"Eio4SSKrSSKrSSKrSSKrSSKrSSKJrJr SSKJ r J r SSK J r SSK Jr SSKJrJr SSKJrJr SSKJr SSKrSSKJs Jr SS KJr SS KJrJ r J!r!J"r"J#r# SS K$J%r% SS K&J'r' SS K(J)r)J*r* SSK+J,r,J-r- /SQr.Sr/\0"5r1\2S4Sjr3"SS\Rh5r5Sr6"SS\5r7"SS5r8"SS\ 5r9SSSS.S\:S\:S\:S \:S!\:S"\:S#\ 4S$jjr;S%\:S#\94S&jr<S/S'\=S(\=S-S)\>4S*jjr?"S+S,\85r@"S-S.\75rAg)0N)ABCabstractmethod)CallableIterable)Enum)partial)AnyOptional) deprecatedSelf)warn)_get_privateuse1_backend_name)_add_execution_trace_observer!_disable_execution_trace_observer _enable_execution_trace_observer_ExperimentalConfig _remove_execution_trace_observer) is_fbcode)1profiler_allow_cudagraph_cupti_lazy_reinit_cuda12)kineto_availableProfilerActivity) MemoryProfileMemoryProfileTimeline)supported_activitiesProfilerActionscheduletensorboard_trace_handlerprofileExecutionTraceObserver ProfilerStepcXU[;a [RU5 [XUS9 gg)N)category stacklevel)_WARNINGS_SHOWNaddr )msgr#r$s T/home/mande/repo/quber/.venv/lib/python3.13/site-packages/torch/profiler/profiler.py _warn_oncer),s' /!C  S ;"c\rSrSrSrSrSrg) _NumpyEncoder2zr Json encoder for numpy types (np.int, np.float, np.array etc.) Returns default encoder if numpy is not available cSSKn[ XR 5(a [U5$[ XR5(a [U5$[ XR5(aUR5$[RR X5$![a" [RR X5s$f=f)zEncode NumPy types to JSONrN) numpy ImportErrorjson JSONEncoderdefault isinstanceintegerintfloatingfloatndarraytolist)selfobjnps r(r3_NumpyEncoder.default8s 7  c:: & &s8O [[ ) ):  ZZ ( (::< ##++D6 6 7##++D6 6 7sB)CCN)__name__ __module__ __qualname____firstlineno____doc__r3__static_attributes__r?r*r(r,r,2s  7r*r,c>[RR5$)a Returns a set of supported profiler tracing activities. Note: profiler uses CUPTI library to trace on-device CUDA kernels. In case when CUDA is enabled but CUPTI is not available, passing ``ProfilerActivity.CUDA`` to profiler results in using the legacy CUDA profiling code (same as in the legacy ``torch.autograd.profiler``). This, in turn, results in including CUDA time in the profiler table output, but not in the JSON trace. )torchautograd_supported_activitiesr?r*r(rrHs >> / / 11r*cH\rSrSrSr\S5r\S5r\S5rSr g)_ITraceObserverVzZAbstract interface for a Trace observer. This satisfies 3 methods: start, stop and cleanupcgNr?r;s r(start_ITraceObserver.startZ r*cgrNr?rOs r(stop_ITraceObserver.stop^rRr*cgrNr?rOs r(cleanup_ITraceObserver.cleanupbrRr*r?N) r@rArBrCrDrrPrTrWrEr?r*r(rKrKVsC9      r*rKc\rSrSrSrSSSSSSSSSSSS. S\\S-S\S\S \S \S \S \S-S \ S-S\S\ /\ 4S-S\ S-SS4Sjjr S/SjrS/SjrS/SjrS/SjrS/SjrS\ 4SjrS0S\ S\ 4SjjrS\S\\SS4SjrS1S\S\S \4S!jjrS"rS#\ S$\ SS4S%jrS#\ S$\ SS4S&jrS#\ S$\ SS4S'jrS(rS\4S)jr\"S*\ S+9S2S\ S,\ S-SS4S-jj5r!S.r"g)3_KinetoProfilegaw Low-level profiler wrap the autograd profile Args: activities (iterable): list of activity groups (CPU, CUDA) to use in profiling, supported values: ``torch.profiler.ProfilerActivity.CPU``, ``torch.profiler.ProfilerActivity.CUDA``, ``torch.profiler.ProfilerActivity.XPU``. Default value: ProfilerActivity.CPU and (when available) ProfilerActivity.CUDA or (when available) ProfilerActivity.XPU. record_shapes (bool): save information about operator's input shapes. profile_memory (bool): track tensor memory allocation/deallocation (see ``export_memory_timeline`` for more details). with_stack (bool): record source information (file and line number) for the ops. with_flops (bool): use formula to estimate the FLOPS of specific operators (matrix multiplication and 2D convolution). with_modules (bool): record module hierarchy (including function names) corresponding to the callstack of the op. e.g. If module A's forward call's module B's forward which contains an aten::add op, then aten::add's module hierarchy is A.B Note that this support exist, at the moment, only for TorchScript models and not eager mode models. experimental_config (_ExperimentalConfig) : A set of experimental options used by profiler libraries like Kineto. Note, backward compatibility is not guaranteed. execution_trace_observer (ExecutionTraceObserver) : A PyTorch Execution Trace Observer object. `PyTorch Execution Traces `__ offer a graph based representation of AI/ML workloads and enable replay benchmarks, simulators, and emulators. When this argument is included the observer start() and stop() will be called for the same time window as PyTorch profiler. acc_events (bool): Enable the accumulation of FunctionEvents across multiple profiling cycles post_processing_timeout_s (float): Optional timeout in seconds for post-processing profiler results. In this context, post-processing happens after the profiling itself has finished. If specified, event parsing will stop after this duration and return partial results. Useful for handling large traces that may take too long to process. .. note:: This API is experimental and subject to change in the future. Enabling shape and stack tracing results in additional overhead. When record_shapes=True is specified, profiler will temporarily hold references to the tensors; that may further prevent certain optimizations that depend on the reference count and introduce extra tensor copies. NF activities record_shapesprofile_memory with_stack with_flops with_modulesexperimental_configexecution_trace_observer acc_eventscustom_trace_id_callbackpost_processing_timeout_sr]r^r_r`rarbrcrdrerfrgreturnc U(a [U5O [5UlX lXPlX0lX@lX`lXplXl Xl Xl Xl SUl SUlSUlSUl["R$UR;aSUlO["R&UR;aSUlOy["R(UR;aSUlOS["R*UR;aSUlO-["R,UR;a[/5Ul0Ulg)NFcudaxpumtiahpu)setrr]r^rar_r`rbrcrdrerfrgprofilerhas_cudagraphsmem_tl use_devicerCUDAXPUMTIAHPU PrivateUse1rpreset_metadata) r;r]r^r_r`rarbrcrdrerfrgs r(__init___KinetoProfile.__init__s.8#j/=Q=S*$,$(#6 (@%$(@%)B&-1 #48   DOO 3$DO  ! !T__ 4#DO  " "doo 5$DO  ! !T__ 4#DO  ) )T__ <;=DO02r*cDUR5 UR5 grN) prepare_trace start_tracerOs r(rP_KinetoProfile.starts  r*c$UR5 grN) stop_tracerOs r(rT_KinetoProfile.stops  r*c~[[S5(a$SSKJs Jn UR R UlURbUR(d[R"[RUR;URUR UR"UR$UR&UR(SUR*URUR,UR.S9 UlURbUR(d [1S5 URR35 g)N _inductorrT) use_cpurrr^rar_r`rb use_kinetorcrerfrgzWarning: Profiler clears events at the end of each cycle.Only events from the current cycle will be reported.To keep events across cycles, set acc_events=True.)hasattrrGtorch._inductor.configrconfigtriton cudagraphsrproreprofrrCPUr]rrr^rar_r`rbrcrfrgr)_prepare_trace)r;inductor_configs r(r|_KinetoProfile.prepare_traces 5+ & & < <"1"8"8"C"CD  MM !4?? LL)--@??"00??#22??!..$($<$<??)-)F)F*.*H*H DM MM % E  $$&r*clUR(aURR5 URc [S5eURR 5 UR (aUR SS5 UR(aUR SS5 UR(aUR SS5 UR(aUR SS5 UR(aUR SS5 [5(GaUR5nU(a)UR S[R"U[S 95 Sn[!["S 5(a'S S KJn U"[)["R*S S55nUR,(aTU(aUS:d[/5(d8S[0R2S'UR SS5 S[0R2S'UR4R75HupEUR XE5 M gg)Nz2Profiler must be initialized before starting tracer_1r`r^rbradistributedInfoclsversionr) TorchVersionrjz0.0z12.6DISABLE_CUPTI_LAZY_REINIT0TEARDOWN_CUPTI)rdrProAssertionError _start_tracer_add_metadata_jsonr`r^rbrar_get_distributed_infor1dumpsr,rrGtorch.torch_versionrgetattrrrprosenvironrxitems)r; dist_info cuda_versionrkvs r(r}_KinetoProfile.start_traces  ( (  ) ) / / 1 ==  !UV V ""$     " "#3S 9 ??  " "< 5     " "?C 8     " ">3 7 ??  " "< 5   224I&&%tzz)'O Lui((<+GEMM65,QR "","7HJJ:= 67&&'BCH 03 +,,,224&&q,55 r*cUR(aURR5 URc [S5eURR SSS5 g)Nz2Profiler must be initialized before stopping trace)rdrTror__exit__rOs r(r_KinetoProfile.stop_tracesI  ( (  ) ) . . 0 ==  !UV V tT40r*pathc&URc [S5eURS5(a[R"SSS9nURR UR 5n[UR S5n[R"US5nURU5 SSS5 SSS5 SSS5 U$URR U5$!,(df  N;=f!,(df  ND=f!,(df  W$=f) zs Exports the collected trace in Chrome JSON format. If kineto is enabled, only last cycle in schedule is exported. Nz:Profiler must be initialized before exporting chrome trace.gzw+b.jsonsuffixrbwb) rorendswithtempfileNamedTemporaryFileexport_chrome_tracenameopengzip writelines)r;rfpretvaluefinfouts r(r"_KinetoProfile.export_chrome_traces ==  L  ==  ,,U7Cr==< :D DmetricchURc [S5eURRX5$)zSave stack traces to a file Args: path (str): save stacks file to this location; metric (str): metric to use: "self_cpu_time_total" or "self_cuda_time_total" z4Profiler must be initialized before exporting stacks)ror export_stacks)r;rrs r(r_KinetoProfile.export_stacks/s/ ==  !WX X}}**488r*enablecVURcgURRX5 g)aToggle collection of activities on/off at any point of collection. Currently supports toggling Torch Ops (CPU) and CUDA activity supported in Kineto Args: activities (iterable): list of activity groups to use in profiling, supported values: ``torch.profiler.ProfilerActivity.CPU``, ``torch.profiler.ProfilerActivity.CUDA`` Examples: .. code-block:: python with torch.profiler.profile( activities=[ torch.profiler.ProfilerActivity.CPU, torch.profiler.ProfilerActivity.CUDA, ] ) as p: code_to_profile_0() // turn off collection of all CUDA activity p.toggle_collection_dynamic(False, [torch.profiler.ProfilerActivity.CUDA]) code_to_profile_1() // turn on collection of all CUDA activity p.toggle_collection_dynamic(True, [torch.profiler.ProfilerActivity.CUDA]) code_to_profile_2() print(p.key_averages().table( sort_by="self_cuda_time_total", row_limit=-1)) N)rotoggle_collection_dynamic)r;rr]s r(r(_KinetoProfile.toggle_collection_dynamic:s#: ==  //Cr*group_by_input_shapegroup_by_stack_ngroup_by_overload_namecjURc [S5eURRXU5$)zAverages events, grouping them by operator name and (optionally) input shapes, stack and overload name. .. note:: To use shape/stack functionality make sure to set record_shapes/with_stack when creating profiler context manager. z8Profiler must be initialized before getting key averages)ror key_averages)r;rrrs r(r_KinetoProfile.key_averages[s= ==  J }})) 4J  r*c^URc [S5eURR$)zw Returns the list of unaggregated profiler events, to be used in the trace callback or after the profiling is finished z4Profiler must be initialized before accessing events)rorfunction_eventsrOs r(events_KinetoProfile.eventsps* ==  !WX X}},,,r*keyvaluecrSURSS5-S-n[RRX5 g)zW Adds a user defined metadata with a string key and a string value into the trace file "z\"N)replacerGrH_add_metadata_json)r;rr wrapped_values r( add_metadata_KinetoProfile.add_metadatays0 emmC77#=  ))#=r*cB[RRX5 g)z[ Adds a user defined metadata with a string key and a valid json value into the trace file N)rGrHrr;rrs r(r _KinetoProfile.add_metadata_jsons ))#5r*c X RU'g)z Preset a user defined metadata when the profiler is not started and added into the trace file later. Metadata is in the format of a string key and a valid json value N)rxrs r(preset_metadata_json#_KinetoProfile.preset_metadata_jsons %*S!r*cSSKJn UR5(aUR5(dgUR 5nUUR 5UR 5UR5URR5S.nUS:XaC[RRR5nSRSU55US'U$)Nr)backendrank world_sizepg_count pg_confignccl.c38# UHn[U5v M g7frN)str).0rs r( 7_KinetoProfile._get_distributed_info..s0NAQs nccl_version)torch.distributed distributed is_availableis_initialized get_backendget_rankget_world_size get_pg_countdistributed_c10d_get_all_pg_configsrGrjrrjoin)r;distrrrs r(r$_KinetoProfile._get_distributed_infos(  ""$*=*=*?*?""$MMO--/))+..BBD   f  ::??224L(+0N0N(NIn %r*cBSnUVs/sHn[X5(aMUS3PM nnU(a[SRU5S35eURbURRc [ S5e[ URR5$s snf)N)r^r_r`z=Truez, z required for memory profiling.zDProfiler and kineto_results must be initialized for memory profiling)r ValueErrorrrokineto_resultsrr)r;requiredimissings r(_memory_profile_KinetoProfile._memory_profilesD(0I18H;aS;I  ' 233RST T == DMM$@$@$H V T]]99::Js B Bz`export_memory_timeline` is deprecated and will be removed in a future version. Please use `torch.cuda.memory._record_memory_history` and `torch.cuda.memory._export_memory_snapshot` instead.)r#devicecUcXUR(a URS:waURS-nO'[RR5(aSOSn[ UR 55UlURS5(aUR RX5 gURS5(a[R"SS S 9nURS 5(a'UR RURU5 O&UR RURU5 [UR5n[R"US 5nUR!U5 SSS5 SSS5 SSS5 gUR RX5 g!,(df  N;=f!,(df  ND=f!,(df  g=f) aExport memory event information from the profiler collected tree for a given device, and export a timeline plot. There are 3 exportable files using ``export_memory_timeline``, each controlled by the ``path``'s suffix. - For an HTML compatible plot, use the suffix ``.html``, and a memory timeline plot will be embedded as a PNG file in the HTML file. - For plot points consisting of ``[times, [sizes by category]]``, where ``times`` are timestamps and ``sizes`` are memory usage for each category. The memory timeline plot will be saved a JSON (``.json``) or gzipped JSON (``.json.gz``) depending on the suffix. - For raw memory points, use the suffix ``.raw.json.gz``. Each raw memory event will consist of ``(timestamp, action, numbytes, category)``, where ``action`` is one of ``[PREEXISTING, CREATE, INCREMENT_VERSION, DESTROY]``, and ``category`` is one of the enums from ``torch.profiler._memory_profiler.Category``. Output: Memory timeline written as gzipped JSON, JSON, or HTML. .. deprecated:: ``export_memory_timeline`` is deprecated and will be removed in a future version. Please use ``torch.cuda.memory._record_memory_history`` and ``torch.cuda.memory._export_memory_snapshot`` instead. Nrjz:0zcuda:0cpuz.htmlrw+trrz raw.json.gzwt)rrrGrjrrrrqrexport_memory_timeline_htmlrrexport_memory_timeline_rawrexport_memory_timelinerrr)r;rrrrrs r(r %_KinetoProfile.export_memory_timelinesCB >4??f#<4/%*ZZ%<%<%>%>E,D,@,@,BC  == ! ! KK 3 3D A ]]5 ! !,,U7Cr==//KK::277FKKK66rwwG"'']c499T4+@DOOC(,A] DC KK . .t <,A+@]] DCs=A9GG(F/:GG/ F= 9G G G G)rer]rfrdrcrprqrgrxr_ror^rrrarbr`rhN)self_cpu_time_total)FrFrN)#r@rArBrCrDrrboolrrKrrr8ryrPrTr|r}rrrrr6rrrrrrrrr FutureWarningr rEr?r*r(rZrZgs)\9=#$  ":>;? =A26/2-.5/2 /2  /2  /2/2/2147/2#2D"8/2/2#+2s7"3d":/2$)4</2 /2b'8--^1;;$ 9# 9s 9DD(01A(BD DF&+ !',  "  !%  *->>C>D>6S666**C*D*( ; ; y 2=32=d 2=d2=  2=r*rZc(\rSrSrSrSrSrSrSrSr g) riz? Profiler actions that can be taken at the specified intervals rr!r?N) r@rArBrCrDNONEWARMUPRECORDRECORD_AND_SAVErEr?r*r(rrs D F FOr*r)repeat skip_firstskip_first_waitwaitwarmupactiverrrrhc ^^^^^^S[S[4UUUUUU4SjjnTS:dTS:dTS::d TS:dTS:a[STSTSTSTS TS 3 5eTS:Xa [S S S 9 U$)a Returns a callable that can be used as profiler ``schedule`` argument. The profiler will skip the first ``skip_first`` steps, then wait for ``wait`` steps, then do the warmup for the next ``warmup`` steps, then do the active recording for the next ``active`` steps and then repeat the cycle starting with ``wait`` steps. The optional number of cycles is specified with the ``repeat`` parameter, the zero value means that the cycles will continue until the profiling is finished. The ``skip_first_wait`` parameter controls whether the first ``wait`` stage should be skipped. This can be useful if a user wants to wait longer than ``skip_first`` between cycles, but not for the first profile. For example, if ``skip_first`` is 10 and ``wait`` is 20, the first cycle will wait 10 + 20 = 30 steps before warmup if ``skip_first_wait`` is zero, but will wait only 10 steps if ``skip_first_wait`` is non-zero. All subsequent cycles will then wait 20 steps between the last active and warmup. steprhc|>US:a[SUS35eUT:a[R$UT-nTS:waUT- nTT-T-nTS:aX- T:a[R$X-nUT:a[R$UTT-:a[R$X!S- :a[R$[R $)NrzStep must be non-negative. Got rr)rrrrrr) r! num_stepsmod_steprrrrrrs r( schedule_fnschedule..schedule_fn s !8 #B4&!JK K * !&& & J D a  DLD6MF* A:$*f4!&& &# d?!&& & v %!(( (!m+%% $33 r*rz.Invalid profiler schedule arguments. Got wait=z (need >= 0), warmup=z (need >= 0), active=z (need > 0), repeat=z (need >= 0), skip_first=z (need >= 0).z>Profiler won't be using warmup, this can skew profiler resultsr!r$)r6rrr )rrrrrrr%s`````` r(rrs0#.2 ax6A:1 j1n[RRT5(d[R"TSS9 T(d(TR 5S[R"53mTSTR5S3nT(aUS-nUR"[RRTU55 g![an[ ST-5UeSnAff=f)NT)exist_okzCan't create directory: r(rz.pt.trace.jsonr) rrisdirmakedirs Exception RuntimeError gethostnamegetpidtime_nsrr)re file_namer+sockettimer-r,s r( handler_fn-tensorboard_trace_handler..handler_fnGsww}}X&& Q Ht4#//12!BIIK=AK"m1T\\^$4NC !E)I   h !BC Q"#=#HIqP QsB>> CCCr)r:r;)r+r,r-r<r:r;s``` @@r(rr;s D D r*cZ^\rSrSrSrSSSSSSSSSSSSSSS.S\\S-S\\/\ 4S-S\S \ 4S-S \ S \ S \ S \ S\ S\ S-S\ S-S\ S\ S-S\/\4S-S\S-SS4U4SjjjrSrSrS"SjrS"SjrS"SjrS"SjrSrS"SjrS"SjrS\R6S-4S jrS!rU=r$)#riYaBProfiler context manager. Args: activities (iterable): list of activity groups (CPU, CUDA) to use in profiling, supported values: ``torch.profiler.ProfilerActivity.CPU``, ``torch.profiler.ProfilerActivity.CUDA``, ``torch.profiler.ProfilerActivity.XPU``. Default value: ProfilerActivity.CPU and (when available) ProfilerActivity.CUDA or (when available) ProfilerActivity.XPU. schedule (Callable): callable that takes step (int) as a single parameter and returns ``ProfilerAction`` value that specifies the profiler action to perform at each step. on_trace_ready (Callable): callable that is called at each step when ``schedule`` returns ``ProfilerAction.RECORD_AND_SAVE`` during the profiling. record_shapes (bool): save information about operator's input shapes. profile_memory (bool): track tensor memory allocation/deallocation. with_stack (bool): record source information (file and line number) for the ops. with_flops (bool): use formula to estimate the FLOPs (floating point operations) of specific operators (matrix multiplication and 2D convolution). with_modules (bool): record module hierarchy (including function names) corresponding to the callstack of the op. e.g. If module A's forward call's module B's forward which contains an aten::add op, then aten::add's module hierarchy is A.B Note that this support exist, at the moment, only for TorchScript models and not eager mode models. experimental_config (_ExperimentalConfig) : A set of experimental options used for Kineto library features. Note, backward compatibility is not guaranteed. execution_trace_observer (ExecutionTraceObserver) : A PyTorch Execution Trace Observer object. `PyTorch Execution Traces `__ offer a graph based representation of AI/ML workloads and enable replay benchmarks, simulators, and emulators. When this argument is included the observer start() and stop() will be called for the same time window as PyTorch profiler. See the examples section below for a code sample. acc_events (bool): Enable the accumulation of FunctionEvents across multiple profiling cycles post_processing_timeout_s (float): Optional timeout in seconds for post-processing profiler results. If specified, event parsing will stop after this duration and return partial results. Useful for handling large traces that may take too long to process. use_cuda (bool): .. deprecated:: 1.8.1 use ``activities`` instead. .. note:: Use :func:`~torch.profiler.schedule` to generate the callable schedule. Non-default schedules are useful when profiling long training jobs and allow the user to obtain multiple traces at the different iterations of the training process. The default schedule simply records all the events continuously for the duration of the context manager. .. note:: Use :func:`~torch.profiler.tensorboard_trace_handler` to generate result files for TensorBoard: ``on_trace_ready=torch.profiler.tensorboard_trace_handler(dir_name)`` After profiling, result files can be found in the specified directory. Use the command: ``tensorboard --logdir dir_name`` to see the results in TensorBoard. For more information, see `PyTorch Profiler TensorBoard Plugin `__ .. note:: Enabling shape and stack tracing results in additional overhead. When record_shapes=True is specified, profiler will temporarily hold references to the tensors; that may further prevent certain optimizations that depend on the reference count and introduce extra tensor copies. Examples: .. code-block:: python with torch.profiler.profile( activities=[ torch.profiler.ProfilerActivity.CPU, torch.profiler.ProfilerActivity.CUDA, ] ) as p: code_to_profile() print(p.key_averages().table(sort_by="self_cuda_time_total", row_limit=-1)) Using the profiler's ``schedule``, ``on_trace_ready`` and ``step`` functions: .. code-block:: python # Non-default profiler schedule allows user to turn profiler on and off # on different iterations of the training loop; # trace_handler is called every time a new trace becomes available def trace_handler(prof): print( prof.key_averages().table(sort_by="self_cuda_time_total", row_limit=-1) ) # prof.export_chrome_trace("/tmp/test_trace_" + str(prof.step_num) + ".json") with torch.profiler.profile( activities=[ torch.profiler.ProfilerActivity.CPU, torch.profiler.ProfilerActivity.CUDA, ], # In this example with wait=1, warmup=1, active=2, repeat=1, # profiler will skip the first step/iteration, # start warming up on the second, record # the third and the forth iterations, # after which the trace will become available # and on_trace_ready (when set) is called; # the cycle repeats starting with the next step schedule=torch.profiler.schedule(wait=1, warmup=1, active=2, repeat=1), on_trace_ready=trace_handler, # on_trace_ready=torch.profiler.tensorboard_trace_handler('./log') # used when outputting for tensorboard ) as p: for iter in range(N): code_iteration_to_profile(iter) # send a signal to the profiler that the next iteration has started p.step() The following sample shows how to setup up an Execution Trace Observer (`execution_trace_observer`) .. code-block:: python with torch.profiler.profile( ... execution_trace_observer=( ExecutionTraceObserver().register_callback("./execution_trace.json") ), ) as p: for iter in range(N): code_iteration_to_profile(iter) p.step() You can also refer to test_execution_trace_with_kineto() in tests/profiler/test_profiler.py. Note: One can also pass any object satisfying the _ITraceObserver interface. NF)r]ron_trace_readyr^r_r`rarbrcrdreuse_cudarfrgr]rr?.r^r_r`rarbrcrdrer@rfrgrhc8 >U(a [U5O [5nU bi[S[SS9 U (a UR [ R 5 O3[ R U;aUR[ R 5 [U5S:Xa [S5e[TU]-UUUUUUU U (aU O[R5U U US9 U(aX lSUlO[ UlSUlX0lSUlURUR$5UlSUl0[*R,[*R,4/_[*R,[*R.4UR0/_[*R,[*R24UR0UR4/_[*R,[*R64UR0UR4/_[*R.[*R,4[9[S 5UR4UR:/_[*R.[*R.4/_[*R.[*R24UR4/_[*R.[*R64UR4/_[*R2[*R,4[9[S 5UR:/_[*R2[*R.4[9[S 5UR:/_[*R2[*R24/_[*R2[*R64/_[*R6[*R,4UR:UR</_[*R6[*R.4UR:UR<UR0/_[*R6[*R24UR:UR<UR0UR4/_[*R6[*R64UR:UR<UR0UR4/_[*R.S4UR4UR:/_[*R2S4UR:UR</[*R6S4UR:UR</0EUl[@RBRE[F5 g) Nz;`use_cuda` is deprecated, use `activities` argument insteadr!r'rz"No valid profiler activities foundr\TFz+Incorrect schedule: WARMUP followed by NONEz+Incorrect schedule: RECORD followed by NONEz-Incorrect schedule: RECORD followed by WARMUP)$rnrr rr&rrsremovelenrsuperryr"build_execution_trace_obs_from_envr record_stepsr*r?step_numcurrent_action step_rec_fnrrrr|rr}rrr _trace_ready action_maprKinetoStepTrackerinit_step_countPROFILER_STEP_NAME)r;r]rr?r^r_r`rarbrcrdrer@rfrgactivities_set __class__s r(ryprofile.__init__s&-7Z'JJL!%=&?  $M $D 0DM %D , "mmDMM:8<:  ."5"5 6:  ."7"7 84;M;M:N:  ."7"7 8""  ; :  ."@"@ A""  D:  " "N$7$7 8KL  ;: " " "N$9$9 :B#: $ " "N$9$9 :T=M=M ::>>-r 2 2 KKBJJNN+CRHH  " " 1 12D E   #33#dmm"44 D     & & ( r*cXlg)z@ Sets a callback to be called when a new trace ID is generated. N)rf)r;callbacks r(set_custom_trace_id_callback$profile.set_custom_trace_id_callbacks )1%r*cJURcgURR$)z Returns the current trace ID. N)rotrace_idrOs r( get_trace_idprofile.get_trace_ids! == }}%%%r*cJUR(aURU5 ggrN)r?rOs r(rJprofile._trace_readys       % r*clURRX45nU(aUH nU"5 M ggrN)rKrf)r;rhrH action_listactions r(r]profile._transit_actions0oo));*GH %& r*cJURcgURR$rN)ro_statsrOs r(ryprofile._statss == }}###r*)rKrHrfr?rFrrGrIr)r@rArBrCrDrrrr6rr rrrKrr8ryrSrrPrTr!rlrprJr]r_ProfilerStatsryrE __classcell__)rPs@r(rrYsCP9=;?48#$  ":>;? $=A26#|C-.5|CC5.01D8 |C !c*T1 |C  |C|C|C|C|C147|C#2D"8|C|C+|C #+2s7"3d":!|C"$)4<#|C$ %|C|C|4 )8 ),1 && $++d2$$r*rc\rSrSrSrSSjrSSjr\S\S4Sj5r SSjr S \ S\ 4S jr SS\ S-4S jjr\SS \S\ S-4S jj5rSSjr\S5rSrSSjrSSjrSSjrS\ S-4SjrSSjrSrg)riafExecution Trace Observer Each process can have a single ExecutionTraceObserver instance. The observer can be added to record function callbacks via calling register_callback() explicitly. Without calling unregister_callback(), repeated calls to register_callback() will not add additional observers to record function callbacks. Once an ExecutionTraceObserver is created, the start() and stop() methods control when the event data is recorded. Deleting or calling unregister_callback() will remove the observer from the record function callbacks, finalize the output file, and will stop incurring any overheads. rhNcXSUlSUlSUlSUlSUlSUlg)z! Initializes the default states. FrdN) _registered_execution_trace_runningextra_resources_collection resources_diroutput_file_pathoutput_file_path_observerrOs r(ryExecutionTraceObserver.__init__s4!(-%*/'"$%'.0&r*c$UR5 gz? Calls unregister_callback() to make sure to finalize outputs. Nunregister_callbackrOs r(__del__ExecutionTraceObserver.__del__   "r*c[RRSS5S:Xa[R"SSSS9nUR nSSS5 [5nURW5 [RRS S5S:XaURS 5 U$URS5 U$g!,(df  Nt=f![ an[S U3S S 9 SnAgSnAff=f)a Returns an ExecutionTraceObserver instance if the environment variable ENABLE_PYTORCH_EXECUTION_TRACE is set to 1, otherwise returns None. Configures the observer to also collect extra resources if the environment variable ``ENABLE_PYTORCH_EXECUTION_TRACE_EXTRAS=1``. These are resources such as generated kernels, index tensor data etc. that are required to make the Execution Trace replayable. ENABLE_PYTORCH_EXECUTION_TRACErrr z.et.jsonFrdeleteNzTExecution trace will not be recorded. Exception on creating default temporary file: r!r'%ENABLE_PYTORCH_EXECUTION_TRACE_EXTRAST) rrrfrrrr3r rregister_callbackset_extra_resource_collection)rfilenamer8ets r(rE9ExecutionTraceObserver.build_execution_trace_obs_from_envs ::>>:C @C G 00*U!wwH()B   *zz~~EsKsR006I007I% jkljmn   s4C B7 C7 CCC C) C$$C)cPXlUR(aURSS9 g)a Collects extra resources such as generated kernels, index tensor data, and any other metadata that is required to complete the Execution Trace content. The caller should call this method with val=True after calling register_callback() if they want to collect the extra resources. T) can_createN)rget_resources_dir)r;vals r(r4ExecutionTraceObserver.set_extra_resource_collections'+.'  * *  " "d " 3r*rcS[4SjnUR(d9XlURS5(aU"5nXl[ U5UlU$)z^ Adds ET observer to record function callbacks. The data will be written to output_file_path. rhc|[R"SSSS9nURsSSS5 $!,(df  g=f)NrrFr)rrr)rs r(get_temp_uncompressed_fileLExecutionTraceObserver.register_callback..get_temp_uncompressed_files+,,U75QUWwwRQQs - ;r)rrrrrr)r;rrs r(r(ExecutionTraceObserver.register_callbacksR  C $4 !((//#=#? -= *<=MND  r*cUR(dgUR(a UR$[RURUS9nU(dgX lUR$)a Generates the resources directory for the generated kernels, or index tensor data or any other metadata that is required to complete the Execution Trace content. The directory is created right where the ET file is being output. Only works if the observer has called set_extra_resource_collection(val=True). Returns None if the observer is not configured with extra resource collection. N) create_dir)rrrget_resources_dir_for_et_pathr)r;rgenerated_paths r(r(ExecutionTraceObserver.get_resources_dirs`..   %% %/MM  ! !jN +!!!r*rc[RRU5up#[RRU[RR U5SS-5n[RR U5(d!U(a[R "U5 U$gU$![a [SU3SS9 gf=f)Nr _resourcesz(Execution trace exception when creating r!r') rrsplitrsplitextexistsmkdirr3r ) trace_pathrwork_dirr9 resource_dirs r(r4ExecutionTraceObserver.get_resources_dir_for_et_path!s!ggmmJ7ww|| bgg&&y1!4|C ww~~l++ HH\*! B<.Q#$  sB,,CCct^S U4SjjnS[S[SS4SjnTR(agTR5 U"5 [ 5 TR RS 5(aU"TRTR 5 S Tlgg![an[ SU3SS 9 SnANlSnAff=f) z5 Removes ET observer from record function callbacks. rhNc>TR5nU(dgSSKJn UR Vs/sHn[ USS5cMURPM! nnUH\nUcM[RRU5n[RRX5n[R"XW5 M^ g![an[SU3SS9 SnAgSnAff=fs snf)Nz>Execution trace exception when generating resource directory: r!r'r) PyCodeCache__file__)rr3r torch._inductor.codecachermodulesrrrrbasenamershutilcopyfile) rr8rr kernel_files kernel_filerdstr;s r(_save_triton_kernelsHExecutionTraceObserver.unregister_callback.._save_triton_kernels<s #557   >%,,,A1j$/ ,  , &ww'' 4ggll<6 1 ,% TUVTWX   s"B:CC: C CCuncompressed_file output_filec"[SUSU35 [US5n[R"US5nURU5 SSS5 SSS5 [R "U5 g!,(df  N-=f!,(df  N6=f)NzExecution Trace: compressing z to rr)printrrrrrB)rrrrs r( _save_gz_fileAExecutionTraceObserver.unregister_callback.._save_gz_fileXsn 12C1DD V W'.#YY{D1TOOC(2/ II' (21/.s"BA/B/ A= 9B Bz(Execution trace failed to save kernels: r!r'gzFr) rrrTr3r rrrr)r;rrr8s` r(r*ExecutionTraceObserver.unregister_callback7s  28 )S )s )t )    IIK S$& - .$$--d33d<S>ST$D    S?sCPQR SsB B7 B22B7cUR$)zN Returns True if the execution trace observer is registered, otherwise False. )rrOs r( is_registered$ExecutionTraceObserver.is_registeredms r*cUR$)z; Returns True if the observer is running, otherwise False. )rrOs r( is_running!ExecutionTraceObserver.is_runningts,,,r*cUR(a4UR(d"[5 SUlUR5 ggg)z Starts to capture. TN)rrr_record_pg_configrOs r(rPExecutionTraceObserver.startzs8   D$A$A , .,0D )  " " $%B r*cJUR(a[5 SUlgg)z Stops to capture. FN)rrrOs r(rTExecutionTraceObserver.stops   ( ( - /,1D ) )r*c$UR5 grrrOs r(rWExecutionTraceObserver.cleanuprr*c>UR(a UR$g)z' Returns the output file name or None. N)rrOs r(get_output_file_path+ExecutionTraceObserver.get_output_file_paths  (( (r*cUR(a[RR5(a[RR 5(af[RR R Rn[RRS[R"U[S95 gggg)Nz## process_group:init ##r) rrGrrrr_worldpg_config_inforH _record_function_with_args_enterr1rr,)r;rs r(r(ExecutionTraceObserver._record_pg_configs   !!..00!!0022"..??FFUUN NN ; ;* >}= 31 r*)rrrrrrr)F)r@rArBrCrDryr staticmethodr rErrr rrrrrpropertyrrrPrTrWrrrEr?r*r(rrs  1# 9Q0R> #$$"S4Z"4', $ t*4%l  - %2# cDj r*r)NF)Brr1rrrabcrrcollections.abcrrenumr functoolsrtypingr r typing_extensionsr r warningsr rGtorch.autograd.profilerrHrortorch._Crtorch._C._profilerrrrrrtorch._environmentrtorch._utils_internalrtorch.autogradrrtorch.profiler._memory_profilerrr__all__rNrnr% UserWarningr)r2r,rrKrZrr6rr*rrrrrr?r*r(rsX #. . &&2)S=P $%)Q< 7D$$7, 2 c "@=@=F T ; ; ;  ;  ;  ;;;|!C!N!EJ #d =A