g^(hddlZddlZddlZddlZddlZddlZddlmZmZda ddZ dZ dZ dZ dZd Zgad ed efd Z dd Zy)N)profileProfilerActivitycy)Nr^/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/torch/_functorch/benchmark_utils.py synchronizer src |dg}|dgk7r8tjjrtjja|i}|i}|5tjdt dD]} ||fi|ttjdt j} t |D]} ||fi|tt j} ddd  z } tdd|i|5} |5ttjdt |D]} ||fi|t dddddd j|| S#1swYxYw#1swY0xYw#1swY4xYw)a0 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 Ncudacpui9 activitiesr) torchr is_availabler manual_seedrangetime perf_counterrexport_chrome_trace)finputtrace_filename optimize_ctxrnum_runsdevices kwargs_for_fkwargs_for_profiler_t0t1timingprofs rdump_chrome_tracer#s*(5'ejj557jj,,  "   ! $q A e $| $ M  $    x A e $| $ M     !"WF  >J >*= >$   M   d #8_ %(<(    ^, M- ! !  s2BE+FAE7F+E47F <FF cPt|}tj|}|d}|S)N traceEvents)openjsonload)filenamerdataeventss rget_chrome_trace_eventsr,Ks' XA 99Q.is_mm_conv_eventtsT eFm # (v& (E&M) (%-'  r)r=r8)r+rO gpu_events sorted_eventsr3s rget_sorted_gpu_mm_conv_eventsrRssH 'v.JM$& U#$ rr) total_lengthct|}ga|D]3}d|vr|ddk(sd|ddvstj|d5|dz}t|}t ||z }t |}t ||z }||fS)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) rJprocess_labelsGPUargslabelsr.g.A)r,r1r8r=rGrR)r)rSr+r3r< utilizationsorted_gpu_mm_conv_eventsmm_conv_utilizations rcompute_utilizationr\s%X .FH*    =, ,%-:Q1Q OOE%L ) *  #%L-f501L@K =f E&'@ALP + ++rc ftjj|}|s#tj|t d|z|t j }tjj||dz}t||||tjg|dg}t||\} } | | fS)a 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 )rr) ospathexistsmakedirsprint contextlib nullcontextjoinr#rCUDAr\) rr trace_folderrtrace_file_namerisExistchrome_trace_file_namerSrYr[s rbenchmark_utilizationrksNggnn\*G  L! -.!--/ WW\\,'8QR$     L(; ($K$ + ++r)r@NNN)Ntmp_chrome_tracer@)rcr'r:r^rrtorch.profilerrrr r#r,r4r=rGrRr1strfloatr\rkrrrrpsz  4  7tD $ ,#,U,L&  =,r