5up4[ U[5(aUR5n[SUSUS35e[U5S:Xag[U5S:XaUS$U$)aCheck a cuda.bindings (cuda-python) call result for errors. All cuda.bindings runtime calls return ``(error, *outputs)``. This helper unpacks the tuple, raises on non-success, and returns the outputs (``None`` for zero outputs, scalar for one, tuple otherwise). zcuda.bindings is not availabler/z ()rN) _HAS_CUDA_BINDINGSr4_cuda_bindings_runtime cudaError_t cudaSuccesscudaGetErrorString isinstancebytesr3len)r,errout_r5s r#_check_cuda_bindingsrGIs  ;<<IC ! - - 9 9 : # 5 5   gu % %nn&G\#b ;<< 3x1} 3x1}1v Jr'cSSKn[R"[UR S5S55nURUlURUlUR"UlUR&UlUR*UlUR.UlUR2UlUR6UlUR:UlUR>Ul U$![ [ 4a [RS:XahSRS[RRSS[RRS/5n[R"SUS35nGN8[R"S5nGNQf=f) Nrhiprtcr 0r z libhiprtc.so)!r r rrrrrrrjoinrrrhiprtcGetErrorStringnvrtcGetErrorStringhiprtcCreateProgramnvrtcCreateProgramhiprtcDestroyProgramnvrtcDestroyProgramhiprtcCompileProgramnvrtcCompileProgramhiprtcGetCodeSizenvrtcGetCUBINSize hiprtcGetCode nvrtcGetCUBINhiprtcGetProgramLogSizenvrtcGetProgramLogSizehiprtcGetProgramLognvrtcGetProgramLoghiprtcAddNameExpressionnvrtcAddNameExpressionhiprtcGetLoweredNamenvrtcGetLoweredName)r r" version_strs r#_get_hiprtc_libraryrcfsA .kk#h55h?BCD"66C 44C!66C!66C11C))C!$!UT:war[R"5nTRU[R"U55 URbURR 5OSn[ SU35eg)Nr.r/)r r0rOr1r2r3r4)r,r5r7 NVRTC_SUCCESSlibnvrtcs r# check_nvrtc#_nvrtc_compile..check_nvrtcso ] "oo'G  ( (g1F G==, $$&)  m_=> > #r'utf-8z--offload-arch=z--gpu-architecture=sm_) include_pathsrhz-Iz12.8zPCH requires CUDA 12.8+, got z--pchz.cuzKernel compilation failed: rJ)' torch.cudarprgencoderhget_device_propertiescurrent_devicerr gcnArchNamemajorminorappendrtrrAssertionErrorryrurCr r0c_void_prQr1r_rUc_size_tr[create_string_bufferr2r]r4r3rWrYrarSraw)rzr{r|r}r~rrr source_bytespropsoptionsrcuda_include_paths cuda_path directoryoptionnvrtc_compatible_flagsrw num_options options_arrayprog c_kernel_namereslog_sizelog binary_sizebinaryc_mangled_name mangled_namerrs @@r#_nvrtc_compilers0$%HM ?C ?D ? ?!''0L! 001J1J1LM ==  $)$5$5#6 $)KK= !> G }});(<=DDFG/0B/CDKKMN8&v.' I;'..01(*I NNR {+224 5+ u}}!! "V + #@ASAS@T!UV V  LG$"F NN6==1 2#;< NN5KL5KTDKK(5KLMg,K__{2W=M ?? D## LL  m3  & & (      &&w/M//mDE  & &t[- HC m??$''fll8.DE))(..9##D#.9#)):J:J:L9MNOO//#K**4k1JKL  ( ():): ;F&&tV45__&N$$T=&,,~:VW'%++224     d!34 ::| ##oMs&R,cL\rSrSrS\R SS4SjrS\SS4SjrS r g) _CudaModuleiImodulerNcXl0Ulgr*)_module_kernels)selfrs r#__init___CudaModule.__init__Js 02 r'name _CudaKernelc XR;aURU$SSKJn U"5n[R"5n[ UR [R"U5URURS555 [X@R5nXPRU'U$![an[SUS35UeSnAff=f)Nr)r+rzNo kernel named 'z' in this module) rtorch.cuda._utilsr+r rr8rr1rrrr4AttributeError)rrr+r6funckernelrDs r# __getattr___CudaModule.__getattr__Ns == ==& & ?*,  V ++LL& dkk'6J  !||4F"(MM$ M V #4TF:J!KLRU U VsA-B00 C:C  C)rr) __name__ __module__ __qualname____firstlineno__r rrrr__static_attributes__r&r'r#rrIs/3v343VV Vr'rc \rSrSrSrS\R S\R SS4SjrSS\\ \ \ 4S \\ \ \ 4S \ S-S \ S \ S-SS4 S jjr S\ SS4Sjr Srg)rigzL Represents a compiled CUDA kernel that can be called with PyTorch tensors. rrrNc*XlX lSUlg)Nr)rr_max_shared_mem_bytes)rrrs r#r_CudaKernel.__init__ls  %&"r'gridblockargs shared_memstreamcSSKnURRR5nU(d/n/n/n UGHvn [ XR 5(aU R (d1U R(aU R5(d [S5e[R"U R55n URU 5 U R[R"U 55 M[ U [5(a>[R "U 5n U R[R"U 55 GM[ U ["5(a>[R$"U 5n U R[R"U 55 GMb['S[)U 535e [R[+U 5-"5n[-U 5H,up[R."U [R5X'M. UcSSKnURR35nUS:aSUR4S:XdX@R4:a4UR4S:XaSOSUR4S3n[7S US US 35e[9UR;UR<USUS US USUS US UUR>US5 5 g)a Call the compiled CUDA kernel Args: grid (tuple): Grid dimensions (grid_x, grid_y, grid_z) block (tuple): Block dimensions (block_x, block_y, block_z) args (list): List of arguments to pass to the kernel. PyTorch tensor arguments will be automatically converted to pointers. shared_mem (int): Shared memory size in bytes stream (torch.cuda.Stream): CUDA stream to use. If None, uses current stream. rNz?All tensor arguments must be CUDA tensors or pinned CPU tensorszUnsupported argument type: znot configuredzonly z bytes configuredzKernel requires z' bytes of shared memory (>= 48KB), but ze. Call kernel.set_shared_memory_config(shared_mem) after compilation and before launching the kernel.r;rL) rrh_utilsr+rATensoris_cudais_cpu is_pinned ValueErrorr rdata_ptrrr1rgc_intfloatc_double TypeErrortyperC enumeratecastrcurrent_streamrr4r8rr_as_parameter_)rrrrrrrr6processed_argsc_argsargptrrr c_args_arrayiconfigured_msgs r#__call___CudaKernel.__call__qsT& **##<<>D13C#||,,{{CJJ3==??$Yooclln5%%c* fll3/0C%% S) fll512C''!??3/ fll845"=d3i[ IJJ+0#f+58 'FA$kk#v?LO( > ZZ..0F  "  & &! +z!!/ 01GG 783  & & ;   &6"r')rrr)r;r;r;rNrN)rrrr__doc__r rrtuplerglistrrrrr&r'r#rrgs'V__'foo'$'&/&/ ! _ CcM"_ S#s]#_ Tk _  _ d _  _ B(6(6(6r'rptx kernel_namesc \SSKn[5n[U[5(aUR S5n[ R "5nURR5nU [UR[ R"U5U55 SSS5 U(d [U5$0nUHcn[ R "5n[UR[ R"U5XGR S555 [X5Xg'Me U$!,(df  N=f)a Loads a CUDA module from PTX code and returns a module object that can access kernels. Args: ptx (bytes or str): The PTX code to load kernel_names (list, optional): List of kernel names to extract from the module. If None, will return a module object with __getattr__. Returns: object: If kernel_names is None, returns a module object with __getattr__ to access kernels. If kernel_names is provided, returns a dict mapping kernel names to _CudaKernel objects. rNr)rr+rArrr rrhrr8rr1rrr) rrrr6rrkernelsrrs r#_cuda_load_modulers '(G#sjj!__ F ZZ & & (F G,,V\\&-A3GH  6""G   ' ' T"FKK,@  $D1  N! s &0D D+deviceoptional allow_cpuc[U[5(aU$[U[5(a[R"U5n[U[R5(aDU(aUR S;a[ SU35eOUR S:wa[ SU35e[RR5(d5[U[RR5(a UR$[XU5$)aGet the device index from :attr:`device`, which can be a torch.device object, a Python integer, or ``None``. If :attr:`device` is a torch.device object, returns the device index if it is a CUDA device. Note that for a CUDA device without a specified index, i.e., ``torch.device('cuda')``, this will return the current default CUDA device if :attr:`optional` is ``True``. If :attr:`allow_cpu` is ``True``, CPU devices will be accepted and ``-1`` will be returned in this case. If :attr:`device` is a Python integer, it is returned as is. If :attr:`device` is ``None``, this will return the current default CUDA device if :attr:`optional` is ``True``. )rhcpuz(Expected a cuda or cpu device, but got: rhz!Expected a cuda device, but got: ) rArgrrrrrjit is_scriptingrhidx_torch_get_device_index)rrrs r#rr-s &# &#f%&%,,'' {{/1 #KF8!TUU2 [[F "@IJ J 99 ! ! # # fejj// 0 0::  "6Y ??r')NNNFr*)FF)"r rtypingrr cuda.bindingsrr=r<r torch._utilsrrrr$r(r+rgr8rGrcrnrprrryboolrrBrrrdictrr&r'r#rs   F&++.+6;;+#&++# 7 7 7:V[[:6FKK6&$fkk$tCy6&*%) $ O$O$O$d O$d{ O$ + O$  O$ 5#:O$dVV