+ j4a0t$^RIHt^RIt^RIt^RIt^RIt^RIt^RIHt^RI H t H t ^RI H t ^RIt^RIHtHt] 'd ^RIHtHt] !R4tR R lsRR R lltR RltRRltRRltRRltRRlt.sR]R&RRltRRRllt R#)) annotationsN)AbstractContextManager)Any TYPE_CHECKING)TypeVar)profileProfilerActivity)CallableSequence_RcV^8dQhRR/#)returnNone)formats"a/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/_functorch/benchmark_utils.py __annotate__rs  T cR#)Nrrrr synchronizersrc@V^8dQhRRRRRRRRR R R R R RRRRRRR/ #)rfCallable[[tuple[Any, ...]], _R]input_tuple[Any, ...]trace_filenamestr optimize_ctxzAbstractContextManager[Any] activitieszSequence[ProfilerActivity]num_runsintdeviceszlist[str] | None kwargs_for_fzdict[str, Any] | Nonekwargs_for_profilerrfloatr)rs"rrrsl77&7 77. 7 + 7  77(7/7 7rc VfR.pVR.8wd?\PP4'd\PPsVf/pVf/pT;_uu_4\P!R4\ ^4Fp V!V3/VB\4K \P!R4\ P!4p \ V4Fp V!V3/VB\4K \ P!4p RRR4X X , p \RRV/VB;_uu_4p T;_uu_4\4\P!R4\ V4Fp V!V3/VB\4K RRR4RRR4X PV4V # +'giL;i +'giL>;i +'giLI;i)a Output the chrome trace of running f(input_, **kwargs_for_f) with [optimize_ctx] [num_runs] times to [trace_filename]. [activities] are the activities that the profiler will record, e.g. ProfilerActivity.CUDA. Return total runtime without the profiler Outputs to trace_filename Ncudacpui9r r) torch accelerator is_availabler manual_seedrangetime perf_counterrexport_chrome_trace)rrrrr r!r#r$r%_t0t1timingprofs&&&&&&&&& rdump_chrome_tracer7so*(5'e//<<>>''33  "   $qA f % % M $    xA f % % M!    "WF  >J >*= > >$ \ M   d #8_&)L) % ? ^, M- \ ? >s2$B!F+ G6AF2 <G F/ 2 G =G G c V^8dQhRRRR/#)rfilenamerrlist[dict[str, Any]]r)rs"rrrUsc.Brc\V4;_uu_4p\P!V4pRRR4XR,pV# +'giL;i)N traceEvents)openjsonload)r9rdataeventss& rget_chrome_trace_eventsrBUs6 h1yy|  - F M s > A c V^8dQhRRRR/#reventzdict[str, Any]rboolr)rs"rrr\s4rc~RV9;'d2VR,\9;'dRV9;'d VR,R8H#)pidphX)gpu_pidsrEs&ris_gpu_compute_eventrM\sO    %LH $   EM   $K3  rc V^8dQhRRRR/#rrAr:rr)rs"rrrfs DD"6D;ODrc.pVF'p\V4'gKVPV4K) \V\P!R4R7#)ts)key)rMappendsortedoperator itemgetter)rAsorted_gpu_eventsrEs& rget_sorted_gpu_eventsrXfsI.0#E**   ' #)<)CWrcRRlp\V4p.pVF$pV!V4'gKVPV4K& V#)c V^8dQhRRRR/#rDr)rs"rr3get_sorted_gpu_mm_conv_events..__annotate__~s   4 rcRV9;'dIRVR,9;'g5RVR,9;'g!RVR,9;'g RVR,9#)namegemmconvcutlasswgradrrLs&ris_mm_conv_event7get_sorted_gpu_mm_conv_events..is_mm_conv_event~sh  eFm # ( (v& ( (E&M) ( (%-'  r)rXrS)rArn gpu_events sorted_eventsrEs& rget_sorted_gpu_mm_conv_eventsrr}sF 'v.J*,M&& U# rz list[Any]rKc$V^8dQhRRRRRR/#)rr9r total_lengthr&rtuple[float, float]r)rs"rrrs",,#,U,?R,rcR\V4p.sVFOpRV9dK VR,R8XgKRVR,R,9gK3\PVR,4KQ VR,p\V4p\ V4V, p\ V4p\ V4V, pWW3#)a Process the chrome traces outputs by the pytorch profiler to compute GPU Utilization and percent of times spent on matmul and convolution Args: filename(str): Name of chrome traces file produced by pytorch profiler total_length(float): total length of the process without profiler in second Return: tuple: (GPU Utilization, percent of time spent on matmul and convolution) riprocess_labelsGPUargslabelsrHg.A)rBrKrSrXrcrr)r9rtrArErW utilizationsorted_gpu_mm_conv_eventsmm_conv_utilizations&& rcompute_utilizationr~s%X .FH    =, ,%-:Q1Q OOE%L )   #%L-f501L@K =f E&'@ALP  ++rc4V^8dQhRRRRRRRRR RR R R R /#)rrrrr trace_folderrrz"AbstractContextManager[Any] | Nonetrace_file_namer!r"rrur)rs"rrrsWA,A,&A, A,A,5 A,  A,  A,A,rc \PPV4pV'g)\P!V4\ RV,4Vf\ P !4p\PPW$R,4p\VVVV\P.VR.R7p\Wx4wrW3#)au Benchmark the GPU Utilization and percent of time spent on matmul and convolution operations of running f(input_, **kwargs_for_f) with [optimize_ctx] [num_runs] times. It will produce a chrome trace file in trace_folder/trace_file_name.json Example: ``` def f(a): return a.sum() a = torch.rand(2**20, device="cuda") utilization, mm_conv_utilization = benchmark_utilization( f, a, "tmp", trace_file_name="tmp_chrome_trace" ) ``` Args: f: function to benchmark input_: input to :attr:`f` trace_folder: name of the folder to store the chrome trace optimize_ctx: the context in which f will run trace_file_name: name of the dumped chrome trace file, default to "tmp_chrome_trace" num_runs: number of times to run f, excluding the warm-up runs, default to 1. Return: tuple: (GPU Utilization, percent of time spent on matmul and convolution) zcreate folder z.jsonr()r!r#) ospathexistsmakedirsprint contextlib nullcontextjoinr7r CUDAr~) rrrrrr!isExistchrome_trace_file_namertr{r}s &&&&&& rbenchmark_utilizationrsVggnn\*G  L! -.!--/ WW\\,'8QR$    L(;($K  ++r)r\NNN)Ntmp_chrome_tracer\)!__conditional_annotations__ __future__rrr>rUrr/rtypingrrtyping_extensionsrr*torch.profilerrr collections.abcr r r rr7rBrMrXrcrrrK__annotations__r~r)rs@rrs""  -%% 42 T] 7tD $),DA,A,r