tjUUddlZddlZddlZddlZddlmZmZddlmZddl m Z ddl m Z m Z ddlmZddlZddlmZgdZGd d eZd Zd Zd ZdZGddZGddZedgdZGdde ZdedefdZdede fdZ!idde"fdde"fddej#fd d ej#fd!d!e"fd"d#e$fd$d%e$fd&d'e$fd(d(ej#fd)d)e"fd*d+e"fd,d-e"fd.d.e"fd/d/e"fd0d0e"fd1d1e"fd2d2e"fid3d4e$fd5d6efd7d7efd8d9e"fd:d;e"fdd?efd@dAe"fdBdCe"fdDdEe"fdFdGefdHdIefdJdKefdLdMe"fdNdOe"fdPdQe"fdRdSe"fdTdUe!fiZ%e&ee'eeege fffe(dV<dWZ)GdXdYeZ*GdZd[eZ+Gd\d]eZ,Gd^d_Z-d`Z.daZ/dbZ0dcZ1dldeZ2 dmdjZ3dkZ4dS)nN) defaultdict namedtuple)Callable) attrgetter)Any NamedTuple) deprecated) DeviceType) EventList EventMetadataFormattedTimesMixinIntervalKernel FunctionEventFunctionEventAvg StringTable MemRecordsAccceZdZdZfdZdZdZdZdZdZ e dZ ddZ dZ dZdedefdZ ddZdZxZS)r a> A list of profiling events with helper methods for analysis and visualization. EventList extends the standard Python list to provide specialized methods for working with profiling events (FunctionEvent or FunctionEventAvg objects). It includes utilities for aggregating statistics, formatting output tables, and exporting profiling data. This class is typically returned by profiler methods and should not be instantiated directly by users. Args: *args: Standard list arguments. use_device (str, optional): Device type for profiling ("cuda", "xpu", etc.). profile_memory (bool, optional): Whether memory profiling was enabled. Default: False. with_flops (bool, optional): Whether to include FLOP counts. Default: False. Attributes: _use_device (str): Device type being profiled. _profile_memory (bool): Whether memory profiling is enabled. _with_flops (bool): Whether FLOP counting is enabled. _tree_built (bool): Whether the event tree structure has been built. Key Methods: table(...): Format events as a table string for display. export_chrome_trace(path): Export to Chrome tracing format. export_stacks(path, metric): Export stack traces with metrics. key_averages(...): Compute averaged statistics grouped by operation name. total_average(): Compute aggregate totals across all events (sums, not averages). Properties: self_cpu_time_total: Sum of self CPU time across all events. Example:: import torch from torch.profiler import profile, ProfilerActivity with profile(activities=[ProfilerActivity.CPU]) as prof: x = torch.randn(100, 100) y = torch.matmul(x, x) # EventList is returned by prof.events() events = prof.events() # Display as formatted table print( events.table( sort_by="cpu_time_total", row_limit=20, top_level_events_only=False ) ) # Export to Chrome tracing format events.export_chrome_trace("trace.json") # Get averaged statistics avg_events = events.key_averages() print(avg_events.table()) # Export stack traces events.export_stacks("stacks.txt", "self_cpu_time_total") See Also: - :class:`FunctionEvent`: Individual profiling event - :class:`FunctionEventAvg`: Averaged profiling statistics - :meth:`table`: Format events as a readable table - :meth:`key_averages`: Aggregate events by operation name c|dd}|dd}|dd}tj|i|||_||_d|_||_dS)N use_deviceprofile_memoryF with_flops)popsuper__init__ _use_device_profile_memory _tree_built _with_flops)selfargskwargsrrr __class__s Z/var/www/html/CA-Chatbot/venv/lib/python3.11/site-packages/torch/autograd/profiler_util.pyrzEventList.__init__bsZZ d33 $4e<<ZZ e44 $)&)))%- %c|||d|_dS)NT)_populate_cpu_children_remove_dup_nodes_set_backward_stacktracesrr s r$ _build_treezEventList._build_treemsF ##%%%     &&(((r%c*|SN)tabler*s r$__str__zEventList.__str__sszz||r%c ttt|D]}||j||jj||jkrt||jjdkrq||j||j_||j||j_||jD]}||j|_|tdkrdSfdt|D}| | |J)NTrc"g|] \}}|v | Sr3).0indev to_deletes r$ z/EventList._remove_dup_nodes..s'RRRwsBS =Q=Q=Q=Q=Qr%) setrangelen cpu_parentname cpu_childrenkernelsadd enumerateclearextend)r idxchnew_evtsr7s @r$r(zEventList._remove_dup_nodesvs1 "ISYY'' ' 'I(4S ,1T#Y^CCDI0=>>!CC8.s=   < %(Oz~$E$E $E$E$Er%threadkeyc|j|jfSr-)rNnode_idevents r$z2EventList._populate_cpu_children..su|U]&Cr%c4|jj|jj gSr-) time_rangestartendrSs r$rUz2EventList._populate_cpu_children..s5#3#9E.bw_parents/{ta  y000r%z1Expected fwd_thread to be set for backward parent)stack sequence_nrrN fwd_threadr_get)r fwd_stacksrMtprns @r$r)z#EventList._set_backward_stacktracess 1 1 1 1 1  . .Cy~~%#)*?_cj1J&&$'IJqM 2 2C #A}<'(K]AL1&NN1b11  2 2r%c4td|DS)Nc3$K|] }|jV dSr-)self_cpu_time_totalr4rTs r$ z0EventList.self_cpu_time_total..s%??5,??????r%)sumr*s r$rxzEventList.self_cpu_time_totals??$??????r%NdK7PFc Jt||||||||j|j|| S)aPrint an EventList as a nicely formatted table. Args: sort_by (str, optional): Attribute used to sort entries. By default they are printed in the same order as they were registered. Valid keys include: ``cpu_time``, ``cuda_time``, ``xpu_time``, ``cpu_time_total``, ``cuda_time_total``, ``xpu_time_total``, ``cpu_memory_usage``, ``cuda_memory_usage``, ``xpu_memory_usage``, ``self_cpu_memory_usage``, ``self_cuda_memory_usage``, ``self_xpu_memory_usage``, ``count``. top_level_events_only(bool, optional): Boolean flag to determine the selection of events to display. If true, the profiler will only display events at top level like top-level invocation of python `lstm`, python `add` or other functions, nested events like low-level cpu/cuda/xpu ops events are omitted for profiler result readability. time_unit(str, optional): A time unit to be used for all values in the table. Valid options are: ``s``, ``ms`` and ``us``. Returns: A string containing the table. ) sort_by row_limitmax_src_column_widthmax_name_column_widthmax_shapes_column_widthheaderrrtop_level_events_only time_unit) _build_tablerr) r rrrrrrrrs r$r.zEventList.tablesC@ !5"7$;/'"7    r%cddl}|jsdn|j}t|d5}d}|d|D]}|j |d|j|jj|j|j s|j nd|j d|j d |j D]@}|d |jd |jjd |j d |d|d |dz }At|dkrD||dz |j||dddddS#1swxYwYdS)zExport an EventList as a Chrome tracing tools file. The checkpoint can be later loaded and inspected under ``chrome://tracing`` URL. Args: path (str): Path where the trace will be written. rNcudaw[zc{{"name": "{}", "ph": "X", "ts": {}, "dur": {}, "tid": {}, "pid": "CPU functions", "args": {{}}}}, z " node_id:z , thread_id:z "z {"name": "z", "ph": "s", "ts": z , "tid": z , "pid": "CPU functions", "id": z, "cat": "cpu_to_z", "args": {}}, r1])osropenwrite trace_nameformatrWrX elapsed_us is_remoterNrRr?r;seektellSEEK_SETtruncate)r pathr device_namefnext_idrM_s r$export_chrome_tracezEventList.export_chrome_trace s  $($4Jff$:J $__+ G GGCLLL! !! !>)' (.v,1133"}R Q#+QQ3:QQQ (( !!AGG(cn((!$!5((#&*(( ") (( +6 (((   qLGG!4yy1}}qvvxx!|R[111 GGCLLLW+ + + + + + + + + + + + + + + + + + sE E;;E?E?c gdS)N)rxself_cuda_time_totalself_xpu_time_totalself_privateuse1_time_totalr3r*s r$supported_export_stacks_metricsz)EventList.supported_export_stacks_metricsXs    r%rmetricc ||vr1tdt|ztdd}t |d5}|D]}|jrt |jdkrt||ddddd d}t|dkrxd }t|jD]}|| |z }|d z } |dd d ztt|z}| |dz ddddS#1swxYwYdS)Nzmetric should be one of: z ; ____rrrdevicexpu privateuse1;rZ  ) r ValueErrorstr maketransrror;getattrreplaceintreversed translater) r rrtranslate_tablerrM metric_value stack_strentrys r$ export_stackszEventList.export_stacks`s ==?? ? ?+d::<<==> --&99 $__ 2 2 29 2SY!!3!3#*vx88 11 99 $$L <((1,,$& %-ci%8%8--E%)I)III%,II$-crcNS$83s.get_keysEIEM""E%&&EO$$E,-- C& . 3,---$ 4 3u122333!##u{#4$4#455:: r%rrrNr)rr_rrrrr@r valuesrrrrorr)r rrrstatsravg_listrMs @r$ key_averageszEventList.key_averagesxs&.  G :EEU9V9V  38_      $  C .0@BX c#hhhh LLNN'/'      ' 'C "3#3"34CI( &#% ) '$&!r%cRt}|D]}||z }d|_d|_|S)a,Compute aggregate statistics across all events. Accumulates statistics from all events into a single FunctionEventAvg object. This is primarily useful for computing total metrics (total CPU time, total memory usage, etc.) across the entire profiling session, regardless of operation type. Note: This sums up times and counts across ALL different operations, so the "average" metrics (like cpu_time) represent the average time per operation call across the entire session, mixing all operation types together. For per-operation averages, use :meth:`key_averages` instead. Returns: FunctionEventAvg: A single aggregate object with key="Total" containing accumulated statistics. NTotal)rrP)r total_statrMs r$ total_averagezEventList.total_averages>&&''  " "C # J!JNN r%)Nr|r}r~rNFN)FrF)__name__ __module__ __qualname____doc__rr+r/r(r'r)propertyrxr.rrrrrr __classcell__)r#s@r$r r sUBBH & & & & &   """0D-D-D-L2224@@X@  "#, , , , \666p   2#2s22224$$ BBBBHr%r cRd}d}||kr ||z ddS||kr ||z ddS|ddS)+Define how to format time in FunctionEvent..A@@.3fsmsusr3)time_us US_IN_SECONDUS_IN_MSs r$ _format_timersb"LH,L(/////(H$,,,,,    r%cZ|dkr|dkrtd|dS|dz|z ddS)rrzExpected time_us == 0 but got NaNgY@.2f%)r_)r total_time_uss r$_format_time_sharersM a<< !K'!K!KLL Luo - 4 4 4 44r%cd}d|z}d|z}t||kr |dz|z ddSt||kr |dz|z ddSt||kr |dz|z ddSt|dzS)z&Return a formatted memory size string.i?rz GBz MBz KBz B)absr)nbytesKBMBGBs r$_format_memoryrs B B B 6{{b3,#,,,,, V  3,#,,,,, V  3,#,,,,,6{{T!!r%c(tfdS)Nc>tt|Sr-)rr)r r=s r$rUz!_attr_formatter..sgdD.A.A!B!Br%)rr=s`r$_attr_formatterrs BBBB C CCr%ceZdZdZedZedZedZedZedZ edZ e dZ e d Z e ed e d Zd S)r z{Helpers for FunctionEvent and FunctionEventAvg. The subclass should define `*_time_total` and `count` attributes. cpu_time device_timecpu_time_totaldevice_time_totalrxself_device_time_totalc@|jdkrdnd|jz|jz SNrgr)countrr*s r$rzFormattedTimesMixin.cpu_time s%jAooss31D+Dtz+QQr%c@|jdkrdnd|jz|jz Sr)rrr*s r$rzFormattedTimesMixin.device_times%jAooss31G+G$*+TTr%z<`cuda_time` is deprecated, please use `device_time` instead.categoryc|jSr-)rr*s r$ cuda_timezFormattedTimesMixin.cuda_times r%N)rrrrr cpu_time_strdevice_time_strcpu_time_total_strdevice_time_total_strself_cpu_time_total_strself_device_time_total_strrrrr FutureWarningrr3r%r$r r s #?:..L%om44O()9::+O,?@@-o.CDD!01I!J!J RRXRUUXUZF  X    r%r ceZdZdZdZdS)rc"||_||_dSr-)rXrY)r rXrYs r$rzInterval.__init__s r%c |j|jz S)z4 Returns the length of the interval )rYrXr*s r$rzInterval.elapsed_us!sx$*$$r%N)rrrrrr3r%r$rrs2%%%%%r%rr)r=rdurationcheZdZUedzed<edzed<eedzed<eedzed<edzed<edzed<edzed<eee fdzed <edzed <edzed <edzed <edzed <edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed<edzed <edzed!<edzed"<edzed#<e dzed$<dS)%r Nregisters_per_thread shared_memorygridblockpriority blocks_per_sm warps_per_sm occupancyest_occupancy_pctqueuedgraph_id graph_node_idstreamcontextchannel channel_typebytesbandwidth_gb_scollective_namedtype in_msg_nelemsout_msg_nelems in_split_sizeout_split_sizeglobal_rank_startglobal_rank_stride group_sizeprocess_group_nameprocess_group_desc group_ranksranksrc_rankdst_rankseqrI) rrrr__annotations__listfloatdictrrboolr3r%r$r r +sR*$$$: s)d  9t Dj4<$,CH~$$$$t|### $JDj: $J 4Z 4Z* :DL   4Z ::$J:$JTz!!!d """d d """d """t *DjDj tOOOTkr%r vrc,|dS)N"stripr8s r$_to_strr>Ts 773<<r%c |dvS)N)1truer3r=s r$_to_boolrBXs r%zregisters per threadrz shared memoryrrrrz blocks per SMrz warps per SMrzest. achieved occupancy %rrrzgraph idrz graph node idrrrrr r!zmemory bandwidth (GB/s)r"zCollective namer#r$z In msg nelemsr%zOut msg nelemsr&z In split sizer'zOut split sizer(zGlobal rank startr)zGlobal rank strider*z Group sizer+zProcess Group Namer,zProcess Group Descriptionr-zProcess Group Ranksr.Rankr/zSrc Rankr0zDst Rankr1Seqr2zIs asynchronized oprI_EVENT_METADATA_KEYSci}d}tD]5\}\}}||}|||||<d}0d||<6|r tdi|ndS)NFTr3)rEitemsrrr ) extra_metafields any_populated kineto_key field_nameconvertr8s r$_build_metadatarNsFM-A-G-G-I-I&&) )Z NN: & & =!(F:  MM!%F:  &3 == " "6 " " "=r%c$eZdZdZddddddddddddejdddddddddddddddddddddddf$dZdZdZd Z e d Z e d Z e e d e dZe e de dZe dZe dZe dZe e de dZe dZe e de dZe dZdZdS)raProfiling information about a single function. FunctionEvent records the execution of a single operation during profiling. These events are obtained from the profiler/kineto and contain detailed timing and memory usage information. .. note:: FunctionEvent objects are typically created by the profiler/kineto and should not be instantiated directly by users. Access them through the profiler's output. Attributes: id (int): Unique identifier for this event. node_id (int): Node identifier for distributed profiling (-1 if not applicable). name (str): Name of the profiled function/operator. overload_name (str): Overload name for the operator (requires _ExperimentalConfig(capture_overload_names=True) set). trace_name (str): Same as name, just changes ProfilerStep* to ProfilerStep# time_range (Interval): Time interval containing start and end timestamps in microseconds. thread (int): Thread ID where the operation started. fwd_thread (int): Thread ID of the corresponding forward operation. kernels (List[Kernel]): List of device kernels launched by this operation. count (int): Number of times this event was called (usually 1). cpu_children (List[FunctionEvent]): Direct CPU child operations. cpu_parent (FunctionEvent): Direct CPU parent operation. input_shapes (List[List[int]]): Shapes of input tensors (requires record_shapes=True). For plain tensor inputs, each entry is a list of dimensions (e.g. ``[16, 16]``). TensorList inputs are represented as an empty list ``[]``; use ``structured_input_shapes`` to get per-element shapes for TensorList inputs. concrete_inputs (List[Any]): Concrete input values (requires record_shapes=true). kwinputs (Dict[str, Any]): Keyword arguments (requires record_shapes=true). stack (List[str]): Python stack trace where the operation was called (requires with_stack=true). scope (int): at::RecordScope identifier (0=forward, 1=backward, etc.). use_device (str): Device type being profiled ("cuda", "xpu", etc.). cpu_memory_usage (int): CPU memory allocated in bytes. device_memory_usage (int): Device memory allocated in bytes. is_async (bool): Whether this is an asynchronous operation. is_remote (bool): Whether this operation occurred on a remote node. sequence_nr (int): Sequence number for autograd operations. device_type (DeviceType): Type of device (CPU, CUDA, XPU, PrivateUse1, etc.). device_index (int): Index of the device (e.g., GPU 0, 1, 2). device_resource_id (int): Resource ID on the device (ie. stream ID). is_legacy (bool): Whether this is from the legacy profiler. flops (int): Estimated floating point operations. is_user_annotation (bool): Whether this is a user-annotated region. metadata_json (str): Deprecated. Use event_metadata instead. event_metadata (EventMetadata): Additional metadata in structured format. structured_input_shapes (List[List[int] | List[List[int]]]): Like ``input_shapes`` but distinguishes TensorList inputs. Plain tensor inputs are ``List[int]``; TensorList inputs are ``List[List[int]]`` containing one shape per tensor in the list. Matches the ``"Input Dims"`` field in the Chrome trace JSON. structured_input_strides (List[List[int] | List[List[int]]]): Strides of input tensors in the same format as ``structured_input_shapes`` (requires record_shapes=True). Properties: cpu_time_total (float): Total CPU time in microseconds. device_time_total (float): Total device (CUDA/XPU/etc) time in microseconds. self_cpu_time_total (float): CPU time excluding child operations. self_device_time_total (float): Device time excluding child operations. self_cpu_memory_usage (int): CPU memory usage excluding child operations. self_device_memory_usage (int): Device memory usage excluding child operations. cpu_time (float): Average CPU time per call. device_time (float): Average device time per call. key (str): Key used for grouping events (usually same as name). See Also: - :class:`torch.profiler.profile`: Context manager for profiling - :class:`EventList`: List container for FunctionEvent objects with helper methods - :class:`FunctionEventAvg`: Averaged statistics over multiple FunctionEvent objects NrFrZc*||_||_||_||_||_t |||_||_||_g|_ d|_ g|_ d|_ ||_ ||_||_| |_| |_| |_| |_| |_||_||_||_||_||_||n||_||_||_||_||_||_d|_ d|_!d|_"||_#||_$||_%| |_&|!|_'|"|_(|#rtS|#nd|_*|$|_+|%|_,|&|_-|'|_.|(|_/|)|_0dS)Nr1rZ)1idrRr=rrrrWrNrqr?rr>r<rconcrete_inputskwinputsrormrcpu_memory_usagedevice_memory_usagerIrrprJ device_indexdevice_resource_idrflopsris_python_function activity_typeself_cpu_percenttotal_cpu_percenttotal_device_percent_metadata_jsonflow_id flow_type flow_start external_idlinked_correlation_idrNevent_metadatastructured_input_shapesstructured_input_strides input_dtypes python_idpython_parent_idpython_module_id)*r rQr=rNstart_usend_usrrqrrormrrTrUrIrrprRrJrVrWrrXrrRrSrrYrZ metadata_jsonr_r`rarbrcrHrerfrgrhrirjs* r$rzFunctionEvent.__init__sX#  "/)$,Xv$>$>! &0%'  1304-9*9(0    &0%5(; & ( +'2!-(0FF6H  )!& /A(:)6 "!#$&!+#* %.'1 +*?"+5 ?OJ ' ' '4 .E$.F%'3'%5%5r%c|jtjkrtd|jt |||dS)NExpected device_type to be CPU)rJr rKr_r?rar)r r=rrs r$ append_kernelzFunctionEvent.append_kernelAsI  z~ - - !ABB B F4::;;;;;r%c|jtjkrtdt |t std|jtjkrtd|j|dS)zAppend a CPU child of type FunctionEvent. One is supposed to append only direct children to the event to have correct self cpu time being reported. roz$Expected child to be a FunctionEventz$Expected child device_type to be CPUN)rJr rKr_ isinstancerr>ra)r childs r$r^zFunctionEvent.append_cpu_childFs  z~ - - !ABB B%// I !GHH H    . . !GHH H   '''''r%c|jtjkrtdt |t std|jtjkrtd||_dS)a$Set the immediate CPU parent of type FunctionEvent. One profiling FunctionEvent should have only one CPU parent such that the child's range interval is completely inside the parent's. We use this connection to determine the event is from top-level op or not. roz%Expected parent to be a FunctionEventz%Expected parent device_type to be CPUN)rJr rKr_rrrr<)r ris r$r`zFunctionEvent.set_cpu_parentTso  z~ - - !ABB B&-00 J !HII I   / / !HII I r%c|js|jtjkrdS|jt d|jDz S)Nrc3$K|] }|jV dSr-)rTr4rss r$rzz6FunctionEvent.self_cpu_memory_usage..is6+ + ',E "+ + + + + + r%)rIrJr rKrTr{r>r*s r$self_cpu_memory_usagez#FunctionEvent.self_cpu_memory_usagees[ = D, >>1$s+ + 040A+ + + ( (   r%c|js|jtjkrdS|jt d|jDz S)Nrc3$K|] }|jV dSr-)rUrws r$rzz9FunctionEvent.self_device_memory_usage..qs6. . */E %. . . . . . r%)rIrJr rKrUr{r>r*s r$self_device_memory_usagez&FunctionEvent.self_device_memory_usagems[ = D, >>1'#. . 373D. . . + +   r%z<`metadata_json` is deprecated. Use `event_metadata` instead.rc|jSr-)r^r*s r$rmzFunctionEvent.metadata_jsonus ""r%zO`self_cuda_memory_usage` is deprecated. Use `self_device_memory_usage` instead.c|jSr-r{r*s r$self_cuda_memory_usagez$FunctionEvent.self_cuda_memory_usage}s ,,r%cb|jtjkr|jSdSr)rJr rKrWrr*s r$rzFunctionEvent.cpu_time_totals+  z~ - -?--// /1r%c|js|jtjkrdS|jt d|jDz S)Nrc3$K|] }|jV dSr-)rrws r$rzz4FunctionEvent.self_cpu_time_total..s6) ) %*E ) ) ) ) ) ) r%)rIrJr rKrr{r>r*s r$rxz!FunctionEvent.self_cpu_time_totals[ = D, >>1"S) ) .2.?) ) ) & &   r%c|js|jsdS|jtjkrb|js=t d|jDt d|jDzSt d|jDS|jtj tj tj tj tj fvrtd|j|jS)Nrc3$K|] }|jV dSr-rr4kinfos r$rzz2FunctionEvent.device_time_total..$DDe5>DDDDDDr%c3$K|] }|jV dSr-r)r4rEs r$rzz2FunctionEvent.device_time_total..s>KK-/B(KKKKKKr%c3$K|] }|jV dSr-rrs r$rzz2FunctionEvent.device_time_total..rr%HExpected device_type to be CUDA, PrivateUse1, MTIA, HPU or XPU, but got )rIrrJr rKrr{r?r>CUDA PrivateUse1MTIAHPUXPUr_rWrr*s r$rzFunctionEvent.device_time_totals =  1  z~ - -> EDDt|DDDDDsKK373DKKKHH DDt|DDDDDD& (%q_c_oqq?--// /r%zA`cuda_time_total` is deprecated. Use `device_time_total` instead.c|jSr-rr*s r$cuda_time_totalzFunctionEvent.cuda_time_totals %%r%cT|js|jsdS|jtjkr&|jt d|jDz S|jtjtj tj tj tj fvrtd|j|jS)Nrc3$K|] }|jV dSr-rrws r$rzz7FunctionEvent.self_device_time_total..s600,1'000000r%r)rIrrJr rKrr{r>rrrrrr_r*s r$rz$FunctionEvent.self_device_time_totals =  1  z~ - -)C00595F000-- & (%q_c_oqq) )r%zK`self_cuda_time_total` is deprecated. Use `self_device_time_total` instead.c|jSr-rr*s r$rz"FunctionEvent.self_cuda_time_totals **r%c|jSr-rr*s r$rPzFunctionEvent.keys yr%c2|j}|j}|j}dgd|jd|jd|jd|jd|jd|j d|j j d |j j d td |jDd |d |d|jd|jdt|jd|jd |d|d|jd|jd|jd|jdS)Nrz.s I I Ie I I Ir%r_time=z thread= input_shapes= cpu_memory_usage=_memory_usage=z is_async=z is_remote=z seq_nr=z is_legacy=>)rrrUjoinrQr=rrJrRrrWrXrYrr>rNrrTrIrrpr)r rrrUs r$__repr__zFunctionEvent.__repr__so * "6 y y y y y y y y y y  y y y y$J\ y y y y+ y y y y6:l y y y yNRN_ y y y y- y y y y7;7J y y y y I It7H I I IJJ y y y yNY y y y yal y y y yI  y y y y(,{  y y y yCFdFWBXBX  y y y y !% 5  y y y y 9D  y y y y Tg  y y y y   y y y y 37.  y y y y KOJZ  y y y y hlgu  y y y y y y r%)rrrrr rKrrpr^r`rrxr{r r rmrrrxrrrrrPrr3r%r$rrsDDZN   $!%Ug6g6g6g6R<<< ( ( ( ! ! !"  X   X ZF## X #ZY-- X -X   X 00X02ZK&& X &**X*(ZU++ X +X      r%rc,eZdZdZddZdZdZdZdS) ra Averaged profiling statistics over multiple FunctionEvent objects. FunctionEventAvg aggregates statistics from multiple FunctionEvent objects with the same key (typically same operation name). This is useful for getting average performance metrics across multiple invocations of the same operation. This class is typically created by calling :meth:`EventList.key_averages()` on a profiler's event list. Attributes: key (str): Grouping key for the events (typically operation name). count (int): Total number of events aggregated. node_id (int): Node identifier for distributed profiling (-1 if not applicable). is_async (bool): Whether the operations are asynchronous. is_remote (bool): Whether the operations occurred on a remote node. use_device (str): Device type being profiled ("cuda", "xpu", etc.). cpu_time_total (int): Accumulated total CPU time in microseconds. device_time_total (int): Accumulated total device time in microseconds. self_cpu_time_total (int): Accumulated self CPU time (excluding children) in microseconds. self_device_time_total (int): Accumulated self device time (excluding children) in microseconds. input_shapes (List[List[int]]): Input tensor shapes (requires record_shapes=true). overload_name (str): Operator overload name (requires _ExperimentalConfig(capture_overload_names=True) set). stack (List[str]): Python stack trace where the operation was called (requires with_stack=true). scope (int): at::RecordScope identifier (0=forward, 1=backward, etc.). cpu_memory_usage (int): Accumulated CPU memory usage in bytes. device_memory_usage (int): Accumulated device memory usage in bytes. self_cpu_memory_usage (int): Accumulated self CPU memory usage in bytes. self_device_memory_usage (int): Accumulated self device memory usage in bytes. cpu_children (List[FunctionEvent]): CPU child events. cpu_parent (FunctionEvent): CPU parent event. device_type (DeviceType): Type of device (CPU, CUDA, XPU, PrivateUse1, etc.). is_legacy (bool): Whether from legacy profiler. flops (int): Total floating point operations. is_user_annotation (bool): Whether this is a user-annotated region. Properties: cpu_time (float): Average CPU time per invocation. device_time (float): Average device time per invocation. See Also: - :class:`EventList.key_averages`: Method that creates FunctionEventAvg objects - :class:`FunctionEvent`: Individual profiling event - :class:`EventList`: Container for profiling events rNc\d|_d|_d|_d|_d|_d|_d|_d|_d|_d|_ d|_ d|_ d|_ d|_ d|_d|_d|_d|_d|_d|_t(j|_d|_d|_dS)NrF)rPrrRrIrrrrrxrrrrormrTrUrxr{r>r<r rKrJrrXr*s r$rzFunctionEventAvg.__init__s#  # $&*#$&'() +,#48)-"& !% %&() *+"-.%8<04'1~$ r%c|j|j|_|j|_|j|_|j|_|j|_|j|_|j|_|j|_|j|_|j |_ |j |_ |j |_ |j |_ |j |_ t|tt fst#d|j|jkrt#d|jd|j|xj|jz c_|xj|jz c_|xj|jz c_|xj|jz c_|xj|jz c_|xj|jz c_|xj|jz c_|xj|jz c_|xj|jz c_|j |j|_n|j|xj|jz c_|S)Nz8Expected other to be a FunctionEvent or FunctionEventAvgz Expected keys to match, but got z vs )rPrRrIrr<r>rrrormrJrrrrrrrr_rrrxrrTrUrxr{rrXr others r$r@zFunctionEventAvg.add-s 8 yDH =DL!NDM"_DN#.DO % 2D !&!4D  % 2D DJDJ$0D "_DN#.DO&+&>D #%-1A!BCC  J  9  L59LL$(LL  u33 %"99   E$==   ##u'CC## !77   E$==   ""e&AA"" %%)GG%% ek! : DJJ [ $ JJ%+ %JJ r%c,||Sr-)r@rs r$__iadd__zFunctionEventAvg.__iadd__Zsxxr%c|jsdn|j}|j}|j}|j}d|jd|jd|jd|d|d|d|dt|jd |j d|d |d S) NrzZ d dfjfw d d  d d(8 d d;F d dNY d dilmqm~ii d d $ 5 d d8C d dS` d d d r%)rN)rrrrrr@rrr3r%r$rrsa++Z2+++Z      r%rceZdZdZdS)rct|dkrtj|n|||<||Srl)r;torch_C _demangle)r rPs r$ __missing__zStringTable.__missing__js:033xx!||EH&&s+++S Cyr%N)rrrrr3r%r$rris#r%rceZdZdZdZdZdS)rz=Acceleration structure for accessing mem_records in interval.c||_g|_g|_t|dkr>t dt |D}t |\|_|_dSdS)NrcLg|]!\}}|d|f"Sr)start_ns)r4irs r$r8z*MemRecordsAcc.__init__..zs-RRR41a1Q4==??A.RRRr%) _mem_records _start_nses_indicesr;r[rAzip)r mem_recordstmps r$rzMemRecordsAcc.__init__uso'&(#% {  a  RR9[;Q;QRRRSSC.13i +D dmmm r%c#Ktj|j|}tj|j|}t ||D]}|j|j|VdS)z: Return all records in the given interval N)bisect bisect_leftr bisect_rightr:rr)r rend_ns start_idxend_idxrs r$ in_intervalzMemRecordsAcc.in_interval}ss&t'7BB %d&6??y'** 6 6A#DM!$45 5 5 5 5 6 6r%N)rrrrrrr3r%r$rrrs8GG88866666r%rcBgd}tfd|DS)N))autograd/__init__ _make_grads)rbackward)z torch/tensorr)_internal/common_utils prof_callable)rprof_func_call)rprof_meth_callc3BK|]}|dvo |dv VdS)rr1Nr3)r4rrs r$rzz&_filter_stack_entry..s;OOAaDEM3adem4OOOOOOr%)all)rfiltered_entriess` r$_filter_stack_entryrs; OOOO>NOOO O OOr%z[memory]z [OutOfMemory]c2ttddddddg}||vS)Nz profiler::_record_function_enterz$profiler::_record_function_enter_newzprofiler::_record_function_exitz aten::is_leafzaten::output_nrzaten::_version)MEMORY_EVENT_NAMEOUT_OF_MEMORY_EVENT_NAME)r=filtered_out_namess r$ _filter_namers2  *.)  % %%r%Fcdt}||}|r|drd}|S)Nz ProfilerStep#z ProfilerStep*)r startswith)r= with_wildcard string_tables r$ _rewrite_namers;==L  D# ??? + + #"D Kr%r|r}r~rc 01234t|dkrdStd|D} td|D} |dj} | s| rtdtd|D}td|D}&t t |fd d | || }t d |Ddz}|t||}t d|Ddz}|t||}d}|}d}d|D}t|dk}|r.t d|Ddz}|t||}dg}|r|d|gdz }| | nd}| r&| d|d|d|d|dg|r9| ddg| r | r| |dd|dg|dtd|D}|r|d d!0dg4dg10 g2d80124fd# }d$}|||r |||d%|zdD] }|||r |d&|||r"|d'||d()|r`d*|D}t|dkr?|t|\}}|d+|||nd,}4d} 1d}!2d}"d}g33fd-}#d}$d}%|D]y}&|$|&j z }$|&j tjkr|&jr |%|&jz }%3|&j tjtjtjtjfvr|&js |%|&jz }%z||#d.|"z|#|| r|#d.|"z|#d/|#|!|#| j||#|!d0}'d1}(d})|D]}&|)|krnv| r|&j|)d%z })|&j}*|&t|*|d2z kr|*d|d2z d3z}*t3|&j |$|&_|&jst3|&j|$nd|&_|*g}+|r5|&j},|&t|,|d2z kr|,d|d2z d3z},|+|,gz }+|+|&j|(|&j |&j| |&j|(|&j|&j | |(|&j!|&j"| gz }+| rwt3|&j|%|&_#|+ |(|&j|&j$| |&j#|(|&j%|&j&| |(|&j'|&j(| g|rz|+ tS|&j*tS|&j+g| r=| r;|+ tS|&j,tS|&j-g|+|&j.|r|+|&j/|r/|+ta|&j1d||r@|&j2dkr|+d4n|+|&j2|zd5|rFd}-t|&j3dkr|'|&j3d|}-|+|-|#| j|+|rqdgt|d%z z}.|&j3d%dD]#}/|#| j|.|'|/|gz$|.d|#| j|.|#|!|#d6|(|$ti|$| | r?|#d| | ndd7|(|%ti|%| d53S)9zUPrint a summary of events (which can be a list of FunctionEvent or FunctionEventAvg).rrc3,K|]}|jdkVdSrNrrys r$rzz_build_table..s*OOu%6:OOOOOOr%c3,K|]}|jdkVdSrr~rys r$rzz_build_table..s*PP7!;PPPPPPr%z9use_device is None, but there is device performance data.c3XK|]%}|jduot|jdkV&dSr)rr;rys r$rzz_build_table..sS   4 ' GC0B,C,Ca,Gr%c3XK|]%}|jduot|jdkV&dSr)rr;rys r$rzz_build_table..sS   D ( IS1D-E-E-Ir%Nct|ddddddS)Nrrrr)rr)rMrs r$rUz_build_table..sCOOFH55WUH--W]H55 !!r%T)rPreverserc3>K|]}t|jVdSr-)r;rPrLs r$rzz_build_table..s*;;SCLL;;;;;;r%c3XK|]%}tt|jV&dSr-)r;rrrLs r$rzz_build_table..s5KKSc#c&6"7"788KKKKKKr% cZg|](}|j t|jdk!|j)Sr)ror;rLs r$r8z _build_table..s;sy'<SYRSASAS ASASASr%c3HK|]}td|DVdS)c34K|]}t|VdSr-r;)r4rs r$rzz)_build_table...s(225CJJ222222r%N)max)r4ros r$rzz_build_table..s9GGu22E22222GGGGGGr%Namez Overload Name)z Self CPU %zSelf CPUz CPU total %z CPU totalz CPU time avgNonezSelf z %z totalz time avgzCPU Memz Self CPU Memz Memz # of Callsc3,K|]}|jdkVdS)rZN)rRrLs r$rzz_build_table..$s)==s*======r%zNode IDrrcdxxd|zt|zdzdzzz cc<dxxd|zdzzz cc<dxx|zz cc<dS)Nrz{: }r-)r)paddingtext_dir SPACING_SIZEheader_sep_lstline_length_lstrow_format_lsts r$ add_columnz _build_table..add_column-sq H s7|| +c 1S<5G H  qS7]cL.@AAg 44r%c gd}|dkrtd|tdttj|dz t t |dz }|dkr|t |ks"tdt |d|tdtj|d z|t|fS) N)FLOPsKFLOPsMFLOPsGFLOPsTFLOPsPFLOPsrz'Expected flops to be positive, but got r1z&Expected log_flops to be in range [0, z ), but got g) r_rminmathlog10r5r;powfloorr)rX flop_headers log_flopss r$auto_scale_flopsz&_build_table..auto_scale_flops4s    A:: !R5!R!RSS S3tz%0014eC .Ws!BBB3CIMMSYMMMr%zTotal Fc\|ddS)Nr)ra)rresults r$raz_build_table..appendfs- a dr%=z1This report only display top-level ops statisticsct||kr.trim_pathsX t99' ' 'YY!11F=D4yy1}}tABBx' r%cbd}d}|dkr ||z ddS|dkr ||z ddS|dkr|ddS|S)Nrrrrrrr3)r default_strrrrs r$override_time_unitz(_build_table..override_time_unitss&     ,3333 3 $  (0000 0 $  %%%% % r%rrz--z8.3fzSelf CPU time total: z time total: )r)6r;anyr RuntimeErrorr r[rrraupperrCrxrJr rKrrrrrrrrr<rPrr[rIrr\rr rrrr]r rrrrrrTrxrUr{rrRrrrXrorr)5rcrrrrrrrrrrhas_device_timehas_device_memrhas_input_shapeshas_overload_namesname_column_widthshapes_column_widthDEFAULT_COLUMN_WIDTHflops_column_widthrstacks has_stackheadersrappend_node_idrr r raw_flops flops_scale flops_header row_format header_sep line_lengthrasum_self_cpu_time_totalsum_self_device_time_totalrMrr event_limitr= row_valuesr src_field empty_headersrrrrrrs5 ` @@@@@r$rrs(  6{{arOOOOOOOOPPPPPPPN%J X/XVWWW       ")!    ;;F;;;;;a?( 13HIIKKFKKKKKaO*!"57NOO-#FF aIK GGGGG G G! K  +"#35IJJ hG(''' G)3(>*""$$$FK % %%' '''&&&)))           .  NN"(((-K---     NN<   ==f=====N"y!!!LTNTN$}oO555555555WWW&J !!!& $%%% Q++-- .)) '(((((~&&& &'''3())) #c2222BB&BBB y>>Q  *:*:3y>>*J*J '[, NN2L22 3 3 3 J) * * * *J"J"J!!$KJ F !"EE3#:: ?jn , , , &#*D D & & O&   *  '#*D D &s[ !!!vDs[ !!!BCCC F: F : g &''' F:   Kp6p6 ) # # E  S^%?  1 Kw ,T>SVW>W1W1W501456>D1  #%<   <  s13J K K K  V  *-M%1 &&*?!*CCC -.K1F1J.K Lu T =/ )J   ')Di    !  "C$:I     c.       '9*,F((C $   &&26! ,&&-s/H)'&)??&s'CDD   I     +   ck * * *  K   c#"2334H5H4HI J J J  EyA~~!!$''''!!SY%<"C"CDDD  )I39~~!!%Icil4DEE   i ( ( ( z *-...  6DCLL1$45M122  %J%'99U.0s-IIIqqwwyyIIIIr%rcZg|](}t|hd&|)S)>r^~)r9issubsetr9s r$r8z1_canonicalize_profiler_events..1s3QQQaA0P0PQQQQr%rZNtsr) event_namer6r7 start_timec|dS)NrAr3)xs r$rUz/_canonicalize_profiler_events..>s !L/r%rOzevent=r@z node=z stack_trace=)rrsplitrasortr) rcevents_with_tracesrTr@r6r7lines source_linesrMs r$_canonicalize_profiler_eventsrIs   YYvr** &M%%k266 Fm'' r:: JIK$5$5d$;$;IIIQQ5QQQ *6>l2&&B !!("o&*#iia00        9 9:::E!   aS& a ac+.> a aSQ^M_ a a     99U  r%)F) NNr|r}r~rFFFN)5rr\jsonr collectionsrrcollections.abcroperatorrtypingrrtyping_extensionsr rtorch.autogradr __all__r4r rrrrr rrr rr>r7rBrloadsr5rEr6rr3rNrrrrrrrrrrrIr3r%r$rSs:  ////////$$$$$$""""""""(((((( %%%%%%   wwwwwwwwt 555 " " "DDD        < % % % % % % % % H<<< = =&&&&&J&&&Rss $E3S9$Eos+$E VTZ $E gtz " $E S! $E ou- $E^U+$E "5u!=$E+tz*$E xo$ES!$Eos+$E xo$E 3$E 3$E ^S)!$E" gs^#$E$E$ 0%8%$E&)73'$E( gw )$E*os++$E,'--$E.ow//$E0'11$E2-s33$E4/55$E6<%7$E8/99$E: "6!@;$E<M73=$E> VSM?$E@S!A$EBS!C$ED E3 > >R R R R R 'R R R j @ @ @ @ @ *@ @ @ F+66666666* P P P* & & &&  ggggV )))))r%