o ,hF@sUdZddlZddlZddlmZddlmZmZmZm Z ddl Z ddl Z ddl m Z ddlmZmZddlmZdd lmZmZd aeZeZgaeeegdfeefed <e e j!d d dZ"e e ee#dfZ$ea%dZ&ee j!j'ed<de(fddZ)e)re j!j*Z*e j!j+Z,e j!j-Z.nedZ*de#de#fddZ,de#de#fddZ.eddde#fddZ/de(fddZ0d[d e(de(fd!d"Z1d#d$Z2d%d&Z3d'd(Z4d)d*Z5Gd+d,d,Z6Gd-ddZ Gd.d/d/e Z7de$ddfd0d1Z8d\dee$defd2d3Z9edd\dee$de:eeffd4d5Z;d\dee$de*fd6d7ZGded?de?fd@d>Z@dAdBZAd>efdCdDZBd\dee$defdEdFZC d\dGe#dee$defdHdIZDd\de$ddfdJdKZEdeefdLdMZFdefdNdOZGde j de j!j'fdPdQZH Rd]dSe#de e#ee j fddfdTdUZId]de e#ee j fde#fdVdWZJddXlKmLZLmMZMmNZNmOZOmPZPmQZQmRZRmSZSmTZTmUZUddYlVmWZWmXZXmYZYmZZZm[Z[m\Z\m]Z]m^Z^m_Z_gdZZ`dS)^z 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) lru_cache)AnyCallableOptionalUniondevice) _dummy_type_LazySeedTracker)_get_device_index)EventStreamF _queued_calls_xpu_isInBadForkcCsdSNFrrrM/var/www/html/scripts/venv/lib/python3.10/site-packages/torch/xpu/__init__.pysrrdefault_generatorsreturncCstjjS)z(Return true if compile with XPU support.)torch_C_has_xpurrrr _is_compiled#sr_XpuDevicePropertiesrcCtdNz(PyTorch was compiled without XPU supportNotImplementedErrorrrrr_exchange_device0r cCrrrrrrr_maybe_exchange_device3r!r")maxsizecCstsdStjS)z*Return the number of XPU device available.r)rrr_xpu_getDeviceCountrrrr device_count7s r%cCs tdkS)z7Return a bool indicating if XPU is currently available.r)r%rrrr is_available?s r&Tincluding_emulationcCstsdS|p tjjS)zKReturn a bool indicating if the current XPU device supports dtype bfloat16.F)r&rxpuget_device_propertieshas_bfloat16_conversions)r'rrris_bf16_supportedEs  r+cCs tot S)z8Return whether PyTorch's XPU state has been initialized.) _initialized_is_in_bad_forkrrrris_initializedOs r.cKsftr|dS|ddrt|tdS|ddr(t|tdSt|tfdS)Nseed_allFseed) r.get_lazy_seed_trackerqueue_seed_all traceback format_stack queue_seedrappend)callablekwargsrrr _lazy_callTs   r:cCs 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_initrrrrinitbs r<c Cststtdr dStqtr WddStr tdts'tdtj dt_ t ddtDz1t D]'\}}z|Wq>tye}zdt|dd |}t||d}~wwWttdnttdwdaWddS1swYdS) 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 enabledTcss|]}|r|VqdSNr).0callsrrr sz_lazy_init..z5XPU call failed lazily at initialization with error: z' XPU call was originally invoked at: )r.hasattr_tls_initialization_lockr- RuntimeErrorrAssertionErrorrr _xpu_initr=rextendr2 get_calls Exceptionstrjoindelattrr,) queued_callorig_tracebackemsgrrrr;ks>      "r;c@s8eZdZdefddZddZdededefd d Zd S) _DeviceGuardindexcCs||_d|_dSN)idxprev_idx)selfrTrrr__init__s z_DeviceGuard.__init__cCtj|j|_dSr>rr(r rWrXrYrrr __enter__z_DeviceGuard.__enter__typevaluer4cCtj|j|_dSrrr(r"rXrWrYr`rar4rrr__exit__z_DeviceGuard.__exit__N)__name__ __module__ __qualname__intrZr^rrerrrrrSsrSc@s<eZdZdZdefddZddZdededefd d Zd S) rzContext-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``. cCst|dd|_d|_dS)NToptionalrV)r rWrX)rYrrrrrZs zdevice.__init__cCr[r>r\r]rrrr^r_zdevice.__enter__r`rar4cCrbrrcrdrrrrerfzdevice.__exit__N)rgrhri__doc__rrZr^rerrrrrs cs eZdZdZfddZZS) 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. cs"|jr|nd}t|dSrU)is_xpu get_devicesuperrZ)rYobjrW __class__rrrZszdevice_of.__init__)rgrhrirmrZ __classcell__rrrsrrns rncCs*tt|}|dkrtj|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)r;r rr_xpu_setDevicerrrr set_devices rwcCs 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 )r)namerrrrget_device_names rycst|fddtDS)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 cs"i|] }|ds|t|qS))__ _pybind11_) startswithgetattr)r?proppropsrr s  z)get_device_capability..)r)dirrrrrget_device_capabilitys rcCstt|dd}t|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: _XpuDeviceProperties: the properties of the device Trk)r;r _get_device_propertiesrrrrr)s r)cCsttjS)z0Return the index of a currently selected device.)r;rr_xpu_getDevicerrrrcurrent_device s rcCs2t|tr t|}|St|trtd|}|S)zReturn the torch.device type object from the passed in device. Args: device (torch.device or int or str): selected device. r() isinstancerLrrrjrrrr _get_devices   rc@sNeZdZUdZeded<dedfddZddZd ed ed efd d 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_streamstreamcCs*||_tdd|_|jdurd|_dSdS)NTrV)rr rW)rYrrrrrZ)s    zStreamContext.__init__cCs|j}|dus |jdkrdStjd|_|jj|jkr9t|jtj|j|_Wdn1s4wYtj|dSrU) rrWrr(current_streamsrc_prev_streamrdst_prev_stream set_stream)rYrrrrr^/s zStreamContext.__enter__r`rar4cCsJ|j}|dus |jdkrdS|jj|jkrtj|jtj|jdSrU)rrWrrrr(rr)rYr`rar4rrrrre;s zStreamContext.__exit__N) rgrhrirmr__annotations__rZr^rrerrrrrs   rrrcCst|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``. )rrrrrrFscCstjj|||ddS)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_setStreamrrrr_set_stream_by_idOs  rcCs*|durdStt|j|j|jddS)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)r;rrrrrrrrr]s  rcCs4ttjt|dd}t|d|d|ddS)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). Trkrr r)r;rr_xpu_getCurrentStreamr r)r streamdatarrrrps  rdata_ptrcCs6ttj|t|dd}t|d|d|ddS)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. Trkrr rr)r;rr_xpu_getStreamFromExternalr r)rrrrrrget_stream_from_externals rcCstt|dd}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). Trk)r;r rr_xpu_synchronizerrrr synchronizes  rcCs(tsgStj}|durgS|S)zr)r?archrrrrAsz$get_gencode_flags..)rlenrM) arch_listrrrget_gencode_flagss rcCs |j}|dur t}tjj|S)zuReturn the XPU Generator object for the given device. Args: device (torch.device): selected device. N)rTrrr(r)rrWrrr_get_generators rr(offsetcs"t|fdd}t|dS)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: ``'xpu'`` (i.e., ``torch.device('xpu')``, the current XPU device). cst}|dSr>)r set_offset)default_generator final_devicerrrcbsz!_set_rng_state_offset..cbN)rr:)rrrrrr_set_rng_state_offsets  rcCstt|}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. )r;rr get_offset)rrrrrr_get_rng_state_offsets r) empty_cachemax_memory_allocatedmax_memory_reserved mem_get_infomemory_allocatedmemory_reserved memory_statsmemory_stats_as_nested_dictreset_accumulated_memory_statsreset_peak_memory_stats) get_rng_stateget_rng_state_all initial_seed manual_seedmanual_seed_allr0r/ set_rng_stateset_rng_state_all)+r rrrrrrrnr%rrrryr)rrrrr<rr&r+r.rrrrrrrrrrrr0r/rwrrrrstreamsr)Tr>)r()arm threadingr4 functoolsrtypingrrrrrtorch._Cr_device torch._utilsr r _utilsr rr rr,localrDLockrErlisttuplerLrr}rr-rj _device_tr2r Generatorboolrr_xpu_exchangeDevicer _xpu_maybeExchangeDevicer"r%r&r+r.r:r<r;rSrnrwrydictrr)rrrrrrrrrrrrrrmemoryrrrrrrrrrrrandomrrrrrr0r/rr__all__rrrrs      (  "  *      0,