¯wgg hUddlZddlZddlZddlZddlZddlZddlZddlZddlZddl Z ddl Z ddl Z ddl Z ddl Z ddlmZddlZddlZddlZddlmZddlmZddlmZddlmZddlmZmZmZmZm Z dd l!m"Z"m#Z#dd l$m%Z%e jLd k(Z'e jLjQd Z)e jLjQd Z*e'rdndZ+e'rdndZ,e'rdndZ-e'rdndZ.e'rdndZ/ej`jce2Z3ej`jiej`jie3Z5ej`jme5dZ7e'rdndZ8dZ9dZ:e e e;dfe e;dffZe9dfe9dfe9dfe9dfdddddZ?e>e@d<d ZAeAdfeAd!feAd!feAd"feAd"feAd#feAd#fd$ZBe>e@d%<gd&ZCd'eee=d(ee=fd)ZDd(ee=fd*ZEd(ee=fd+ZFd(e=fd,ZGd-ZHd.ZId/ZJd0ZKd1ZLeFZMeMreGd2ndZNeMejjd3nd4ZQdZRejj5eSd5ejjjd6dd7DZRejjreEndZWejjd8xsejjd9ZZejd:Z\gd;Z]gd<Z^gd=Z_gd>Z`gd?ZaeZbd@dAdBZcdCZdd(eefdDZfd(ee=fdEZgdFZhd(e=fdGZidHe=d(eefdIZjd(e eee"ffdJZkdKe=dLe"d(dfdMZlGdNdOe%ZmdPZndQZodvdReed(ee=fdSZpdvdReed(ee=fdTZq dwdUee=ee=fdVeeefdWZrdXZsdYZtdZZu dvd[Zvd\Zw dxd]Zx dyd^e=d_eedVeeed(dfd`ZydUee=d^e=d_eedVeeed(df daZz dvdUee=d^e=d_eedVeeedbeed(df dcZ{ddZ|deZ}dfZ~dzdgeee=d(ee=fdhZdzdgeee=d(ee=fdiZdje=d_eed(e=fdkZd_eed(ee;fdlZd^e=d_eedme=d(dfdnZdoZdpZ d{dqZ d{drZd(e=fdsZdte=d(eefduZy)|N)Path) FileBaton)ExtensionVersioner) hipify_python)GeneratedFileCleaner)DictListOptionalUnionTuple) TorchVersionVersion) build_extwin32darwinlinuxz.pydz.soz.exelibz.dllz/DLLz-shared)oem)rr)ri^.) r r))rr r)z11.011.111.211.311.411.511.611.7CUDA_GCC_VERSIONS)r(rr) r)r)r r!r"r#r$r%r&CUDA_CLANG_VERSIONS)get_default_build_rootcheck_compiler_ok_for_platform*get_compiler_abi_compatibility_and_versionBuildExtension CppExtension CUDAExtension include_paths library_pathsload load_inlineis_ninja_availableverify_ninja_availability&remove_extension_h_precompiler_headersget_cxx_compilercheck_compiler_is_gccargsreturncH|sgS|Dcgc]}d|vrd|dn|c}Scc}w)zQuote command-line arguments for DOS/Windows conventions. Just wraps every argument which contains blanks in double quotes, and returns a new argument list.  "r)r;args ^/home/mcse/projects/flask/flask-venv/lib/python3.12/site-packages/torch/utils/cpp_extension.py_nt_quote_argsrBRs3  9= >##*auAJ# - >> >scZtjjdxstjjd}|tjd}|=tj j tj j |}nUtr,tjd}t|dk(rd}n|d}nd}tj j|sd}|r=tjjstd |d tj |S) zFind the CUDA install path. CUDA_HOME CUDA_PATHNnvccz7C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v*.*rrz/usr/local/cudaz+No CUDA runtime is found, using CUDA_HOME=''file)osenvirongetshutilwhichpathdirname IS_WINDOWSgloblenexiststorchcuda is_availableprintsysstderr) cuda_home nvcc_path cuda_homess rA_find_cuda_homer^]s {+Jrzz~~k/JILL(   (BCI!YYMO z?a' "I *1 I- 77>>), 002 ;I;aH::  ctjjdxstjjd}|tjd}|tj j tj j tj j|}tj j|dk(rCtj j |}n#d}tj j|r|}|r9tjjtd|dtj|S) zFind the ROCm install path. ROCM_HOME ROCM_PATHhipcchipz /opt/rocmz+No ROCm runtime is found, using ROCM_HOME='rGrH)rJrKrLrMrNrOrPrealpathbasenamerTrUversionrdrXrYrZ) rocm_home hipcc_path fallback_paths rA_find_rocm_homerkxs {+Jrzz~~k/JI\\'*  !  ,)./Iww *e3GGOOI6 (Mww~~m,) U]]&&. ;I;aH::  r_ct tdtr tdtjj tg|S)z Join paths with ROCM_HOME, or raises an error if it ROCM_HOME is not set. This is basically a lazy way of raising an error for missing $ROCM_HOME only once we need to get any ROCm-specific path. zSROCM_HOME environment variable is not set. Please set it to your ROCm install root.zDBuilding PyTorch extensions using ROCm and Windows is not supported.)raOSErrorrQrJrOjoinpathss rA_join_rocm_homerqsLAB B ;< < 77<< *E **r_a8 !! WARNING !! !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! Your compiler ({}) may be ABI-incompatible with PyTorch! Please use a compiler that is ABI-compatible with GCC 5.0 and above. See https://gcc.gnu.org/onlinedocs/libstdc++/manual/abi.html. See https://gist.github.com/goldsborough/d466f43e8ffc948ff92de7486c5216d6 for instructions on how to install GCC 5 or higher. !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! !! WARNING !! a !! WARNING !! !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! Your compiler ({user_compiler}) is not compatible with the compiler Pytorch was built with for this platform, which is {pytorch_compiler} on {platform}. Please use {pytorch_compiler} to to compile your extension. Alternatively, you may compile PyTorch from source using {user_compiler}, and then you can also use {user_compiler} to compile your extension. See https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md for help with compiling PyTorch from source. !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! !! WARNING !! z The detected CUDA version ({0}) mismatches the version that was used to compile PyTorch ({1}). Please make sure to use the same CUDA versions. zThe detected CUDA version ({0}) has a minor version mismatch with the version that was used to compile PyTorch ({1}). Most likely this shouldn't be a problem.z CUDA was not found on the system, please set the CUDA_HOME or the CUDA_PATH environment variable or add NVCC to your system PATH. The extension compilation will fail. rdTFc#2K|]}t|ywN)int).0vs rA rwsJAQJs. CUDNN_HOME CUDNN_PATHz\d+\.\d+\.\d+\w+\+\w+) z/MDz/wd4819z/wd4251z/wd4244z/wd4267z/wd4275z/wd4018z/wd4190z/wd4624z/wd4067z/wd4068z/EHsc)&base_class_has_different_dll_interfacefield_without_dll_interface#dll_interface_conflict_none_assumed(dll_interface_conflict_dllexport_assumed)z-D__CUDA_NO_HALF_OPERATORS__z-D__CUDA_NO_HALF_CONVERSIONS__z"-D__CUDA_NO_BFLOAT16_CONVERSIONS__z-D__CUDA_NO_HALF2_OPERATORS__z--expt-relaxed-constexpr)-fPICz-D__HIP_PLATFORM_AMD__=1z -DUSE_ROCM=1z -DHIPBLAS_V2)z-DCUDA_HAS_FP16=1z-D__HIP_NO_HALF_OPERATORS__=1z-D__HIP_NO_HALF_CONVERSIONS__=1x86 x86_amd64)rz win-amd64ctr"tjjdd}|Stjjdd}|S)NCXXclc++)rQrJrKrL)compilers rAr9r9s;::>>%. O::>>%/ Or_c^tjtjj Srs)!BUILT_FROM_SOURCE_VERSION_PATTERNmatchrUrg __version__rr_rA_is_binary_buildrs 066u}}7P7PQ QQr_ctrddgSgdS)Nclang++clang)zg++gcczgnu-c++zgnu-ccrr)IS_MACOSrr_rA _accepted_compilers_for_platformrs#+Iw h1hhr_ctjj|r*t|5}|j }ddd|k(ryt|d5}|j |dddy#1swY6xYw#1swYyxYw)z Equivalent to writing the content into the file but will not touch the file if it already had the right content (to avoid triggering recompile). Nw)rJrOrTopenreadwrite)filename new_contentfcontent source_files rA _maybe_writer s~  ww~~h (^ qffhG  k !  h ' +&''  ''sA1A=1A:=Bc|tjjtjj dS)a Return the path to the root folder under which extensions will built. For each extension module built, there will be one folder underneath the folder returned by this function. For example, if ``p`` is the path returned by this function and ``ext`` the name of an extension, the build folder for the extension will be ``p/ext``. This directory is **user-specific** so that multiple users on the same machine won't meet permission issues. torch_extensions)appname)rJrOrerU_appdirsuser_cache_dirrr_rAr,r,s, 77  ENN99BT9U VVr_rc,trytj|ytjj t fdtDrytjj}d|d<tj|dgtj|jt}trt!j"dt j$}t!j&||}t)|d k7rd |vStjj |d j+tjj-d k(rd |vryt fdtDSt.r|j1dSy)a Verify that the compiler is the expected one for the current platform. Args: compiler (str): The compiler executable to check. Returns: True if the compiler is gcc/g++ on Linux or clang/clang++ on macOS, and always True for Windows. TFc3&K|]}|v ywrsrruname compiler_paths rArwz1check_compiler_ok_for_platform..=s PT4= PCLC_ALL-vrZenv^COLLECT_GCC=(.*)$rz clang versionrr gcc versionc3&K|]}|v ywrsrrs rArwz1check_compiler_ok_for_platform..OsXT4=(Xrz Apple clang)rQrMrNrJrOreanyrrKcopy subprocess check_outputSTDOUTdecodeSUBPROCESS_DECODE_ARGSIS_LINUXrecompile MULTILINEfindallrSstriprfr startswith)rrversion_stringpatternresultsrs @rAr-r-*sHLL*MGG$$]3M P-M-O PP **// CCMhZ,,h-=jFWFW]`ahhkABN**12<<@**Wn5 w<1 #n4 4(()9)9);< 77  M *e 3 8WX5U5WXXX((77 r_cts dtdfStjj ddvr dtdfSt |sQt jtj|tdtjdtdfStr dtdfS trOt}t!j"|dd g}|j$t&j)j+d }nt,}t!j"|t j. }t1j2d |j$t&j)}|gd nt5|j7}t=t?t@||k\rdtd jC|fS|dd jC|}t jtDj|dtd jC|fS#t8$rCtj:\}}}t jd|d|dtdfcYSwxYw)a Determine if the given compiler is ABI-compatible with PyTorch alongside its version. Args: compiler (str): The compiler executable name to check (e.g. ``g++``). Must be executable in a shell process. Returns: A tuple that contains a boolean that defines if the compiler is (likely) ABI-incompatible with PyTorch, followed by a `TorchVersion` string that contains the compiler version separated by dots. Tz0.0.0TORCH_DONT_CHECK_COMPILER_ABION1YESTRUEYr) user_compilerpytorch_compilerplatformFz-dumpfullversionz -dumpversionrx)rZz(\d+)\.(\d+)\.(\d+))0rrz$Error checking compiler version for : r>)#rrrJrKrLr-warningswarnWRONG_COMPILER_WARNINGformatrrYrrrMINIMUM_GCC_VERSIONrrrrrsplitMINIMUM_MSVC_VERSIONrrsearchlistgroups Exceptionexc_infotuplemaprtrnABI_INCOMPATIBILITY_WARNING)rminimum_required_version versionstrrg compiler_infor_errors rAr.r.Vs  l7+,, zz~~56:YYl7+,, *( 3 ,33"=?B\\4# $|G,--l7+,, . ': $#00(? < 12 33 .lln 5! =49.4.0 is requiredrrrrz There are no z) version bounds defined for CUDA version zV11.4.48rrxz>=z, > >>/55 54;;K>V>VW^>_':ev33 m_4]^n]opq:NO_:`77 ) ).BFW.W(/ %#&88C5I,J#K (+S:S1T(U% !9 :#>[=\] l+CD D3M?"EUDVW==M.cls_with_optionscF|jt||i|yrs)updatesuper__init__)selfr;kwargs __class__optionss rArz>BuildExtension.with_options..cls_with_options.__init__s! g& $1&1r_)__name__ __module__ __qualname__r __classcell__)rrs@rAcls_with_optionsrs  2 2r_r r)clsrr s ` rA with_optionszBuildExtension.with_optionss 2s 2  r_r<c t||i||jdd|_|jdd|_|jr9d}t s,t j|jdd|_yyy)Nno_python_abi_suffixF use_ninjaTznAttempted to use ninja as the BuildExtension backend but {}. Falling back to using the slow distutils backend.zwe could not find ninja.) rrrLrrr6rrr)rr;rmsgrs rArzBuildExtension.__init__sw $)&)$*JJ/Eu$M!K6 >>KC%' cjj)CDE!&( r_cJt||jrd|_yy)NT)rfinalize_optionsrforce)rrs rArzBuildExtension.finalize_optionss!  " >>DJ r_c j\}}d}tj}t|d}|sO|rM|jD]-}t j j|\}}|dk(s+d}nt|d}|s|rM|rts t||jD]}t|jtr%dD] }||jvsg|j|<"j|ddD]B} ttj d| } | #t"r*j|d| d | d Dj%|j'|d |jvs̉j(rJd |j*d j,xj.gdz c_tj0j2j5r j,xj.dgz c_j,j6dk(rNj,xj8ddgz c_j,j:j,j<nj,j>dfd dddfd } dfd } d dfd } dfd }j,j6dk(r0j(r|j,_nA| j,_n/j(r| j,_n| j,_tAjBy)NF.cuT)cxxrF-DTORCH_API_INCLUDE_EXTENSION_H COMPILER_TYPESTDLIB BUILD_ABI _PYBIND11_ -DPYBIND11_z="r? nvcc_dlinkz;With dlink=True, ninja is required to build cuda extension rx)r.cuh.hipz.mmmsvcr"cjjdk(rdnd}|jddz}tfd|Ds|j |yy)Nr$z/{}:z-{}=stdzc++17c3@K|]}|jywrsr)ruflagcpp_flag_prefixs rArwzZBuildExtension.build_extensions..append_std17_if_no_std_present..5sKDt7K)r compiler_typerrappend)cflagscpp_format_prefixcpp_flagr*rs @rAappend_std17_if_no_std_presentzGBuildExtension.build_extensions..append_std17_if_no_std_present/sY+/--*E*E*OU[ /66u=O&0HKFKK h'Lr_ctddgz|zt|z}tjd}|%t d|Ds|j d|g|S)N--compiler-options'-fPIC'CCc3>K|]}|jdyw))-ccbinz--compiler-bindirNr(rur)s rArwzKBuildExtension.build_extensions..unix_cuda_flags..Bs`QUDOO,KL`r7)COMMON_NVCC_FLAGS_get_cuda_arch_flagsrJgetenvrextend)r._ccbins rAunix_cuda_flagsz8BuildExtension.build_extensions..unix_cuda_flags8sc'+Y783F;.convert_to_absolute_paths_inplaceHsS s5z*=A77==q2#%77??58#<a=!r_cJtj|} jj}t |rut r t ddn tddg} jjd|t|tr|d}t rt|zt|z}n |}nt|tr|d}t r t|z} | |||||| jjd|y# jjdwxYw)NrrcrF compiler_sor)rdeepcopyrrG _is_cuda_fileIS_HIP_EXTENSIONrq_join_cuda_homeset_executable isinstancedictCOMMON_HIPCC_FLAGS_get_rocm_arch_flagsCOMMON_HIP_FLAGS) objsrcextcc_argsextra_postargspp_optsr.original_compilerrFr1original_compilerr?s rAunix_wrap_single_compilezABuildExtension.build_extensions..unix_wrap_single_compileOs]]>2F O$(MM$=$=! %?OOE7;UdejlrUstDMM00E!&$/!''!3f!.unix_wrap_ninja_compilehs (4J .dmm.H.H I ,,Z-97-4nF 3Aw!MM66w}UM"mm77;OC w78I .$/,U3 ">2 .< *; 7# K+ nd3'5f'=$'+N';$#'7:NO_:`'`$'7:L'LO_'_$'67G'H$./?@7BC!u{{1~C C.win_cuda_flagss%1&9: ;r_c tj| _d} fd} | j_ |||||||| j_S# j_wxYw)NcD tjd fd|DDcgc]}|r|jd}}tjd fd|DDcgc]}|r|jd}}tjd fd|DDcgc]}|r|jd}}t|dk\r=t|dk\r.|d }|d }t |rt d d }t jtrjd }n)t jtr j}ng}|d d gz}tD] } d| g|z} tD] } dd| zg|z}|d|d|g|z|z}nxt jtr$tjdz}|||z }n:t jtr tjz}|||z }|Scc}wcc}wcc}w)Nz /T(p|c)(.*)c3@K|]}j|ywrsr)ruelem src_regexs rArwzbBuildExtension.build_extensions..win_wrap_single_compile..spawn..(O4)>(Or+ryz/Fo(.*)c3@K|]}j|ywrsry)rurz obj_regexs rArwzbBuildExtension.build_extensions..win_wrap_single_compile..spawn..r|r+rz ((\-|\/)I.*)c3@K|]}j|ywrsry)rurz include_regexs rArwzbBuildExtension.build_extensions..win_wrap_single_compile..spawn..sHDm11$7Hr+rrrF -std=c++17--use-local-env -Xcompiler-Xcudafe--diag_suppress=z-cz-or) rrrrSrIrKrMr.rNrCOMMON_MSVC_FLAGSMSVC_IGNORE_CUDAFE_WARNINGS)cmdmsrc_listobj_list include_listrSrRrFr.r)ignore_warningrr~r{r1original_spawnrrus @@@rAspawnzOBuildExtension.build_extensions..win_wrap_single_compile..spawns!JJ}5 (O3(O#$GGAJ JJy1 (O3(O#$GGAJ !# ? ; ICH AGGAJ    x=A%#h-1*<"1+C"1+C$S).uf=%dkk48%)[[%8F' T:%)[[F%'F!/!7v #DKK6!2T[[!@6v>v %c**Y   sH%H#H)rrHr.rr)r]rlrmrfrnrorVrprr1rYrrrus rAwin_wrap_single_compilez@BuildExtension.build_extensions..win_wrap_single_compilesg--7DK!N/ +b 5&+ #'V(4e](6A'5 #n #s AA+c jjsjjtjj |}jj jj||||||\}} }} }|xsg} g} |r&| jjjn%| jjj| jt| | z| z} ttt|} t|t r|d}n t#|}|d}d}| rddg}| D]$}|j%d|j%|&t&D]'}|j%d|j%d|z)|j| t|t r|d}n t#|}|}t)| } t)|}| rt)|}t)|}t|t rd|vr |d}nd}t+|| | |||||d | | S) NrrrrrrrFr!Tr\)r initialized initializerJrOrCrfrgr=compile_options_debugcompile_optionsrrrrIrMrNrr-rrBrk)r]rlrmrfrnrorVrprr^rWrqr.rer_rar` common_cflagrrbr1rErrus rAwin_wrap_ninja_compilez?BuildExtension.build_extensions..win_wrap_ninja_compilesH==,, ((*4J .dmm.H.H I ,,Z-97-4nF 3Aw*/RMF dmmAAB dmm;;<  !2 3m+g5FC w78I .$/,U3 ">2 *; 7# K+->? $15L&&|4&&|45'BLN&&z2&&'9N'JKL""7+nd3'5f'=$'+N';$#12B#C #F+F(5K,[9 #12B#C .$/LN4R)7|8T)U&)-& 1''!1'= *# %Nr_r<N)NNNrNNN)" _check_abiiter extensionsnextr]rJrOsplitextrJrrMextra_compile_argsrN_add_compile_flagrrU_CrQ_define_torch_extension_name_add_gnu_cpp_abi_flagrrrsrc_extensionsbackendsmpsis_builtr,_cpp_extensionsrr_compilerbuild_extensions)rrrcuda_extextension_iter extensionsourcerrTrvalrZrsrrr1rErYrr?rus` @@@@@@rArzBuildExtension.build_extensionss.*.//*;' 'doo.. y#++ ))&13%<#H   ^T2I y ,  /? @ wI)66=*?C)">">><> 44S9?  " "9.O P@ Tehh*TF(;<?:**9 D6C5PQ6RS T  - -i 8  & &y 1y;;;~~v)denesesdttu'vv~/ w4 $$(??$ >>   & & ( MM ( (UG 3 ( == & && 0 MM ) )eV_ < )#}}44 !]]00N#}}55  (   = O O404+/15*+2637,0N N ` ; 04+/15*+2637,0C 5C 5L/3*.04)*1526+/T T p == & && 0~~(> %(? %~~(? %)A &""4(r_ct||}|jr-|jd}|dd|ddz}dj |}|S)Nrx)rget_ext_filenamerrrn)rext_name ext_filenameext_filename_parts without_abirs rArzBuildExtension.get_ext_filenamefs]w/9   $ $!-!3!3C!8 ,Sb14Frs4KKK88K0Lr_ct|jdr|jjd}n t}t |\}}t r1dt jvrdt jvr d}t|||fS)N compiler_cxxrVSCMD_ARG_TGT_ARCHDISTUTILS_USE_SDKzIt seems that the VC environment is activated but DISTUTILS_USE_SDK is not set.This may lead to multiple activations of the VC env.Please set `DISTUTILS_USE_SDK=1` and try again.) hasattrrrr9r.rQrJrK UserWarning)rrrrgrs rArzBuildExtension._check_abivsx 4==. 1}}11!4H')H?I 7 ."**>> # xdoctest: +SKIP >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CPP_EXT) >>> from setuptools import setup >>> from torch.utils.cpp_extension import BuildExtension, CppExtension >>> setup( ... name='extension', ... ext_modules=[ ... CppExtension( ... name='extension', ... sources=['extension.cpp'], ... extra_compile_args=['-g'], ... extra_link_args=['-Wl,--no-as-needed', '-lm']) ... ], ... cmdclass={ ... 'build_ext': BuildExtension ... }) rf library_dirs librariesc10rU torch_cpu torch_pythonsleefrlanguage)rLr2r3r-rQ setuptools Extension)rr]r;rrfrrs rAr0r0s:::nb1LMO#L)F>::nb1LMO#L)F> ;+I U W [! ^$!#F;F:   g ? ? ??r_c F|jdg}|tdz }||d<|jdg}|jd|jd|jd|jdtr4|jd |jd |jd n3|jd |jd |jd||d<|jdg}tr t j }t j|||tjj|dg|Dcgc]!}tjj|#c}ddd} t} |D]t} tjj| } | | vr| | j| | jn| } | jtjj| |vt| }|t!dz }||d<d|d<|jdg}|jddxs|}|r|jdi}|jdg}|dgz }||Dcgc]}d| c}z }||Dcgc]}d| c}z }t"j$j&,t)t"j$j&dk\r|dgz }||d<||d<t+j,||g|i|Scc}wcc}wcc}w)a Create a :class:`setuptools.Extension` for CUDA/C++. Convenience method that creates a :class:`setuptools.Extension` with the bare minimum (but often sufficient) arguments to build a CUDA/C++ extension. This includes the CUDA include path, library path and runtime library. All arguments are forwarded to the :class:`setuptools.Extension` constructor. Full list arguments can be found at https://setuptools.pypa.io/en/latest/userguide/ext_modules.html#extension-api-reference Example: >>> # xdoctest: +SKIP >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CPP_EXT) >>> from setuptools import setup >>> from torch.utils.cpp_extension import BuildExtension, CUDAExtension >>> setup( ... name='cuda_extension', ... ext_modules=[ ... CUDAExtension( ... name='cuda_extension', ... sources=['extension.cpp', 'extension_kernel.cu'], ... extra_compile_args={'cxx': ['-g'], ... 'nvcc': ['-O2']}, ... extra_link_args=['-Wl,--no-as-needed', '-lcuda']) ... ], ... cmdclass={ ... 'build_ext': BuildExtension ... }) Compute capabilities: By default the extension will be compiled to run on all archs of the cards visible during the building process of the extension, plus PTX. If down the road a new card is installed the extension may need to be recompiled. If a visible card has a compute capability (CC) that's newer than the newest version for which your nvcc can build fully-compiled binaries, Pytorch will make nvcc fall back to building kernels with the newest version of PTX your nvcc does support (see below for details on PTX). You can override the default behavior using `TORCH_CUDA_ARCH_LIST` to explicitly specify which CCs you want the extension to support: ``TORCH_CUDA_ARCH_LIST="6.1 8.6" python build_my_extension.py`` ``TORCH_CUDA_ARCH_LIST="5.2 6.0 6.1 7.0 7.5 8.0 8.6+PTX" python build_my_extension.py`` The +PTX option causes extension kernel binaries to include PTX instructions for the specified CC. PTX is an intermediate representation that allows kernels to runtime-compile for any CC >= the specified CC (for example, 8.6+PTX generates PTX that can runtime-compile for any GPU with CC >= 8.6). This improves your binary's forward compatibility. However, relying on older PTX to provide forward compat by runtime-compiling for newer CCs can modestly reduce performance on those newer CCs. If you know exact CC(s) of the GPUs you want to target, you're always better off specifying them individually. For example, if you want your extension to run on 8.0 and 8.6, "8.0+PTX" would work functionally because it includes PTX that can runtime-compile for 8.6, but "8.0 8.6" would be better. Note that while it's possible to include all supported archs, the more archs get included the slower the building process will be, as it will build a separate kernel image for each arch. Note that CUDA-11.5 nvcc will hit internal compiler error while parsing torch/extension.h on Windows. To workaround the issue, move python binding logic to pure C++ file. Example use: #include at::Tensor SigmoidAlphaBlendForwardCuda(....) Instead of: #include torch::Tensor SigmoidAlphaBlendForwardCuda(...) Currently open issue for nvcc bug: https://github.com/pytorch/pytorch/issues/69460 Complete workaround code example: https://github.com/facebookresearch/pytorch3d/commit/cb170ac024a949f1f9614ffe6af1c38d972f7d48 Relocatable device code linking: If you want to reference device symbols across compilation units (across object files), the object files need to be built with `relocatable device code` (-rdc=true or -dc). An exception to this rule is "dynamic parallelism" (nested kernel launches) which is not used a lot anymore. `Relocatable device code` is less optimized so it needs to be used only on object files that need it. Using `-dlto` (Device Link Time Optimization) at the device code compilation step and `dlink` step help reduce the protentional perf degradation of `-rdc`. Note that it needs to be used at both steps to be useful. If you have `rdc` objects you need to have an extra `-dlink` (device linking) step before the CPU symbol linking step. There is also a case where `-dlink` is used without `-rdc`: when an extension is linked against a static lib containing rdc-compiled objects like the [NVSHMEM library](https://developer.nvidia.com/nvshmem). Note: Ninja is required to build a CUDA Extension with RDC linking. Example: >>> # xdoctest: +SKIP >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CPP_EXT) >>> CUDAExtension( ... name='cuda_extension', ... sources=['extension.cpp', 'extension_kernel.cu'], ... dlink=True, ... dlink_libraries=["dlink_lib"], ... extra_compile_args={'cxx': ['-g'], ... 'nvcc': ['-O2', '-rdc=true']}) rT)rVrrrUrramdhip64c10_hip torch_hipcudartc10_cuda torch_cudarf*)project_directoryoutput_directoryheader_include_dirsincludes extra_files show_detailedis_pytorch_extensionhipify_extra_files_onlyrrdlink_librariesdlinkFrr!z-dlink-Lz-lr!z-dlto)rLr3r-rJrJgetcwdrhipifyrOrnrCset hipified_pathaddrelpathrr2rUrgrVrrr)rr]r;rrrrf build_dirs hipify_resulthipified_sourcesrs_abshipified_s_absrrrextra_compile_args_dlinkxs rAr1r1sL::nb1LMt,,L)F> ;+I U W [! ^$$#%"$&#F;::nb1LIIK %,,'& ,ggll9c235<=+=!%$(  5 MFGGOOF+EEJmE[+E2@@L,E2@@SX   !K L  M'(Mt,,L)F>F:jj!2B7O JJw & 9/E #ZZ(:%E$Gs&L L LrVctjjtd}|tjj|ddddtjj|dtjj|dg}|rQtrK|j tjj|d|j t d|S|rtd}|dk7r|j |tjjd d x}r|dk7r|j |t3|j tjjtd|S) z Get the include paths required to build a C++ or CUDA extension. Args: cuda: If `True`, includes CUDA-specific include paths. Returns: A list of include path strings. includerUcsrcapiTHTHCTHHz /usr/include CUDA_INC_PATHN) rJrOrn _TORCH_PATHrJr-rqrKrKrLrz)rV lib_includerpcuda_home_include cuda_inc_paths rAr2r2{s'',,{I6K  ['65)D  [$'  [%( E   RWW\\+u56 _Y/0 L +I6  . LL* + ZZ^^OTB BM B/ LL '  ! LLj)< = Lr_crtg}|r]trWd}|jt|t3|jt j jtd|S|rtr!t j jdd}nTd}t j jt|s*t j jtdrd}|jt|t3|jt j jt||S)z Get the library paths required to build a C++ or CUDA extension. Args: cuda: If `True`, includes CUDA-specific library paths. Returns: A list of library path strings. rx64lib64) TORCH_LIB_PATHrJr-rqHIP_HOMErJrOrnrQrTrKrz)rVrplib_dirs rAr3r3s E   _W-.   LLh6 7 L  ggll5%0GGGGNN?7#;<GGNN?5#9:  _W-.  ! LLj': ; Lr_r]rec rt|t|tr|gn||||||xs t||||| | |  S)a Load a PyTorch C++ extension just-in-time (JIT). To load an extension, a Ninja build file is emitted, which is used to compile the given sources into a dynamic library. This library is subsequently loaded into the current Python process as a module and returned from this function, ready for use. By default, the directory to which the build file is emitted and the resulting library compiled to is ``/torch_extensions/``, where ```` is the temporary folder on the current platform and ```` the name of the extension. This location can be overridden in two ways. First, if the ``TORCH_EXTENSIONS_DIR`` environment variable is set, it replaces ``/torch_extensions`` and all extensions will be compiled into subfolders of this directory. Second, if the ``build_directory`` argument to this function is supplied, it overrides the entire path, i.e. the library will be compiled into that folder directly. To compile the sources, the default system compiler (``c++``) is used, which can be overridden by setting the ``CXX`` environment variable. To pass additional arguments to the compilation process, ``extra_cflags`` or ``extra_ldflags`` can be provided. For example, to compile your extension with optimizations, pass ``extra_cflags=['-O3']``. You can also use ``extra_cflags`` to pass further include directories. CUDA support with mixed compilation is provided. Simply pass CUDA source files (``.cu`` or ``.cuh``) along with other sources. Such files will be detected and compiled with nvcc rather than the C++ compiler. This includes passing the CUDA lib64 directory as a library directory, and linking ``cudart``. You can pass additional flags to nvcc via ``extra_cuda_cflags``, just like with ``extra_cflags`` for C++. Various heuristics for finding the CUDA install directory are used, which usually work fine. If not, setting the ``CUDA_HOME`` environment variable is the safest option. Args: name: The name of the extension to build. This MUST be the same as the name of the pybind11 module! sources: A list of relative or absolute paths to C++ source files. extra_cflags: optional list of compiler flags to forward to the build. extra_cuda_cflags: optional list of compiler flags to forward to nvcc when building CUDA sources. extra_ldflags: optional list of linker flags to forward to the build. extra_include_paths: optional list of include directories to forward to the build. build_directory: optional path to use as build workspace. verbose: If ``True``, turns on verbose logging of load steps. with_cuda: Determines whether CUDA headers and libraries are added to the build. If set to ``None`` (default), this value is automatically determined based on the existence of ``.cu`` or ``.cuh`` in ``sources``. Set it to `True`` to force CUDA headers and libraries to be included. is_python_module: If ``True`` (default), imports the produced shared library as a Python module. If ``False``, behavior depends on ``is_standalone``. is_standalone: If ``False`` (default) loads the constructed extension into the process as a plain dynamic library. If ``True``, build a standalone executable. Returns: If ``is_python_module`` is ``True``: Returns the loaded PyTorch extension as a Python module. If ``is_python_module`` is ``False`` and ``is_standalone`` is ``False``: Returns nothing. (The shared library is loaded into the process as a side effect.) If ``is_standalone`` is ``True``. Return the path to the executable. (On Windows, TORCH_LIB_PATH is added to the PATH environment variable as a side effect.) Example: >>> # xdoctest: +SKIP >>> from torch.utils.cpp_extension import load >>> module = load( ... name='extension', ... sources=['extension.cpp', 'extension_kernel.cu'], ... extra_cflags=['-O2'], ... verbose=True) keep_intermediates) _jit_compilerMr_get_build_directory) rr] extra_cflagsextra_cuda_cflags extra_ldflagsextra_include_pathsrcrdreis_python_module is_standaloners rAr4r4sPx  - 7>/g>- / /r_cg}dD]A}ttjd|}|#tr*|j d|d|dC|S)Nrrr z=\"z\")rrUrrQr-) abi_cflagspnamepvals rA_get_pybind11_abi_build_flagsr0s^J9Buxx:eW!56  J    E7$tfC@ AB r_cfdtttjjzg}|Sr)rrtrUrr)glibcxx_abi_cflagss rA_get_glibcxx_abi_build_flagsrDs*5C@_@_<`8aab r_ctsytjj}d|d< t j |dgtj |jt}tjdtj}tj||}t|dk7rytjj!|d j#}tjj%|d k(rd |vry y#t$r`} t j |dgtj |jt}n#t$r}Yd}~Yd}~yd}~wwxYwYd}~d}~wwxYw) NFrrrrrrrrrrT)rrJrKrrrrrrrrrrrrSrOrerrf)rrrerrrs rAr:r:Hs=  **// CCMl00(D1A*J[J[adelloEFjj-r||r)rstringelements rA listToStringzF_check_and_build_extension_h_precompiler_headers..listToString|s7 9M &G w} %F & r_cxtjddd|d|d|d|d|d|d|djS)Nz[ \n]+r>z z -x c++-header z -o z )rsubr)r head_file head_file_pchrqtorch_include_dirsr r s rAformat_precompiler_header_cmdzW_check_and_build_extension_h_precompiler_headers..format_precompiler_header_cmdszvv   /)DqI[H\\]^q]rrsuAtBBCDQCRR   %'  r_c*|jdd}|S)Nr>r)replace)r signatures rAcommand_to_signaturezN_check_and_build_extension_h_precompiler_headers..command_to_signaturesKKS) r_ctjj|}|duryt|5}|j }||k(cdddS#1swYyxYw)NF)rJrOisfilerr) file_pathr)b_existrIrs rAcheck_pch_signature_in_filezU_check_and_build_extension_h_precompiler_headers..check_pch_signature_in_filesP''..+ e  )_ (iikG'  ( ( (s AActjj|s t|j ddyy#t $r6}|j t jk7rtd||Yd}~yd}~wwxYw)NT)parentsexist_okzFail to create path ) rJrOrTrmkdirrmerrnoEEXISTr)path_direxcs rA_create_if_not_existzN_check_and_build_extension_h_precompiler_headers.._create_if_not_existsrww~~h' SX$$TD$A( S99 ,&)=hZ'HIsR- Ss? A>,A99A>ctjj|t|d5}|j ||j dddy#1swYyxYw)Nr)rJrOrPrrclose)r-pch_signrr8s rAwrite_pch_signature_to_filezU_check_and_build_extension_h_precompiler_headers..write_pch_signature_to_filesNRWW__Y78 )S ! Q GGH  GGI   s "AA'c tj|dtjy#tj$r}t d||d}~wwxYw)NT)shellrZz)Compile PreCompile Header fail, command: )rrrCalledProcessErrorr)pch_cmdrs rAbuild_precompile_headerzQ_check_and_build_extension_h_precompiler_headers..build_precompile_headersN ]  # #G4 @Q@Q R,, ]!J7)TU[\ \ ]s&)AA  Ar>-Iz-I z-I {}rrrrrT) rr9r:rJrOrnrr sysconfigget_pathrrr,)r r rrb_is_gccr#r$head_file_signaturer r&r*r/r<rAextra_cflags_strrextra_include_paths_strrr%torch_include_dirs_strrqrcommon_cflags_strr@r; b_same_signr8s @rA0_check_and_build_extension_h_precompiler_headersrL`s8 !H$X.H5 [)WmLIGGLLiBSTM'',,{IwHZ[  (S ] $L1!hh?R':;G2gY;XZ'',,{I6K k]y)))45rww||K&%ST **<=M ;<< lG,,M&C&EFfFFM&B&DEfEEM$]3+HiPacy|LNefG#G,H ww~~m$D0(#$7B12ExP %  #G , '(;X F ? <&GEs H> H H$cd}tjjtddd}tjjtddd}||||y)Ncntjj|rtj|yyrs)rJrOrTremove) path_files rA_remove_if_file_existszFremove_extension_h_precompiler_headers.._remove_if_file_existss# 77>>) $ IIi  %r_rrUrr)rJrOrnr)rQr$rFs rAr8r8sM!GGLLiBSTM'',,{IwHZ[=)./r_c>|xs t|| }t|tr|g}|xsg}t|tr|g}|jdd|dur t ||n t |g}|j dt|tr|g}t|tr|Dcic]}||}}n't|tstdt||jD]>\}}| r|j d|d|d|d $|j d|d |d |d @|j d ||z }tjj|d }t|dj||g}|r|jdd|jdd|jddtjj|d}t|dj||j |t!|||||||| | | d|  Scc}w)aE Load a PyTorch C++ extension just-in-time (JIT) from string sources. This function behaves exactly like :func:`load`, but takes its sources as strings rather than filenames. These strings are stored to files in the build directory, after which the behavior of :func:`load_inline` is identical to :func:`load`. See `the tests `_ for good examples of using this function. Sources may omit two required parts of a typical non-inline C++ extension: the necessary header includes, as well as the (pybind11) binding code. More precisely, strings passed to ``cpp_sources`` are first concatenated into a single ``.cpp`` file. This file is then prepended with ``#include ``. Furthermore, if the ``functions`` argument is supplied, bindings will be automatically generated for each function specified. ``functions`` can either be a list of function names, or a dictionary mapping from function names to docstrings. If a list is given, the name of each function is used as its docstring. The sources in ``cuda_sources`` are concatenated into a separate ``.cu`` file and prepended with ``torch/types.h``, ``cuda.h`` and ``cuda_runtime.h`` includes. The ``.cpp`` and ``.cu`` files are compiled separately, but ultimately linked into a single library. Note that no bindings are generated for functions in ``cuda_sources`` per se. To bind to a CUDA kernel, you must create a C++ function that calls it, and either declare or define this C++ function in one of the ``cpp_sources`` (and include its name in ``functions``). See :func:`load` for a description of arguments omitted below. Args: cpp_sources: A string, or list of strings, containing C++ source code. cuda_sources: A string, or list of strings, containing CUDA source code. functions: A list of function names for which to generate function bindings. If a dictionary is given, it should map function names to docstrings (which are otherwise just the function names). with_cuda: Determines whether CUDA headers and libraries are added to the build. If set to ``None`` (default), this value is automatically determined based on whether ``cuda_sources`` is provided. Set it to ``True`` to force CUDA headers and libraries to be included. with_pytorch_error_handling: Determines whether pytorch error and warning macros are handled by pytorch instead of pybind. To do this, each function ``foo`` is called via an intermediary ``_safe_foo`` function. This redirection might cause issues in obscure cases of cpp. This flag should be set to ``False`` when this redirect causes issues. Example: >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CPP_EXT) >>> from torch.utils.cpp_extension import load_inline >>> source = """ at::Tensor sin_add(at::Tensor x, at::Tensor y) { return x.sin() + y.sin(); } """ >>> module = load_inline(name='inline_extension', ... cpp_sources=[source], ... functions=['sin_add']) .. note:: By default, the Ninja backend uses #CPUS + 2 workers to build the extension. This may use up too many resources on some systems. One can control the number of workers by setting the `MAX_JOBS` environment variable to a non-negative number. rz#include Tz*PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {z3Expected 'functions' to be a list or dict, but was zm.def("z", torch::wrap_pybind_function(z), "z");z", z, "}zmain.cpp z#include rz#include ryz#include zcuda.cuF)rr)r rMrinsertrLr8r-rrNrtypeitemsrJrOrnrr)r cpp_sources cuda_sources functionsr r r r rcrdrerwith_pytorch_error_handlingruse_pch module_defr function_name docstringcpp_source_pathr]cuda_source_paths rAr5r5s9l&L)=dG)LO+s#"m %2L,$$~ q89$8GZ[.0  FG i %" I i &'01!A1I1It,RSWXaSbRcde e(1(9 ` $M9*!!GM?:YZgYhhlmvlwwz"{|!!GM?#m_CPY{Z]"^_  ` #z! ggll?J?O$))K"89GA9:A23A:;77<<C%tyy'>?'(  - / /52s& Hrcrdc p| r | r td|ttt|}td|xsgD} tj |} tj ||||||g||| | }|dkDr:|| k7r.|r,td|dd|d|d |d ztj |d |}ttjj|d }|jrW || k7rt| 5}t r|s| rt#j$||||ng|Dcgc]!}tjj'|#c}t)dtjjt*dg||d| }t-}|D]E}tjj'|}|j/||vr||j0n|Gt3|}t5|||xsg|xsg|xsg|xsg||||  dddn!|rtd|dtj |j7n|j9|rtd|d tj | r t;||St=||| Scc}w#1swYkxYw#|j7wxYw)Nz>`is_python_module` and `is_standalone` are mutually exclusive.c3$K|]}d|v yw)cudnnNr)rurs rArwz_jit_compile..s?aW\?s)build_argumentsrcrerrrz*The input conditions for extension module z have changed. zBumping to version z and re-building as _v...rHlockrrT) rrrrignoresr show_progressr clean_ctx) rr]r r r r rcrdrerz9No modifications detected for re-loaded extension module z, skipping build step...zLoading extension module )rrrrIJIT_EXTENSION_VERSIONER get_versionbump_version_if_changedrXrYrZrrJrOrn try_acquirerrJrrrCrqrrrrr#_write_ninja_file_and_build_libraryreleasewait_get_exec_path_import_module_from_library)rr]r r r r rcrdrerrr with_cudnn old_versionrgbatonrkrrrrrs rArr}sMYZZM734 ?=+>B??J)55d;K%== %'8-I\]')#>G{ k !g >tfOT'y0DTF"WIUXYZzz #r'# bggll?F; N OOQ)N 55H MMOs70J#%J'&J B>J *J#JJ J##J5c |tt} t| | tt t |} t jj|d} |rtd| dtjt| |||||||dd|  |rtdtjt||dy) N build.ninjaEmitting ninja build file rgrH rOr.r_r`rarbr]r^ldflagslibrary_targetrezCompiling objects...z%Error compiling objects for extension error_prefix)r7r9r.rrrIrJrOrnrXrYrZ_write_ninja_file_run_ninja_build) r]r^r.r_r`rarbrcrdrerbuild_file_paths rArkrks!H.x8M734 ggll?MBO *?*;3?cjjQ )5  $3::6= >r_rc tt} t| |tt t |}t |xsg||| }tjj|d} |rtd| dtjt| |||xsg|xsg|xsg|xsg||  |rtd|dtjt||d|d y) NryrzrgrH) rOrr]r r r r rerzBuilding extension module zError building extension 'rGr~)r7r9r.rrrI_prepare_ldflagsrJrOrnrXrYrZ"_write_ninja_file_to_build_libraryr) rr]r r r r rcrdrerrrs rArprps!H.x8M734 $ M ggll?MBO *?*;3?cjjQ'  !'R+1r#)r/52# % *4&43::F1$q9;r_cj tjdjy#t$rYywxYw)zxReturn ``True`` if the `ninja `_ build system is available on the system, ``False`` otherwise.zninja --versionTF)rrrrrr_rAr6r6/s8 1 7 7 9: s #& 22c.ts tdy)zRaise ``RuntimeError`` if `ninja `_ build system is not available on the system, does nothing otherwise.z(Ninja is required to load C++ extensionsN)r6rrr_rAr7r79s  EFF r_c:trtjjtj d}|j d|r|j d|j d|r"|j d|j d|j d|j dt|s|j d |j d|n|j d t|j d |r|j trd nd |j d|r|j trdnd|j d|s|j d|r|j dt|r|rtdtjtrn|j dtdd|j dt7|j dtjjtdd|Stsd}tjjt|s*tjjtdrd}|j d t||j dt6|j d tjjtd|Str.|j d td|j d|S)Nlibszc10.libz c10_cuda.libz torch_cpu.libztorch_cuda.libz"-INCLUDE:?warp_size@cuda@at@@YAHXZz torch.libz /LIBPATH:ztorch_python.librz-lc10z -lc10_hipz -lc10_cudaz -ltorch_cpuz -ltorch_hipz -ltorch_cudaz-ltorchz-ltorch_pythonz -Wl,-rpath,z%Detected CUDA files, patching ldflagsrHrrz cudart.librz-lcudartz -lamdhip64)rQrJrOrnrYbase_exec_prefixr-rrJrXrZrKrzrTrq)r rerdrpython_lib_path extra_lib_dirs rArr?s'',,s';';VDY'    0_-   !1 2  !E F[)y(89:  !3 4  9_,=!> ? r.!123W%   0@l S]+   2B WY'  !1 2   ;~.>!? @  9 K   9_UE-J,K!L M   .%$$yj%QV1W0X%YZ "#MGGNN?=#ABGGNN?5#9:!&  2om&D%E!F G   ,%$$r"'',,z7*K)L%MN   2oe&<%=!> ?   . r_r.c ||D]}d|vrd|vs gcStjgd}gd}||Dcgc]}|dz c}z}tjj dd}|st j dg}ttjjD]}tjj|} tjjD cgc]#} d | vrt| jd d %} } td | D} t!| | } | d d| d } | |vs|j#| t%|}|dxxdz cc<nM|j'dd}|j)D]\} }|j'| |}|jd}g}|D]r} | |vrt+d| d| d | ddjdd z}|j#d|d|| j-ds\|j#d|d|tt%t/|Scc}wcc} w)ad Determine CUDA arch flags to use. For an arch, say "6.1", the added compile flag will be ``-gencode=arch=compute_61,code=sm_61``. For an added "+PTX", an additional ``-gencode=arch=compute_xx,code=compute_xx`` is added. See select_compute_arch.cmake for corresponding named and supported arches when building with CMake. NTORCH_EXTENSION_NAMEarch) )z Kepler+Tesla3.7)Keplerz3.5+PTX)z Maxwell+Tegra5.3)Maxwellz 5.0;5.2+PTX)Pascalz 6.0;6.1+PTX)z Volta+Tegra7.2)Voltaz7.0+PTX)Turingz7.5+PTX)z Ampere+Tegra8.7)Amperez 8.0;8.6+PTX)Adaz8.9+PTX)Hopperz9.0+PTX)z3.5rz5.0z5.2rz6.0z6.1z6.2z7.0rz7.5z8.0z8.6rz8.9z9.0z9.0az+PTXTORCH_CUDA_ARCH_LISTzTORCH_CUDA_ARCH_LIST is not set, all archs for visible cards are included for compilation. If this is not desired, please set os.environ['TORCH_CUDA_ARCH_LIST'].sm_rrc30K|]}|dz|dzfyw)rNr)rusms rArwz'_get_cuda_arch_flags..s"N2B"Hb2g#6"Nsrrxrr>;zUnknown CUDA arch (z) or GPU not supportedry+z-gencode=arch=compute_z ,code=sm_z,code=compute_) collections OrderedDictrJrKrLrrrArUrV device_countget_device_capability get_arch_listrtrmaxminr-sortedr(rWrendswithr)r.r) named_archessupported_archesrvalid_arch_strings _arch_list arch_listrD capabilityr supported_smmax_supported_sm named_archarchvalflagsnums rAr;r;zs D%-~  ** , LX)AQ,RAQZ,RR  6=J   U V uzz..01 'A99!rz--offload-arch=)rJrKrLrUr_cuda_getArchFlagsrr()r.r)_archs archFlagsarchsrrs rArPrPs (D$&.D*@&'' ( ZZ^^/ 6F HH//1 OO%EEsC(..s327 8$tf % 8E 8 n E L 9s B(rctjjd}|t}tj j dn,dtj j jdd}dtjjtjj}|d|}tjj||}|rtd|d tj tjj||}tjj!|s8|rtd |d tj tj"|d |S)NTORCH_EXTENSIONS_DIRcpucurxrpyrzUsing z as PyTorch extensions root...rHzCreating extension directory rgT)r2)rJrKrLr,rUrgrVr(rY version_informinorrOrnrXrZrTmakedirs)rrdroot_extensions_directorycu_strpython_version build_folderrcs rAr r s( " /E F ($:$<! ==--5%u}}))11#r:;< c..445c6F6F6L6L5MN()6(3 $&GGLL %|%5! 011OPWZWaWabggll#>/ *  1/1B#FSZZ X Od3 r_ctjjd}|<|jr,|rt d|dt j t|S|rt dt j y)NMAX_JOBSzUsing envvar MAX_JOBS (z) as the number of workers...rHzqAllowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N))rJrKrLisdigitrXrYrZrt)rdmax_jobss rA_get_num_workersrskzz~~j)H 0 0 2  +H:5RSzz #8} M::  r_rcddg}t|}||jdt|gtjj }t rd|vrddlm}|jj}t|}|jj|} | jD cic]\} } | j| } } } |jD]\} } | j} | | vs| | | <!| } t j"j%t j&j%d} t)j*||r| nt(j,t(j.|d| ycc} } w#t(j0$rf}t!j2\}}}|}t5|d r/|j6r#|d |j6j8t:z }t=||d}~wwxYw) Nninjarz-jrr) distutilsrT)stdoutrZcwdcheckroutputr)rr=rrJrKrrQrrutil get_platformPLAT_TO_VCVARS _msvccompiler _get_vc_envrWupperrYrflushrZrrunPIPErr?rrrrrr)rcrdrcommand num_workersrr plat_name plat_specvc_envkrvuk stdout_filenorrrmessages rArrsoG"7+Kc+./0 **// C*#5(NN//1 "9- ((44Y?+1<<>:41a!'')Q,::IIK DAqBr   +    $+=$$  -;:  ( ( +lln 5! 5( # /ELL//1GHIJ JG7#* +s -E:A9FG9A!G44G9ctrttjddj dvrgt dtjddj dD}|s0tdtjddtj d<tjj||tS)NPATHrrc3K|]I}tjj|xr$tjj|tKywrs)rJrOrTsamefiler)rups rArwz!_get_exec_path..Ms>  GGNN1  E"''"2"21n"E E sAA) rQrrJr<rrrKrOrnEXEC_EXT) module_namerOtorch_lib_in_paths rArsrsKsnBIIfb,A,G,G,LL YYvr*005  !$2#31RYYvr5J4K!LBJJv  77<< hZ8 99r_ctjj||t}|rtj j ||}|Jtj j|}t|jtjjsJ|jj||Stjj||Srs)rJrOrnLIB_EXT importlibrspec_from_file_locationmodule_from_specrMloaderabcLoader exec_modulerUops load_library)rrOrfilepathspecmodules rArtrtVsww||D[M'";K|]}|jdyw)z-std=Nr(r8s rArwz5_write_ninja_file_to_build_library..sGDtw/Gr9r5r7rr<ctjjtjj|d}t |r r|d}|S|d}|S)Nrz.cuda.oz.o)rJrOrrfrI)r file_nametargetres rAobject_file_pathz<_write_ninja_file_to_build_library..object_file_paths]GG$$RWW%5%5k%BCAF  %)"{'*F "{"%F r_z-undefined dynamic_lookupr{)rrJrOrCr2rCrDrQr-rrirjrrrBrJrQrOrPr:r;rrr<r SHARED_FLAGrrrr)rOrr]r r r r rerr)rI user_includessystem_includespython_include_pathrqrrr. cuda_flagsrcc_envrrSr^r|rTr}s ` rArres.::TDJJL:L:2CD$DD.;>?&C&EFfFFM 8WXWBwi.XX ]S'Bu{{7345SS O\Iekk'&:%;<\\ &B&DEfEEM!22l^ClR''  ?"= \(*<~*MNQ[[  \#|n4J' 3J .):; ;J /; ;J + +JGJGG!!,/YYt_F!&/*<   c c 188$8G8"r FG23  )#(CvcU^N #% q;D<HL GYS\EX9s@M MMM:&M M#2 M( M-2M2# M7M<c  d} | |}| |}| |}| |}| |}| |}t|t|k(sJt|dkDsJt} dg} | jd| | s|r[dtjvrtj d}nt r tdd}n tdd}| jd |t r t|z}d d j|g}|jd d j|| rF|jd d j||jdd j||jdd j||jdd j||Dcgc]!}tjj|#}}dg}tr#|jd|jdn3|jd|jd|jd| rqdg}d}tjj =tj dddk7r$|jd|jdd}|jd|dg}t#||D]\}}t%|xr| }|rd nd!}tr$|j'd"d#}|j'd"d#}|j'd d$}|j'd d$}|jd%|d&|d ||rstjjtjj)|dd'}d(g}|jd)d%|d*d j|g}||gz }ngg}}| d+g}trt+j,d,d-gj.t0j3d.}t|d/k\r3tjj)|dj'd"d#}n t5d0|jd1|d2n|jd3d%| d4d j|g}d5| g} nggg} }}| ||g}!| r|!j|!|||||| gz }!d6jd7|!D}"|"d8z }"t7||"ycc}w)9aWrite a ninja file that does the desired compiling and linking. `path`: Where to write this file `cflags`: list of flags to pass to $cxx. Can be None. `post_cflags`: list of flags to append to the $cxx invocation. Can be None. `cuda_cflags`: list of flags to pass to $nvcc. Can be None. `cuda_postflags`: list of flags to append to the $nvcc invocation. Can be None. `sources`: list of paths to source files `objects`: list of desired paths to objects, one per source. `ldflags`: list of flags to pass to linker. Can be None. `library_target`: Name of the output library. Can be None; in that case, we do no linking. `with_cuda`: If we should be compiling with CUDA. cP|gS|Dcgc]}|jc}Scc}wrs)r)rr)s rAsanitize_flagsz)_write_ninja_file..sanitize_flagss& =I-23TDJJL3 33s#rzninja_required_version = 1.3zcxx = PYTORCH_NVCCrrcrFznvcc = z cflags = r>zpost_cflags = zcuda_cflags = zcuda_post_cflags = zcuda_dlink_post_cflags = z ldflags = z rule compilez@ command = cl /showIncludes $cflags -c $in /Fo$out $post_cflagsz deps = msvczD command = $cxx -MMD -MF $out.d $cflags -c $in -o $out $post_cflagsz depfile = $out.dz deps = gcczrule cuda_compilerN*TORCH_EXTENSION_SKIP_NVCC_GEN_DEPENDENCIESrrz?--generate-dependencies-with-compile --dependency-output $out.dz command = $nvcc z. $cuda_cflags -c $in -o $out $cuda_post_cflags cuda_compiler:z$:z$ zbuild rzdlink.ozrule cuda_devlinkz5 command = $nvcc $in -o $out $cuda_dlink_post_cflagsz: cuda_devlink z rule linkwhererz rz'MSVC is required to load C++ extensionsz command = "z)/link.exe" $in /nologo $ldflags /out:$outz% command = $cxx $in $ldflags -o $outz: link zdefault z c3>K|]}dj|yw)rTN)rn)rubs rArwz$_write_ninja_file..c s71$))A,7r9rT)rSr9r-rJrKr<rJrqrKrQrnrOrCrQrUrgrVziprIr(rPrrrrrrr)#rOr.r_r`rarbr]r^r|r}rerrconfigrFrrI compile_rulecuda_compile_rule nvcc_gendepsbuildr object_fileis_cuda_sourcerule devlink_out devlink_ruledevlink link_rulecl_pathscl_pathlinkdefaultblocksrs# rArrs24 F #F -K -K%&67+,BCW%G w<3w< '' ' w#((;"7!89: ~chh{&;%<=> *3884D+E*FGH LL,SXX6L-M,NOP LL:chhw/012299rwwt$9G9##L N PO, R T01N+01  ==   )bii8dfi.jnq.q  $ $%9 :  $ $^ 4]L   .\ ] _ E$'$9C [&{3A !/~Y %--c48K%--c48K!))#t4 !))#t4  vk]"TF!K=ABCggll277??71:#> J +, STK=8I7JKLK=  "Bg ! M >z..040677=v?UWW\W\]cWd 8}!''//(1+6>>sDI"#LMM   }WI5^_ `   D E(0A/BCDn-./#%r24 e\ *F '( |YwgFFFkk777G tOGw[:s6&S#clt tdtjjtg|S)z Join paths with CUDA_HOME, or raises an error if it CUDA_HOME is not set. This is basically a lazy way of raising an error for missing $CUDA_HOME only once we need to get any CUDA-specific path. zSCUDA_HOME environment variable is not set. Please set it to your CUDA install root.)rDrmrJrOrnros rArKrKh s7AB B 77<< *E **r_rOcddg}tr|jdtjj |d|vS)Nrr"r#r)rJr-rJrOr)rO valid_exts rArIrIu s;I 77  D !! $ 11r_)F) NNNNNFNTFT) NNNNNNNFNTTTF)Trsr)rrRr importlib.abcrJrrirMrrrYrCrrpathlibrr4rUtorch._appdirs file_batonr_cpp_extension_versionerrrrhipify.hipify_pythonrtypingr r r r r torch.torch_versionrrsetuptools.command.build_extrrrQrrrrr CLIB_PREFIXCLIB_EXTrrOrC__file___HERErPrrnrrrrrt VersionRanger VersionMapr'__annotations__MINIMUM_CLANG_VERSIONr+__all__rBr^rkrqrrrrrrarrgrdrJ ROCM_VERSIONrrrV _is_compiledrDrKrLrzrrrrr:rQrOrlrr9boolrrrr,r-r.rr/r0r1r2r3r4rrr:rLr8r5rrkrpr6r7rr;rPr rrrsrtrrrKrIrr_rAr@s    !8!65552 \\W $ << " "8 , << " "7 +&E6RbE 6U"f  !ggoobggooe45 k51&0R%U38_eCHo56 #|# $ !' * ' * ' * ' *  !: " "G , "G , "G , "G , "G , "G , "G ,#Z  ?$s), ?c ?#6#0 +s + "v   %.?5 !D&29J9J9V4^c == J):):)@)@)Ebq)IJJL!&!8!8!:O  ZZ^^L ) IRZZ^^L-I %/BJJ/G$H!S -. R$Ri$s)i' W W)S)T)X14E$ BTc(>(>(>D>(>'+(>j$,;c,;,;,;D>,;,;)-,;^G 8vZ$s)!4ZS Zz$s)!4S ,sTc2 d x} 4+c4+D4+4+PT4+n: l:>l^R %)R h +s +222r_