a 8SicwS@s"ddlZddlZddlZddlZddlmZddlmZddlm Z m Z m Z m Z m Z mZmZddlmZddlZddlmmZddlmZmZddlmZdd ZGd d d eZGd d d eZddde e e e e e dddZ!e edddZ"de#ee#e$dddZ%GdddeZ&dS)N)Enum)partial)AnyCallableDictIterableListOptionalTuple)warn)ProfilerActivitykineto_available)_ExperimentalConfigcCs tjS)a Returns a set of supported profiler tracing activities. Note: profiler uses CUPTI library to trace on-device CUDA kernels. In case when CUDA is enabled but CUPTI is not available, passing ``ProfilerActivity.CUDA`` to profiler results in using the legacy CUDA profiling code (same as in the legacy ``torch.autograd.profiler``). This, in turn, results in including CUDA time in the profiler table output, but not in the JSON trace. )torchautograd_supported_activitiesrrS/var/www/html/django/DPS/env/lib/python3.9/site-packages/torch/profiler/profiler.pysupported_activitiess rc @seZdZdZddddddddeeeeeeeeeedddZ ddZ d d Z d d Z d dZ ddZedddZd%eedddZd&eedddZddZeeddd Zeedd!d"Zd#d$ZdS)'_KinetoProfileaLow-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``. Default value: ProfilerActivity.CPU and (when available) ProfilerActivity.CUDA. 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 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. .. note:: This API is an experimental and subject to change in 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_configcCsB|r t|nt|_||_||_||_||_||_||_d|_ dSN) setrrrrrrrrprofiler)selfrrrrrrrrrr__init__<s z_KinetoProfile.__init__cCs||dSr) prepare_trace start_tracer!rrrstartOsz_KinetoProfile.startcCs |dSr) stop_tracer%rrrstopSsz_KinetoProfile.stopc CsHtjtj|jvtj|jv|j|j|j|j |j d|j d |_ |j dS)NT) use_cudause_cpurrrrr use_kinetor)profprofiler CUDArCPUrrrrrrr _prepare_tracer%rrrr#Vs   z_KinetoProfile.prepare_tracecCs@|jdusJ|jtr<|}|r<|dt|dS)NZdistributedInfo)r _start_tracer _get_distributed_infoadd_metadata_jsonjsondumps)r!Z dist_inforrrr$ds  z_KinetoProfile.start_tracecCs"|jdusJ|jddddSr)r __exit__r%rrrr'msz_KinetoProfile.stop_trace)pathc Cs|js J|drtjdddd}||j|j}t|jF}t|d}| |Wdn1sr0YWdn1s0Yt |j|S|j|SdS)zD Exports the collected trace in Chrome JSON format. .gzzw+tz.jsonF)suffixdeletewtN) r endswithtempfileNamedTemporaryFilecloseexport_chrome_tracenameopengzip writelinesosremove)r!r7fpZretvaluefinfoutrrrr@qs   F z"_KinetoProfile.export_chrome_traceself_cpu_time_total)r7metriccCs|js J|j||S)aSave stack traces in a file in a format suitable for visualization. Args: path (str): save stacks file to this location; metric (str): metric to use: "self_cpu_time_total" or "self_cuda_time_total" .. note:: Example of using FlameGraph tool: - git clone https://github.com/brendangregg/FlameGraph - cd FlameGraph - ./flamegraph.pl --title "CPU time" --countname "us." profiler.stacks > perf_viz.svg )r export_stacks)r!r7rKrrrrLs z_KinetoProfile.export_stacksr)group_by_input_shapegroup_by_stack_ncCs|js J|j||S)aAverages events, grouping them by operator name and (optionally) input shapes and stack. .. note:: To use shape/stack functionality make sure to set record_shapes/with_stack when creating profiler context manager. )r key_averages)r!rMrNrrrrOs z_KinetoProfile.key_averagescCs|js J|jjS)z Returns the list of unaggregated profiler events, to be used in the trace callback or after the profiling is finished )r function_eventsr%rrreventss z_KinetoProfile.events)keyvaluecCs&d|ddd}tj||dS)zo Adds a user defined metadata with a string key and a string value into the trace file "z\"N)replacerr_add_metadata_json)r!rRrSZ wrapped_valuerrr add_metadatasz_KinetoProfile.add_metadatacCstj||dS)zs Adds a user defined metadata with a string key and a valid json value into the trace file N)rrrV)r!rRrSrrrr3sz _KinetoProfile.add_metadata_jsoncCs8ddlm}|r|s dS|||dS)Nr)backendrank world_size)Ztorch.distributed distributed is_availableis_initialized get_backendget_rankget_world_size)r!distrrrr2s z$_KinetoProfile._get_distributed_info)rJ)Fr)__name__ __module__ __qualname____doc__r rr boolrr"r&r(r#r$r'strr@rLintrOrQrWr3r2rrrrrs:     rc@s eZdZdZdZdZdZdZdS)ProfilerActionzG Profiler actions that can be taken at the specified intervals rN)rbrcrdreNONEWARMUPRECORDRECORD_AND_SAVErrrrris ri)repeat skip_first)waitwarmupactiverqrrreturncs`ttdfdd }dkrDdkrDdkrDdkrDdksLJddkr\td|S)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. )steprvcs|dks J|krtjS|8}}dkrH||krHtjS||}|kr^tjS|krptjS||dkrtjStjSdS)Nrrj)rirmrnrorp)rwZ num_stepsZmod_steprurqrrrsrtrr schedule_fns   zschedule..schedule_fnrz#Invalid profiler schedule argumentsz>Profiler won't be using warmup, this can skew profiler results)rhrir )rsrtrurqrrryrrxrschedulesrz)_rvcCstjS)zy Default profiler behavior - immediately starts recording the events, keeps doing it on every profiler step. )riro)r{rrr_default_schedule_fnsr|F)dir_name worker_nameuse_gzipcs8ddlddlddlddfdd }|S)a  Outputs tracing files to directory of ``dir_name``, then that directory can be directly delivered to tensorboard as logdir. ``worker_name`` should be unique for each worker in distributed scenario, it will be set to '[hostname]_[pid]' by default. rN)rvcsjs>zjddWnty<tdYn0sZdtdt  d}r~|d}| j |dS)NT)exist_okzCan't create directory: z{}_{}z{}.{}.pt.trace.jsonir8) r7isdirmakedirs Exception RuntimeErrorformat gethostnamerggetpidrhtimer@join)r, file_namer}rEsocketrrr~rr handler_fns  z-tensorboard_trace_handler..handler_fn)rErr)r}r~rrrrrtensorboard_trace_handlers  rc seZdZdZddddddddddd eeeeeege feede fe e e e e ee ee d fddZ dd Zd d Zd d ZddZddZddZddZZS)r-aProfiler 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``. Default value: ProfilerActivity.CPU and (when available) ProfilerActivity.CUDA. 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. 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, # 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), 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() NF) rrzon_trace_readyrrrrrrr).c 's.|r t|nt} | durJtd| r4| tjntj| vrJ| tjt| dks^Jdtj ||||||| d|r||_ d|_ n t |_ d|_ ||_ d|_| |j|_d|_tjtjfgtjtjf|jgtjtjf|j|jgtjtjf|j|jgtjtjfttd|j|jgtjtjfgtjtjf|jgtjtjf|jgtjtjfttd|jgtjtjfttd |jgtjtjfgtjtjfgtjtjf|j|jgtjtjf|j|j|jgtjtjf|j|j|j|jgtjtjf|j|j|j|jgtjdf|j|jgtjdf|j|jgtjdf|j|jgi|_dS) Nz7use_cuda is deprecated, use activities argument insteadrz"No valid profiler activities foundrTFz+Incorrect schedule: WARMUP followed by NONEz+Incorrect schedule: RECORD followed by NONEz-Incorrect schedule: RECORD followed by WARMUP)rrr addr r.rFlensuperr"rz record_stepsr|rstep_numcurrent_action step_rec_fnrirmrnr#ror$rprr' _trace_ready action_map) r!rrzrrrrrrrr)Zactivities_set __class__rrr"zs            zprofile.__init__cCs ||Sr)r&r%rrr __enter__szprofile.__enter__cCs |dSr)r()r!exc_typeexc_valexc_tbrrrr6szprofile.__exit__cCs:|tj|j|jr6tdt|j|_ |j dS)N ProfilerStep#) _transit_actionrirmrrr,record_functionrgrrrr%rrrr&sz profile.startcCs.|jr|jr|jddd||jddSr)rrr6rrr%rrrr(s z profile.stopcCs|jr|jr|jddd|j}|j}|jd7_||j|_|||jt|jr~t dt ||_|j dS)zP Signals the profiler that the next profiling step has started. Nrjr) rrr6rrrzrr, kineto_steprrgr)r! prev_actionZcur_steprrrrws z profile.stepcCs|jr||dSr)rr%rrrrszprofile._trace_readycCs(|j||f}|r$|D] }|qdSr)rget)r!rr action_listactionrrrrszprofile._transit_action)rbrcrdrer rr rrhrirrfrr"rr6r&r(rwrr __classcell__rrrrr- s<o Ur-)NF)'rCr4rEr=enumr functoolsrtypingrrrrrr r warningsr rtorch.autograd.profilerrr r,torch.autogradr r Ztorch._C._autogradrrobjectrrirhrzr|rgrfrr-rrrrs&  $  $