DŽgGUdZddlZddlZddlZddlZddlZddlmZddlm Z m Z m Z m Z m Z mZmZddlZddlZddlmZddlmZmZmZddlmZdd lmZdd lmZdd lmZm Z m!Z!m"Z"m#Z#dd l$m%Z%m&Z&m'Z' dd lm(Z(da*ejVZ,ejZZ.ga/e ee gdfe e0fe1d<e2ejfddZ4eee0e5dfZ6dZ7dZ8 ddlm9Z: e:jvsddlZddde e6de0fd?Zedde e6dee5e5ffd@Zfde6deBfdAZgde6dBe6deHfdCZhGdDdEZidFe dGdeifdHZjdIZkdFe'fdJZldee e5e e0ffdKZmde5fdLZnde5fdMZode e e0fdNZpde e e0fdOZqdPe e0dQe e0de e5fdRZrde5fdSZsde5fdTZtde ee5efde5fdUZudave e5e1dV<de5fdWZwde e0fdXZxde0fdYZyde5fdZZzdde6ddfd[Z{d\Z|dde e6de'fd]Z}dde e6de'fd^Z~d_Zd`ee5e0fddfdaZde5fdbZdde eee5ffdcZdde eee5ffddZde ee5efde5fdeZdde eee5fde5fdfZdde eee5fde5fdgZdde eee5fde5fdhZdde eee5fde5fdiZdde eee5fde5fdjZdde eee5fde5fdkZdde eee5fde5fdlZdde eee5fde5fdmZdde eee5fde5fdnZdde eee5fde5fdoZdee5e0ej$fdej$fdpZdej$dejfjfdqZ ddre5dee5e0ej$fddfdsZddee5e0ej$fde5fdtZdduldduledvZGdwdxZddylmZmZGdzd{eZGd|d}eZGd~deZGddeZGddeZGddeZGddeZGddeZGddeZGddeZGddeZGddeZGddeZ[[ejPjSeejPjSeejPjSeejPjSeejPjSeejPjSeejPjSeejPjSeejPjSeejPjSeejPjSeejPjSeGddZdZeWeddlmZmZmZmZmZmZgdZy#e)$rdZ(YwxYw#e>$rYSwxYw#[:wxYw#e)$r Z?e?Z8YdZ?[?hdZ?[?wwxYw)aM This package adds support for CUDA tensor types. It implements the same function as CPU tensors, but they utilize GPUs for computation. It is lazily initialized, so you can always import it, and use :func:`is_available()` to determine if your system supports CUDA. :ref:`cuda-semantics` has more details about working with CUDA. N) lru_cache)AnyCallablecastListOptionalTupleUniondevice) _dummy_type_LazySeedTracker classproperty)Device)gds)_get_device_index) CUDAGraphgraphgraph_pool_handleis_current_stream_capturingmake_graphed_callables)EventExternalStreamStream)_cudartF _queued_calls_cuda_isInBadForkcyNFr![/home/mcse/projects/flask_80/flask-venv/lib/python3.12/site-packages/torch/cuda/__init__.pyr$3sr")versionT_CudaDeviceProperties_cuda_exchangeDevicer returnc$|dkrytdNrz)PyTorch was compiled without CUDA support RuntimeErrorr s r#_exchange_devicer.U A:FGGr"_cuda_maybeExchangeDevicec$|dkrytdr*r,r s r#_maybe_exchange_devicer2_r/r"has_half has_magmar!default_generatorsc6ttjdS)z)Return true if compile with CUDA support._cuda_getDeviceCount)hasattrtorch_Cr!r"r# _is_compiledr;ks 5883 44r"c2tjddk(S)NPYTORCH_NVML_BASED_CUDA_CHECK1)osgetenvr!r"r#_nvml_based_availrAps 994 5 <~!! xx,,.22r"including_emulationcxtjjrytsytjj }tjj}|Mt |jdddk\r-tjj|jdk\ry|syt|S)zQReturn a bool indicating if the current CUDA/ROCm device supports dtype bfloat16.TF.r ) r9r%hiprDcudacurrent_deviceintsplitget_device_propertiesmajor_check_bf16_tensor_supported)rEr cuda_versions r#is_bf16_supportedrSs }} > ZZ & & (F==%%L  ""3'* +r 1 JJ , ,V 4 : :a ?  ( //r")maxsizecr tjdgtj|y#t$rYywxYw)Ng?)dtyper TF)r9tensorbfloat16 Exceptionr s r#rQrQs2 cU%..@ s '* 66cBtjj|yN)r9r: _cuda_sleep)cycless r#_sleepr_s HH r" arch_stringcl|jdd}|jdr|dd}t|S)z4Extracts the architecture string from a CUDA version_raNr+)rNendswithrM)r`bases r#_extract_arch_versionrfs8   S !! $D }}SCRy t9r"c d}d}tjjtjj }t t D]}t|}|d}|d}t|}|dz|z}tdtjjDd} || ksatj|||||| dz| dzfzyy) Nz Found GPU%d %s which requires CUDA_VERSION >= %d to work properly, but your PyTorch was compiled with CUDA_VERSION %d. Please install the correct PyTorch binary using instructions from https://pytorch.org z Found GPU%d %s which is of cuda capability %d.%d. PyTorch no longer supports this GPU because it is too old. The minimum cuda capability supported by this library is %d.%d. rr c32K|]}t|ywr\)rf).0archs r# z$_check_capability..sT&t,Ts#)default) r9r%rKr:_cuda_getCompiledVersionrangerCget_device_capabilityget_device_namemin get_arch_listwarningswarn) incorrect_binary_warn old_gpu_warn CUDA_VERSIONd capabilityrPminorname current_archmin_archs r#_check_capabilityrsL  }}%xx88: |~& A.q1JqMEqME"1%D 2:-LT9Q9Q9STHh&  $uh"nhmLM &r"c  d}tjjyt}t |dk(ry|Dcgc]}d|vst |}}t tD]p}t|\ }t fd|D}|r(t|} dz|z}tj|j||dj||rycc}w)Na  {} with CUDA capability sm_{} is not compatible with the current PyTorch installation. The current PyTorch install supports CUDA capabilities {}. If you want to use the {} GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/ rsm_c3.K|] }|dzk(yw)rhNr!)rjsm cap_majors r#rlz _check_cubins..sE"bI-Esrh )r9r%rKrtlenrfrprCrqanyrrrurvformatjoin) incompatible_device_warn arch_listrk supported_smidx cap_minor supported device_namer{rs @r# _check_cubinsrs   }}!I 9~CZB[]04!; < D+ ,_00sS FB2F*F E1D4E1F4 E.=,E))E..E11FFFc"ttS)a Retrieves the CUDA runtime API module. This function initializes the CUDA runtime environment if it is not already initialized and returns the CUDA runtime API module (_cudart). The CUDA runtime API module provides access to various CUDA runtime functions. Args: ``None`` Returns: module: The CUDA runtime API module (_cudart). Raises: RuntimeError: If CUDA cannot be re-initialized in a forked subprocess. AssertionError: If PyTorch is not compiled with CUDA support or if libcudart functions are unavailable. Example of CUDA operations with profiling: >>> import torch >>> from torch.cuda import cudart, check_error >>> import os >>> >>> os.environ['CUDA_PROFILE'] = '1' >>> >>> def perform_cuda_operations_with_streams(): >>> stream = torch.cuda.Stream() >>> with torch.cuda.stream(stream): >>> x = torch.randn(100, 100, device='cuda') >>> y = torch.randn(100, 100, device='cuda') >>> z = torch.mul(x, y) >>> return z >>> >>> torch.cuda.synchronize() >>> print("====== Start nsys profiling ======") >>> check_error(cudart().cudaProfilerStart()) >>> with torch.autograd.profiler.emit_nvtx(): >>> result = perform_cuda_operations_with_streams() >>> print("CUDA operations completed.") >>> check_error(torch.cuda.cudart().cudaProfilerStop()) >>> print("====== End nsys profiling ======") To run this example and save the profiling information, execute: >>> $ nvprof --profile-from-start off --csv --print-summary -o trace_name.prof -f -- python cudart_test.py This command profiles the CUDA operations in the provided script and saves the profiling information to a file named `trace_name.prof`. The `--profile-from-start off` option ensures that profiling starts only after the `cudaProfilerStart` call in the script. The `--csv` and `--print-summary` options format the profiling output as a CSV file and print a summary, respectively. The `-o` option specifies the output file name, and the `-f` option forces the overwrite of the output file if it already exists. )rrr!r"r#cudartrXslL Nr"c*eZdZUdZeed<dZeed<y) cudaStatusrSUCCESS"ERROR_NOT_READYN)rrrrrM__annotations__rr!r"r#rrsGSOSr"rc(eZdZdeddffd ZxZS) CudaErrorcoder(Nctjtj|}t||d|dy)Nz ())rcudaGetErrorString cudaErrorsuper__init__)selfrr __class__s r#rzCudaError.__init__s8(():):4)@A C54&*+r")rrrrMr __classcell__rs@r#rrs,S,T,,r"rrescT|tjjk7r t|yr\)rrsuccessr)rs r# check_errorrs% g'''n(r"c2eZdZdefdZdZdededefdZy) _DeviceGuardindexc ||_d|_yNr+)rprev_idx)rrs r#rz_DeviceGuard.__init__s r"c`tjj|j|_yr\r9rKr.rrrs r# __enter__z_DeviceGuard.__enter__ 33DHH= r"typevaluerc`tjj|j|_yr r9rKr2rrrrrrs r#__exit__z_DeviceGuard.__exit__::44T]]Cr"N)rrrrMrrrrr!r"r#rrs-c>Sr"rc6eZdZdZdefdZdZdededefdZy) r zContext-manager that changes the selected device. Args: device (torch.device or int): device index to select. It's a no-op if this argument is a negative integer or ``None``. c6t|d|_d|_y)NToptionalr+)rrr)rr s r#rzdevice.__init__s$Vd; r"c`tjj|j|_yr\rrs r#rzdevice.__enter__rr"rrrc`tjj|j|_yr rrs r#rzdevice.__exit__rr"N)rrr__doc__rrrrr!r"r#r r s2s>Sr"c"eZdZdZfdZxZS) device_ofaContext-manager that changes the current device to that of given object. You can use both tensors and storages as arguments. If a given object is not allocated on a GPU, this is a no-op. Args: obj (Tensor or Storage): object allocated on the selected device. c`|jr|jnd}t| |yr)is_cuda get_devicerr)robjrrs r#rzdevice_of.__init__s$"%++cnn2 r")rrrrrrrs@r#rrsr"rcdt|}|dk\r tjj|yy)a=Set the current device. Usage of this function is discouraged in favor of :any:`device`. In most cases it's better to use ``CUDA_VISIBLE_DEVICES`` environmental variable. Args: device (torch.device or int): selected device. This function is a no-op if this argument is negative. rN)rr9r:_cuda_setDevicer s r# set_devicers,v &F {   (r"c,t|jS)aGet the name of a device. Args: device (torch.device or int or str, optional): device for which to return the name. This function is a no-op if this argument is a negative integer. It uses the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). Returns: str: the name of the device )rOr}r s r#rrrrs ! ( - --r"cHt|}|j|jfS)aGet the cuda capability of a device. Args: device (torch.device or int or str, optional): device for which to return the device capability. This function is a no-op if this argument is a negative integer. It uses the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). Returns: tuple(int, int): the major and minor cuda capability of the device )rOrPr|)r props r#rqrqs! ! (D ::tzz !!r"ctt|d}|dks |tk\r tdt |S)zGet the properties of a device. Args: device (torch.device or int or str): device for which to return the properties of the device. Returns: _CudaDeviceProperties: the properties of the device TrrInvalid device id)rrrCr_get_device_propertiesr s r#rOrOs<L v 5F zV|~-011 !& ))r" peer_devicectt|d}t|}|dks |tk\r td|dks |tk\r tdtj j ||S)z5Check if peer access between two devices is possible.TrrrzInvalid peer device id)rrrCrr9r:_cuda_canDeviceAccessPeer)r rs r#can_device_access_peerrslL v 5F#K0K zV|~-011Q+7566 88 - -fk BBr"cNeZdZUdZeded<dedfdZdZdeded efd Z y ) StreamContextaContext-manager that selects a given stream. All CUDA kernels queued within its context will be enqueued on a selected stream. Args: Stream (Stream): selected stream. This manager is a no-op if it's ``None``. .. note:: Streams are per-device. torch.cuda.Stream cur_streamstreamc||_tdd|_tjj s|jd|_tjj sdntj jd|_tjj sd|_ ytj jd|_ y)NTr+) rrrr9jit is_scriptingrKdefault_streamsrc_prev_streamdst_prev_stream)rrs r#rzStreamContext.__init__+s $T40yy%%'xx ..0Dejj6O6OPT6U  ..0D 6;jj6O6OPT6U r"c|j}||jdk(rytjj d|_|j j |j k7rLt |j 5tjj |j |_dddtjj|y#1swY)xYwr) rrr9rKcurrent_streamr r r  set_stream)rrs r#rzStreamContext.__enter__9s[[  R $zz88>    & &**;*; ; ))* T',zz'@'@ARAR'S$ T j) T Ts ;/CCrrrc*|j}||jdk(ry|jj|jk7r)tj j |jtj j |jyr)rrr r r9rKrr )rrrrrs r#rzStreamContext.__exit__Hsj[[  R     & &**;*; ; JJ ! !$"6"6 7 d223r"N) rrrrrrrrrrr!r"r#rrsF ,--  x(;<   * 4S 4 4 4r"rrrct|S)aMWrap around the Context-manager StreamContext that selects a given stream. Arguments: stream (Stream): selected stream. This manager is a no-op if it's ``None``. ..Note:: In eager mode stream is of type Stream class while in JIT it is an object of the custom class ``torch.classes.cuda.Stream``. )rrs r#rrVs   r"cHtjj|||y)zset stream specified by the stream id, device index and device type Args: stream_id (int): stream id in stream pool device_index (int): device index in topo device_type (int): enum device type  stream_id device_index device_typeN)r9r:_cuda_setStreamrs r#_set_stream_by_idrbs$ HH!r"cb|yt|j|j|jy)aSet the current stream.This is a wrapper API to set the stream. Usage of this function is discouraged in favor of the ``stream`` context manager. Args: stream (Stream): selected stream. This function is a no-op if this argument is ``None``. Nr)rrrrrs r#rrqs/~""((&&r"cLtjd}tjjrtjd}||}|t t dSdtdtfd}dtdtdttfd }|jd r ||d S|jd r ||d Sg}|jd D]N}||j}||vrtttgcS|d kr|S|j|P|S)z0Parse CUDA_VISIBLE_DEVICES environment variable.CUDA_VISIBLE_DEVICESHIP_VISIBLE_DEVICES@sr(c|syt|D]7\}}|js |dk(r|dvsn|dzt|k(s3|dz }9dkDrt|d|SdS)z:Return -1 or positive integer sequence string starts with.r+rz+-rN) enumerateisdigitrrM)r rcs r#_strtoulz(_parse_visible_devices.._strtoulsml FCIIKC1HdQw#a& q   #Qws1Tc7|.B.r"lstprefixcg}|jdD]D}||vrtttgcS|j |s|S|j |F|S)N,)rNrrr startswithr)r&r'rcselems r#parse_list_with_prefixz6_parse_visible_devices..parse_list_with_prefixs_IIcN Ds{DIr**??6*  JJt   r"zGPU-MIG-r)r)r?r@r9r%rJlistrprrMrr*rNstriprr)var hip_devicesr%r-rcr,xs r#_parse_visible_devicesr5s ))* +C }}ii 56  "C {E"I /C /C / C  c  ~~f%c622 ~~f%c622B # TZZ\ " 7S 2& & q5  I !  Ir"ctsy tjtj}t|S#tj$r,}t j d|j Yd}~yd}~wwxYw)Nr+z&Can't initialize amdsmi - Error code: ) _HAS_PYNVMLamdsmi amdsmi_initAmdSmiExceptionrurverr_codeamdsmi_get_processor_handlesr)rsocket_handless r#_raw_device_count_amdsmir>sg 88:N ~   ! ! >qzzlKLs<A;"A66A;c ddlm}m}m}|d}|j }|dk7rt j dy|d}|j||}|dk7rt j dy~|jS)zgReturn number of devices as reported by NVML or negative value if NVML discovery/initialization failed.r)byrefc_intCDLLlibnvidia-ml.so.1Can't initialize NVMLr+Can't get nvml device count) ctypesr@rArBnvmlInitrurvnvmlDeviceGetCount_v2r)r@rArBnvml_hr3 dev_counts r#_raw_device_count_nvmlrKsx)) % &F  B Qw -.b I  % %eI&6 7B Qw 34 ??r"cddlm}m}m}m}m}t sy tj tj}t|}g}t|D]J} tj|} tj| } |j!t#| L|S#tj$rtjdYywxYw#tj$rtjdYywxYw#tj$rtjdYywxYw#tj$rtjdYywxYw)Nrr@rAc_void_prBcreate_string_bufferzCan't initialize amdsmizCan't get amdsmi device countzCannot get amd device handlerzCannot get uuid for amd device)rFr@rArNrBrOr7r8r9r:rurvr<rrpamdsmi_get_gpu_device_uuidrr) r@rArNrBrOr=rJuuidsrhandleruuids r#_raw_device_uuid_amdsmirTs6II <<>' EY   99;C@G 44W=D  SY   L-  ! ! /0  ! ! 56%%  MM9 :  %%  MM: ; sFB'CD4D2'(CC(D?D(D/.D/2(EEcddlm}m}m}m}m}|d}|j }|dk7rtjdy|d}|j||}|dk7rtjdyg}t|jD]} |} |j| || }|dk7rtjdyd } || } |j| | | }|dk7rtjd y|j| jj!d j#d ~|S) z^Return list of device UUID as reported by NVML or None if NVM discovery/initialization failed.rrMrCrDNr+rEzCan't get device handle`zCan't get device UUIDascii)rFr@rArNrBrOrGrurvrHrprnvmlDeviceGetHandleByIndex_v2nvmlDeviceGetUUIDrrawdecoder0) r@rArNrBrOrIr3rJrQrdev_idbuf_lenbufs r#_raw_device_uuid_nvmlr`s$II % &F  B Qw -.b I  % %eI&6 7B Qw 34EY__% :  1 1#uV} E 7 MM3 4"7+  % %fc7 ; 7 MM1 2 SWW^^G,22489 :  Lr" candidatesrQcdtdttdtfd}g}|D]A}|||}|dkr|S||vrtttgcS|j |C|S)zqGiven the set of partial uuids and list of known uuids builds a set of ordinals excluding ambiguous partials IDs. candidaterQr(cfd}t|D] \}}|j|s|dk7ry|}"|Sr)r"r*)rcrQ best_matchrrSs r#uuid_to_orinalz3_transform_uuid_to_ordinals..uuid_to_orinalsF "5) IC??9-RJ  r"r)rrrMrr)rarQrfr3rcrs r#_transform_uuid_to_ordinalsrgs # d3i C B Y. 7  I "9S 2& & # Ir"ct}|sy t|dturyt}|dkr|St |D]\}}t t ||k\s|cS t|S#t$rYyt$rYywxYw)Nrr+) r5rrr>r"rrMOSErrorAttributeErrorr)visible_devicesraw_cntrvals r#_device_count_amdsmirn2s,.O  " #s *.0G!|%o6 SS>W,J     s-A+A+$A+A+A++ B6BBct}|sy t|dturD|djdryt }|yt t tt||}n;t}|dkr|St|D]\}}t t||k\s|cSt|S#t$rYyt$rYywxYw)zReturn number of devices as reported by NVML taking CUDA_VISIBLE_DEVICES into account. Negative value is returned if NVML discovery or initialization has failed. rr.r+)r5rrr*r`rgrrrKr"rMrirjr)rkrQrlrrms r#_device_count_nvmlrpHs -.O  " #s *q!,,V4)+E}9T#Y0%O-.G!|%o6 SS>W,J     s4(B- B-2B-8$B-B-!B-- C8CCcJt|d}t}t|dtur8t }| t dt ttt||}ttt|}|dks|t|k\rt d|d|d||S)zNReturn the NVML index of the device, taking CUDA_VISIBLE_DEVICES into account.TrrzCan't get device UUIDsdevice z& is not visible (CUDA_VISIBLE_DEVICES=r) rr5rrr`r-rgrrrMr)r rrkrQs r#_get_nvml_device_indexrsjs FT 2C,.O OA 3&%' =78 85 cO ,e 49o6O Qw#_--cU@@QQR S   3 r"_cached_device_countctsyttStjjr t n t }|dkrtjjn|}tr|a|S)z$Return the number of GPUs available.r) r;rtr9r%rJrnrpr:r7r) nvml_countrs r#rCrCs] >'##+0==+<+<%'BTBVJ+5>%%'zA  Hr"c~tsgStjj}|gS|j S)z=Return list CUDA architectures this library was compiled for.)rDr9r:_cuda_getArchFlagsrN) arch_flagss r#rtrts8 > ,,.J    r"c t}t|dk(ry|Dcgc]}|jd}}dj|Dcgc]\}}d|d|d|c}}Scc}wcc}}w)z9Return NVCC gencode flags this library was compiled with.rrrbrz-gencode compute=compute_z,code=)rtrrNr)rrk arch_list_kinds r#get_gencode_flagsr~s}I 9~.78d$**S/8J8 88!+ t(vVD64& A  9 s A%A* cRttjjS)z0Return the index of a currently selected device.)rr9r:_cuda_getDevicer!r"r#rLrLsL 88 # # %%r"cttjj|5tjj cdddS#1swYyxYw)a,Wait for all kernels in all streams on a CUDA device to complete. Args: device (torch.device or int, optional): device for which to synchronize. It uses the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). N)rr9rKr r:_cuda_synchronizer s r# synchronizers@L   6 ",xx))+,,,s AAcRttjjS)axForce collects GPU memory after it has been released by CUDA IPC. .. note:: Checks if any sent CUDA tensors could be cleaned from the memory. Force closes shared memory file used for reference counting if there is no active counters. Useful when the producer process stopped actively sending tensors and want to release unused memory. )rr9r:_cuda_ipc_collectr!r"r# ipc_collectrsL 88 % % ''r"cttjjt |d}t |d|d|dS)aSReturn the currently selected :class:`Stream` for a given device. Args: device (torch.device or int, optional): selected device. Returns the currently selected :class:`Stream` for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). Trrrr)rr9r:_cuda_getCurrentStreamrrr streamdatas r#rrJL00&40J Q-jmTU r"cttjjt |d}t |d|d|dS)a=Return the default :class:`Stream` for a given device. Args: device (torch.device or int, optional): selected device. Returns the default :class:`Stream` for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). Trrrrr)rr9r:_cuda_getDefaultStreamrrrs r#r r rr"cRttjjS)z6Return cublasHandle_t pointer to current cuBLAS handle)rr9r:_cuda_getCurrentBlasHandler!r"r#current_blas_handlersL 88 . . 00r" debug_modectt|tr#|dk(rd}n|dk(rd}n|dk(rd}n tdtj j |y) aSet the debug mode for cuda synchronizing operations. Args: debug_mode(str or int): if "default" or 0, don't error or warn on synchronizing operations, if "warn" or 1, warn on synchronizing operations, if "error" or 2, error out synchronizing operations. Warning: This is an experimental feature, and not all synchronizing operations will trigger warning or error. In particular, operations in torch.distributed and torch.sparse namespaces are not covered yet. rnrrvrerrorrzGinvalid value of debug_mode, expected one of `default`, `warn`, `error`N)r isinstancerr-r9r:_cuda_set_sync_debug_mode)rs r#set_sync_debug_moders^L*c"  "J 6 !J 7 "JY  HH&&z2r"cRttjjS)zEReturn current value of debug mode for cuda synchronizing operations.)rr9r:_cuda_get_sync_debug_moder!r"r#get_sync_debug_modersL 88 - - //r"ctstdtddlm} tj t|}tj|}|S#|$r}t d|d}~wwxYw)Nz=pynvml does not seem to be installed or it can't be imported.r)NVMLError_DriverNotLoadedz-cuda driver can't be loaded, is cuda enabled?) r7ModuleNotFoundError _PYNVML_ERRpynvmlrrGr-rsnvmlDeviceGetHandleByIndex)r rrhandles r#_get_pynvml_handlerrsq ! K  1S$F +F  . .v 6F M %SJKQRRSsAA* A%%A*ctstdt tjt|}tj|}|S#tj $r}t d|d}~wwxYw)Nz=amdsmi does not seem to be installed or it can't be imported.z>amdsmi driver can't be loaded, requires >=ROCm5.6 installation) r7rrr8r9r:r-_get_amdsmi_device_indexr<)r rrs r#_get_amdsmi_handlerr&sx ! K   &f -F  0 0 26 :F M  ! ! L  sAA4# A//A4ct|d}t}t|dtur t dt t ttt|}||vrt d|d|d||S)zKReturn the amdsmi index of the device, taking visible_devices into account.Trrz5HIP_VISIBLE_DEVICES should be indices and not stringsrrz% is not visible (HIP_VISIBLE_DEVICES=r) rr5rrr-dictr"rrrM)r rrkidx_maps r#rr6s FT 2C,.O OA 3&RSS9T$s)_=>?G 'cU??PPQ R   3<r"c\t}t|}tj|dS)N vram_used)rrr8amdsmi_get_gpu_vram_usager rs r#_get_amdsmi_memory_usagerDs* "F %f -F  + +F 3K @@r"ct}t|}tj|}tj|dS)N gfx_activity)rrr8r<amdsmi_get_gpu_activityrs r#_get_amdsmi_utilizationrJs= "F %f -F  0 0 26 :F  ) )& 1. AAr"ct|}tj|tjjtj j Sr\)rr8amdsmi_get_temp_metricAmdSmiTemperatureTypeJUNCTIONAmdSmiTemperatureMetricCURRENTrs r#_get_amdsmi_temperaturerQs@  (F  ( ($$--&&.. r"ct|}tj|d}|dk7r|Stj|dS)Naverage_socket_powerzN/Acurrent_socket_power)rr8amdsmi_get_power_info)r r socket_powers r#_get_amdsmi_power_drawrZsG  (F//78NOLu++F34JKKr"ct|}tj|tjj}d|vr|dS|dS)Ncur_clkclk)rr8amdsmi_get_clock_info AmdSmiClkTypeGFX)r r clock_infos r#_get_amdsmi_clock_ratercsH  (F--ff6J6J6N6NOJJ)$$%  r"ctjjsIt}t |}t j |}t j|jSt|S)aReturn the percent of time over the past sample period during which global (device) memory was being read or written as given by `nvidia-smi`. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). Warning: Each sample period may be between 1 second and 1/6 second, depending on the product being queried. ) r9r%rJrrsrrnvmlDeviceGetUtilizationRatesmemoryrrs r# memory_usagerlsU ==  $&'/226:33F;BBB'//r"ctjjsJt|}t |}t j |}t j|jSt|S)aReturn the percent of time over the past sample period during which one or more kernels was executing on the GPU as given by `nvidia-smi`. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). Warning: Each sample period may be between 1 second and 1/6 second, depending on the product being queried. ) r9r%rJrrsrrrgpurrs r# utilizationrsW ==  $V,'/226:33F;???&v..r"ctjjs!t|}t j |dSt |S)a Return the average temperature of the GPU sensor in Degrees C (Centigrades). The average temperature is computed based on past sample period as given by `nvidia-smi`. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). Warning: Each sample period may be between 1 second and 1/6 second, depending on the product being queried. r)r9r%rJrrnvmlDeviceGetTemperaturerrs r# temperaturers9 ==  $V,..vq99&v..r"ctjjs t|}t j |St |S)a Return the average power draw of the GPU sensor in mW (MilliWatts) over the past sample period as given by `nvidia-smi` for Fermi or newer fully supported devices. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). Warning: Each sample period may be between 1 second and 1/6 second, depending on the product being queried. )r9r%rJrrnvmlDeviceGetPowerUsagerrs r# power_drawrs7 ==  $V,--f55%f--r"ctjjs!t|}t j |dSt |S)aReturn the clock speed of the GPU SM in Hz Hertz over the past sample period as given by `nvidia-smi`. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). Warning: Each sample period may be between 1 second and 1/6 second, depending on the product being queried. r)r9r%rJrrnvmlDeviceGetClockInforrs r# clock_raters9 ==  $V,,,VQ77%f--r"ct|trtj|}|St|trtjd|}|S)zReturn the torch.device type object from the passed in device. Args: device (torch.device or int): selected device. rK)rrr9r rMr s r# _get_devicersD &#f% M FC ff- Mr"cl|j}| t}tjj|S)zvReturn the CUDA Generator object for the given device. Args: device (torch.device): selected device. )rrLr9rKr5)r rs r#_get_generatorrs/ ,,C { :: ( ( --r"offsetc@t|fd}t|y)a'Set the random number generator state offset of the specified GPU. Args: offset (int): The desired offset device (torch.device or int, optional): The device to set the RNG state. Default: ``'cuda'`` (i.e., ``torch.device('cuda')``, the current CUDA device). c>t}|jyr\)r set_offset)default_generator final_devicers r#cbz!_set_rng_state_offset..cbs*<8$$V,r"N)rr)rr rrs` @r#_set_rng_state_offsetrsv&L-rNr"cbtt|}t|}|jS)aPReturn the random number generator state offset of the specified GPU. Args: device (torch.device or int, optional): The device to return the RNG state offset of. Default: ``'cuda'`` (i.e., ``torch.device('cuda')``, the current CUDA device). .. warning:: This function eagerly initializes CUDA. )rrr get_offset)r rrs r#_get_rng_state_offsetrs-Lv&L&|4  ' ' ))r")*cDttt||g|i|Sr\)rr _CudaBase__new__clsargsrs r# _lazy_newrs%L C ( >t >v >>r"c*eZdZdZdZfdZeZxZS)rTFct|j5t| |i|cdddS#1swYyxYwr\)r rrr)rrrrs r#rz_CudaBase.type!s:DOO% & 17<00 1 1 1s4=) rrrr is_sparserrrrrs@r#rrsGI1Gr"r)_LegacyStorage_warn_typed_storage_removalcDeZdZedZedZeddddZy)_CudaLegacyStoragec,ttd)Nz+from_buffer: Not available for CUDA storage)rr-rs r# from_bufferz_CudaLegacyStorage.from_buffer/s#%HIIr"ctd)Nz2_new_with_weak_ptr: Not available for CUDA storager,rs r#_new_with_weak_ptrz%_CudaLegacyStorage._new_with_weak_ptr4sOPPr"N)r rWctd)Nz4_new_shared_filename: Not available for CUDA storager,)rmanagerrsizer rWs r#_new_shared_filenamez'_CudaLegacyStorage._new_shared_filename8sQRRr")rrr classmethodrrrr!r"r#rr.sGJJQQ@DDSSr"rc,eZdZedZedZy) ByteStoragec.t|jSr\r_dtypers r#rWzByteStorage.dtype>#%{{r"c"tjSr\)r9uint8rs r#rzByteStorage._dtypeC {{r"NrrrrrWrr!r"r#rr=(r"rc,eZdZedZedZy) DoubleStoragec.t|jSr\rrs r#rWzDoubleStorage.dtypeIrr"c"tjSr\)r9doublers r#rzDoubleStorage._dtypeN ||r"Nrr!r"r#r r H(r"r c,eZdZedZedZy) FloatStoragec.t|jSr\rrs r#rWzFloatStorage.dtypeTrr"c"tjSr\)r9floatrs r#rzFloatStorage._dtypeYrr"Nrr!r"r#rrSrr"rc,eZdZedZedZy) HalfStoragec.t|jSr\rrs r#rWzHalfStorage.dtype_rr"c"tjSr\)r9halfrs r#rzHalfStorage._dtyped zzr"Nrr!r"r#rr^(r"rc,eZdZedZedZy) LongStoragec.t|jSr\rrs r#rWzLongStorage.dtypejrr"c"tjSr\)r9longrs r#rzLongStorage._dtypeorr"Nrr!r"r#rrirr"rc,eZdZedZedZy) IntStoragec.t|jSr\rrs r#rWzIntStorage.dtypeurr"c"tjSr\)r9rMrs r#rzIntStorage._dtypezs yyr"Nrr!r"r#r"r"ts(r"r"c,eZdZedZedZy) ShortStoragec.t|jSr\rrs r#rWzShortStorage.dtyperr"c"tjSr\)r9shortrs r#rzShortStorage._dtyperr"Nrr!r"r#r&r&rr"r&c,eZdZedZedZy) CharStoragec.t|jSr\rrs r#rWzCharStorage.dtyperr"c"tjSr\)r9int8rs r#rzCharStorage._dtyperr"Nrr!r"r#r+r+rr"r+c,eZdZedZedZy) BoolStoragec.t|jSr\rrs r#rWzBoolStorage.dtyperr"c"tjSr\)r9boolrs r#rzBoolStorage._dtyperr"Nrr!r"r#r0r0rr"r0c,eZdZedZedZy)BFloat16Storagec.t|jSr\rrs r#rWzBFloat16Storage.dtyperr"c"tjSr\)r9rYrs r#rzBFloat16Storage._dtypes ~~r"Nrr!r"r#r5r5s(r"r5c,eZdZedZedZy)ComplexDoubleStoragec.t|jSr\rrs r#rWzComplexDoubleStorage.dtyperr"c"tjSr\)r9cdoublers r#rzComplexDoubleStorage._dtypes }}r"Nrr!r"r#r9r9s(r"r9c,eZdZedZedZy)ComplexFloatStoragec.t|jSr\rrs r#rWzComplexFloatStorage.dtyperr"c"tjSr\)r9cfloatrs r#rzComplexFloatStorage._dtyperr"Nrr!r"r#r>r>rr"r>ceZdZdZdZdZy)_WrappedTritonKernelzBJust a simple wrapper to store some metadata for testing purposes.c ||_d|_yr kernelkernel_invoked)rrFs r#rz_WrappedTritonKernel.__init__s #r"c8|j|i|}d|_|S)NTrE)rrrrs r#__call__z_WrappedTritonKernel.__call__s$dkk4*6*" r"N)rrrrrrIr!r"r#rCrCsL$r"rCc.tjrytd}td}tjj ddu}|rEtj jdd|dtj jdd|dyy) Nc"ddlm}||ddi|S)Nr) bsr_dense_mm skip_checksT)torch.sparse._triton_opsrL)rrrLs r# kernel_implz-_register_triton_kernels..kernel_impls9T>t>v>>r"c"ddlm}||ddi|S)Nr)bsr_dense_addmmrMT)rNrQ)rrrQs r#addmm_kernel_implz3_register_triton_kernels..addmm_kernel_impls<A$A&AAr"triton_triton_bsr_dense_mm_outzS_triton_bsr_dense_mm_out(Tensor bsr, Tensor dense, *, Tensor(a!) out) -> Tensor(a!) SparseCsrCUDA_triton_bsr_dense_addmm_outz_triton_bsr_dense_addmm_out(Tensor input, Tensor bsr, Tensor dense, *, Scalar beta, Scalar alpha, Tensor(a!) out) -> Tensor(a!))r9_running_with_deployrC importlibutil find_spec_TritonLibrary registerOp)rOrR has_tritons r#_register_triton_kernelsr^s !!#?? BB ))(34?J '' & a     '' )O   r")amp jiteratornvtxprofilersparsetunable)rr5BFloat16Tensorr0 BoolTensorr ByteTensorr+ CharTensorr9r>r  DoubleTensorr FloatTensorr HalfTensorr" IntTensorr LongTensorr& ShortTensorrrrrrrrr_caching_allocator_alloccaching_allocator_deleterrrrrrLrr5r r rCr empty_cacheget_allocator_backendCUDAPluggableAllocatorchange_current_allocatorrtrqrrrOr~ get_rng_stateget_rng_state_allrrrgraphsr3r4r initial_seedrrDrSrrr`list_gpu_processesr manual_seedmanual_seed_allmax_memory_allocatedmax_memory_cachedmax_memory_reserved mem_get_informemory_allocated memory_cachedmemory_reservedmemory_snapshot memory_statsmemory_stats_as_nested_dictmemory_summaryrMemPoolMemPoolContext use_mem_poolrrrncclrarbrandomreset_accumulated_memory_statsreset_max_memory_allocatedreset_max_memory_cachedreset_peak_memory_statsrrrset_per_process_memory_fraction set_rng_stateset_rng_state_allrrrcrstreamsrrdr)Tr\)rK)rrXr? threadingrru functoolsrtypingrrrrrr r r9torch._Cr _device torch._utilsr rr torch.typesrrr_utilsrrwrrrrrrrrrr ImportErrorrlocalrLockrrrrgetattrr:rrM _device_tr7rr%_versionrJrr8rerrrr8r&r'r.r0r2r3r3 _has_magmar4r5 Generatorr;rArDrSrQr_rfrrrrrZrOutOfMemoryErrorrrrrr-rrrrrrrrqrOrrrrrr5r>rKrTr`rgrnrprsrtrCrtr~rLrrrr rrrrrrrrrrrrrrrrrrrrrr staticmethodrr torch.storagerrrrr rrrr"r&r+r0r5r9r>_storage_classesaddrCr^r_r`rarbrcrd__all__r!r"r#rs   DDD #EE%32  y%y~~'t (2t8 d3i '(%(($7G '3T) *   ) ||    &' 588,-!HH::'(?@ 588+,xx44HHH  58801"XX??HsHsH $((%% 4%02E%((,,-25d5 =4= 3d 3 040< 2!sB42 G"  = I 88,, 4n7t , , ST  (   )y )T ) .HY/ .3 ."(9"5"sCx""*)*0E*" C9 C9 C C5454p !8/0 !] ! v$3d3ic&: ;3l # $$s)!4@xS 2@DId3iDQTI6 c , C D 8E#v+,>#? C &'+hsm* c $tCy 3 && , ,T , (8I.&$8I.&$1 3E#s(O3360S0 vs{); <" vs{); <  XeCK.@%A c AXeFCK.@%AASA BHU63;-?$@BCBHU63;-?$@CL8E&#+,>#?L3L!8E&#+,>#?!3!0% "4500*/vs{!34//*/vs{!34//*.xfck 23.s.&.xfck 23.s.$ c3 45 %,,  .5<< .EHH,>,> .:@ sC56 & *%S%,,(>"? *S * ??  F S S$&%$$#%$$(-,=)<(;':&<(;';';';'?+/0./  ! H #$>=t .G0     KsZ3\2]]5]2\=<\=] ] ]  ] ]]]$]]$