js3UddlZddlmZddlmZmZmZmZmZm Z m Z m Z ddl Z eee ge fed<e jrddlmZndZe e jee dfZiZeeefed<iZeee fed<Gdd ZGd d ZGd d eZeee ge fed<e jrddlmZndZGddeZeGddZ GddeZ!iZ"eee efed<da#de ee jfde efdZ$de ee jfde efdZ%dee ee effdZ&dZ'dS)N) dataclass)AnyCallableDictIterableOptionalTupleTypeUnionget_cuda_stream)_cuda_getCurrentRawStream caching_worker_device_propertiescaching_worker_current_devicesc^eZdZdZGddZGddZGddZGdd Zed Z ede fd Z ede d e fd Z ede d e fdZedZed efdZedejfdZedZedejfdZede de de fdZede d e fdZed#de fdZed#de fdZed#de fdZed$d efd!Zed#de d e fd"ZdS)%DeviceInterfacez This is a simple device runtime interface for Inductor. It enables custom backends to be integrated with Inductor in a device-agnostic semantic. ceZdZdefdZdS)DeviceInterface.devicedevicectNNotImplementedErrorclsrs OC:\Projects\Historical\venv\Lib\site-packages\torch/_dynamo/device_interface.py__new__zDeviceInterface.device.__new__s% %N)__name__ __module__ __qualname__ _device_trrrrrs/ & & & & & & &rrceZdZdZdS)DeviceInterface.Eventc td)NzYEvent should be inherited from torch.Event, otherwise, it couldn't be captured by dynamo.rrargskwargss rrzDeviceInterface.Event.__new__!s%k rNrrr rr"rrEventr$ #     rr*ceZdZdZdS)DeviceInterface.Streamc td)Nz[Stream should be inherited from torch.Stream, otherwise, it couldn't be captured by dynamo.rr&s rrzDeviceInterface.Stream.__new__'s%m rNr)r"rrStreamr-&r+rr/cheZdZdZedefdZedefdZeddefdZ dS) DeviceInterface.Workera Worker API to query device properties that will work in multi processing workers that cannot use the GPU APIs (due to processing fork() and initialization time issues). Properties are recorded in the main process before we fork the workers. rctrrrs r set_devicez!DeviceInterface.Worker.set_device4% %rreturnctrrr"rrcurrent_devicez%DeviceInterface.Worker.current_device8r5rNctrrr3s rget_device_propertiesz,DeviceInterface.Worker.get_device_properties<r5rr) rrr __doc__ staticmethodintr4r8r!r:r"rrWorkerr1,s    &s & & &  &  & & & &  &  & &) & & &  & & &rr>ctrrr"rrr8zDeviceInterface.current_device@!!rctrrr3s rr4zDeviceInterface.set_deviceDr@rr6ctrrr3s rmaybe_exchange_devicez%DeviceInterface.maybe_exchange_deviceHr@rctrrr3s rexchange_devicezDeviceInterface.exchange_deviceLr@rctrrr"rr device_countzDeviceInterface.device_countPr@rctrrr"rr is_availablezDeviceInterface.is_availableTr@rstreamctrrrJs rrJzDeviceInterface.streamXr@rctrrr"rrcurrent_streamzDeviceInterface.current_stream\r@rctrrrLs r set_streamzDeviceInterface.set_stream`r@r stream_id device_index device_typectrr)rQrRrSs r_set_stream_by_idz!DeviceInterface._set_stream_by_iddr@r device_idxctrrrVs rget_raw_streamzDeviceInterface.get_raw_streamhr@rNctrrr3s r synchronizezDeviceInterface.synchronizelr@rc6|j|Sr)r>r:rs rr:z%DeviceInterface.get_device_propertiespsz//777rctrrr3s rget_compute_capabilityz&DeviceInterface.get_compute_capabilitytr@rFincluding_emulationctrrr_s ris_bf16_supportedz!DeviceInterface.is_bf16_supportedxr@rctrrr3s rmemory_allocatedz DeviceInterface.memory_allocated|r@rrF)rrr r;rr*r/r>r<r8r!r4r=rCrErGboolrItorchrJrNrPrUrYr[ classmethodr:r^rbrdr"rrrrsQ &&&&&&&&  &&&&&&&&(""\""9"""\""c"c"""\""""""\"""\""$"""\""u|"""\"""\""5<"""\""S""#"""\""3"3"""\"""I"""\"889888[8""y"""\"""t"""\""""c"""\"""rrcXeZdZdZdeedeeddfdZdZ de d e d e fd Z dS) DeviceGuarda_ This class provides a context manager for device switching. This is a stripped down version of torch.{device_name}.device. The context manager changes the current device to the given device index on entering the context and restores the original device on exiting. The device is switched using the provided device interface. device_interfaceindexr6Nc0||_||_d|_dS)N)rkidxprev_idx)selfrkrls r__init__zDeviceGuard.__init__s!1 rc`|j&|j|j|_dSdSr)rorkrErp)rqs r __enter__zDeviceGuard.__enter__s0 8  1AA$(KKDMMM rtypevalue tracebackc\|j$|j|j|_dS)NF)rorkrCrp)rqrurvrws r__exit__zDeviceGuard.__exit__s) 8 ,BB4=QQDHur) rrr r;r rrr=rrrtrryr"rrrjrjs $_ 5>Fsm LLLSrrjceZdZejjZejjZejjZGddZe ejj Z e ejj Z e ejj Z e ejj Z e ejjZe ejjZe ejjZe ejjZe ejjZe eZe ejjZe ejjZe ejjZe ejjZe defdZe ddefdZdS) CudaInterfacecdeZdZedefdZedefdZeddefdZdS)CudaInterface.Workerrc|td<dSNcudarr3s rr4zCudaInterface.Worker.set_devices5; *6 2 2 2rr6cjdtvr tdStjSr)rrgrr8r"rrr8z#CudaInterface.Worker.current_devices-7775f==:,,.. .rNc|Wt|tr!tj|}|jdksJt|tjr|j}|t j}dtvr?dttj D}|td<td|S)NrcLg|]!}tj|"Sr")rgrr:.0is r z>CudaInterface.Worker.get_device_properties..s8J44Q77r) isinstancestrrgrrurlr{r>r8rrangerrGr device_props rr:z*CudaInterface.Worker.get_device_propertiess!fc**1"\&11F!;&0000fel33*#\F~&-<<>>==="5:#:#:#<#<== r}s  r6c>tjSr)rgrrIr"rrrIzCudaInterface.is_availablesz&&(((rNrctjj*tj|\}}|dz|zStj|jdddS)N :r)rgversionhiprget_device_capabilityr: gcnArchNamesplit)rmajormins rr^z$CudaInterface.get_compute_capabilityse =  $99&AAJE32:# #:33F;;GMMcSTUUVWX Xrr)rrr rgrrr*r/r>r<r8r4rGrJrNrPrUr[r:r rY_exchange_devicerE_maybe_exchange_devicerCrdrbrfrIr!r^r"rrr{r{s Z F J E Z FDDDDDDDD>"\%*";<??O(L)JKK#|EJ$?@@$ UZ%ABB)$)))\)YYyYYY\YYYrr{get_xpu_stream)_xpu_getCurrentRawStreamceZdZejjZejjZejjZGddZe ejj Z e ejj Z e ejj Z e ejj Z e ejjZe ejjZe ejjZe ejjZe ejjZe eZe ejjZe ejjZe ejjZe defdZe d defdZe d d edefd ZdS) XpuInterfacecdeZdZedefdZedefdZeddefdZdS)XpuInterface.Workerrc|td<dSNxpurr3s rr4zXpuInterface.Worker.set_devices4: *5 1 1 1rr6cjdtvr tdStjSr)rrgrr8r"rrr8z"XpuInterface.Worker.current_devices-6665e<<9++-- -rNc|Wt|tr!tj|}|jdksJt|tjr|j}|t j}dtvr?dttj D}|td<td|S)NrcLg|]!}tj|"Sr")rgrr:rs rrz=XpuInterface.Worker.get_device_properties..s8I33A66r) rrrgrrurlrr>r8rrrrGrs rr:z)XpuInterface.Worker.get_device_propertiess!fc**0"\&11F!;%////fel33*#\F~%,;;==<<<"59#9#9#;#;<< ;F073E:6B Brrrr"rrr>rs  ;s ; ; ;  ;  . . . .  .  C C) C C C  C C Crr>r6c>tjSr)rgrrIr"rrrIzXpuInterface.is_availablesy%%'''rNrcDtj|}|Sr)rgrr)rccs rr^z#XpuInterface.get_compute_capabilitys Y , ,V 4 4 rFr_c>tjSr)rgrrbras rrbzXpuInterface.is_bf16_supported#sy**,,,rrre)rrr rgrrr*r/r>r<r8r4rGrJrNrPrUr[r:rrYrrErrCrdrfrIr!r^rbr"rrrrs Y F IOE Y FCCCCCCCC>"\%)":;;Nei233J< 677L \%)* + +F!\%)":;;Nei233J$ UY%@AA,uy455K(L)HII!\.11N"l59#=>>O(L)IJJ#|EI$>??($(((\(y\--t----\---rrceZdZUeed<dS)CpuDevicePropertiesmulti_processor_countN)rrr r=__annotations__r"rrrr(srrceZdZGddejZedefdZed dede fdZ ede fdZ ed Z ed defd ZGd d ZdS) CpuInterfacec*eZdZddZdefdZddZdS) CpuInterface.EventTcd|_dS)Ngtime)rq enable_timings rrrzCpuInterface.Event.__init__/s DIIIrr6c&|j|jz dzS)Nir)rq end_events r elapsed_timezCpuInterface.Event.elapsed_time2sNTY.$6 6rNc6tj|_dSr)r perf_counter)rqrJs rrecordzCpuInterface.Event.record5s)++DIIIr)Tr)rrr rrfloatrrr"rrr*r.sW     7U 7 7 7 7 , , , , , ,rr*r6cdS)NTr"r"rrrIzCpuInterface.is_available8strNrcdS)Nr"r3s rr^z#CpuInterface.get_compute_capability<srrcdSNrr"rXs rrYzCpuInterface.get_raw_stream@qrcdSrr"r"rrr8zCpuInterface.current_deviceDrrcdSrr"r3s rr[zCpuInterface.synchronizeHs rc,eZdZeddefdZdS)CpuInterface.WorkerNrcPddl}|}t|Sr)multiprocessing cpu_countr)rrrs rr:z)CpuInterface.Worker.get_device_propertiesMs- " " " "'1133I&y11 1rr)rrr r<r!r:r"rrr>rLs>  2 2) 2 2 2  2 2 2rr>r)rrr rgr*r<rfrIr!rr^r=rYr8r[r>r"rrrr-s0,,,,, ,,,$\yC\c\\  I   \ 2222222222rrdevice_interfacesFrrkc\t|tjr|j}|t|<dSr)rrgrrur)rrks rregister_interface_for_devicerYs0&%,'' 0frr6ct|tjr|j}tst |t vr t |Std|)NzNo interface for device )rrgrru_device_initializedinit_device_regrrr3s rget_interface_for_deviceras_&%,''  """ (( AAA B BBrc^tsttSr)rrritemsr"rr get_registered_device_interfacesrks(   " " $ $$rctdtttjD]}td|ttdt ttjD]}td|t tdtda dS)Nrzcuda:rzxpu:cpuT) rr{rrgrrGrrrr)rs rrrqs!&-888 5:**,, - -BB%kakk=AAAA!%666 59))++ , ,@@%jQjj,????!%666r)(r dataclassesrtypingrrrrrr r r rgr=rr _is_compiledtorch._Cr r rrr!rrrrjr{rrrrrrrrrrrrr"rrrsH !!!!!!NNNNNNNNNNNNNNNNNNNN (C5#:.////:EEEEEEEO %,S$. / 46 $sCx.55513S#X333h"h"h"h"h"h"h"h"V6AYAYAYAYAYOAYAYAYH3%*-....9CCCCCCCN>->->->->-?>->->-B  %2%2%2%2%2?%2%2%2P794T/2238881 #u|# $18<_8M1111CU3 +<%=C$BWCCCC%(5d?>S9S3T*U%%%%     r