9idSSKrSSKrSSKrSSKrSSKrSSKrSSKJr SSKJr SSK J r SSK J r SSK Jr SSKJr \R"R%\R"R'\55r\R"R+\S5/r\R"R+\S 5rS /rSq\R4"5S 5r\R4"5S 5r"S S\5rSrSSSSSS.r SSSSSS.r!Sr"\#"\"5r$Sr%\&"S\'"S555r(S\(S'S\(S'S\(S'Sr)S r*"S!S"\5r+"S#S$\5r,g)%N)Path)knobs)compile_module_from_src) _allocation) GPUTarget) GPUDriverincludelibcudac ([RR=n(aU/$[R"SS/5R SS9nUR 5Vs/sHnSU;dM UR5SPM nnUVs/sH"n[RRU5PM$ nn[R"S5nU(alU(deURS5Vs/sHIn[RR[RRUS55(dMGUPMK nnS nU(aUS [U5-- nUS - nO US - nUS - n[SU55(dU5eU$s snfs snfs snf)Nz/sbin/ldconfigz-pignore)errors libcuda.so.1LD_LIBRARY_PATH:zlibcuda.so cannot found! z!Possible files are located at %s.z:Please create a symlink of libcuda.so to any of the files.zlibcuda_dirs..)s/Sddrww~~bggll4@AAdsA A )rnvidia libcuda_path subprocess check_outputdecode splitlinessplitrrdirnamegetenvrrstrany) env_libcuda_pathlibslinelocslocdirsenv_ld_library_pathdirmsgs r libcuda_dirsr0sd <<4444 !!  " "$4d#; < C C8 C TD*.): U):nPT>T DJJL ):D U,0 1DSBGGOOC DD 1))$564288=s=PRPWPWP\P\]`bpPqAr=s &C  2SY>> KK MM >> SdS S SXUXX S K V 1ts F*F)F )AF3Fc$[/[5Q$N) libdevice_dirr0r library_dirsr6-s  +LN ++r5c.^\rSrSrU4SjrSrSrU=r$) CudaUtils7cn>[US5(d[[U]U5UlUR$)Ninstance)hasattrsuperr8__new__r;)cls __class__s rr>CudaUtils.__new__9s-sJ'' C8=CL||r5cr[[[RR [ S55R 5S[5[[S9nURq URUl URUl URUl URUlURUlg)Nzdriver.c cuda_utilssrcnamer6 include_dirs libraries)rrrrrr# read_textr6rGrH PyCUtensorMap load_binaryget_device_propertiescuOccupancyMaxActiveClustersset_printf_fifo_sizefill_tma_descriptor)selfmods r__init__CudaUtils.__init__>s%RWW\\':67AAC%%  )) ??%(%>%>",/,L,L)$'$<$<!#&#:#: r5)rMrOrLrKrN)__name__ __module__ __qualname____firstlineno__r>rR__static_attributes__ __classcell__r@s@rr8r87s ;;r5r8cUSS:XagURS5(ag0SS_SS_S S _S S _S S_SS_SS_SS_SS_SS_SS_SS_SS_SS_SS_SS_U$)Nr* CUdeviceptr tensordesc CUtensorMapi1int8_ti8i16int16_ti32int32_ti64int64_tu1uint8_tu8u16uint16_tu32uint32_tu64uint64_tfp16doublebf16fp32f32fp64 nvTmaDesc) startswith)tys r ty_to_cppr{Ts !u| }}\""  h  h  y  y   y   i   i  z  z  z        x    ]! " #  r5rmrorq)rrrtrurvrw pack_fp16 pack_bf16 pack_fp32 pack_fp64 iiiKKppOOOOOOc|^^^^U4SjnU4SjmU4SjmU4SjmU"UR55n[U5VVs0sHupVXV_M nnnSRUR5Vs/sH nT"U5PM sn5n[U-n /n UR5H n T"X5 M [U 5VVs0sHupVXV_M nnn[ U5S:a)SSRSUR 555-OSn /n UR 5HUupWUS :XaM U[ ;aU R[ US U35 M6U R[U5S U35 MW SRU 5n/nUR 5H~upWUSS :XaURS US 35 M%U[ ;aURSUS35 MFUS:XaURSU35 MbUS :wdMjURSU35 M [[ U55nSnUR 5VVs/sHupWUSS :XdMSUSUSUSUS3 PM! nnnUR 5VVs/sHupWUS:XdM SUSUSUS3PM nnnUR 5VVs/sH-upWU[ ;dM[ USUS[USUS3PM/ nnnUR 5VVs/sHupWUS :wdM SU3PM nnnURS 5 URS!5 S"[ U5S:aSU-OSS#SRU5S$URUR 5VVs/sHupWT"U5SUS%3PM snn5S&U S'U S(URU5SURU5SURU5S)[ U5S:aSSRU5-OSS*3nU$s snnfs snfs snnfs snnfs snnfs snnfs snnfs snnf)+Nc>/nSnUGHKn[U[5(Ga URS5(Ga T (aT UOSnUS- n[R"SU5nUR S5nUR S5nUR S5S-nUcLURSU-5 [SU-5Hn URS5 M URS 5 OURS 5 [U5Hn URS 5 M [U5Hn URS5 M GM:URU5 GMN T (aU[T 5:XdeU$) Nrr^ztensordesc<([^[>]*)\[([^]]*)\],r\rgr`rxre) isinstancer%ryrematchgroupcountappendrangelen) signatureoutputtensordesc_idxsigmetardtypeshapendim_tensordesc_metas r_expand_signature(make_launcher.._expand_signatures5C#s##|(D(D:I~6t!#!CSI A A{{3'!+<MM#+. #1t8_ e,-MM$'MM+.tAMM%(%tAMM%(% c"9<#nO8L&LLL r5cv>[U[5(aUH nT"X!5 M gURU5 gr2)rtupler)rrx_flatten_signatures rr)make_launcher.._flatten_signatures0 c5 ! !"1- MM# r5c>[U[5(a!SR[TU55nSUS3$USS:XagUS;ag[ U5$)Nr[]rr\z PyObject* constexprrx)rrrmapr{)rzval_extracted_types rr&make_launcher.._extracted_typesW b% ((334Cse1:  a5C< + +}r5c >[U[5(a!SR[TU55nSUS3$USS:XagUS;agUR S5(agS S S S S SSSSSS. [ U5$)N()rr\Orr^dlbhiLBHIK) rslongrardrfrhrjrmrorq)rrrrryr{)rzr format_ofs rr make_launcher..format_ofs b% ''#i,-Cse1:  a5C< + + == & &   B-  r5rrz, c30# UH upSU3v M g7f)z&_argNr4)rrrzs rr make_launcher..s L:K5:Ksrz argr\ptr_infoz.dev_ptr_arg_storagerxz*tma_ptrz zDevicePtrInfo ptr_infoz = getPointer(_argz); if (!ptr_infoz.valid) return NULL;zCUtensorMap* tma_ptrz = getTmaDesc(_argz); if (!tma_ptrz) return NULL;z _argz _storage = z(_argz);z&argz&global_scratchz&profile_scratcha #include "cuda.h" #include #include #include #define PY_SSIZE_T_CLEAN #include typedef struct { PyObject_HEAD; _Alignas(128) CUtensorMap tensorMap; } PyCUtensorMapObject; static inline void gpuAssert(CUresult code, const char *file, int line) { if (code != CUDA_SUCCESS) { const char* prefix = "Triton Error [CUDA]: "; const char* str; cuGetErrorString(code, &str); char err[1024] = {0}; strcat(err, prefix); strcat(err, str); PyGILState_STATE gil_state; gil_state = PyGILState_Ensure(); PyErr_SetString(PyExc_RuntimeError, err); PyGILState_Release(gil_state); } } #define CUDA_CHECK(ans) { gpuAssert((ans), __FILE__, __LINE__); } typedef CUresult (*cuLaunchKernelEx_t)(const CUlaunchConfig* config, CUfunction f, void** kernelParams, void** extra); static cuLaunchKernelEx_t getLaunchKernelExHandle() { // Open the shared library void* handle = dlopen("libcuda.so.1", RTLD_LAZY); if (!handle) { PyErr_SetString(PyExc_RuntimeError, "Failed to open libcuda.so.1"); return NULL; } // Clear any existing error dlerror(); cuLaunchKernelEx_t cuLaunchKernelExHandle = (cuLaunchKernelEx_t)dlsym(handle, "cuLaunchKernelEx"); // Check for errors const char *dlsym_error = dlerror(); if (dlsym_error) { PyErr_SetString(PyExc_RuntimeError, "Failed to retrieve cuLaunchKernelEx from libcuda.so.1"); return NULL; } return cuLaunchKernelExHandle; } static void _launch(int gridX, int gridY, int gridZ, int num_warps, int num_ctas, int launch_cooperative_grid, int launch_pdl, int clusterDimX, int clusterDimY, int clusterDimZ, int shared_memory, CUstream stream, CUfunction function, CUdeviceptr global_scratch, CUdeviceptr profile_scratchz) { void *params[] = { a }; if (gridX*gridY*gridZ > 0) { // 4 attributes that we can currently pass maximum CUlaunchAttribute launchAttr[4]; static cuLaunchKernelEx_t cuLaunchKernelExHandle = NULL; if (cuLaunchKernelExHandle == NULL) { cuLaunchKernelExHandle = getLaunchKernelExHandle(); } CUlaunchConfig config; config.gridDimX = gridX; config.gridDimY = gridY; config.gridDimZ = gridZ; if (num_ctas != 1) { config.gridDimX *= clusterDimX; config.gridDimY *= clusterDimY; config.gridDimZ *= clusterDimZ; } config.blockDimX = 32 * num_warps; config.blockDimY = 1; config.blockDimZ = 1; config.sharedMemBytes = shared_memory; config.hStream = stream; config.attrs = launchAttr; int num_attrs = 0; if (launch_pdl != 0) { CUlaunchAttribute pdlAttr = { .id = CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION, .value = 1}; launchAttr[num_attrs] = pdlAttr; ++num_attrs; } if (launch_cooperative_grid != 0) { CUlaunchAttribute coopAttr = { .id = CU_LAUNCH_ATTRIBUTE_COOPERATIVE, .value = 1}; launchAttr[num_attrs] = coopAttr; ++num_attrs; } if (num_ctas != 1) { CUlaunchAttribute clusterAttr = {}; clusterAttr.id = CU_LAUNCH_ATTRIBUTE_CLUSTER_DIMENSION; clusterAttr.value.clusterDim.x = clusterDimX; clusterAttr.value.clusterDim.y = clusterDimY; clusterAttr.value.clusterDim.z = clusterDimZ; launchAttr[num_attrs] = clusterAttr; ++num_attrs; CUlaunchAttribute clusterSchedulingAttr = {}; clusterSchedulingAttr.id = CU_LAUNCH_ATTRIBUTE_CLUSTER_SCHEDULING_POLICY_PREFERENCE; clusterSchedulingAttr.value.clusterSchedulingPolicyPreference = CU_CLUSTER_SCHEDULING_POLICY_SPREAD; launchAttr[num_attrs] = clusterSchedulingAttr; ++num_attrs; } config.numAttrs = num_attrs; CUDA_CHECK(cuLaunchKernelExHandle(&config, function, params, 0)); } } typedef struct _DevicePtrInfo { CUdeviceptr dev_ptr; bool valid; } DevicePtrInfo; static PyObject* data_ptr_str = NULL; static PyObject* py_tensor_map_type = NULL; static inline DevicePtrInfo getPointer(PyObject *obj, int idx) { DevicePtrInfo ptr_info; ptr_info.dev_ptr = 0; ptr_info.valid = true; if (PyLong_Check(obj)) { ptr_info.dev_ptr = PyLong_AsUnsignedLongLong(obj); return ptr_info; } if (obj == Py_None) { // valid nullptr return ptr_info; } PyObject *ret = PyObject_CallMethodNoArgs(obj, data_ptr_str); if (!ret) { PyErr_SetString(PyExc_TypeError, "Pointer argument must be either uint64 or have data_ptr method"); ptr_info.valid = false; goto cleanup; } if (!PyLong_Check(ret)) { PyErr_SetString(PyExc_TypeError, "data_ptr method of Pointer object must return 64-bit int"); ptr_info.valid = false; goto cleanup; } ptr_info.dev_ptr = PyLong_AsUnsignedLongLong(ret); if(!ptr_info.dev_ptr) return ptr_info; uint64_t dev_ptr; int status = cuPointerGetAttribute(&dev_ptr, CU_POINTER_ATTRIBUTE_DEVICE_POINTER, ptr_info.dev_ptr); if (status == CUDA_ERROR_INVALID_VALUE) { PyErr_Format(PyExc_ValueError, "Pointer argument (at %d) cannot be accessed from Triton (cpu tensor?)", idx); ptr_info.valid = false; } else if (status != CUDA_SUCCESS) { CUDA_CHECK(status); // Catch any other cuda API errors ptr_info.valid = false; } ptr_info.dev_ptr = dev_ptr; cleanup: Py_XDECREF(ret); return ptr_info; } static inline CUtensorMap* getTmaDesc(PyObject *obj) { if (sizeof(CUtensorMap*) != 8) { PyErr_SetString(PyExc_SystemError, "getTmaDesc() requires 64-bit compilation"); return NULL; } if (Py_TYPE(obj) != (PyTypeObject*)py_tensor_map_type) { PyErr_Format(PyExc_TypeError, "object must be of type PyCUtensorMap, got %s", Py_TYPE(obj)->tp_name); return NULL; } CUtensorMap* map = &((PyCUtensorMapObject*)obj)->tensorMap; uintptr_t align_128 = (uintptr_t)map & (128 - 1); if (align_128 != 0) { PyErr_Format(PyExc_ValueError, "CUtensorMap must be aligned to 128B, but got (&map) mod 128 = %ld", align_128); return NULL; } return map; } static void ensureCudaContext() { CUcontext pctx; CUDA_CHECK(cuCtxGetCurrent(&pctx)); if (!pctx) { // Ensure device context. CUdevice device; CUDA_CHECK(cuDeviceGet(&device, 0)); CUDA_CHECK(cuDevicePrimaryCtxRetain(&pctx, device)); CUDA_CHECK(cuCtxSetCurrent(pctx)); } } static uint16_t pack_fp16(double f) { uint16_t result; // from https://github.com/python/pythoncapi-compat #if 0x030600B1 <= PY_VERSION_HEX && PY_VERSION_HEX <= 0x030B00A1 && !defined(PYPY_VERSION) _PyFloat_Pack2(f, (unsigned char*)&result, 1); #else PyFloat_Pack2(f, (unsigned char*)&result, 1); #endif return result; } static uint16_t pack_bf16(double f) { float f32 = (float)f; uint32_t u32 = *(uint32_t*)&f32; return (uint16_t)(u32 >> 16); } static uint32_t pack_fp32(double f) { float f32 = (float)f; return *(uint32_t*)&f32; } static uint64_t pack_fp64(double f) { return *(uint64_t*)&f; } static PyObject* launch(PyObject* self, PyObject* args) { // ensure cuda context is valid before calling any CUDA APIs, e.g. before getPointer calls cuPointerGetAttributes ensureCudaContext(); int gridX, gridY, gridZ; uint64_t _stream; uint64_t _function; int launch_cooperative_grid; int launch_pdl; PyObject *launch_enter_hook = NULL; PyObject *launch_exit_hook = NULL; PyObject *kernel_metadata = NULL; PyObject *launch_metadata = NULL; PyObject *global_scratch_obj = NULL; PyObject *profile_scratch_obj = NULL; ;z if(!PyArg_ParseTuple(args, "aM", &gridX, &gridY, &gridZ, &_stream, &_function, &launch_cooperative_grid, &launch_pdl, &global_scratch_obj, &profile_scratch_obj, &kernel_metadata, &launch_metadata, &launch_enter_hook, &launch_exit_hookaT)) { return NULL; } int num_warps, num_ctas, shared_memory, clusterDimX, clusterDimY, clusterDimZ; if (!PyArg_ParseTuple(kernel_metadata, "iiiiii", &num_warps, &num_ctas, &shared_memory, &clusterDimX, &clusterDimY, &clusterDimZ)) { PyErr_SetString(PyExc_TypeError, "kernel_metadata must be a tuple"); return NULL; } // extract launch metadata if (launch_enter_hook != Py_None){ PyObject* ret = PyObject_CallOneArg(launch_enter_hook, launch_metadata); if (!ret) return NULL; Py_DECREF(ret); } CUdeviceptr global_scratch = 0; if (global_scratch_obj != Py_None) { DevicePtrInfo global_scratch_info = getPointer(global_scratch_obj, -1); if (!global_scratch_info.valid) { return NULL; } global_scratch = global_scratch_info.dev_ptr; } CUdeviceptr profile_scratch = 0; if (profile_scratch_obj != Py_None) { DevicePtrInfo profile_scratch_info = getPointer(profile_scratch_obj, -1); if (!profile_scratch_info.valid) { return NULL; } profile_scratch = profile_scratch_info.dev_ptr; } // raise exception asap z Py_BEGIN_ALLOW_THREADS; _launch(gridX, gridY, gridZ, num_warps, num_ctas, launch_cooperative_grid, launch_pdl, clusterDimX, clusterDimY, clusterDimZ, shared_memory, (CUstream)_stream, (CUfunction)_function, global_scratch, profile_scratchap); Py_END_ALLOW_THREADS; if (PyErr_Occurred()) { return NULL; } if(launch_exit_hook != Py_None){ PyObject* ret = PyObject_CallOneArg(launch_exit_hook, launch_metadata); if (!ret) return NULL; Py_DECREF(ret); } Py_RETURN_NONE; } static PyMethodDef ModuleMethods[] = { {"launch", launch, METH_VARARGS, "Entry point for all kernels with this signature"}, {NULL, NULL, 0, NULL} // sentinel }; static struct PyModuleDef ModuleDef = { PyModuleDef_HEAD_INIT, "__triton_launcher", NULL, //documentation -1, //size ModuleMethods }; PyMODINIT_FUNC PyInit___triton_launcher(void) { data_ptr_str = PyUnicode_InternFromString("data_ptr"); if(data_ptr_str == NULL) { return NULL; } PyObject* driver_mod = PyImport_ImportModule("triton.backends.nvidia.driver"); if (driver_mod == NULL) { return NULL; } py_tensor_map_type = PyObject_GetAttrString(driver_mod, "PyCUtensorMap"); if (py_tensor_map_type == NULL) { return NULL; } PyObject *m = PyModule_Create(&ModuleDef); if(m == NULL) { return NULL; } PyModule_AddFunctions(m, ModuleMethods); return m; } ) values enumerater_BASE_ARGS_FORMATritemsFLOAT_STORAGE_TYPErr{rFLOAT_PACK_FUNCTION) constantsrrrexpand_signaturersrz args_formatformatflat_signaturer args_list arg_decl_list arg_declsinternal_args_listparamsnewline ptr_decls tma_declsfloat_storage_declsrErrrs ` @@@r make_launcherrs%N.))9)9);<"+,<"=>"=$!"=I>''93C3C3EF3ER9R=3EFGK  ,FN!3/""+N";<";$!";I "FG__&&EA a5C< f #5aS1#=MaSPde& foetetevev\a\]   X qc!3A3oaSWev __&&EA # # Z b ! "%s+6I"6M5NeTUSVVXY& '0oo&7 M&7UQ2;LjQCj&7F M MM#$ MM$%5cjx{|ExFIJxJdhktdtPRcSSyy()y*r << @QR@QuqOB'(aS2@QRSTU &x(QR[P[%\J <<  <<  <<#$%&YBEFXBY\]B]Z^aeajajk}a~Z~ceYf2fyN C^ JI ?F =8  NdSsHP/PPP P  P&P&6P, P, P2 P2P8c#(# UHoU4v M g7fr2r4)rrs rrr[s: 1A s  c RUcLUR/URQURQURS:HPURQURQ$USnUSnUSnUSnUSnURnURnUSS:XdeURS:XaSOS n U(a[ U5nUS==S -ss'[ R RRRRURR5UU[UUUUU 5n U /UQUQ$) Nnanswizzle elem_size elem_type block_size fp4_paddedrrrr) baserstridespaddinglisttritonruntimedriveractiveutilsrOdata_ptrTMA_DTYPE_DEVICE_TO_HOST) argmetadatarrrrrrrr cu_tensor_maps rmake_tensordesc_argras2c399cs{{cCKK54Hc399cWZWbWbccy!G%I%I,'J,'J IIEkkG 2;!  ;;%'aQGU  b Q NN))0066JJ  +  M  ,E ,G ,,r5c ^^^[SUR555nU(dT$[[UR55VVs/sH6upE[ U[ 5(dMUR S5(dM4UPM8 snn5mT(a[T5[T5:XdeT(dS/[T5-mUUU4SjnU$s snnf)Nc3r# UH-n[U[5=(a URS5v M/ g7f)r^N)rr%ry)rrs rr)wrap_handle_tensordesc..s+r_qX[jc2Ss~~l7SS_qs57r^c>[US[5nSn[U[S5HAup4UT;a%UR[ UTU55 US- nM0UR U5 MC T"U6$)Nrr)r_BASE_ARGS_FORMAT_LENrextendrr)args final_argsrrrlaunchertensordesc_indicesrs rinner%wrap_handle_tensordesc..inners~$5 567 %:%; <=FA&&!!"5c?>;Z"[\!#!!#& > $$r5)r&rsetrrr%ryr)rrrhas_tensor_desc_argrrrrs` ` @rwrap_handle_tensordescrsr_h_o_o_qrr "9#3#3#56p6vq*S#:NSVSaSabnSo6pr#o"6#>P:Q"QQ Q &3'9#:: % L! qsC -C C c \rSrSrSrSrSrg) CudaLauncheric^[TS5(a TRO [5nU4SjnUR5VVs0sHupVU"U5U_M nnnTRR5VVs0sHupVXV_M nnn[ USS5n[ X7U5m[TS[5[[S9n [R"[RURS5Ul[#U R$Xx5UlUR&UlUR(UlUR*UlUR,UlUR.UlUR0Ulgs snnfs snnf)Nrc~>[U[5(a&TRRR U54$U$r2)rr%fn arg_namesindex)rrEs r'CudaLauncher.__init__..s2Z3=O=OSVV--33A69VUVVr5r__triton_launcherrDr)r<rdictrrgetattrrrr6rGrH functoolsreduceoperatormul cluster_dimsnum_ctasrlaunchglobal_scratch_sizeglobal_scratch_alignprofile_scratch_sizeprofile_scratch_alignlaunch_cooperative_grid launch_pdl) rPrErrarg_idxidxvaluerrrQs ` rrRCudaLauncher.__init__s6%,S+%>%>CMMDF V;D??;LM;LZSWS\5(;L M25--2E2E2GH2GJCSZ2G H!(,=tDI/B%$%%  "((x7L7LaP ,SZZT #+#?#? $,$A$A!$,$A$A!%-%C%C"'/'G'G$"--'NHs E57E;c .^^^^^UUUUU4SjnU"TRTR[R5nU"TRTR [R 5n TR"TTTTUTRTRX/ UQ76 g)Ncx>US:a3TT-T-nUT R-U-nUR5nU"XAT 5$gNr)rget) sizealign allocator grid_size alloc_sizealloc_fngridXgridYgridZrPstreams rallocate_scratch/CudaLauncher.__call__..allocate_scratchsFax!EME1 &6= $==? 6::r5) rrr _allocatorrr_profile_allocatorrrr) rPr'r(r)r*functionrr+global_scratchprofile_scratchs ````` r__call__CudaLauncher.__call__s  *$*B*BDD]D]_j_u_uv*4+D+DdF`F`+6+I+IK E5%4;W;WY]YhYh" <6: [5Ul[Ul[TU]5 gr2)r8rr launcher_clsr=rR)rPr@s rrRCudaDriver.__init__s[ ( r5c|UR5nURU5nUSS-US-nSn[SX#5$)Nrrr r )get_current_deviceget_device_capabilityr)rPdevice capability warp_sizes rget_current_targetCudaDriver.get_current_targetsI((*//7 ]R'*Q-7  77r5cJSSKnURSUR55$)Nrr )torchr=r;rPrCs rget_active_torch_device"CudaDriver.get_active_torch_devices||FD$;$;$=>>r5c"SSKnUR$r)rCr rDs rget_device_interfaceCudaDriver.get_device_interfaceszzr5cSSKnURR5=(a URRSL$![ a gf=f)NrF)rCr is_availableversionhip ImportError)rCs r is_activeCudaDriver.is_activesC  ::**,L%--2C2Ct2K L  srpsu  8&., ''//"''**84 5 Wi01  We, H   . ,,;;: 4            $-.Wv :b ::  $-N2(<6(