# mypy: allow-untyped-defsimportuuidfromcollectionsimportdefaultdictfromdataclassesimportdataclassfromtimeimportperf_counter_nsfromtypingimportAny,Dict,Iterable,List,Optionalfromwarningsimportwarnimporttorchimporttorch.cudafromtorch._Cimport_get_privateuse1_backend_namefromtorch._C._profilerimport_ExperimentalConfigfromtorch.autogradimport(_disable_profiler,_enable_profiler,_kineto_step,_prepare_profiler,_ProfilerResult,_supported_activities,_toggle_collection_dynamic,DeviceType,kineto_available,ProfilerActivity,ProfilerConfig,ProfilerState,)fromtorch.autograd.profiler_utilimport(_filter_name,_filter_stack_entry,_rewrite_name,EventList,FunctionEvent,MEMORY_EVENT_NAME,MemRecordsAcc,OUT_OF_MEMORY_EVENT_NAME,)fromtorch.futuresimportFuture__all__=["profile","record_function","emit_itt","emit_nvtx","load_nvprof","EnforceUnique","parse_nvprof_trace","KinetoStepTracker","EventList","FunctionEvent","MemRecordsAcc",]try:# Available in Python >= 3.2fromcontextlibimportContextDecoratoras_ContextDecoratorexceptImportError:importfunctoolsclass_ContextDecorator:# type: ignore[no-redef]def__enter__(self):raiseNotImplementedErrordef__exit__(self,exc_type,exc_val,exc_tb):raiseNotImplementedErrordef__call__(self,func):@functools.wraps(func)defwrapped(*args,**kwargs):withself:returnfunc(*args,**kwargs)returnwrapped# global python state - whether profiler is currently enabled# useful for fast python checks to reduce latency_is_profiler_enabled:bool=Falsedef_set_is_profiler_enabled(enable:bool):global_is_profiler_enabled_is_profiler_enabled=enabledef_run_on_profiler_start():_set_is_profiler_enabled(True)def_run_on_profiler_stop():_set_is_profiler_enabled(False)@dataclassclass_ProfilerStats:"Profiler timing and stats used by developers to catch issues/regressions"profiling_window_duration_sec:float=0number_of_events:int=0profiler_prepare_call_duration_us:int=0profiler_enable_call_duration_us:int=0profiler_disable_call_duration_us:int=0parse_kineto_call_duration_us:int=0function_events_build_tree_call_duration_us:int=0
[docs]classprofile:"""Context manager that manages autograd profiler state and holds a summary of results. Under the hood it just records events of functions being executed in C++ and exposes those events to Python. You can wrap any code into it and it will only report runtime of PyTorch functions. Note: profiler is thread local and is automatically propagated into the async tasks Args: enabled (bool, optional): Setting this to False makes this context manager a no-op. use_cuda (bool, optional): Enables timing of CUDA events as well using the cudaEvent API. (will be deprecated) use_device (str, optional): Enables timing of device events. Adds approximately 4us of overhead to each tensor operation when use cuda. The valid devices options are 'cuda', 'xpu', 'mtia' and 'privateuseone'. record_shapes (bool, optional): If shapes recording is set, information about input dimensions will be collected. This allows one to see which dimensions have been used under the hood and further group by them using prof.key_averages(group_by_input_shape=True). Please note that shape recording might skew your profiling data. It is recommended to use separate runs with and without shape recording to validate the timing. Most likely the skew will be negligible for bottom most events (in a case of nested function calls). But for higher level functions the total self cpu time might be artificially increased because of the shape collection. with_flops (bool, optional): If with_flops is set, the profiler will estimate the FLOPs (floating point operations) value using the operator's input shape. This allows one to estimate the hardware performance. Currently, this option only works for the matrix multiplication and 2D convolution operators. profile_memory (bool, optional): track tensor memory allocation/deallocation. with_stack (bool, optional): record source information (file and line number) for the ops. 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. use_kineto (bool, optional): experimental, enable profiling with Kineto profiler. use_cpu (bool, optional): profile CPU events; setting to ``False`` requires ``use_kineto=True`` and can be used to lower the overhead for GPU-only profiling. experimental_config (_ExperimentalConfig) : A set of experimental options used by profiler libraries like Kineto. Note, backward compatibility is not guaranteed. acc_events (bool): Enable the accumulation of FunctionEvents across multiple profiling cycles .. warning: Enabling memory profiling or source attribution incurs additional profiler overhead .. warning: This context managers should not be called recursively, i.e. no nested instances are allowed .. warning: Due to some CUDA multiprocessing limitations (multiprocessing-cuda-note_), one cannot use the profiler with ``use_device = 'cuda'`` to benchmark DataLoaders with ``num_workers > 0``. If you wish to benchmark data loading, please use ``use_device = None`` or ``num_workers = 0``. Example: >>> # xdoctest: +SKIP >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD_PROFILER) >>> x = torch.randn((1, 1), requires_grad=True) >>> with torch.autograd.profiler.profile() as prof: >>> for _ in range(100): # any normal python code, really! >>> y = x ** 2 >>> y.backward() >>> # NOTE: some columns were removed for brevity >>> print(prof.key_averages().table(sort_by="self_cpu_time_total")) ----------------------------------- --------------- --------------- --------------- Name Self CPU total CPU time avg Number of Calls ----------------------------------- --------------- --------------- --------------- mul 32.048ms 32.048ms 200 pow 27.041ms 27.041ms 200 PowBackward0 9.727ms 55.483ms 100 torch::autograd::AccumulateGrad 9.148ms 9.148ms 100 torch::autograd::GraphRoot 691.816us 691.816us 100 ----------------------------------- --------------- --------------- --------------- """def__init__(self,enabled=True,*,use_cuda=False,# Deprecateduse_device=None,record_shapes=False,with_flops=False,profile_memory=False,with_stack=False,with_modules=False,use_kineto=False,use_cpu=True,experimental_config=None,acc_events=False,custom_trace_id_callback=None,):self.enabled:bool=enabledifnotself.enabled:returnself.use_cuda=use_cudaifself.use_cuda:warn("The attribute `use_cuda` will be deprecated soon, ""please use ``use_device = 'cuda'`` instead.",FutureWarning,stacklevel=2,)self.use_device:Optional[str]="cuda"else:self.use_device=use_device# TODO Consider changing _function_events into data structure with size capself._function_events:Optional[EventList]=Noneself._old_function_events:Optional[EventList]=None# Function event processing is done lazilyself._needs_processing=Falseself.entered=Falseself.record_shapes=record_shapesself.with_flops=with_flopsself.record_shapes|=self.with_flopsself.profile_memory=profile_memoryself.with_stack=with_stackself.with_modules=with_modulesself.use_cpu=use_cpuself.acc_events=acc_eventsifexperimental_configisNone:experimental_config=_ExperimentalConfig()self.experimental_config=experimental_configself.kineto_results:Optional[_ProfilerResult]=Noneself.profiling_start_time_ns=0self.profiling_end_time_ns=0self._stats=_ProfilerStats()self.custom_trace_id_callback=custom_trace_id_callbackself.trace_id=""ifnotself.use_cpu:assert(use_kineto),"Device-only events supported only with Kineto (use_kineto=True)"ifself.use_deviceisnotNone:VALID_DEVICE_OPTIONS=["cuda","xpu","mtia"]if_get_privateuse1_backend_name()!="privateuseone":VALID_DEVICE_OPTIONS.append(_get_privateuse1_backend_name())ifself.use_devicenotinVALID_DEVICE_OPTIONS:warn(f"The {self.use_device} is not a valid device option.")self.use_device=Noneifself.use_device=="cuda"andnottorch.cuda.is_available():warn("CUDA is not available, disabling CUDA profiling")self.use_cuda=Falseself.use_device=Noneifself.use_device=="xpu"andnottorch.xpu.is_available():warn("XPU is not available, disabling XPU profiling")self.use_device=Noneself.kineto_activities=set()ifself.use_cpu:self.kineto_activities.add(ProfilerActivity.CPU)self.profiler_kind=ProfilerState.KINETOifself.use_device=="cuda":ifnotuse_kinetoorProfilerActivity.CUDAnotin_supported_activities():assertself.use_cpu,"Legacy CUDA profiling requires use_cpu=True"self.profiler_kind=ProfilerState.KINETO_GPU_FALLBACKelse:self.kineto_activities.add(ProfilerActivity.CUDA)elifself.use_device=="xpu":assert(use_kinetoandProfilerActivity.XPUin_supported_activities()),"Legacy XPU profiling is not supported. Requires use_kineto=True on XPU devices."self.kineto_activities.add(ProfilerActivity.XPU)elifself.use_device=="mtia":assert(use_kinetoandProfilerActivity.MTIAin_supported_activities()),"Legacy MTIA profiling is not supported. Requires use_kineto=True on MTIA devices."self.kineto_activities.add(ProfilerActivity.MTIA)elifself.use_deviceisnotNoneandself.use_device!="privateuseone":if(notuse_kinetoorProfilerActivity.PrivateUse1notin_supported_activities()):assert(self.use_cpu),"Legacy custombackend profiling requires use_cpu=True"self.profiler_kind=ProfilerState.KINETO_PRIVATEUSE1_FALLBACKelse:self.kineto_activities.add(ProfilerActivity.PrivateUse1)assert(len(self.kineto_activities)>0),"No activities specified for the profiler"defdefault_trace_id(self):# Generate a UUIDuuid_raw=uuid.uuid4()returnf"{uuid_raw.int:032X}"defcreate_trace_id(self):ifself.custom_trace_id_callback:returnself.custom_trace_id_callback()returnself.default_trace_id()defconfig(self,create_trace_id=False):# only need to generate new trace id upon prepare trace not start traceifcreate_trace_id:trace_id=self.create_trace_id()self.trace_id=trace_idreturnProfilerConfig(self.profiler_kind,self.record_shapes,self.profile_memory,self.with_stack,self.with_flops,self.with_modules,self.experimental_config,self.trace_id,)def__enter__(self):ifnotself.enabled:returnifself.entered:raiseRuntimeError("Profiler context manager is not reentrant")self._prepare_trace()self._start_trace()returnselfdef_prepare_trace(self):self.entered=Truet0=perf_counter_ns()_prepare_profiler(self.config(create_trace_id=True),self.kineto_activities)t1=perf_counter_ns()self._stats.profiler_prepare_call_duration_us=int((t1-t0)/1000)def_start_trace(self):self.entered=True_run_on_profiler_start()t0=perf_counter_ns()_enable_profiler(self.config(create_trace_id=False),self.kineto_activities)t1=perf_counter_ns()self._stats.profiler_enable_call_duration_us=int((t1-t0)/1000)self.profiling_start_time_ns=t1def__exit__(self,exc_type,exc_val,exc_tb):ifnotself.enabled:returnifself.use_deviceandhasattr(torch,self.use_device):device_module=getattr(torch,self.use_device)ifhasattr(device_module,"synchronize"):device_module.synchronize()ifself._function_eventsandself.acc_events:self._old_function_events=self._function_eventsself._function_events=Noneself._needs_processing=Truet0=perf_counter_ns()self.kineto_results=_disable_profiler()t1=perf_counter_ns()self._stats.profiler_disable_call_duration_us=int((t1-t0)/1000)self.profiling_end_time_ns=t0_run_on_profiler_stop()self._stats.profiling_window_duration_sec=((self.profiling_end_time_ns-self.profiling_start_time_ns)*1.0/1e9)# If we plan to accumulate events we should post process the function events# right away to retain the state across mulitple start/stop callsifself.acc_events:self._ensure_function_events()returnFalsedef__repr__(self):ifself._needs_processing:self._ensure_function_events()ifself._function_eventsisNone:return"<unfinished torch.autograd.profile>"returnrepr(self._function_events)def__str__(self):ifself._needs_processing:self._ensure_function_events()ifself._function_eventsisNone:return"<unfinished torch.autograd.profile>"returnstr(self._function_events)def_ensure_function_events(self):"""Process function events lazily if required"""ifself._function_eventsisnotNone:returnself._needs_processing=Falset0=perf_counter_ns()parsed_results=[]ifself.kineto_results:parsed_results=self._parse_kineto_results(self.kineto_results)t1=perf_counter_ns()self._stats.parse_kineto_call_duration_us=int((t1-t0)/1000)self._function_events=EventList(parsed_results,use_device=self.use_device,profile_memory=self.profile_memory,with_flops=self.with_flops,)t0=perf_counter_ns()self._function_events._build_tree()t1=perf_counter_ns()self._stats.function_events_build_tree_call_duration_us=int((t1-t0)/1000)self._stats.number_of_events=len(self._function_events)ifself._old_function_eventsandself.acc_events:forevtinself._old_function_events:self._function_events.append(evt)self._old_function_events=Noneifself._function_eventsisNone:raiseRuntimeError("Profiler didn't finish running")@propertydeffunction_events(self):ifself._function_eventsisNoneorself._needs_processing:self._ensure_function_events()returnself._function_eventsdeftable(self,sort_by=None,row_limit=100,max_src_column_width=75,max_name_column_width=55,max_shapes_column_width=80,header=None,top_level_events_only=False,):self._ensure_function_events()assertself._function_eventsisnotNonereturnself._function_events.table(sort_by=sort_by,row_limit=row_limit,max_src_column_width=max_src_column_width,max_name_column_width=max_name_column_width,max_shapes_column_width=max_shapes_column_width,header=header,top_level_events_only=top_level_events_only,)table.__doc__=EventList.table.__doc__
[docs]defexport_chrome_trace(self,path):""" Exports the collected trace in Chrome JSON format. If kineto is enabled, only last cycle in schedule is exported. """ifkineto_available():self.kineto_results.save(path)# type: ignore[union-attr]else:self._ensure_function_events()returnself._function_events.export_chrome_trace(path)# type: ignore[union-attr]
export_chrome_trace.__doc__=EventList.export_chrome_trace.__doc__defexport_stacks(self,path:str,metric:str="self_cpu_time_total"):self._ensure_function_events()assertself._function_eventsisnotNone,"Expected profiling results"assertself.with_stack,"export_stacks() requires with_stack=True"returnself._function_events.export_stacks(path,metric)deftoggle_collection_dynamic(self,enabled:bool,activities:Iterable[ProfilerActivity]):""" Toggles the collection of activities for the current profiler instance. """return_toggle_collection_dynamic(enabled,set(activities))
total_average.__doc__=EventList.total_average.__doc__@propertydefself_cpu_time_total(self):"""Returns total time spent on CPU. The total time is a sum of all self times across all the events. """self._ensure_function_events()assertself._function_eventsisnotNonereturnself._function_events.self_cpu_time_totaldef_parse_kineto_results(self,result:_ProfilerResult):# result.events() has most of the events - PyTorch op-level and device-level eventstrace_start_ns=result.trace_start_ns()mem_records=[[evt,False]forevtinresult.events()ifevt.name()==MEMORY_EVENT_NAME]oom_records=[evtforevtinresult.events()ifevt.name()==OUT_OF_MEMORY_EVENT_NAME]mem_records_acc=MemRecordsAcc(mem_records)def_cpu_memory_usage(mem_record):return(mem_record.nbytes()ifmem_record.device_type()in[DeviceType.CPU,DeviceType.MKLDNN,DeviceType.IDEEP]else0)def_device_memory_usage(mem_record):return(mem_record.nbytes()ifmem_record.device_type()in[DeviceType.CUDA,DeviceType.PrivateUse1,DeviceType.HIP]else0)# Create and return FunctionEvent list, which contains all function events# Here 2 function events are created:# all_function_events contains all events associated with each kineto event from resultall_function_events=[]# frontend_function_events contains the events in aten or torch frontend level,# whose correlation id is 0frontend_function_events=[]device_corr_map:Dict[int,List[FunctionEvent]]={}max_evt_id=0forkineto_eventinresult.events():if_filter_name(kineto_event.name()):continuerel_start_ns=kineto_event.start_ns()-trace_start_nsrel_end_ns=kineto_event.end_ns()-trace_start_nsabs_end_ns=kineto_event.end_ns()cpu_memory_usage=0device_memory_usage=0ifkineto_event.device_type()==DeviceType.CPU:# find the corresponding memory allocation eventsformem_recordinmem_records_acc.in_interval(kineto_event.start_ns()/1000,abs_end_ns/1000):cpu_memory_usage+=_cpu_memory_usage(mem_record[0])device_memory_usage+=_device_memory_usage(mem_record[0])mem_record[1]=Trueis_async=kineto_event.is_async()or(kineto_event.start_thread_id()!=kineto_event.end_thread_id())fe=FunctionEvent(id=kineto_event.correlation_id(),name=_rewrite_name(name=kineto_event.name(),with_wildcard=True),trace_name=_rewrite_name(name=kineto_event.name(),with_wildcard=False),thread=kineto_event.start_thread_id(),start_us=rel_start_ns/1000,end_us=rel_end_ns/1000,fwd_thread=kineto_event.fwd_thread_id(),input_shapes=kineto_event.shapes(),concrete_inputs=kineto_event.concrete_inputs(),kwinputs=kineto_event.kwinputs(),stack=[entryforentryinkineto_event.stack()if_filter_stack_entry(entry)],scope=kineto_event.scope(),use_device=self.use_device,cpu_memory_usage=cpu_memory_usage,device_memory_usage=device_memory_usage,is_async=is_async,sequence_nr=kineto_event.sequence_nr(),device_type=kineto_event.device_type(),device_index=kineto_event.device_index(),device_resource_id=kineto_event.device_resource_id(),flops=kineto_event.flops(),is_user_annotation=kineto_event.is_user_annotation(),)max_evt_id=max(max_evt_id,fe.id)iffe.device_type==DeviceType.CPUandnotfe.is_async:ifself.use_device=="privateuseone":privateuse1_time=kineto_event.privateuse1_elapsed_us()ifprivateuse1_time>0:fe.append_kernel(fe.name,fe.device_index,privateuse1_time)fe.is_legacy=Trueelifself.use_device=="cuda":# Check if we have CUDA time as a fallbackcuda_time=kineto_event.cuda_elapsed_us()ifcuda_time>0:fe.append_kernel(fe.name,fe.device_index,cuda_time)fe.is_legacy=Trueall_function_events.append(fe)corr_id=kineto_event.linked_correlation_id()ifcorr_id>0:ifcorr_idnotindevice_corr_map:device_corr_map[corr_id]=[]device_corr_map[corr_id].append(fe)elifcorr_id==0:frontend_function_events.append(fe)else:raiseRuntimeError(f"Got negative correlation id {corr_id} in profiler post processing")# associate device kernels and device runtime (CPU) with CPU eventsforfeinfrontend_function_events:if(fe.device_type==DeviceType.CPUandnotfe.is_asyncandfe.idindevice_corr_map):forf_evtindevice_corr_map[fe.id]:if(f_evt.device_type==DeviceType.CUDAorf_evt.device_type==DeviceType.PrivateUse1):fe.append_kernel(f_evt.name,f_evt.device_index,f_evt.time_range.end-f_evt.time_range.start,)eliff_evt.device_type==DeviceType.CPU:# make sure that 'thread' of a CPU Kineto (e.g. Device Runtime) event is associated# with the 'thread' of the corresponding linked PyTorch event to properly track# parents and childrenf_evt.thread=fe.threaddefcreateFunctionEventForMemoryEvents(evt):rel_start_ns=evt.start_ns()-trace_start_nsfe=FunctionEvent(id=max_evt_id,name=evt.name(),trace_name=None,# not outputting in the tracethread=evt.start_thread_id(),start_us=rel_start_ns/1000,end_us=rel_start_ns/1000,# no durationfwd_thread=evt.start_thread_id(),input_shapes=[],stack=[],scope=0,# RecordScope::FUNCTIONuse_device=self.use_device,cpu_memory_usage=_cpu_memory_usage(evt),device_memory_usage=_device_memory_usage(evt),is_async=False,sequence_nr=-1,device_type=DeviceType.CPU,device_index=0,)returnfe# output top-level memory eventsformem_recordinmem_records:ifnotmem_record[1]:max_evt_id+=1fe=createFunctionEventForMemoryEvents(mem_record[0])all_function_events.append(fe)foroom_recordinoom_records:max_evt_id+=1fe=createFunctionEventForMemoryEvents(oom_record)all_function_events.append(fe)all_function_events.sort(key=lambdaevt:[evt.time_range.start,-evt.time_range.end])returnall_function_events
[docs]classrecord_function(_ContextDecorator):"""Context manager/function decorator that adds a label to a code block/function when running autograd profiler. Label will only appear if CPU activity tracing is enabled. It is useful when tracing the code profile. Args: name (str): Label assigned to the block of code. node_id (int): ID of node, for distributed profiling. Unset in non-distributed cases. Example: >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD_PROFILER) >>> x = torch.randn((1, 1), requires_grad=True) >>> with torch.autograd.profiler.profile() as prof: ... y = x ** 2 ... with torch.autograd.profiler.record_function("label-z"): # label the block ... z = y ** 3 ... y.backward() ... >>> # xdoctest: +IGNORE_WANT >>> # NOTE: some columns were removed for brevity >>> print(prof.key_averages().table(sort_by="self_cpu_time_total")) ----------------------------------- --------------- --------------- --------------- Name Self CPU total % CPU time avg Number of Calls ----------------------------------- --------------- --------------- --------------- pow 60.77% 47.470us 3 mul 21.73% 25.465us 2 PowBackward0 12.03% 121.891us 1 torch::autograd::AccumulateGrad 2.70% 6.324us 1 label-z 2.13% 12.421us 1 torch::autograd::GraphRoot 0.64% 1.503us 1 ----------------------------------- --------------- --------------- --------------- Self CPU time total: 234.344us CUDA time total: 0.000us """def__init__(self,name:str,args:Optional[str]=None):self.name:str=nameself.args:Optional[str]=args# Whether or not we should run record function's end callbacks when exiting.self.run_callbacks_on_exit:bool=True# TODO: TorchScript ignores standard type annotation here# self.record: Optional["torch.classes.profiler._RecordFunction"] = Noneself.record=torch.jit.annotate(Optional["torch.classes.profiler._RecordFunction"],None)def__enter__(self):self.record=torch.ops.profiler._record_function_enter_new(self.name,self.args)returnselfdef__exit__(self,exc_type:Any,exc_value:Any,traceback:Any):ifnotself.run_callbacks_on_exit:return# Local variable is needed by TorchScript to refine Optional[T] to Trecord=self.recordassertrecordisnotNone# TODO: Too slow with __torch_function__ handling enabled# See https://github.com/pytorch/pytorch/issues/76410ifnottorch.jit.is_scripting():withtorch._C.DisableTorchFunctionSubclass():torch.ops.profiler._record_function_exit._RecordFunction(record)else:torch.ops.profiler._record_function_exit(record)def_call_end_callbacks_on_future(self,fut:Future[Any])->Future[Any]:"""Use for profiling async calls that return a future. Calling this function will extend recording beyond this scope, until the future is satisfied. It is useful for profiling the end to end time of asynchronous calls. This function should only be called once to attach the callback onto the future, and will throw if called multiple times. Args: fut: (torch._C.Future): future for which to schedule callback for. Returns: A future that completes with the value of the passed in future when the profiling callbacks have ran. """# Throw if we have already attached a callback onto the future.ifnotself.run_callbacks_on_exit:raiseRuntimeError("_call_end_callbacks_on_future can only be called once.")# We are scheduling to run this RecordFunction's end callbacks when the# passed in future completes, so don't run end callbacks on exit.self.run_callbacks_on_exit=False# Local variable is needed by TorchScript to refine Optional[T] to Trecord=self.recordassertrecordisnotNone# TODO: Too slow with __torch_function__ handling enabled# See https://github.com/pytorch/pytorch/issues/76410ifnottorch.jit.is_scripting():withtorch._C.DisableTorchFunctionSubclass():profiled_future=(torch.ops.profiler._call_end_callbacks_on_jit_fut._RecordFunction(record,fut))else:profiled_future=torch.ops.profiler._call_end_callbacks_on_jit_fut(record,fut)returnprofiled_future
[docs]classemit_itt:"""Context manager that makes every autograd operation emit an ITT range. It is useful when running the program under Intel(R) VTune Profiler:: vtune <--vtune-flags> <regular command here> The Instrumentation and Tracing Technology (ITT) API enables your application to generate and control the collection of trace data during its execution across different Intel tools. This context manager is to annotate Intel(R) VTune Profiling trace. With help of this context manager, you will be able to see labled ranges in Intel(R) VTune Profiler GUI. .. warning: This context manager should not be called recursively, i.e. at most one instance should be enabled at any given time. Args: enabled (bool, optional): Setting ``enabled=False`` makes this context manager a no-op. Default: ``True``. record_shapes (bool, optional): If ``record_shapes=True``, the itt range wrapping each autograd op will append information about the sizes of Tensor arguments received by that op, in the following format: ``[[arg0.size(0), arg0.size(1), ...], [arg1.size(0), arg1.size(1), ...], ...]`` Non-tensor arguments will be represented by ``[]``. Arguments will be listed in the order they are received by the backend op. Please note that this order may not match the order in which those arguments were passed on the Python side. Also note that shape recording may increase the overhead of itt range creation. Default: ``False`` Example: >>> # xdoctest: +SKIP("Undefined variables") >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD_PROFILER) >>> with torch.autograd.profiler.emit_itt(): ... model(x) """def__init__(self,enabled=True,record_shapes=False):self.enabled=enabledself.entered=Falseself.record_shapes=record_shapesdef__enter__(self):ifnotself.enabled:returnifself.entered:raiseRuntimeError("ITT annotation context manager is not reentrant")self.entered=True_run_on_profiler_start()_enable_profiler(ProfilerConfig(ProfilerState.ITT,self.record_shapes,False,False,False,False,_ExperimentalConfig(),),set(),)returnselfdef__exit__(self,exc_type,exc_val,exc_tb):ifnotself.enabled:return_disable_profiler()_run_on_profiler_stop()returnFalse
[docs]classemit_nvtx:"""Context manager that makes every autograd operation emit an NVTX range. It is useful when running the program under nvprof:: nvprof --profile-from-start off -o trace_name.prof -- <regular command here> Unfortunately, there's no way to force nvprof to flush the data it collected to disk, so for CUDA profiling one has to use this context manager to annotate nvprof traces and wait for the process to exit before inspecting them. Then, either NVIDIA Visual Profiler (nvvp) can be used to visualize the timeline, or :func:`torch.autograd.profiler.load_nvprof` can load the results for inspection e.g. in Python REPL. .. warning: This context manager should not be called recursively, i.e. at most one instance should be enabled at any given time. Args: enabled (bool, optional): Setting ``enabled=False`` makes this context manager a no-op. Default: ``True``. record_shapes (bool, optional): If ``record_shapes=True``, the nvtx range wrapping each autograd op will append information about the sizes of Tensor arguments received by that op, in the following format: ``[[arg0.size(0), arg0.size(1), ...], [arg1.size(0), arg1.size(1), ...], ...]`` Non-tensor arguments will be represented by ``[]``. Arguments will be listed in the order they are received by the backend op. Please note that this order may not match the order in which those arguments were passed on the Python side. Also note that shape recording may increase the overhead of nvtx range creation. Default: ``False`` Example: >>> # xdoctest: +SKIP("undefined variables") >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD_PROFILER) >>> with torch.cuda.profiler.profile(): ... model(x) # Warmup CUDA memory allocator and profiler ... with torch.autograd.profiler.emit_nvtx(): ... model(x) **Forward-backward correlation** When viewing a profile created using :class:`emit_nvtx` in the Nvidia Visual Profiler, correlating each backward-pass op with the corresponding forward-pass op can be difficult. To ease this task, :class:`emit_nvtx` appends sequence number information to the ranges it generates. During the forward pass, each function range is decorated with ``seq=<N>``. ``seq`` is a running counter, incremented each time a new backward Function object is created and stashed for backward. Thus, the ``seq=<N>`` annotation associated with each forward function range tells you that if a backward Function object is created by this forward function, the backward object will receive sequence number N. During the backward pass, the top-level range wrapping each C++ backward Function's ``apply()`` call is decorated with ``stashed seq=<M>``. ``M`` is the sequence number that the backward object was created with. By comparing ``stashed seq`` numbers in backward with ``seq`` numbers in forward, you can track down which forward op created each backward Function. Any functions executed during the backward pass are also decorated with ``seq=<N>``. During default backward (with ``create_graph=False``) this information is irrelevant, and in fact, ``N`` may simply be 0 for all such functions. Only the top-level ranges associated with backward Function objects' ``apply()`` methods are useful, as a way to correlate these Function objects with the earlier forward pass. **Double-backward** If, on the other hand, a backward pass with ``create_graph=True`` is underway (in other words, if you are setting up for a double-backward), each function's execution during backward is given a nonzero, useful ``seq=<N>``. Those functions may themselves create Function objects to be executed later during double-backward, just as the original functions in the forward pass did. The relationship between backward and double-backward is conceptually the same as the relationship between forward and backward: The functions still emit current-sequence-number-tagged ranges, the Function objects they create still stash those sequence numbers, and during the eventual double-backward, the Function objects' ``apply()`` ranges are still tagged with ``stashed seq`` numbers, which can be compared to `seq` numbers from the backward pass. .. warning: The sequence number is thread-local, and some forward functions don't create an associated backward Function object (instead delegating that to sub-functions further down the call chain). For these reasons, the correspondence of stashed sequence numbers in backward Function ``apply()`` ranges with `seq` numbers in forward-pass ranges is not guaranteed to be 1 to 1. The sequence numbers alone may not be enough to fully disambiguate which forward function created which backward Function object. You may need to make a judgment based on analytic knowledge of what the expected correspondence should be. """def__init__(self,enabled=True,record_shapes=False):self.enabled=enabledself.entered=Falseself.record_shapes=record_shapesdef__enter__(self):ifnotself.enabled:returnifself.entered:raiseRuntimeError("NVTX annotation context manager is not reentrant")self.entered=Truetorch.cuda.synchronize()_run_on_profiler_start()_enable_profiler(ProfilerConfig(ProfilerState.NVTX,self.record_shapes,False,False,False,False,_ExperimentalConfig(),),set(),)returnselfdef__exit__(self,exc_type,exc_val,exc_tb):ifnotself.enabled:returntorch.cuda.synchronize()_disable_profiler()_run_on_profiler_stop()returnFalse
[docs]defload_nvprof(path):"""Open an nvprof trace file and parses autograd annotations. Args: path (str): path to nvprof trace """returnEventList(parse_nvprof_trace(path))
[docs]classEnforceUnique:"""Raises an error if a key is seen more than once."""def__init__(self):self.seen=set()
[docs]defsee(self,*key):r""" Observe a key and raise an error if it is seen multiple times. """ifkeyinself.seen:raiseRuntimeError("duplicate key: "+str(key))self.seen.add(key)
[docs]defparse_nvprof_trace(path):importsqlite3conn=sqlite3.connect(path)conn.row_factory=sqlite3.Row# Parse strings tablestrings={}forrinconn.execute("SELECT _id_ as id, value FROM StringTable"):strings[r["id"]]=torch._C._demangle(r["value"])# First, find all functions and create FunctionEvents for themmarker_query=""" SELECT start.id AS marker_id, start.name, start.timestamp AS start_time, end.timestamp AS end_time FROM CUPTI_ACTIVITY_KIND_MARKER AS start INNER JOIN CUPTI_ACTIVITY_KIND_MARKER AS end ON start.id = end.id WHERE start.name != 0 AND end.name = 0 """functions=[]functions_map={}unique=EnforceUnique()forrowinconn.execute(marker_query):unique.see(row["marker_id"])evt=FunctionEvent(id=row["marker_id"],node_id=0,# missing a node_id when calling FunctionEvent. This is just to ensure# that pytorch doesn't crash when creating a FunctionEvent() objectname=strings[row["name"]],start_us=row["start_time"],end_us=row["end_time"],thread=0,)# TODO: find in sqlite databasefunctions.append(evt)functions_map[evt.id]=evt# Now, correlate all kernels with FunctionEventskernel_query=""" SELECT start.id AS marker_id, start.name, start.timestamp, end.timestamp, runtime._id_ AS runtime_id, runtime.cbid, runtime.start AS runtime_start, runtime.end AS runtime_end, kernel.start AS kernel_start, kernel.end AS kernel_end, kernel.name AS kernel_name FROM CUPTI_ACTIVITY_KIND_MARKER AS start INNER JOIN CUPTI_ACTIVITY_KIND_MARKER AS end ON start.id = end.id INNER JOIN CUPTI_ACTIVITY_KIND_RUNTIME as runtime ON (start.timestamp < runtime.start AND runtime.end < end.timestamp) INNER JOIN CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL AS kernel ON kernel.correlationId = runtime.correlationId """unique=EnforceUnique()forrowinconn.execute(kernel_query):unique.see(row["marker_id"],row["runtime_id"])# 211 is cudaKernelLaunch for cuda >= 9.2assertrow["cbid"]==211evt=functions_map[row["marker_id"]]evt.append_kernel(row["kernel_name"],0,row["kernel_end"]-row["kernel_start"])functions.sort(key=lambdaevt:evt.time_range.start)returnfunctions
[docs]classKinetoStepTracker:"""Provides an abstraction for incrementing the step count globally. Previously, we only had one place to mark that a step() has occurred in the program via pytorch profiler step(). We will now add step hooks in the Optimizer class https://github.com/pytorch/pytorch/issues/88446 - This could mean programs that already call profiler.step() every iteration can end up double incrementing step count. - If a model uses multiple optimizers we can also have double or more counting of the step. We fix this by adding a layer of abstraction before calling step() to the kineto library. The idea is to maintain steps per requester in a dict: .. code-block:: { "ProfilerStep": 100, # triggered by profiler step() call "Optimizer1Step": 100, # Optimizer 1 or 2 are just examples, could be SGD, Adam etc "Optimizer2Step": 100, } To figure out the global step count just take the max of dict values (100). If one of the count increments the max will go up. .. code-block:: { "ProfilerStep": 100, "Optimizer1Step": 101, # Optimizer1 got incremented first say "Optimizer2Step": 100, } Then global step count is 101 We only call the kineto step() function when global count increments. NOTE: Please do not use the KinetoStepTracker in modules beside the Optimizer for now. The result could be incorrect increments of the step count. """_current_step=0_step_dict:Dict[str,int]=defaultdict(int)
[docs]@classmethoddefinit_step_count(cls,requester:str):r""" Initialize for a given requester. """cls._step_dict[requester]=cls._current_step
[docs]@classmethoddeferase_step_count(cls,requester:str)->bool:r""" Remove a given requester. """returncls._step_dict.pop(requester,None)isnotNone
[docs]@classmethoddefincrement_step(cls,requester:str)->int:"""Increments the step count for the requester. Additionally if the max over all step counts has incremented then trigger the _kineto_step() returns global step count """ifrequesternotincls._step_dict:cls.init_step_count(requester)cls._step_dict[requester]+=1new_step=max(cls._step_dict.values())ifnew_step>cls._current_step:delta=new_step-cls._current_stepifdelta>1:warn("Profiler step count has increased more than 1 - "f"current_step = {cls._current_step} step dict = {cls._step_dict}")for_inrange(0,delta):_kineto_step()cls._current_step=new_stepreturncls._current_step
[docs]@classmethoddefcurrent_step(cls)->int:r""" Get the latest step for any requester """returncls._current_step
Docs
Access comprehensive developer documentation for PyTorch
To analyze traffic and optimize your experience, we serve cookies on this site. By clicking or navigating, you agree to allow our usage of cookies. As the current maintainers of this site, Facebook’s Cookies Policy applies. Learn more, including about available controls: Cookies Policy.