/jTUdZddlZddlZddlmZddlmZddlmZm Z m Z ddl Z ddl Z ddl mZmZddlmZdd lmZdd lmZmZmZmZmZdd lmZmZd aejZej Z!ga"e#e$egdfe#e%fe&d <e'e j(ddZ)ea*dZ+e$e j(j,e&d<de-fdZ.e.r%e j(j/Z/e j(j0Z1e j(j2Z3nedZ/de4de4fdZ1de4de4fdZ3edde4fdZ5de-fdZ6dFde-de-fdZ7de-fdZ8dZ9dGd Z:dGd!Z;dGd"Z<Gd#d$Z=Gd%dZ>Gd&d'e>Z?deddfd(Z@dHdede%fd)ZAeddHdedeBe%effd*ZC dHdede/fd+ZDde4fd,ZEde4e%ze j>zde j>fd-ZFded.ede-fd/ZGGd0d1ZHd2e d3deHfd4ZIdGd5ZJd2eddfd6ZKdHdedefd7ZLdHd8e4dedefd9ZMdHdeddfd:ZNde#e%fd;ZOde%fd<ZPde j>de j(j,fd=ZQ dId?e4de4e%ze j>zddfd@ZRdIde4e%ze j>zde4fdAZSddBlTmUZUmVZVmWZWmXZXmYZYmZZZm[Z[m\Z\m]Z]m^Z^m_Z_m`Z`maZambZbmcZcmdZdmeZeddClfmgZgmhZhmiZimjZjmkZkmlZlmmZmmnZnmoZoe dDe$e4e4fZpgdEZqdS)Jz This package introduces support for the XPU backend, specifically tailored for Intel GPU optimization. This package is lazily initialized, so you can always import it, and use :func:`is_available()` to determine if your system supports XPU. N)Callable) lru_cache)AnyNewTypeOptional) _dummy_type_LazySeedTracker)Device)_get_device_index)graphgraph_pool_handleis_current_stream_capturingmake_graphed_callablesXPUGraph)EventStreamF _queued_calls_xpu_isInBadForkcdSNFrW/home/longshao/multi-rider-rag/.venv/lib/python3.11/site-packages/torch/xpu/__init__.pyr&srrdefault_generatorsreturnc$tjjS)z(Return true if compile with XPU support.)torch_C_has_xpurrr _is_compiledr"+s 8 r_XpuDevicePropertiesdevicec tdNz(PyTorch was compiled without XPU supportNotImplementedErrorr$s r_exchange_devicer*8!"LMMMrc tdr&r'r)s r_maybe_exchange_devicer-;r+r)maxsizec^tsdStjS)z*Return the number of XPU device available.r)r"rr _xpu_getDeviceCountrrr device_countr1?s) >>q 8 ' ' ) ))rc&tdkS)z7Return a bool indicating if XPU is currently available.r)r1rrr is_availabler3Gs >>A rTincluding_emulationcltsdS|p"tjjS)zKReturn a bool indicating if the current XPU device supports dtype bfloat16.F)r3rxpuget_device_propertieshas_bfloat16_conversions)r4s ris_bf16_supportedr9Ms5 >>u F 9 * * , , ErchtsdStjjS)zGReturn a bool indicating if the current XPU device supports dtype tf32.F)r3rr6r7'has_subgroup_matrix_multiply_accumulaterrris_tf32_supportedr<Ws, >>u 9 * * , , TTrc.tot S)z8Return whether PyTorch's XPU state has been initialized.) _initialized_is_in_bad_forkrrris_initializedr@bs  1 1 111rc tr |dS|ddr.t|t jdS|ddr.t|t jdSt|t jfdS)Nseed_allFseed) r@get_lazy_seed_trackerqueue_seed_all traceback format_stack queue_seedrappend)callablekwargss r _lazy_callrMgs G  ::j% ( ( G  - -h 8N8P8P Q Q Q Q Q ZZ & & G  ) )(I4J4L4L M M M M M  (I,B,D,D!E F F F F Frc"tdS)zInitialize PyTorch's XPU state. This is a Python API about lazy initialization that avoids initializing XPU until the first time it is accessed. Does nothing if the XPU state is already initialized. N) _lazy_initrrrinitrPus LLLLLrc 2tsttdrdSt5tr ddddSt rt dt stdtj dt_ t dtD tD]Z\}} |#t $r=}dt#|dd|}t!||d}~wwxYw t'tdn#t'tdwxYwdaddddS#1swxYwYdS) Nis_initializingzuCannot re-initialize XPU in forked subprocess. To use XPU with multiprocessing, you must use the 'spawn' start methodz#Torch not compiled with XPU enabledTc3K|]}||V dSNr).0callss r z_lazy_init..s(XXuRWXUXXXXXXrz5XPU call failed lazily at initialization with error: z' XPU call was originally invoked at: )r@hasattr_tls_initialization_lockr? RuntimeErrorr"AssertionErrorrr _xpu_initrRrextendrE get_calls Exceptionstrjoindelattrr>) queued_callorig_tracebackemsgs rrOrO~s074):;; !!      !!!!!!!!    I ~~ H !FGG G  $XX0B0L0L0N0NXXXXXX -/< 0 0+ ^0KMMMM 000\PSTUPVPV\\BD''.BYBY\\$C..a/ 0 0 D+ , , , ,GD+ , , , , C!!!!!!!!!!!!!!!!!!sTF  BF * E#7 DE# E  8EE  E# F #E::F  FFc8eZdZdeddfdZdZdededefd ZdS) _DeviceGuardindexrNc"||_d|_dSN)idxprev_idx)selfrks r__init__z_DeviceGuard.__init__s rcXtj|j|_dSrTrr6r*rorprqs r __enter__z_DeviceGuard.__enter__ 2248<< rtypevaluerGcXtj|j|_dSrrr6r-rprorqrxryrGs r__exit__z_DeviceGuard.__exit__ 933DMBBur)__name__ __module__ __qualname__intrrrvrr}rrrrjrjsncd===Srrjc<eZdZdZdeddfdZdZdededefd ZdS) r$zContext-manager that changes the selected device. Args: device (torch.device or int or str): device index to select. It's a no-op if this argument is a negative integer or ``None``. rNc@t|d|_d|_dS)NToptionalrn)r rorp)rqr$s rrrzdevice.__init__s!$Vd;;; rcXtj|j|_dSrTrtrus rrvzdevice.__enter__rwrrxryrGcXtj|j|_dSrr{r|s rr}zdevice.__exit__r~r)rrr__doc__rrrrvr}rrrr$r$sxst===Src$eZdZdZdfd ZxZS) 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 XPU, this is a no-op. Args: obj (Tensor or Storage): object allocated on the selected device. rNc|jr|nd}t|dSrm)is_xpu get_devicesuperrr)rqobjro __class__s rrrzdevice_of.__init__s<"%*4cnn" rrN)rrrrrr __classcell__)rs@rrrsGrrctt|}|dkr!tj|dSdS)zSet the current device. Args: device (torch.device or int or str): selected device. This function is a no-op if this argument is negative. rN)rOr rr _xpu_setDevicer)s r set_devicersFLLL v & &F {{ '''''{rc*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.xpu.current_device`, if :attr:`device` is ``None`` (default). Returns: str: the name of the device )r7namer)s rget_device_namers ! ( ( --rct|ttttt dt ttffdtDS)aGet the xpu 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.xpu.current_device`, if :attr:`device` is ``None`` (default). Returns: dict[str, Any]: the xpu capability dictionary of the device Nci|]:}|dtt|x7|;S)__) startswith isinstancegetattr)rUkeypropsserializable_typesrys r z)get_device_capability..sa    ~~d##  !4!447I J J   U   r) r7rfloatboolrbrxlisttupledictdir)r$rrrys @@@rget_device_capabilityrsl "& ) )EudCdT5$O      u::   rc^tt|d}t|S)aGet the properties of a device. Returns _XpuDeviceProperties containing the following device properties: - ``name`` (str): device name. - ``platform_name`` (str): SYCL platform name. - ``vendor`` (str): device vendor. - ``device_id`` (int): device identifier (product ID). - ``driver_version`` (str): driver version. - ``version`` (str): runtime version. - ``max_compute_units`` (int): number of parallel compute units. - ``gpu_eu_count`` (int): number of EUs (Execution Unit). - ``max_work_group_size``: (int): maximum number of work-items permitted in a work-group. - ``max_num_sub_groups`` (int): maximum number of sub-groups supported in a work-group. - ``sub_group_sizes``: (list[int]): a list of supported sub-group sizes. - ``local_mem_size`` (int): device local memory capacity that can be allocated per work-group in bytes. - ``has_fp16`` (bool): whether float16 dtype is supported. - ``has_fp64`` (bool): whether float64 dtype is supported. - ``has_atomic64`` (bool): whether 64-bit atomic operations are supported. - ``has_bfloat16_conversions`` (bool): whether bfloat16 conversions are supported. - ``has_subgroup_matrix_multiply_accumulate`` (bool): whether DPAS (Dot Product Accumulate Systolic) is supported. - ``has_subgroup_matrix_multiply_accumulate_tensor_float32`` (bool): whether DPAS with tf32 inputs is supported. - ``has_subgroup_2d_block_io`` (bool): whether 2D block I/O for efficient matrix multiplication is supported. - ``total_memory`` (int): device global memory in bytes. - ``gpu_subslice_count`` (int): number of subslice. - ``architecture`` (int): device architecture identifier (experimental). - ``type`` (str): device type, e.g. 'cpu', 'gpu', accelerator', 'host', 'unknown'. - ``uuid`` (Any): device UUID (Universal Unique ID), 16 bytes. Args: device (torch.device or int or str): device for which to return the properties of the device. Returns: _XpuDeviceProperties: the properties of the device Tr)rOr _get_device_propertiesr)s rr7r7 s.JLLL v 5 5 5F !& ) ))rcZttjS)z0Return the index of a currently selected device.)rOrr _xpu_getDevicerrrcurrent_devicer5sLLL 8 " " $ $$rct|trtj|}n*t|trtjd|}|S)zReturn the torch.device type object from the passed in device. Args: device (torch.device or int or str): selected device. r6)rrbrr$rr)s r _get_devicer;sO &#-f%% FC -eV,, Mrpeerctt|d}t|d}tj||S)a/Query whether a device can access a peer device's memory. Args: device (torch.device or int or str): selected device. peer (torch.device or int or str): peer device to query access to. Returns: bool: ``True`` if ``device`` can access ``peer``, ``False`` otherwise. Tr)rOr rr _xpu_canDeviceAccessPeer)r$rs rcan_device_access_peerrHsILLL v 5 5 5F TD 1 1 1D 8 , ,VT : ::rc`eZdZUdZeded<dedddfdZdZd ed ed efd Z dS) StreamContextaContext-manager that selects a given stream. All XPU 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.xpu.Stream cur_streamstreamrNc^||_tdd|_|j d|_dSdS)NTrn)rr ro)rqrs rrrzStreamContext.__init__fs5 $T400 8 DHHH  rc|j}| |jdkrdStjd|_|jj|jkrUt |j5tj|j|_dddn #1swxYwYtj|dSrm) rrorr6current_streamsrc_prev_streamr$dst_prev_stream set_stream)rqrs rrvzStreamContext.__enter__ls[  R F$y77==   &**; ; ; )** S S',y'?'? @Q'R'R$ S S S S S S S S S S S S S S S Z(((((s$*BB!BrxryrGc|j}| |jdkrdS|jj|jkr$tj|jtj|jdSrm)rrorr$rr6rr)rqrxryrGrs rr}zStreamContext.__exit__xsl[  R F   &**; ; ; I !5 6 6 6 T122222r) rrrrr__annotations__rrrvrr}rrrrrXs  +,,,,x(:; ) ) )3S33333333rrrrc t|S)zWrap 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``. )rrs rrrs   rcJtj|||dS)aset stream specified by the stream id, device index and device type Args: stream_id (int): not visible to the user, used to assigned to the specific stream. device_index (int): selected device index. device_type (int): selected device type.  stream_id device_index device_typeN)rr _xpu_setStreamrs r_set_stream_by_idrs6 H!rcl|dStt|j|j|jdS)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)rOrrrrrs rrrsJ~LLL"(&rcttjt |d}t |d|d|dS)aRReturn 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.xpu.current_device`, if :attr:`device` is ``None`` (default). Trrr r)rOrr _xpu_getCurrentStreamr r)r$ streamdatas rrrsaLLL//&4000J Q-jmTU   rdata_ptrcttj|t |d}t |d|d|dS)a;Return a :class:`Stream` from an external SYCL queue. This function is used to wrap SYCL queue created in other libraries in order to facilitate data exchange and multi-library interactions. .. note:: This function doesn't manage the queue life-cycle, it is the user responsibility to keep the referenced queue alive while this returned stream is being used. The different SYCL queue pointers will result in distinct :class:`Stream` objects, even if the SYCL queues they dereference are equivalent. Args: data_ptr(int): Integer representation of the `sycl::queue*` value passed externally. device(torch.device or int, optional): the device where the queue was originally created. It is the user responsibility to ensure the device is specified correctly. Trrr rr)rOrr _xpu_getStreamFromExternalr r)rr$rs rget_stream_from_externalrsd LLL44#FT:::J Q-jmTU   rc~tt|d}tj|S)a*Wait for all kernels in all streams on a XPU device to complete. Args: device (torch.device or int, optional): device for which to synchronize. It uses the current device, given by :func:`~torch.xpu.current_device`, if :attr:`device` is ``None`` (default). Tr)rOr rr _xpu_synchronizer)s r synchronizers5LLL v 5 5 5F 8 $ $V , ,,rctsgStj}|gS|S)z> ++--J     rct}t|dkrdSddd|DS)zIReturn XPU AOT(ahead-of-time) build flags this library was compiled with.rrXz-device ,c3K|]}|VdSrTr)rUarchs rrWz$get_gencode_flags..s"::t::::::r)rlenrc) arch_lists rget_gencode_flagsrsJI 9~~r .cbs+*<88$$V,,,,,rr)rrM)rr$rrs` @r_set_rng_state_offsetrsGv&&L-------rNNNNNrctt|}t|}|S)aLReturn the random number generator state offset of the specified GPU. Args: device (torch.device or int, optional): The device to return the RNG state offset of. Default: ``'xpu'`` (i.e., ``torch.device('xpu')``, the current XPU device). .. warning:: This function eagerly initializes XPU. )rOrr get_offset)r$rrs r_get_rng_state_offsetrs:LLLv&&L&|44  ' ' ) ))r)change_current_allocator empty_cacheget_per_process_memory_fractionmax_memory_allocatedmax_memory_reserved mem_get_infomemory_allocatedmemory_reservedmemory_snapshot memory_statsmemory_stats_as_nested_dictMemPoolreset_accumulated_memory_statsreset_peak_memory_statsset_per_process_memory_fraction use_mem_poolXPUPluggableAllocator) get_rng_stateget_rng_state_all initial_seed manual_seedmanual_seed_allrCrB set_rng_stateset_rng_state_all _POOL_HANDLE)9rrrrrrrrrrr$rr1rrrrr7rrrrrr rrPrr3r9rr@r<rrrrrrrrrrrrrrrrCrBrrrrrrstreamsr)TrrT)r6)rr threadingrGcollections.abcr functoolsrtypingrrrrtorch._C torch._utilsrr torch.typesr _utilsr graphsr rrrrr rrr>localrZLockr[rrrrbrrr r?rEr Generatorrr"r#_xpu_exchangeDevicer*_xpu_maybeExchangeDevicer-rr1r3r9r<r@rMrPrOrjr$rrrrrr7rrrrrrrrrrrrrrrmemoryrrrrrrrrrrrrrrrrrrandomrrrrrrCrBrrr __all__rrrrs9$$$$$$)))))))))) 66666666%%%%%%#""""""" y%y~''t (2t8 d3i '('%($6 FF%%''02E%(,-222d <>> N 88x3"X>';'=>>NNNNNNNsNsNNNN 1*c****d 44U4UUUU222 G G G G%%%%P        (         (v ($ ( ( ( ( . .F .c . . . . 4&DcN2'*'* '*'*'*'*'*T%%%%% c EL0 U\     ;6 ; ;D ; ; ; ; (3(3(3(3(3(3(3(3V!8./!M!!!!    v$&6V$sFf2 - - -$ - - - -tCy=3==== -5< -EH,> - - - -5: sU\1 & * *#)el": *s * * * *"&                      w~uS#X77 : : : r