`iD dZddlmZddlZddlZddlZddlZddlZddlZddl m Z m Z ddl m Z ddlmZddlmZddlmZmZmZdd lmZmZdd lmZejd kZ dPdQdZdRd!ZejdSd"Z dSd#Z!dTd+Z" dPdUd.Z#dVd0Z$e%&d1d2d3d4d5d6Z'dWd8Z(dXd;Z)dYd>Z* dZd[dBZ+ddddddddddC d\dGZ,dddddddddd?dH d]dKZ-ddddddddLd^dMZ.dddddddd?dNd_dOZ/dS)`zBBuild and load C++/CUDA sources into a tvm_ffi Module using Ninja.) annotationsN)MappingSequence) nullcontext)Path)Any)find_dlpack_include_pathfind_include_pathfind_libtvm_ffi)Module load_module)FileLockwin32 cpp_source str | None cuda_source cpp_filesSequence[str] | None cuda_files functions!Sequence[str] | Mapping[str, str] extra_cflags Sequence[str]extra_cuda_cflags extra_ldflagsextra_include_paths embed_cubinMapping[str, bytes] | Nonereturnstrc  tj d fd  |||t|nd|t|nd|||||| f  ddS) z;Generate a unique hash for the given sources and functions.objrrNonec|ddSt|tr?d|ddSt|tr,d|dSt|t rYdt |D] }||}||!dSt|tr'd|D] }|dStdt|)NsNonesstrutf-8sbytessMappingsSequencezUnsupported type: ) update isinstancer encodebytesrsortedkeysr ValueErrortype)r"keyitem_hashms i/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/tvm_ffi/cpp/extension.pyr0z_hash_sources.._hash6s~ ; HHW      S ! ! ? HHV    HHSZZ(( ) ) ) ) ) U # # ? HHX    HHSMMMMM W % % ? HHZ chhjj))  3xc d   X & & ? HH[ ! ! !  d   =$s))==>> >N)r"rrr#)hashlibsha256r* hexdigest) rrrrrrrrrrr0r1s @@r2 _hash_sourcesr8's A???????, E  !*!6F9   D","8F:   d           ;;==" r3pathcontentr#cbt|}|rH|5}|}dddn #1swxYwY||krdS|d5}||ddddS#1swxYwYdS)zIWrite content to path if it does not already exist with the same content.Nw)rexistsopenreadwrite)r9r:pfexisting_contents r2 _maybe_writerD^s T Axxzz VVXX ( vvxx  ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( w & & F  s#AA AB$$B(+B(cDtjdptjd}|tjd}|'t t |jj}ntrit d}t| d}t|dkrtdt |d}nd }t | std|S) zFind the CUDA install path. CUDA_HOME CUDA_PATHNnvccz2C:/Program Files/NVIDIA GPU Computing Toolkit/CUDAzv*.*rzLCould not find CUDA installation. Please set CUDA_HOME environment variable.z/usr/local/cuda)osenvirongetshutilwhichr rparent IS_WINDOWSlistgloblen RuntimeErrorr=) cuda_home nvcc_path cuda_root cuda_homess r2_find_cuda_homerXjs {++Jrz~~k/J/JIL((  DOO29::II . !UVV !).."8"899 z??a''&f 1 .. -  ??))++ "b r3c dtjvrtjd}g}|D]q}t|ddkrt d||d\}}|d||d||rd|S tjgdd d }|j d  d d }|d\}}d||d||S#t$rYdSwxYw)z&Get the CUDA target architecture flag.TVM_FFI_CUDA_ARCH_LIST.zInvalid CUDA architecture: z-gencode=arch=compute_z ,code=sm_ )z nvidia-smiz--query-gpu=compute_capz--format=csv,noheaderT)argscapture_outputcheckr% rz#-gencode=arch=compute_70,code=sm_70) rIrJsplitrRr,appendjoin subprocessrunstdoutdecodestrip Exception) arch_listflagsarchmajorminorstatus compute_caps r2_get_cuda_targetrrs2:--J78>>@@  Y YD4::c??##q(( !Et!E!EFFF::c??LE5 LLW%WWWWPUWW X X X Xxx 9^WWW#F !-..w77==??EEdKKANK&,,S11LE5QEQ5QQ5Q%QQ Q 9 9 9888 9s?BE EEr^ list[str]cwdstr | os.PathLike[str]r_boolsubprocess.CompletedProcessc V tttjdddz dz dz }t|st dtj|ddd d d d gd d d j }|st dtt|dz dz dz }t|st d|d |d |}tj|d||d S#t tj f$r;}td d ||d}~wwxYw)zOLocates the Developer Command Prompt and runs a command within its environment.zProgramFiles(x86)zC:\Program Files (x86)zMicrosoft Visual Studio Installerz vswhere.exezvswhere.exe not found.z-latestz -prereleasez -products*z -propertyinstallationPathT)r_textr`z$No Visual Studio installation found.Common7Toolsz VsDevCmd.batzVsDevCmd.bat not found at: z'"{vsdevcmd_path}" -arch=x64 & {command}r]) vsdevcmd_pathcommandF)r`rtr_shellzEFailed to run the following command in MSVC developer environment: {}N)r rrIrJrKr=FileNotFoundErrorrerfrgriformatrdCalledProcessErrorrS)r^rtr_ vswhere_pathvs_install_pathr cmd_commandes r2_run_command_in_dev_promptrs 6  35NOO P P' (    L!!((** >#$<== =%."         L#$JKK KD11I=G.XYY M""))++ S#$Q-$Q$QRR R @FF'$G  ~ u#nTX     z< = S Z Z      sEEF(-6F##F(name with_cudac Vttg} tt} t | j} | j} trgd} gd}dd| | dg}nggd} gd}dd | d g}|rU|tgz }|d t ttd z d gz }| d|Dz}|d|Dz}|d|Dz}| d|Dz}|D]z}| d | dd| d | dd{g}| d| d tjdtrdnd| d d||r| d t ttdz dz | d d|| d d|| d!| d"tr+| d#| d$n?| d%| d&| d'| d!|ri| d(| d%| d&| d)| d!ts|r| d*| d+| d!| d,| d-t"jd.| d!| d/tr| d0n| d1| d!g}t't)|D]]\}}d2|d3}| d4 || dd| |^t't)|D]]\}}d5|d3}| d6 || dd| |^trd7nd8}ts|r d9}d|}| d:|d;|| d!|}t)|D]o}d<|d3} |d=}!| d:| d>|| d?|!| d@|| d!| }p| d:||dA|| d!nGd|}"| d:||dA|"| d!| dB||| d!dC|S)Dzz)_generate_ninja_build..s EEEtzz||EEEr3c6g|]}|Srrrs r2rz)_generate_ninja_build..s (T(T(T$(T(T(Tr3c6g|]}|Srrrs r2rz)_generate_ninja_build..s H H H$ H H Hr3cjg|]0}tt|1Srr rresolve)rr9s r2rz)_generate_ninja_build..s?---&*DJJ   !!---r3z-I{}:z$:zninja_required_version = 1.3zcxx = {}CXXclzc++z cflags = {}r]z nvcc = {}binrHzcuda_cflags = {}z ldflags = {}z rule compilez5 command = $cxx /showIncludes $cflags -c $in /Fo$outz deps = msvcz depfile = $out.dz deps = gccz7 command = $cxx -MMD -MF $out.d $cflags -c $in -o $outzrule compile_cudazm command = $nvcc --generate-dependencies-with-compile --dependency-output $out.d $cuda_cflags -c $in -o $outzrule merge_objectsz command = ld -r -o $out $inzrule embed_cubinz command = z[ -m tvm_ffi.utils.embed_cubin --output-obj $out --input-obj $in --cubin $cubin --name $namez rule linkz- command = $cxx $in /link $ldflags /out:$outz% command = $cxx $in $ldflags -o $outcpp_z.ozbuild {}: compile {}cuda_zbuild {}: compile_cuda {}.dll.soz unified.ozbuild z: merge_objects unified_with_.cubinz: embed_cubin z cubin = z name = z: link zdefault ra)r r rr r rNstemrOrrrrXrcreplacerIrJrKrdsys executable enumerater*r+)#rrrrrrrrrdefault_include_paths tvm_ffi_libtvm_ffi_lib_pathtvm_ffi_lib_namedefault_cflagsdefault_cuda_cflagsdefault_ldflagscflags cuda_cflagsldflags include_pathsr9ninja obj_filesicpp_pathobj_name cuda_pathext unified_obj obj_files_str current_obj cubin_namenext_obj cubin_filelink_files_strs# r2_generate_ninja_buildrs!/002J2L2LM(())K;-.."'!   BAA  *( * * % % %  877JJJ$&=+;&=&={K   $4$6$6#7 7   c$'8'8"9"9G"CDDEE  O EE EEE EF%(T(TBS(T(T(TTK H H- H H HHG)--.A---M CC fmmDLLd$;$;<<===6==c4)@)@AABBBBE LL/000 LL""2:>>%9VQV#W#WXXYYY LL%%chhv&6&677888G [''D1B1B,C,Ce,Kf,T(U(UVVWWW '..sxx /D/DEEFFF LL&&sxx'8'899::: LL LL   P LMMM _%%%% )*** ^$$$ NOOO LL ())) )*** ^$$$ {     R     LL- . . . LL8 9 9 9 LL    LL+ , , , LLKs~KKK    LL    LL> DEEEE <=== LLI  !2!233## 8!<<< +228X=M=McSW=X=XYYZZZ""""!&"4"455## 9 1=== 077)BSBSTWY]B^B^__```"""" )&&EC +! ++  JkJJ=JJKKK R"  !1!1!3!344 # #J5z555H&...J LLG(GG+GG H H H LL2j22 3 3 3 LL1Z11 2 2 2 LL   "KK  =d=C== ==>>> R),, @d@C@@@@AAA R LL'D'#''((( LL 99U  r3 build_dirc&ddg}tjdd}||d|gz }trt ||d}nt jd||d }|jd krd |jg}trd nd }|jr0| d|j ||j r0| d|j |td |dS)z:Build the module in the given build directory using ninja.rz-vMAX_JOBSNz-jT)r^rtr_F)r`r^rtr_rzninja exited with status oemr%zstdout: zstderr: ra)rIrJrKrOrrerf returncodergrcrhstderrrSrd)rr num_workersrpmsgencodings r2 build_ninjars)oG*..T22KD+&&_+iX\]]]e'yY]^^^ A>6+<>>?&355G = E JJC6=#7#7#A#ACC D D D = E JJC6=#7#7#A#ACC D D D499S>>***r3z\\z\"z\nz\rz\t)\"ra  sc6|tS)z2Escape special characters for C++ string literals.) translate_CPP_ESCAPE_TABLE)rs r2_escape_cpp_string_literalrs ;;( ) ))r3sourceMapping[str, str]c.dddddd|g}|D]N\}}|d|d|d |r+t|}|d |d |d O|dd |S)z:Decorate the given source code with TVM FFI export macros.z%#include z#include z#include z$#include z#include rzTVM_FFI_DLL_EXPORT_TYPED_FUNC(z, z);z"TVM_FFI_DLL_EXPORT_TYPED_FUNC_DOC(z, "z");ra)itemsrcrrd)rrsources func_namefunc_doc escaped_docs r2_decorate_with_tvm_ffirs 0$$.' G )00`` 8R RRYRRRSSS  `4X>>K NN^ ^^k^^^ _ _ _ NN2 99W  r3seqSequence[str] | str | NonecX|gSt|tr|gSt|SN)r'r rP)rs r2 _str_seq2listrs1 { C  u Cyyr3Tbuild_directory need_lockc (dt|D} dt|D} t| } t| } | s | s Jd|t|ng}|t|ng}|t|ng}|t|ng}|tjdt td}tdd| | i||||| }t||d|z }n!t| }| dd | rtrtd | r7| D]"\}}||d z }||#t!|| ||||| | | }|rt#t |d z n t%5t't |dz |t)t |trdnd}t |||z cdddS#1swxYwYdS)z&Real implementation of build function.cjg|]0}tt|1SrrrrAs r2rz_build_impl..s0NNNSa**++NNNr3cjg|]0}tt|1Srrrs r2rz_build_impl..s0PPPc$q''//++,,PPPr3z0Either cpp_files or cuda_files must be provided.NTVM_FFI_CACHE_DIR~/.cache/tvm-ffi_Tparentsexist_okz/CUBIN embedding is not yet supported on Windowsr) rrrrrrrrrlockz build.ninjarr)rrvrPrIrJrKr r expanduserr8rmkdirrONotImplementedErrorr write_bytesrrrrDr)rrrrrrrrrr cpp_path_listcuda_path_listwith_cpprextra_ldflags_listextra_cflags_listextra_cuda_cflags_listextra_include_paths_list cache_dir source_hashrr cubin_bytes cubin_path ninja_sourcers r2 _build_implrs1ON]95M5MNNNMPPmJ6O6OPPPNM""H^$$I TyTT"TTT 0=0Im,,,r.:.F\***B8I8UT"3444[](y.data_ptr())[i] = static_cast(x.data_ptr())[i] + 1; } } ''' # compile the cpp source code and load the module lib_path: str = tvm_ffi.cpp.build_inline( name="hello", cpp_sources=cpp_source, functions="add_one_cpu", ) # load the module mod: Module = tvm_ffi.load_module(lib_path) # use the function from the loaded module to perform x = torch.tensor([1, 2, 3, 4, 5], dtype=torch.float32) y = torch.empty_like(x) mod.add_one_cpu(x, y) torch.testing.assert_close(x + 1, y) raNrci|]}|dS)rr)rrs r2 z build_inline..s777Tb777r3rrrTrzmain.cppzcuda.curF) rrrrrrrrrr)rrdrvrPr'r rdictrrIrJrKrrr8rrrrDr)rrr rrrrrrrcpp_source_listrrcuda_source_listrrrrrr function_maprrrcpp_file cuda_files r2 build_inliner"saJ$K00O?++JO$$H$\22)),--K%&&I0=0Im,,,r.:.F\***B8I8UT"3444[](y.data_ptr())[i] = static_cast(x.data_ptr())[i] + 1; } } ''' # compile the cpp source code and load the module mod: Module = tvm_ffi.cpp.load_inline( name="hello", cpp_sources=cpp_source, functions="add_one_cpu", ) # use the function from the loaded module to perform x = torch.tensor([1, 2, 3, 4, 5], dtype=torch.float32) y = torch.empty_like(x) mod.add_one_cpu(x, y) torch.testing.assert_close(x + 1, y) ) rrr rrrrrrrr)r r) rrr rrrrrrrrs r2 load_inlinersQR #%%/' 3+# ,   r3)rrrrrrrc 2t||||||||d S)aCompile and build a C++/CUDA module from source files. This function compiles the given C++ and/or CUDA source files into a shared library. Both ``cpp_files`` and ``cuda_files`` are compiled to object files, and then linked together into a shared library. It's possible to only provide cpp_files or cuda_files. The path to the compiled shared library is returned. Note that this function does not automatically export functions to the tvm ffi module. You need to manually use the TVM FFI export macros (e.g., ``TVM_FFI_DLL_EXPORT_TYPED_FUNC``) in your source files to export functions. This gives you more control over which functions are exported and how they are exported. Extra compiler and linker flags can be provided via the ``extra_cflags``, ``extra_cuda_cflags``, and ``extra_ldflags`` parameters. The default flags are generally sufficient for most use cases, but you may need to provide additional flags for your specific use case. The include dir of tvm ffi and dlpack are used by default for the compiler to find the headers. Thus, you can include any header from tvm ffi in your source files. You can also provide additional include paths via the ``extra_include_paths`` parameter and include custom headers in your source code. The compiled shared library is cached in a cache directory to avoid recompilation. The `build_directory` parameter is provided to specify the build directory. If not specified, a default tvm ffi cache directory will be used. The default cache directory can be specified via the `TVM_FFI_CACHE_DIR` environment variable. If not specified, the default cache directory is ``~/.cache/tvm-ffi``. Parameters ---------- name The name of the tvm ffi module. cpp_files The C++ source files to compile. It can be a list of file paths or a single file path. Both absolute and relative paths are supported. cuda_files The CUDA source files to compile. It can be a list of file paths or a single file path. Both absolute and relative paths are supported. extra_cflags The extra compiler flags for C++ compilation. The default flags are: - On Linux/macOS: ['-std=c++17', '-fPIC', '-O2'] - On Windows: ['/std:c++17', '/MD', '/O2'] extra_cuda_cflags The extra compiler flags for CUDA compilation. The default flags are: - ['-Xcompiler', '-fPIC', '-std=c++17', '-O2'] (Linux/macOS) - ['-Xcompiler', '/std:c++17', '/O2'] (Windows) extra_ldflags The extra linker flags. The default flags are: - On Linux/macOS: ['-shared', '-L', '-ltvm_ffi'] - On Windows: ['/DLL', '/LIBPATH:', '.lib'] extra_include_paths The extra include paths for header files. Both absolute and relative paths are supported. build_directory The build directory. If not specified, a default tvm ffi cache directory will be used. By default, the cache directory is ``~/.cache/tvm-ffi``. You can also set the ``TVM_FFI_CACHE_DIR`` environment variable to specify the cache directory. Returns ------- lib_path: str The path to the built shared library. Example ------- .. code-block:: python import torch from tvm_ffi import Module import tvm_ffi.cpp # Assume we have a C++ source file "my_ops.cpp" with the following content: # ```cpp # #include # #include # #include # #include # #include # # void add_one_cpu(tvm::ffi::TensorView x, tvm::ffi::TensorView y) { # TVM_FFI_ICHECK(x.ndim() == 1) << "x must be a 1D tensor"; # DLDataType f32_dtype{kDLFloat, 32, 1}; # TVM_FFI_ICHECK(x.dtype() == f32_dtype) << "x must be a float tensor"; # TVM_FFI_ICHECK(y.ndim() == 1) << "y must be a 1D tensor"; # TVM_FFI_ICHECK(y.dtype() == f32_dtype) << "y must be a float tensor"; # TVM_FFI_ICHECK(x.size(0) == y.size(0)) << "x and y must have the same shape"; # for (int i = 0; i < x.size(0); ++i) { # static_cast(y.data_ptr())[i] = static_cast(x.data_ptr())[i] + 1; # } # } # # TVM_FFI_DLL_EXPORT_TYPED_FUNC(add_one_cpu, add_one_cpu); # ``` # compile the cpp source file and get the library path lib_path: str = tvm_ffi.cpp.build( name="my_ops", cpp_files="my_ops.cpp", ) # load the module mod: Module = tvm_ffi.load_module(lib_path) # use the function from the loaded module x = torch.tensor([1, 2, 3, 4, 5], dtype=torch.float32) y = torch.empty_like(x) mod.add_one_cpu(x, y) torch.testing.assert_close(x + 1, y) T) rrrrrrrrr)rrrrrrrrrs r2buildrs7|  !+#/'    r3)rrrrrrrrc Ntt|||||||||S)a*Compile, build and load a C++/CUDA module from source files. This function compiles the given C++ and/or CUDA source files into a shared library and loads it as a tvm ffi module. Both ``cpp_files`` and ``cuda_files`` are compiled to object files, and then linked together into a shared library. It's possible to only provide cpp_files or cuda_files. Note that this function does not automatically export functions to the tvm ffi module. You need to manually use the TVM FFI export macros (e.g., :c:macro:`TVM_FFI_DLL_EXPORT_TYPED_FUNC`) in your source files to export functions. This gives you more control over which functions are exported and how they are exported. Extra compiler and linker flags can be provided via the ``extra_cflags``, ``extra_cuda_cflags``, and ``extra_ldflags`` parameters. The default flags are generally sufficient for most use cases, but you may need to provide additional flags for your specific use case. The include dir of tvm ffi and dlpack are used by default for the compiler to find the headers. Thus, you can include any header from tvm ffi in your source files. You can also provide additional include paths via the ``extra_include_paths`` parameter and include custom headers in your source code. The compiled shared library is cached in a cache directory to avoid recompilation. The `build_directory` parameter is provided to specify the build directory. If not specified, a default tvm ffi cache directory will be used. The default cache directory can be specified via the `TVM_FFI_CACHE_DIR` environment variable. If not specified, the default cache directory is ``~/.cache/tvm-ffi``. Parameters ---------- name The name of the tvm ffi module. cpp_files The C++ source files to compile. It can be a list of file paths or a single file path. Both absolute and relative paths are supported. cuda_files The CUDA source files to compile. It can be a list of file paths or a single file path. Both absolute and relative paths are supported. extra_cflags The extra compiler flags for C++ compilation. The default flags are: - On Linux/macOS: ['-std=c++17', '-fPIC', '-O2'] - On Windows: ['/std:c++17', '/MD', '/O2'] extra_cuda_cflags The extra compiler flags for CUDA compilation. The default flags are: - ['-Xcompiler', '-fPIC', '-std=c++17', '-O2'] (Linux/macOS) - ['-Xcompiler', '/std:c++17', '/O2'] (Windows) extra_ldflags The extra linker flags. The default flags are: - On Linux/macOS: ['-shared', '-L', '-ltvm_ffi'] - On Windows: ['/DLL', '/LIBPATH:', '.lib'] extra_include_paths The extra include paths for header files. Both absolute and relative paths are supported. build_directory The build directory. If not specified, a default tvm ffi cache directory will be used. By default, the cache directory is ``~/.cache/tvm-ffi``. You can also set the ``TVM_FFI_CACHE_DIR`` environment variable to specify the cache directory. keep_module_alive Whether to keep the module alive. If True, the module will be kept alive for the duration of the program until libtvm_ffi.so is unloaded. Returns ------- mod: Module The loaded tvm ffi module. See Also -------- :py:func:`tvm_ffi.load_module` Example ------- .. code-block:: python import torch from tvm_ffi import Module import tvm_ffi.cpp # Assume we have a C++ source file "my_ops.cpp" with the following content: # ```cpp # #include # #include # #include # #include # #include # # void add_one_cpu(tvm::ffi::TensorView x, tvm::ffi::TensorView y) { # TVM_FFI_ICHECK(x.ndim() == 1) << "x must be a 1D tensor"; # DLDataType f32_dtype{kDLFloat, 32, 1}; # TVM_FFI_ICHECK(x.dtype() == f32_dtype) << "x must be a float tensor"; # TVM_FFI_ICHECK(y.ndim() == 1) << "y must be a 1D tensor"; # TVM_FFI_ICHECK(y.dtype() == f32_dtype) << "y must be a float tensor"; # TVM_FFI_ICHECK(x.size(0) == y.size(0)) << "x and y must have the same shape"; # for (int i = 0; i < x.size(0); ++i) { # static_cast(y.data_ptr())[i] = static_cast(x.data_ptr())[i] + 1; # } # } # # TVM_FFI_DLL_EXPORT_TYPED_FUNC(add_one_cpu, add_one_cpu); # ``` # compile the cpp source file and load the module mod: Module = tvm_ffi.cpp.load( name="my_ops", cpp_files="my_ops.cpp", ) # use the function from the loaded module x = torch.tensor([1, 2, 3, 4, 5], dtype=torch.float32) y = torch.empty_like(x) mod.add_one_cpu(x, y) torch.testing.assert_close(x + 1, y) rr)r r) rrrrrrrrrs r2loadrsKH  !%/' 3+ ,    r3r)rrrrrrrrrrrrrrrrrrrrrr )r9r r:r rr#)rr )r^rsrtrur_rvrrw)rr rrvrrrrrrrrrrrrrrrr )rr rr#)rr rr )rr rrrr )rrrrs)TN)rr rrrrrrrrrrrrrrrrvrrrr )rr rrr rrr rrrrrrrrrrrrrr )rr rrr rrr rrrrrrrrrrrrrrvrr )rr rrrrrrrrrrrrrrrr )rr rrrrrrrrrrrrrrrrvrr )0__doc__ __future__r functoolsr5rIrLrercollections.abcrr contextlibrpathlibrtypingrtvm_ffi.libinfor r r tvm_ffi.moduler r tvm_ffi.utilsrplatformrOr8rD lru_cacherXrrrrrr maketransrrrrrrrrrrr3r2r+sU"IH""""""  --------""""""XXXXXXXXXX........"""""" \W $ /344444n     899994<<<<P/3gggggT++++,MM  **** 2$.2M;M;M;M;M;f/3/3@D)-.2*.04"&.2M M M M M M f/3/3@D)-.2*.04"&.2"WWWWWWz-1-1)-.2*.04"&HHHHHH\-1-1)-.2*.04"&"PPPPPPPPr3