From 28bc0b4c2789c83bc5d1a407c277793498960eb4 Mon Sep 17 00:00:00 2001 From: Carlos Gutierrez Date: Sun, 16 Nov 2025 21:28:55 +0000 Subject: [PATCH] adding the needed data --- inference_benchmarks/inference_metrics.csv | 200 ++ inference_benchmarks/inference_metrics.json | 2600 +++++++++++++++++++ 2 files changed, 2800 insertions(+) diff --git a/inference_benchmarks/inference_metrics.csv b/inference_benchmarks/inference_metrics.csv index 2523100..2296560 100644 --- a/inference_benchmarks/inference_metrics.csv +++ b/inference_benchmarks/inference_metrics.csv @@ -199,3 +199,203 @@ run_1763326968_non_optimized,1763326968.8236582,False,105,0,0.000249147415161132 run_1763326968_optimized,1763326968.826099,True,105,0,0.0003414154052734375,0,0,0.0,,cuda run_1763326972_non_optimized,1763326972.1689265,False,110,0,0.00025177001953125,0,0,0.0009765625,,cuda run_1763326972_optimized,1763326972.171371,True,110,0,0.0003230571746826172,0,0,0.0,,cuda +run_1763328146_non_optimized,1763328146.8811655,False,53,21,0.38083600997924805,55.14184438899121,18.135048094249907,84.0693359375,,cuda +run_1763328146_optimized,1763328146.9318924,True,53,17,0.04855465888977051,350.1208820886508,2.856156405280618,83.21142578125,,cuda +run_1763328150_non_optimized,1763328150.6292245,False,66,6,0.3474395275115967,17.269192261953368,57.90658791859945,84.06982421875,,cuda +run_1763328150_optimized,1763328150.6806788,True,66,17,0.049269676208496094,345.0398157288582,2.8982162475585938,83.58984375,,cuda +run_1763328154_non_optimized,1763328154.400125,False,60,40,0.35491299629211426,112.7036778531425,8.872824907302856,84.0693359375,,cuda +run_1763328154_optimized,1763328154.4400854,True,60,13,0.03777933120727539,344.1034974567393,2.9061024005596456,83.29736328125,,cuda +run_1763328157_non_optimized,1763328158.1473851,False,61,0,0.3587629795074463,0,0,84.0693359375,,cuda +run_1763328157_optimized,1763328158.16346,True,61,4,0.013820648193359375,289.42202594534916,3.4551620483398438,83.0693359375,,cuda +run_1763328161_non_optimized,1763328161.8730447,False,60,0,0.3552854061126709,0,0,84.0693359375,,cuda +run_1763328161_optimized,1763328161.8809595,True,60,0,0.0031197071075439453,0,0,82.90087890625,,cuda +run_1763328165_non_optimized,1763328165.5648606,False,64,6,0.3446638584136963,17.40826562905318,57.443976402282715,84.0693359375,,cuda +run_1763328165_optimized,1763328165.572956,True,64,1,0.0058612823486328125,170.61112918971688,5.8612823486328125,83.0693359375,,cuda +run_1763328169_non_optimized,1763328169.4239416,False,54,12,0.37118029594421387,32.32930231243613,30.93169132868449,84.0693359375,,cuda +run_1763328169_optimized,1763328169.4591513,True,54,11,0.03293347358703613,334.0066747265317,2.9939521442760118,83.0693359375,,cuda +run_1763328172_non_optimized,1763328173.1438408,False,67,0,0.3334786891937256,0,0,84.06982421875,,cuda +run_1763328172_optimized,1763328173.149282,True,67,0,0.003131389617919922,0,0,83.09716796875,,cuda +run_1763328176_non_optimized,1763328176.8386126,False,63,11,0.3452491760253906,31.86104635103033,31.386288729580965,84.0693359375,,cuda +run_1763328176_optimized,1763328176.8440063,True,63,0,0.0030989646911621094,0,0,82.984375,,cuda +run_1763328180_non_optimized,1763328180.5212622,False,73,15,0.3206193447113037,46.78445093045305,21.374622980753582,84.06982421875,,cuda +run_1763328180_optimized,1763328180.598697,True,73,27,0.07303571701049805,369.6821377973062,2.7050265559443725,84.07080078125,,cuda +run_1763328183_non_optimized,1763328184.3314278,False,46,36,0.3898129463195801,92.3519866127949,10.828137397766113,84.0693359375,,cuda +run_1763328183_optimized,1763328184.429176,True,46,35,0.09540414810180664,366.8603587618767,2.7258328029087613,83.5302734375,,cuda +run_1763328187_non_optimized,1763328188.1023467,False,70,3,0.32543349266052246,9.218473413642966,108.47783088684082,84.06982421875,,cuda +run_1763328187_optimized,1763328188.1267638,True,70,7,0.022072315216064453,317.13936356369766,3.1531878880092075,83.4140625,,cuda +run_1763328191_non_optimized,1763328191.8269985,False,66,25,0.3369431495666504,74.1964928865686,13.477725982666016,84.06982421875,,cuda +run_1763328191_optimized,1763328191.8957512,True,66,23,0.06402778625488281,359.21904137746134,2.783816793690557,83.7685546875,,cuda +run_1763328195_non_optimized,1763328195.6528826,False,50,0,0.38687896728515625,0,0,84.0693359375,,cuda +run_1763328195_optimized,1763328195.7567282,True,50,37,0.10147953033447266,364.60555028240094,2.7426900090398014,83.7080078125,,cuda +run_1763328199_non_optimized,1763328199.4423645,False,66,0,0.3354356288909912,0,0,84.06982421875,,cuda +run_1763328199_optimized,1763328199.4558053,True,66,3,0.011044025421142578,271.6400846250162,3.681341807047526,83.18359375,,cuda +run_1763328202_non_optimized,1763328203.2210126,False,43,26,0.40613532066345215,64.01807150761248,15.620589256286621,84.0693359375,,cuda +run_1763328202_optimized,1763328203.314865,True,43,33,0.09137701988220215,361.1411276329831,2.7690006024909746,83.38427734375,,cuda +run_1763328206_non_optimized,1763328207.059134,False,61,0,0.3586292266845703,0,0,84.0693359375,,cuda +run_1763328206_optimized,1763328207.0646808,True,61,0,0.003091096878051758,0,0,82.9287109375,,cuda +run_1763328210_non_optimized,1763328210.8018575,False,58,3,0.3591897487640381,8.35213145787962,119.72991625467937,84.0693359375,,cuda +run_1763328210_optimized,1763328210.857825,True,58,19,0.053505897521972656,355.1010426878175,2.8160998695775086,83.4130859375,,cuda +run_1763328214_non_optimized,1763328214.5682755,False,66,0,0.3383166790008545,0,0,84.06982421875,,cuda +run_1763328214_optimized,1763328214.5738587,True,66,0,0.003110647201538086,0,0,83.0693359375,,cuda +run_1763328217_non_optimized,1763328218.2879,False,67,4,0.3340427875518799,11.974513891813226,83.51069688796997,84.06982421875,,cuda +run_1763328217_optimized,1763328218.3589163,True,67,24,0.0684976577758789,350.37694396101637,2.8540690739949546,83.828125,,cuda +run_1763328221_non_optimized,1763328222.0590568,False,59,0,0.3572876453399658,0,0,84.0693359375,,cuda +run_1763328221_optimized,1763328222.0646763,True,59,0,0.0031256675720214844,0,0,82.87353515625,,cuda +run_1763328225_non_optimized,1763328225.7766004,False,63,0,0.3473188877105713,0,0,84.0693359375,,cuda +run_1763328225_optimized,1763328225.7822132,True,63,0,0.0031120777130126953,0,0,82.984375,,cuda +run_1763328229_non_optimized,1763328229.5409186,False,53,18,0.3730945587158203,48.245142094688894,20.72747548421224,84.0693359375,,cuda +run_1763328229_optimized,1763328229.5943456,True,53,18,0.05087685585021973,353.79544783568343,2.8264919916788735,83.240234375,,cuda +run_1763328232_non_optimized,1763328233.287513,False,67,0,0.3368711471557617,0,0,84.06982421875,,cuda +run_1763328232_optimized,1763328233.358895,True,67,24,0.06883907318115234,348.63920867794354,2.8682947158813477,83.828125,,cuda +run_1763328236_non_optimized,1763328237.07395,False,64,10,0.34280872344970703,29.170786260540087,34.2808723449707,84.0693359375,,cuda +run_1763328236_optimized,1763328237.1121686,True,64,12,0.035607337951660156,337.0091865977449,2.967278162638346,83.38427734375,,cuda +run_1763328240_non_optimized,1763328240.8759682,False,54,40,0.37175869941711426,107.59667510865668,9.293967485427856,84.0693359375,,cuda +run_1763328240_optimized,1763328240.8816338,True,54,0,0.0030775070190429688,0,0,82.73681640625,,cuda +run_1763328244_non_optimized,1763328244.644577,False,51,32,0.3875398635864258,82.5721506527383,12.110620737075806,84.0693359375,,cuda +run_1763328244_optimized,1763328244.7303627,True,51,30,0.0831613540649414,360.7444868751505,2.772045135498047,83.5302734375,,cuda +run_1763328248_non_optimized,1763328248.4432864,False,60,13,0.3590657711029053,36.2050661639767,27.620443930992714,84.0693359375,,cuda +run_1763328248_optimized,1763328248.5309665,True,60,30,0.08289480209350586,361.9044770281231,2.7631600697835283,83.79736328125,,cuda +run_1763328251_non_optimized,1763328252.2412915,False,56,0,0.3664982318878174,0,0,84.0693359375,,cuda +run_1763328251_optimized,1763328252.246967,True,56,0,0.0030715465545654297,0,0,82.791015625,,cuda +run_1763328255_non_optimized,1763328255.9404693,False,65,8,0.3392612934112549,23.580644639889247,42.40766167640686,84.06982421875,,cuda +run_1763328255_optimized,1763328255.9791915,True,65,12,0.036064863204956055,332.73382826393066,3.0054052670796714,83.4140625,,cuda +run_1763328259_non_optimized,1763328259.683114,False,59,8,0.36543822288513184,21.891525021219906,45.67977786064148,84.0693359375,,cuda +run_1763328259_optimized,1763328259.728382,True,59,14,0.04237031936645508,330.41997816716747,3.02645138331822,83.29736328125,,cuda +run_1763328263_non_optimized,1763328263.4533045,False,54,34,0.37717175483703613,90.14460803060481,11.09328690697165,84.0693359375,,cuda +run_1763328263_optimized,1763328263.459047,True,54,0,0.0030853748321533203,0,0,82.73681640625,,cuda +run_1763328266_non_optimized,1763328267.1686065,False,54,11,0.37029480934143066,29.706060475337207,33.6631644855846,84.0693359375,,cuda +run_1763328266_optimized,1763328267.2656913,True,54,33,0.09436678886413574,349.6992999042453,2.8595996625495683,83.7080078125,,cuda +run_1763328270_non_optimized,1763328270.970244,False,59,7,0.3639066219329834,19.235703826486322,51.98666027614048,84.0693359375,,cuda +run_1763328270_optimized,1763328270.976036,True,59,0,0.0030896663665771484,0,0,82.87353515625,,cuda +run_1763328274_non_optimized,1763328274.6672146,False,65,0,0.3401796817779541,0,0,84.06982421875,,cuda +run_1763328274_optimized,1763328274.673098,True,65,0,0.003112316131591797,0,0,83.04150390625,,cuda +run_1763328278_non_optimized,1763328278.3698728,False,60,30,0.35071659088134766,85.5391526377759,11.690553029378254,84.0693359375,,cuda +run_1763328278_optimized,1763328278.4178205,True,60,15,0.04522132873535156,331.7018853599899,3.0147552490234375,83.35498046875,,cuda +run_1763328281_non_optimized,1763328282.1500068,False,56,12,0.3714306354522705,32.30751277526762,30.952552954355877,84.0693359375,,cuda +run_1763328281_optimized,1763328282.2164667,True,56,22,0.061430931091308594,358.1257781572615,2.792315049604936,83.4423828125,,cuda +run_1763328285_non_optimized,1763328285.9228153,False,59,0,0.35640788078308105,0,0,84.0693359375,,cuda +run_1763328285_optimized,1763328285.9286857,True,59,0,0.003099203109741211,0,0,82.87353515625,,cuda +run_1763328289_non_optimized,1763328289.66304,False,49,39,0.38460493087768555,101.40275609831696,9.86166489429963,84.0693359375,,cuda +run_1763328289_optimized,1763328289.7524455,True,49,31,0.08660292625427246,357.95557195124974,2.7936427823958856,83.50048828125,,cuda +run_1763328293_non_optimized,1763328293.461601,False,57,17,0.3626689910888672,46.87469956821971,21.33347006405101,84.0693359375,,cuda +run_1763328293_optimized,1763328293.507706,True,57,15,0.043273210525512695,346.63478438135326,2.8848807017008467,83.26806640625,,cuda +run_1763328296_non_optimized,1763328297.203139,False,68,0,0.33896493911743164,0,0,84.06982421875,,cuda +run_1763328296_optimized,1763328297.2200859,True,68,4,0.014091014862060547,283.86883692599235,3.5227537155151367,83.26904296875,,cuda +run_1763328300_non_optimized,1763328300.9405568,False,57,18,0.361710786819458,49.76351454230864,20.095043712192112,84.0693359375,,cuda +run_1763328300_optimized,1763328300.9465425,True,57,0,0.0031290054321289062,0,0,82.81884765625,,cuda +run_1763328304_non_optimized,1763328304.6751106,False,57,26,0.36957812309265137,70.35048444542788,14.214543195871206,84.0693359375,,cuda +run_1763328304_optimized,1763328304.73792,True,57,21,0.059926509857177734,350.4292182216033,2.8536433265322727,83.4423828125,,cuda +run_1763328308_non_optimized,1763328308.454106,False,62,0,0.3475470542907715,0,0,84.0693359375,,cuda +run_1763328308_optimized,1763328308.494272,True,62,12,0.03523826599121094,340.5388903924222,2.9365221659342446,83.326171875,,cuda +run_1763328311_non_optimized,1763328312.2665422,False,48,23,0.3990597724914551,57.63547615036164,17.350424890932832,84.0693359375,,cuda +run_1763328311_optimized,1763328312.3281333,True,48,21,0.05866241455078125,357.9804916113929,2.7934483119419644,83.1826171875,,cuda +run_1763328315_non_optimized,1763328316.0431576,False,65,35,0.34833741188049316,100.47729243624202,9.952497482299805,84.06982421875,,cuda +run_1763328315_optimized,1763328316.0491786,True,65,0,0.0031168460845947266,0,0,83.04150390625,,cuda +run_1763328319_non_optimized,1763328319.7818987,False,54,0,0.3795814514160156,0,0,84.0693359375,,cuda +run_1763328319_optimized,1763328319.8439722,True,54,21,0.059159040451049805,354.97533157888375,2.817097164335705,83.35498046875,,cuda +run_1763328323_non_optimized,1763328323.5640802,False,60,6,0.34952616691589355,17.16609675590835,58.254361152648926,84.0693359375,,cuda +run_1763328323_optimized,1763328323.6400027,True,60,26,0.07297134399414062,356.3042500914841,2.8065901536207933,83.677734375,,cuda +run_1763328326_non_optimized,1763328327.3651006,False,56,20,0.3710308074951172,53.90387966708992,18.55154037475586,84.0693359375,,cuda +run_1763328326_optimized,1763328327.4108949,True,56,15,0.04281282424926758,350.36230996268864,2.8541882832845054,83.240234375,,cuda +run_1763328330_non_optimized,1763328331.1276593,False,55,0,0.3657987117767334,0,0,84.0693359375,,cuda +run_1763328330_optimized,1763328331.2517033,True,55,45,0.12110424041748047,371.58071298636867,2.6912053426106772,84.06982421875,,cuda +run_1763328334_non_optimized,1763328334.8939357,False,86,14,0.28943705558776855,48.369756842536205,20.674075399126327,84.06982421875,,cuda +run_1763328334_optimized,1763328334.9002066,True,86,0,0.003253459930419922,0,0,83.64208984375,,cuda +run_1763328338_non_optimized,1763328338.600254,False,87,13,0.28672194480895996,45.340094245879136,22.055534216073845,84.06982421875,,cuda +run_1763328338_optimized,1763328338.6394484,True,87,13,0.03615999221801758,359.51335170704044,2.781537862924429,84.07080078125,,cuda +run_1763328341_non_optimized,1763328342.2768755,False,92,8,0.27281713485717773,29.323671345599582,34.10214185714722,84.06982421875,,cuda +run_1763328341_optimized,1763328342.3042974,True,92,8,0.021982908248901367,363.91909156968865,2.747863531112671,84.07080078125,,cuda +run_1763328345_non_optimized,1763328345.9446712,False,91,6,0.27774763107299805,21.60234446220379,46.29127184549967,84.06982421875,,cuda +run_1763328345_optimized,1763328345.961895,True,91,4,0.014170408248901367,282.27838815512746,3.542602062225342,83.94921875,,cuda +run_1763328349_non_optimized,1763328349.593344,False,96,0,0.25559139251708984,0,0,84.06982421875,,cuda +run_1763328349_optimized,1763328349.5997064,True,96,0,0.0032455921173095703,0,0,83.9384765625,,cuda +run_1763328352_non_optimized,1763328352.9822614,False,100,0,0.00023055076599121094,0,0,73.24609375,,cuda +run_1763328352_optimized,1763328352.9855807,True,100,0,0.0003197193145751953,0,0,73.2470703125,,cuda +run_1763328356_non_optimized,1763328356.5787084,False,98,2,0.24971413612365723,8.00915811594106,124.85706806182861,84.06982421875,,cuda +run_1763328356_optimized,1763328356.5879197,True,98,2,0.006028890609741211,331.73599082532525,3.0144453048706055,84.07080078125,,cuda +run_1763328359_non_optimized,1763328360.2265294,False,95,5,0.26633787155151367,18.77314694629497,53.267574310302734,84.06982421875,,cuda +run_1763328359_optimized,1763328360.2355754,True,95,1,0.005945444107055664,168.19601395516702,5.945444107055664,83.97900390625,,cuda +run_1763328363_non_optimized,1763328363.877551,False,99,1,0.2541792392730713,3.934231618836794,254.1792392730713,84.0283203125,,cuda +run_1763328363_optimized,1763328363.8867087,True,99,0,0.0036110877990722656,0,0,84.029296875,,cuda +run_1763328367_non_optimized,1763328367.2638996,False,107,0,0.00025200843811035156,0,0,73.24609375,,cuda +run_1763328367_optimized,1763328367.2672691,True,107,0,0.0003268718719482422,0,0,73.2470703125,,cuda +run_1763328370_non_optimized,1763328370.609966,False,111,0,0.00025463104248046875,0,0,73.24609375,,cuda +run_1763328370_optimized,1763328370.613386,True,111,0,0.00032639503479003906,0,0,73.2470703125,,cuda +run_1763328373_non_optimized,1763328373.9479587,False,104,0,0.00026726722717285156,0,0,73.24609375,,cuda +run_1763328373_optimized,1763328373.9513137,True,104,0,0.00033593177795410156,0,0,73.2470703125,,cuda +run_1763328377_non_optimized,1763328377.5480406,False,97,3,0.2508835792541504,11.95773756464522,83.62785975138347,84.06982421875,,cuda +run_1763328377_optimized,1763328377.5599065,True,97,3,0.008674144744873047,345.85542301137923,2.891381581624349,84.07080078125,,cuda +run_1763328380_non_optimized,1763328380.9351366,False,102,0,0.0002307891845703125,0,0,73.24609375,,cuda +run_1763328380_optimized,1763328380.9385371,True,102,0,0.0003266334533691406,0,0,73.2470703125,,cuda +run_1763328384_non_optimized,1763328384.280055,False,102,0,0.0002384185791015625,0,0,73.24609375,,cuda +run_1763328384_optimized,1763328384.2835102,True,102,0,0.00032138824462890625,0,0,73.2470703125,,cuda +run_1763328387_non_optimized,1763328387.6202579,False,105,0,0.00023627281188964844,0,0,73.24609375,,cuda +run_1763328387_optimized,1763328387.6237223,True,105,0,0.00032711029052734375,0,0,73.2470703125,,cuda +run_1763328390_non_optimized,1763328391.2143693,False,94,0,0.26361989974975586,0,0,84.06982421875,,cuda +run_1763328390_optimized,1763328391.2352936,True,94,6,0.017615556716918945,340.6080259863301,2.9359261194864907,84.07080078125,,cuda +run_1763328394_non_optimized,1763328394.615491,False,102,0,0.00023412704467773438,0,0,73.24609375,,cuda +run_1763328394_optimized,1763328394.6189466,True,102,0,0.0003268718719482422,0,0,73.2470703125,,cuda +run_1763328397_non_optimized,1763328397.9617507,False,103,0,0.00023031234741210938,0,0,73.24609375,,cuda +run_1763328397_optimized,1763328397.965208,True,103,0,0.0003170967102050781,0,0,73.2470703125,,cuda +run_1763328401_non_optimized,1763328401.3134701,False,107,0,0.0002567768096923828,0,0,73.24609375,,cuda +run_1763328401_optimized,1763328401.3170755,True,107,0,0.00034499168395996094,0,0,73.2470703125,,cuda +run_1763328404_non_optimized,1763328404.659879,False,110,0,0.0002472400665283203,0,0,73.24609375,,cuda +run_1763328404_optimized,1763328404.6633892,True,110,0,0.0003345012664794922,0,0,73.2470703125,,cuda +run_1763328407_non_optimized,1763328407.9927475,False,106,0,0.00025534629821777344,0,0,73.24609375,,cuda +run_1763328407_optimized,1763328407.9964025,True,106,0,0.0003371238708496094,0,0,73.2470703125,,cuda +run_1763328411_non_optimized,1763328411.3210385,False,109,0,0.00023627281188964844,0,0,73.24609375,,cuda +run_1763328411_optimized,1763328411.3245878,True,109,0,0.0003235340118408203,0,0,73.2470703125,,cuda +run_1763328414_non_optimized,1763328414.661873,False,110,0,0.00024056434631347656,0,0,73.24609375,,cuda +run_1763328414_optimized,1763328414.6654513,True,110,0,0.00032258033752441406,0,0,73.2470703125,,cuda +run_1763328417_non_optimized,1763328417.996759,False,114,0,0.00023698806762695312,0,0,73.24609375,,cuda +run_1763328417_optimized,1763328418.000359,True,114,0,0.00033736228942871094,0,0,73.2470703125,,cuda +run_1763328421_non_optimized,1763328421.3330092,False,113,0,0.0002493858337402344,0,0,73.24609375,,cuda +run_1763328421_optimized,1763328421.3366413,True,113,0,0.0003311634063720703,0,0,73.2470703125,,cuda +run_1763328424_non_optimized,1763328424.6846545,False,113,0,0.0002396106719970703,0,0,73.24609375,,cuda +run_1763328424_optimized,1763328424.6882706,True,113,0,0.0003235340118408203,0,0,73.2470703125,,cuda +run_1763328428_non_optimized,1763328428.0342734,False,112,0,0.00023412704467773438,0,0,73.24609375,,cuda +run_1763328428_optimized,1763328428.0378458,True,112,0,0.0003237724304199219,0,0,73.2470703125,,cuda +run_1763328431_non_optimized,1763328431.3756785,False,109,0,0.00023365020751953125,0,0,73.24609375,,cuda +run_1763328431_optimized,1763328431.3793805,True,109,0,0.0003218650817871094,0,0,73.2470703125,,cuda +run_1763328434_non_optimized,1763328434.742513,False,102,0,0.00023174285888671875,0,0,73.24609375,,cuda +run_1763328434_optimized,1763328434.7461731,True,102,0,0.00031876564025878906,0,0,73.2470703125,,cuda +run_1763328438_non_optimized,1763328438.0807612,False,105,0,0.0002357959747314453,0,0,73.24609375,,cuda +run_1763328438_optimized,1763328438.0844612,True,105,0,0.0003268718719482422,0,0,73.2470703125,,cuda +run_1763328441_non_optimized,1763328441.415201,False,109,0,0.00023245811462402344,0,0,73.24609375,,cuda +run_1763328441_optimized,1763328441.4189343,True,109,0,0.00034737586975097656,0,0,73.2470703125,,cuda +run_1763328444_non_optimized,1763328444.7733834,False,103,0,0.00023651123046875,0,0,73.24609375,,cuda +run_1763328444_optimized,1763328444.7771332,True,103,0,0.0003209114074707031,0,0,73.2470703125,,cuda +run_1763328448_non_optimized,1763328448.1191192,False,102,0,0.00023126602172851562,0,0,73.24609375,,cuda +run_1763328448_optimized,1763328448.1228065,True,102,0,0.0003218650817871094,0,0,73.2470703125,,cuda +run_1763328451_non_optimized,1763328451.461982,False,107,0,0.00024175643920898438,0,0,73.24609375,,cuda +run_1763328451_optimized,1763328451.4657404,True,107,0,0.00033092498779296875,0,0,73.2470703125,,cuda +run_1763328454_non_optimized,1763328454.8150704,False,112,0,0.00024199485778808594,0,0,73.24609375,,cuda +run_1763328454_optimized,1763328454.818905,True,112,0,0.0003361701965332031,0,0,73.2470703125,,cuda +run_1763328458_non_optimized,1763328458.150419,False,105,0,0.0002446174621582031,0,0,73.24609375,,cuda +run_1763328458_optimized,1763328458.154374,True,105,0,0.0003304481506347656,0,0,73.2470703125,,cuda +run_1763328461_non_optimized,1763328461.7283993,False,97,3,0.2511739730834961,11.943912672045563,83.7246576944987,84.06982421875,,cuda +run_1763328461_optimized,1763328461.7408895,True,97,3,0.008848905563354492,339.0249764246262,2.9496351877848306,84.07080078125,,cuda +run_1763328465_non_optimized,1763328465.1261904,False,102,0,0.0002315044403076172,0,0,73.24609375,,cuda +run_1763328465_optimized,1763328465.1299899,True,102,0,0.00033164024353027344,0,0,73.2470703125,,cuda +run_1763328468_non_optimized,1763328468.4668736,False,102,0,0.0002491474151611328,0,0,73.24609375,,cuda +run_1763328468_optimized,1763328468.4707043,True,102,0,0.00032591819763183594,0,0,73.2470703125,,cuda +run_1763328471_non_optimized,1763328471.8006985,False,109,0,0.00022912025451660156,0,0,73.24609375,,cuda +run_1763328471_optimized,1763328471.8047862,True,109,0,0.0003237724304199219,0,0,73.2470703125,,cuda +run_1763328475_non_optimized,1763328475.1509132,False,112,0,0.0002551078796386719,0,0,73.24609375,,cuda +run_1763328475_optimized,1763328475.1548111,True,112,0,0.0003528594970703125,0,0,73.2470703125,,cuda +run_1763328478_non_optimized,1763328478.5046594,False,107,0,0.00023484230041503906,0,0,73.24609375,,cuda +run_1763328478_optimized,1763328478.5085387,True,107,0,0.0003216266632080078,0,0,73.2470703125,,cuda +run_1763328481_non_optimized,1763328481.850465,False,106,0,0.000232696533203125,0,0,73.24609375,,cuda +run_1763328481_optimized,1763328481.854363,True,106,0,0.0003254413604736328,0,0,73.2470703125,,cuda +run_1763328485_non_optimized,1763328485.2010415,False,105,0,0.0002512931823730469,0,0,73.24609375,,cuda +run_1763328485_optimized,1763328485.2049637,True,105,0,0.0003311634063720703,0,0,73.2470703125,,cuda +run_1763328488_non_optimized,1763328488.538209,False,102,0,0.00025153160095214844,0,0,73.24609375,,cuda +run_1763328488_optimized,1763328488.5422387,True,102,0,0.0003306865692138672,0,0,73.2470703125,,cuda +run_1763328491_non_optimized,1763328491.8936143,False,106,0,0.0002357959747314453,0,0,73.24609375,,cuda +run_1763328491_optimized,1763328491.8975422,True,106,0,0.0003228187561035156,0,0,73.2470703125,,cuda +run_1763328495_non_optimized,1763328495.2653213,False,105,0,0.00024080276489257812,0,0,73.24609375,,cuda +run_1763328495_optimized,1763328495.2692711,True,105,0,0.0003209114074707031,0,0,73.2470703125,,cuda +run_1763328498_non_optimized,1763328498.633191,False,105,0,0.0002338886260986328,0,0,73.24609375,,cuda +run_1763328498_optimized,1763328498.6372032,True,105,0,0.0003268718719482422,0,0,73.2470703125,,cuda +run_1763328501_non_optimized,1763328501.9915178,False,110,0,0.00023627281188964844,0,0,73.24609375,,cuda +run_1763328501_optimized,1763328501.9954562,True,110,0,0.0003197193145751953,0,0,73.2470703125,,cuda diff --git a/inference_benchmarks/inference_metrics.json b/inference_benchmarks/inference_metrics.json index 58ae38c..4e48239 100644 --- a/inference_benchmarks/inference_metrics.json +++ b/inference_benchmarks/inference_metrics.json @@ -2599,6 +2599,2606 @@ "memory_used_mb": 0.0, "gpu_utilization": null, "device": "cuda" + }, + { + "run_name": "run_1763328146_non_optimized", + "timestamp": 1763328146.8811655, + "optimized": false, + "prompt_length": 53, + "generated_length": 21, + "total_time": 0.38083600997924805, + "tokens_per_second": 55.14184438899121, + "time_per_token": 18.135048094249907, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328146_optimized", + "timestamp": 1763328146.9318924, + "optimized": true, + "prompt_length": 53, + "generated_length": 17, + "total_time": 0.04855465888977051, + "tokens_per_second": 350.1208820886508, + "time_per_token": 2.856156405280618, + "memory_used_mb": 83.21142578125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328150_non_optimized", + "timestamp": 1763328150.6292245, + "optimized": false, + "prompt_length": 66, + "generated_length": 6, + "total_time": 0.3474395275115967, + "tokens_per_second": 17.269192261953368, + "time_per_token": 57.90658791859945, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328150_optimized", + "timestamp": 1763328150.6806788, + "optimized": true, + "prompt_length": 66, + "generated_length": 17, + "total_time": 0.049269676208496094, + "tokens_per_second": 345.0398157288582, + "time_per_token": 2.8982162475585938, + "memory_used_mb": 83.58984375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328154_non_optimized", + "timestamp": 1763328154.400125, + "optimized": false, + "prompt_length": 60, + "generated_length": 40, + "total_time": 0.35491299629211426, + "tokens_per_second": 112.7036778531425, + "time_per_token": 8.872824907302856, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328154_optimized", + "timestamp": 1763328154.4400854, + "optimized": true, + "prompt_length": 60, + "generated_length": 13, + "total_time": 0.03777933120727539, + "tokens_per_second": 344.1034974567393, + "time_per_token": 2.9061024005596456, + "memory_used_mb": 83.29736328125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328157_non_optimized", + "timestamp": 1763328158.1473851, + "optimized": false, + "prompt_length": 61, + "generated_length": 0, + "total_time": 0.3587629795074463, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328157_optimized", + "timestamp": 1763328158.16346, + "optimized": true, + "prompt_length": 61, + "generated_length": 4, + "total_time": 0.013820648193359375, + "tokens_per_second": 289.42202594534916, + "time_per_token": 3.4551620483398438, + "memory_used_mb": 83.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328161_non_optimized", + "timestamp": 1763328161.8730447, + "optimized": false, + "prompt_length": 60, + "generated_length": 0, + "total_time": 0.3552854061126709, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328161_optimized", + "timestamp": 1763328161.8809595, + "optimized": true, + "prompt_length": 60, + "generated_length": 0, + "total_time": 0.0031197071075439453, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.90087890625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328165_non_optimized", + "timestamp": 1763328165.5648606, + "optimized": false, + "prompt_length": 64, + "generated_length": 6, + "total_time": 0.3446638584136963, + "tokens_per_second": 17.40826562905318, + "time_per_token": 57.443976402282715, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328165_optimized", + "timestamp": 1763328165.572956, + "optimized": true, + "prompt_length": 64, + "generated_length": 1, + "total_time": 0.0058612823486328125, + "tokens_per_second": 170.61112918971688, + "time_per_token": 5.8612823486328125, + "memory_used_mb": 83.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328169_non_optimized", + "timestamp": 1763328169.4239416, + "optimized": false, + "prompt_length": 54, + "generated_length": 12, + "total_time": 0.37118029594421387, + "tokens_per_second": 32.32930231243613, + "time_per_token": 30.93169132868449, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328169_optimized", + "timestamp": 1763328169.4591513, + "optimized": true, + "prompt_length": 54, + "generated_length": 11, + "total_time": 0.03293347358703613, + "tokens_per_second": 334.0066747265317, + "time_per_token": 2.9939521442760118, + "memory_used_mb": 83.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328172_non_optimized", + "timestamp": 1763328173.1438408, + "optimized": false, + "prompt_length": 67, + "generated_length": 0, + "total_time": 0.3334786891937256, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328172_optimized", + "timestamp": 1763328173.149282, + "optimized": true, + "prompt_length": 67, + "generated_length": 0, + "total_time": 0.003131389617919922, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 83.09716796875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328176_non_optimized", + "timestamp": 1763328176.8386126, + "optimized": false, + "prompt_length": 63, + "generated_length": 11, + "total_time": 0.3452491760253906, + "tokens_per_second": 31.86104635103033, + "time_per_token": 31.386288729580965, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328176_optimized", + "timestamp": 1763328176.8440063, + "optimized": true, + "prompt_length": 63, + "generated_length": 0, + "total_time": 0.0030989646911621094, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.984375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328180_non_optimized", + "timestamp": 1763328180.5212622, + "optimized": false, + "prompt_length": 73, + "generated_length": 15, + "total_time": 0.3206193447113037, + "tokens_per_second": 46.78445093045305, + "time_per_token": 21.374622980753582, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328180_optimized", + "timestamp": 1763328180.598697, + "optimized": true, + "prompt_length": 73, + "generated_length": 27, + "total_time": 0.07303571701049805, + "tokens_per_second": 369.6821377973062, + "time_per_token": 2.7050265559443725, + "memory_used_mb": 84.07080078125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328183_non_optimized", + "timestamp": 1763328184.3314278, + "optimized": false, + "prompt_length": 46, + "generated_length": 36, + "total_time": 0.3898129463195801, + "tokens_per_second": 92.3519866127949, + "time_per_token": 10.828137397766113, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328183_optimized", + "timestamp": 1763328184.429176, + "optimized": true, + "prompt_length": 46, + "generated_length": 35, + "total_time": 0.09540414810180664, + "tokens_per_second": 366.8603587618767, + "time_per_token": 2.7258328029087613, + "memory_used_mb": 83.5302734375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328187_non_optimized", + "timestamp": 1763328188.1023467, + "optimized": false, + "prompt_length": 70, + "generated_length": 3, + "total_time": 0.32543349266052246, + "tokens_per_second": 9.218473413642966, + "time_per_token": 108.47783088684082, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328187_optimized", + "timestamp": 1763328188.1267638, + "optimized": true, + "prompt_length": 70, + "generated_length": 7, + "total_time": 0.022072315216064453, + "tokens_per_second": 317.13936356369766, + "time_per_token": 3.1531878880092075, + "memory_used_mb": 83.4140625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328191_non_optimized", + "timestamp": 1763328191.8269985, + "optimized": false, + "prompt_length": 66, + "generated_length": 25, + "total_time": 0.3369431495666504, + "tokens_per_second": 74.1964928865686, + "time_per_token": 13.477725982666016, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328191_optimized", + "timestamp": 1763328191.8957512, + "optimized": true, + "prompt_length": 66, + "generated_length": 23, + "total_time": 0.06402778625488281, + "tokens_per_second": 359.21904137746134, + "time_per_token": 2.783816793690557, + "memory_used_mb": 83.7685546875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328195_non_optimized", + "timestamp": 1763328195.6528826, + "optimized": false, + "prompt_length": 50, + "generated_length": 0, + "total_time": 0.38687896728515625, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328195_optimized", + "timestamp": 1763328195.7567282, + "optimized": true, + "prompt_length": 50, + "generated_length": 37, + "total_time": 0.10147953033447266, + "tokens_per_second": 364.60555028240094, + "time_per_token": 2.7426900090398014, + "memory_used_mb": 83.7080078125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328199_non_optimized", + "timestamp": 1763328199.4423645, + "optimized": false, + "prompt_length": 66, + "generated_length": 0, + "total_time": 0.3354356288909912, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328199_optimized", + "timestamp": 1763328199.4558053, + "optimized": true, + "prompt_length": 66, + "generated_length": 3, + "total_time": 0.011044025421142578, + "tokens_per_second": 271.6400846250162, + "time_per_token": 3.681341807047526, + "memory_used_mb": 83.18359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328202_non_optimized", + "timestamp": 1763328203.2210126, + "optimized": false, + "prompt_length": 43, + "generated_length": 26, + "total_time": 0.40613532066345215, + "tokens_per_second": 64.01807150761248, + "time_per_token": 15.620589256286621, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328202_optimized", + "timestamp": 1763328203.314865, + "optimized": true, + "prompt_length": 43, + "generated_length": 33, + "total_time": 0.09137701988220215, + "tokens_per_second": 361.1411276329831, + "time_per_token": 2.7690006024909746, + "memory_used_mb": 83.38427734375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328206_non_optimized", + "timestamp": 1763328207.059134, + "optimized": false, + "prompt_length": 61, + "generated_length": 0, + "total_time": 0.3586292266845703, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328206_optimized", + "timestamp": 1763328207.0646808, + "optimized": true, + "prompt_length": 61, + "generated_length": 0, + "total_time": 0.003091096878051758, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.9287109375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328210_non_optimized", + "timestamp": 1763328210.8018575, + "optimized": false, + "prompt_length": 58, + "generated_length": 3, + "total_time": 0.3591897487640381, + "tokens_per_second": 8.35213145787962, + "time_per_token": 119.72991625467937, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328210_optimized", + "timestamp": 1763328210.857825, + "optimized": true, + "prompt_length": 58, + "generated_length": 19, + "total_time": 0.053505897521972656, + "tokens_per_second": 355.1010426878175, + "time_per_token": 2.8160998695775086, + "memory_used_mb": 83.4130859375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328214_non_optimized", + "timestamp": 1763328214.5682755, + "optimized": false, + "prompt_length": 66, + "generated_length": 0, + "total_time": 0.3383166790008545, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328214_optimized", + "timestamp": 1763328214.5738587, + "optimized": true, + "prompt_length": 66, + "generated_length": 0, + "total_time": 0.003110647201538086, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 83.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328217_non_optimized", + "timestamp": 1763328218.2879, + "optimized": false, + "prompt_length": 67, + "generated_length": 4, + "total_time": 0.3340427875518799, + "tokens_per_second": 11.974513891813226, + "time_per_token": 83.51069688796997, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328217_optimized", + "timestamp": 1763328218.3589163, + "optimized": true, + "prompt_length": 67, + "generated_length": 24, + "total_time": 0.0684976577758789, + "tokens_per_second": 350.37694396101637, + "time_per_token": 2.8540690739949546, + "memory_used_mb": 83.828125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328221_non_optimized", + "timestamp": 1763328222.0590568, + "optimized": false, + "prompt_length": 59, + "generated_length": 0, + "total_time": 0.3572876453399658, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328221_optimized", + "timestamp": 1763328222.0646763, + "optimized": true, + "prompt_length": 59, + "generated_length": 0, + "total_time": 0.0031256675720214844, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.87353515625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328225_non_optimized", + "timestamp": 1763328225.7766004, + "optimized": false, + "prompt_length": 63, + "generated_length": 0, + "total_time": 0.3473188877105713, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328225_optimized", + "timestamp": 1763328225.7822132, + "optimized": true, + "prompt_length": 63, + "generated_length": 0, + "total_time": 0.0031120777130126953, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.984375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328229_non_optimized", + "timestamp": 1763328229.5409186, + "optimized": false, + "prompt_length": 53, + "generated_length": 18, + "total_time": 0.3730945587158203, + "tokens_per_second": 48.245142094688894, + "time_per_token": 20.72747548421224, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328229_optimized", + "timestamp": 1763328229.5943456, + "optimized": true, + "prompt_length": 53, + "generated_length": 18, + "total_time": 0.05087685585021973, + "tokens_per_second": 353.79544783568343, + "time_per_token": 2.8264919916788735, + "memory_used_mb": 83.240234375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328232_non_optimized", + "timestamp": 1763328233.287513, + "optimized": false, + "prompt_length": 67, + "generated_length": 0, + "total_time": 0.3368711471557617, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328232_optimized", + "timestamp": 1763328233.358895, + "optimized": true, + "prompt_length": 67, + "generated_length": 24, + "total_time": 0.06883907318115234, + "tokens_per_second": 348.63920867794354, + "time_per_token": 2.8682947158813477, + "memory_used_mb": 83.828125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328236_non_optimized", + "timestamp": 1763328237.07395, + "optimized": false, + "prompt_length": 64, + "generated_length": 10, + "total_time": 0.34280872344970703, + "tokens_per_second": 29.170786260540087, + "time_per_token": 34.2808723449707, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328236_optimized", + "timestamp": 1763328237.1121686, + "optimized": true, + "prompt_length": 64, + "generated_length": 12, + "total_time": 0.035607337951660156, + "tokens_per_second": 337.0091865977449, + "time_per_token": 2.967278162638346, + "memory_used_mb": 83.38427734375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328240_non_optimized", + "timestamp": 1763328240.8759682, + "optimized": false, + "prompt_length": 54, + "generated_length": 40, + "total_time": 0.37175869941711426, + "tokens_per_second": 107.59667510865668, + "time_per_token": 9.293967485427856, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328240_optimized", + "timestamp": 1763328240.8816338, + "optimized": true, + "prompt_length": 54, + "generated_length": 0, + "total_time": 0.0030775070190429688, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.73681640625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328244_non_optimized", + "timestamp": 1763328244.644577, + "optimized": false, + "prompt_length": 51, + "generated_length": 32, + "total_time": 0.3875398635864258, + "tokens_per_second": 82.5721506527383, + "time_per_token": 12.110620737075806, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328244_optimized", + "timestamp": 1763328244.7303627, + "optimized": true, + "prompt_length": 51, + "generated_length": 30, + "total_time": 0.0831613540649414, + "tokens_per_second": 360.7444868751505, + "time_per_token": 2.772045135498047, + "memory_used_mb": 83.5302734375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328248_non_optimized", + "timestamp": 1763328248.4432864, + "optimized": false, + "prompt_length": 60, + "generated_length": 13, + "total_time": 0.3590657711029053, + "tokens_per_second": 36.2050661639767, + "time_per_token": 27.620443930992714, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328248_optimized", + "timestamp": 1763328248.5309665, + "optimized": true, + "prompt_length": 60, + "generated_length": 30, + "total_time": 0.08289480209350586, + "tokens_per_second": 361.9044770281231, + "time_per_token": 2.7631600697835283, + "memory_used_mb": 83.79736328125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328251_non_optimized", + "timestamp": 1763328252.2412915, + "optimized": false, + "prompt_length": 56, + "generated_length": 0, + "total_time": 0.3664982318878174, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328251_optimized", + "timestamp": 1763328252.246967, + "optimized": true, + "prompt_length": 56, + "generated_length": 0, + "total_time": 0.0030715465545654297, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.791015625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328255_non_optimized", + "timestamp": 1763328255.9404693, + "optimized": false, + "prompt_length": 65, + "generated_length": 8, + "total_time": 0.3392612934112549, + "tokens_per_second": 23.580644639889247, + "time_per_token": 42.40766167640686, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328255_optimized", + "timestamp": 1763328255.9791915, + "optimized": true, + "prompt_length": 65, + "generated_length": 12, + "total_time": 0.036064863204956055, + "tokens_per_second": 332.73382826393066, + "time_per_token": 3.0054052670796714, + "memory_used_mb": 83.4140625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328259_non_optimized", + "timestamp": 1763328259.683114, + "optimized": false, + "prompt_length": 59, + "generated_length": 8, + "total_time": 0.36543822288513184, + "tokens_per_second": 21.891525021219906, + "time_per_token": 45.67977786064148, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328259_optimized", + "timestamp": 1763328259.728382, + "optimized": true, + "prompt_length": 59, + "generated_length": 14, + "total_time": 0.04237031936645508, + "tokens_per_second": 330.41997816716747, + "time_per_token": 3.02645138331822, + "memory_used_mb": 83.29736328125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328263_non_optimized", + "timestamp": 1763328263.4533045, + "optimized": false, + "prompt_length": 54, + "generated_length": 34, + "total_time": 0.37717175483703613, + "tokens_per_second": 90.14460803060481, + "time_per_token": 11.09328690697165, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328263_optimized", + "timestamp": 1763328263.459047, + "optimized": true, + "prompt_length": 54, + "generated_length": 0, + "total_time": 0.0030853748321533203, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.73681640625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328266_non_optimized", + "timestamp": 1763328267.1686065, + "optimized": false, + "prompt_length": 54, + "generated_length": 11, + "total_time": 0.37029480934143066, + "tokens_per_second": 29.706060475337207, + "time_per_token": 33.6631644855846, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328266_optimized", + "timestamp": 1763328267.2656913, + "optimized": true, + "prompt_length": 54, + "generated_length": 33, + "total_time": 0.09436678886413574, + "tokens_per_second": 349.6992999042453, + "time_per_token": 2.8595996625495683, + "memory_used_mb": 83.7080078125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328270_non_optimized", + "timestamp": 1763328270.970244, + "optimized": false, + "prompt_length": 59, + "generated_length": 7, + "total_time": 0.3639066219329834, + "tokens_per_second": 19.235703826486322, + "time_per_token": 51.98666027614048, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328270_optimized", + "timestamp": 1763328270.976036, + "optimized": true, + "prompt_length": 59, + "generated_length": 0, + "total_time": 0.0030896663665771484, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.87353515625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328274_non_optimized", + "timestamp": 1763328274.6672146, + "optimized": false, + "prompt_length": 65, + "generated_length": 0, + "total_time": 0.3401796817779541, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328274_optimized", + "timestamp": 1763328274.673098, + "optimized": true, + "prompt_length": 65, + "generated_length": 0, + "total_time": 0.003112316131591797, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 83.04150390625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328278_non_optimized", + "timestamp": 1763328278.3698728, + "optimized": false, + "prompt_length": 60, + "generated_length": 30, + "total_time": 0.35071659088134766, + "tokens_per_second": 85.5391526377759, + "time_per_token": 11.690553029378254, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328278_optimized", + "timestamp": 1763328278.4178205, + "optimized": true, + "prompt_length": 60, + "generated_length": 15, + "total_time": 0.04522132873535156, + "tokens_per_second": 331.7018853599899, + "time_per_token": 3.0147552490234375, + "memory_used_mb": 83.35498046875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328281_non_optimized", + "timestamp": 1763328282.1500068, + "optimized": false, + "prompt_length": 56, + "generated_length": 12, + "total_time": 0.3714306354522705, + "tokens_per_second": 32.30751277526762, + "time_per_token": 30.952552954355877, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328281_optimized", + "timestamp": 1763328282.2164667, + "optimized": true, + "prompt_length": 56, + "generated_length": 22, + "total_time": 0.061430931091308594, + "tokens_per_second": 358.1257781572615, + "time_per_token": 2.792315049604936, + "memory_used_mb": 83.4423828125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328285_non_optimized", + "timestamp": 1763328285.9228153, + "optimized": false, + "prompt_length": 59, + "generated_length": 0, + "total_time": 0.35640788078308105, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328285_optimized", + "timestamp": 1763328285.9286857, + "optimized": true, + "prompt_length": 59, + "generated_length": 0, + "total_time": 0.003099203109741211, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.87353515625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328289_non_optimized", + "timestamp": 1763328289.66304, + "optimized": false, + "prompt_length": 49, + "generated_length": 39, + "total_time": 0.38460493087768555, + "tokens_per_second": 101.40275609831696, + "time_per_token": 9.86166489429963, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328289_optimized", + "timestamp": 1763328289.7524455, + "optimized": true, + "prompt_length": 49, + "generated_length": 31, + "total_time": 0.08660292625427246, + "tokens_per_second": 357.95557195124974, + "time_per_token": 2.7936427823958856, + "memory_used_mb": 83.50048828125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328293_non_optimized", + "timestamp": 1763328293.461601, + "optimized": false, + "prompt_length": 57, + "generated_length": 17, + "total_time": 0.3626689910888672, + "tokens_per_second": 46.87469956821971, + "time_per_token": 21.33347006405101, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328293_optimized", + "timestamp": 1763328293.507706, + "optimized": true, + "prompt_length": 57, + "generated_length": 15, + "total_time": 0.043273210525512695, + "tokens_per_second": 346.63478438135326, + "time_per_token": 2.8848807017008467, + "memory_used_mb": 83.26806640625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328296_non_optimized", + "timestamp": 1763328297.203139, + "optimized": false, + "prompt_length": 68, + "generated_length": 0, + "total_time": 0.33896493911743164, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328296_optimized", + "timestamp": 1763328297.2200859, + "optimized": true, + "prompt_length": 68, + "generated_length": 4, + "total_time": 0.014091014862060547, + "tokens_per_second": 283.86883692599235, + "time_per_token": 3.5227537155151367, + "memory_used_mb": 83.26904296875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328300_non_optimized", + "timestamp": 1763328300.9405568, + "optimized": false, + "prompt_length": 57, + "generated_length": 18, + "total_time": 0.361710786819458, + "tokens_per_second": 49.76351454230864, + "time_per_token": 20.095043712192112, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328300_optimized", + "timestamp": 1763328300.9465425, + "optimized": true, + "prompt_length": 57, + "generated_length": 0, + "total_time": 0.0031290054321289062, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 82.81884765625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328304_non_optimized", + "timestamp": 1763328304.6751106, + "optimized": false, + "prompt_length": 57, + "generated_length": 26, + "total_time": 0.36957812309265137, + "tokens_per_second": 70.35048444542788, + "time_per_token": 14.214543195871206, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328304_optimized", + "timestamp": 1763328304.73792, + "optimized": true, + "prompt_length": 57, + "generated_length": 21, + "total_time": 0.059926509857177734, + "tokens_per_second": 350.4292182216033, + "time_per_token": 2.8536433265322727, + "memory_used_mb": 83.4423828125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328308_non_optimized", + "timestamp": 1763328308.454106, + "optimized": false, + "prompt_length": 62, + "generated_length": 0, + "total_time": 0.3475470542907715, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328308_optimized", + "timestamp": 1763328308.494272, + "optimized": true, + "prompt_length": 62, + "generated_length": 12, + "total_time": 0.03523826599121094, + "tokens_per_second": 340.5388903924222, + "time_per_token": 2.9365221659342446, + "memory_used_mb": 83.326171875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328311_non_optimized", + "timestamp": 1763328312.2665422, + "optimized": false, + "prompt_length": 48, + "generated_length": 23, + "total_time": 0.3990597724914551, + "tokens_per_second": 57.63547615036164, + "time_per_token": 17.350424890932832, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328311_optimized", + "timestamp": 1763328312.3281333, + "optimized": true, + "prompt_length": 48, + "generated_length": 21, + "total_time": 0.05866241455078125, + "tokens_per_second": 357.9804916113929, + "time_per_token": 2.7934483119419644, + "memory_used_mb": 83.1826171875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328315_non_optimized", + "timestamp": 1763328316.0431576, + "optimized": false, + "prompt_length": 65, + "generated_length": 35, + "total_time": 0.34833741188049316, + "tokens_per_second": 100.47729243624202, + "time_per_token": 9.952497482299805, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328315_optimized", + "timestamp": 1763328316.0491786, + "optimized": true, + "prompt_length": 65, + "generated_length": 0, + "total_time": 0.0031168460845947266, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 83.04150390625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328319_non_optimized", + "timestamp": 1763328319.7818987, + "optimized": false, + "prompt_length": 54, + "generated_length": 0, + "total_time": 0.3795814514160156, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328319_optimized", + "timestamp": 1763328319.8439722, + "optimized": true, + "prompt_length": 54, + "generated_length": 21, + "total_time": 0.059159040451049805, + "tokens_per_second": 354.97533157888375, + "time_per_token": 2.817097164335705, + "memory_used_mb": 83.35498046875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328323_non_optimized", + "timestamp": 1763328323.5640802, + "optimized": false, + "prompt_length": 60, + "generated_length": 6, + "total_time": 0.34952616691589355, + "tokens_per_second": 17.16609675590835, + "time_per_token": 58.254361152648926, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328323_optimized", + "timestamp": 1763328323.6400027, + "optimized": true, + "prompt_length": 60, + "generated_length": 26, + "total_time": 0.07297134399414062, + "tokens_per_second": 356.3042500914841, + "time_per_token": 2.8065901536207933, + "memory_used_mb": 83.677734375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328326_non_optimized", + "timestamp": 1763328327.3651006, + "optimized": false, + "prompt_length": 56, + "generated_length": 20, + "total_time": 0.3710308074951172, + "tokens_per_second": 53.90387966708992, + "time_per_token": 18.55154037475586, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328326_optimized", + "timestamp": 1763328327.4108949, + "optimized": true, + "prompt_length": 56, + "generated_length": 15, + "total_time": 0.04281282424926758, + "tokens_per_second": 350.36230996268864, + "time_per_token": 2.8541882832845054, + "memory_used_mb": 83.240234375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328330_non_optimized", + "timestamp": 1763328331.1276593, + "optimized": false, + "prompt_length": 55, + "generated_length": 0, + "total_time": 0.3657987117767334, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.0693359375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328330_optimized", + "timestamp": 1763328331.2517033, + "optimized": true, + "prompt_length": 55, + "generated_length": 45, + "total_time": 0.12110424041748047, + "tokens_per_second": 371.58071298636867, + "time_per_token": 2.6912053426106772, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328334_non_optimized", + "timestamp": 1763328334.8939357, + "optimized": false, + "prompt_length": 86, + "generated_length": 14, + "total_time": 0.28943705558776855, + "tokens_per_second": 48.369756842536205, + "time_per_token": 20.674075399126327, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328334_optimized", + "timestamp": 1763328334.9002066, + "optimized": true, + "prompt_length": 86, + "generated_length": 0, + "total_time": 0.003253459930419922, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 83.64208984375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328338_non_optimized", + "timestamp": 1763328338.600254, + "optimized": false, + "prompt_length": 87, + "generated_length": 13, + "total_time": 0.28672194480895996, + "tokens_per_second": 45.340094245879136, + "time_per_token": 22.055534216073845, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328338_optimized", + "timestamp": 1763328338.6394484, + "optimized": true, + "prompt_length": 87, + "generated_length": 13, + "total_time": 0.03615999221801758, + "tokens_per_second": 359.51335170704044, + "time_per_token": 2.781537862924429, + "memory_used_mb": 84.07080078125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328341_non_optimized", + "timestamp": 1763328342.2768755, + "optimized": false, + "prompt_length": 92, + "generated_length": 8, + "total_time": 0.27281713485717773, + "tokens_per_second": 29.323671345599582, + "time_per_token": 34.10214185714722, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328341_optimized", + "timestamp": 1763328342.3042974, + "optimized": true, + "prompt_length": 92, + "generated_length": 8, + "total_time": 0.021982908248901367, + "tokens_per_second": 363.91909156968865, + "time_per_token": 2.747863531112671, + "memory_used_mb": 84.07080078125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328345_non_optimized", + "timestamp": 1763328345.9446712, + "optimized": false, + "prompt_length": 91, + "generated_length": 6, + "total_time": 0.27774763107299805, + "tokens_per_second": 21.60234446220379, + "time_per_token": 46.29127184549967, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328345_optimized", + "timestamp": 1763328345.961895, + "optimized": true, + "prompt_length": 91, + "generated_length": 4, + "total_time": 0.014170408248901367, + "tokens_per_second": 282.27838815512746, + "time_per_token": 3.542602062225342, + "memory_used_mb": 83.94921875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328349_non_optimized", + "timestamp": 1763328349.593344, + "optimized": false, + "prompt_length": 96, + "generated_length": 0, + "total_time": 0.25559139251708984, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328349_optimized", + "timestamp": 1763328349.5997064, + "optimized": true, + "prompt_length": 96, + "generated_length": 0, + "total_time": 0.0032455921173095703, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 83.9384765625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328352_non_optimized", + "timestamp": 1763328352.9822614, + "optimized": false, + "prompt_length": 100, + "generated_length": 0, + "total_time": 0.00023055076599121094, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328352_optimized", + "timestamp": 1763328352.9855807, + "optimized": true, + "prompt_length": 100, + "generated_length": 0, + "total_time": 0.0003197193145751953, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328356_non_optimized", + "timestamp": 1763328356.5787084, + "optimized": false, + "prompt_length": 98, + "generated_length": 2, + "total_time": 0.24971413612365723, + "tokens_per_second": 8.00915811594106, + "time_per_token": 124.85706806182861, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328356_optimized", + "timestamp": 1763328356.5879197, + "optimized": true, + "prompt_length": 98, + "generated_length": 2, + "total_time": 0.006028890609741211, + "tokens_per_second": 331.73599082532525, + "time_per_token": 3.0144453048706055, + "memory_used_mb": 84.07080078125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328359_non_optimized", + "timestamp": 1763328360.2265294, + "optimized": false, + "prompt_length": 95, + "generated_length": 5, + "total_time": 0.26633787155151367, + "tokens_per_second": 18.77314694629497, + "time_per_token": 53.267574310302734, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328359_optimized", + "timestamp": 1763328360.2355754, + "optimized": true, + "prompt_length": 95, + "generated_length": 1, + "total_time": 0.005945444107055664, + "tokens_per_second": 168.19601395516702, + "time_per_token": 5.945444107055664, + "memory_used_mb": 83.97900390625, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328363_non_optimized", + "timestamp": 1763328363.877551, + "optimized": false, + "prompt_length": 99, + "generated_length": 1, + "total_time": 0.2541792392730713, + "tokens_per_second": 3.934231618836794, + "time_per_token": 254.1792392730713, + "memory_used_mb": 84.0283203125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328363_optimized", + "timestamp": 1763328363.8867087, + "optimized": true, + "prompt_length": 99, + "generated_length": 0, + "total_time": 0.0036110877990722656, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.029296875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328367_non_optimized", + "timestamp": 1763328367.2638996, + "optimized": false, + "prompt_length": 107, + "generated_length": 0, + "total_time": 0.00025200843811035156, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328367_optimized", + "timestamp": 1763328367.2672691, + "optimized": true, + "prompt_length": 107, + "generated_length": 0, + "total_time": 0.0003268718719482422, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328370_non_optimized", + "timestamp": 1763328370.609966, + "optimized": false, + "prompt_length": 111, + "generated_length": 0, + "total_time": 0.00025463104248046875, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328370_optimized", + "timestamp": 1763328370.613386, + "optimized": true, + "prompt_length": 111, + "generated_length": 0, + "total_time": 0.00032639503479003906, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328373_non_optimized", + "timestamp": 1763328373.9479587, + "optimized": false, + "prompt_length": 104, + "generated_length": 0, + "total_time": 0.00026726722717285156, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328373_optimized", + "timestamp": 1763328373.9513137, + "optimized": true, + "prompt_length": 104, + "generated_length": 0, + "total_time": 0.00033593177795410156, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328377_non_optimized", + "timestamp": 1763328377.5480406, + "optimized": false, + "prompt_length": 97, + "generated_length": 3, + "total_time": 0.2508835792541504, + "tokens_per_second": 11.95773756464522, + "time_per_token": 83.62785975138347, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328377_optimized", + "timestamp": 1763328377.5599065, + "optimized": true, + "prompt_length": 97, + "generated_length": 3, + "total_time": 0.008674144744873047, + "tokens_per_second": 345.85542301137923, + "time_per_token": 2.891381581624349, + "memory_used_mb": 84.07080078125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328380_non_optimized", + "timestamp": 1763328380.9351366, + "optimized": false, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.0002307891845703125, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328380_optimized", + "timestamp": 1763328380.9385371, + "optimized": true, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.0003266334533691406, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328384_non_optimized", + "timestamp": 1763328384.280055, + "optimized": false, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.0002384185791015625, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328384_optimized", + "timestamp": 1763328384.2835102, + "optimized": true, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.00032138824462890625, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328387_non_optimized", + "timestamp": 1763328387.6202579, + "optimized": false, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.00023627281188964844, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328387_optimized", + "timestamp": 1763328387.6237223, + "optimized": true, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.00032711029052734375, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328390_non_optimized", + "timestamp": 1763328391.2143693, + "optimized": false, + "prompt_length": 94, + "generated_length": 0, + "total_time": 0.26361989974975586, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328390_optimized", + "timestamp": 1763328391.2352936, + "optimized": true, + "prompt_length": 94, + "generated_length": 6, + "total_time": 0.017615556716918945, + "tokens_per_second": 340.6080259863301, + "time_per_token": 2.9359261194864907, + "memory_used_mb": 84.07080078125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328394_non_optimized", + "timestamp": 1763328394.615491, + "optimized": false, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.00023412704467773438, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328394_optimized", + "timestamp": 1763328394.6189466, + "optimized": true, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.0003268718719482422, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328397_non_optimized", + "timestamp": 1763328397.9617507, + "optimized": false, + "prompt_length": 103, + "generated_length": 0, + "total_time": 0.00023031234741210938, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328397_optimized", + "timestamp": 1763328397.965208, + "optimized": true, + "prompt_length": 103, + "generated_length": 0, + "total_time": 0.0003170967102050781, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328401_non_optimized", + "timestamp": 1763328401.3134701, + "optimized": false, + "prompt_length": 107, + "generated_length": 0, + "total_time": 0.0002567768096923828, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328401_optimized", + "timestamp": 1763328401.3170755, + "optimized": true, + "prompt_length": 107, + "generated_length": 0, + "total_time": 0.00034499168395996094, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328404_non_optimized", + "timestamp": 1763328404.659879, + "optimized": false, + "prompt_length": 110, + "generated_length": 0, + "total_time": 0.0002472400665283203, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328404_optimized", + "timestamp": 1763328404.6633892, + "optimized": true, + "prompt_length": 110, + "generated_length": 0, + "total_time": 0.0003345012664794922, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328407_non_optimized", + "timestamp": 1763328407.9927475, + "optimized": false, + "prompt_length": 106, + "generated_length": 0, + "total_time": 0.00025534629821777344, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328407_optimized", + "timestamp": 1763328407.9964025, + "optimized": true, + "prompt_length": 106, + "generated_length": 0, + "total_time": 0.0003371238708496094, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328411_non_optimized", + "timestamp": 1763328411.3210385, + "optimized": false, + "prompt_length": 109, + "generated_length": 0, + "total_time": 0.00023627281188964844, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328411_optimized", + "timestamp": 1763328411.3245878, + "optimized": true, + "prompt_length": 109, + "generated_length": 0, + "total_time": 0.0003235340118408203, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328414_non_optimized", + "timestamp": 1763328414.661873, + "optimized": false, + "prompt_length": 110, + "generated_length": 0, + "total_time": 0.00024056434631347656, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328414_optimized", + "timestamp": 1763328414.6654513, + "optimized": true, + "prompt_length": 110, + "generated_length": 0, + "total_time": 0.00032258033752441406, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328417_non_optimized", + "timestamp": 1763328417.996759, + "optimized": false, + "prompt_length": 114, + "generated_length": 0, + "total_time": 0.00023698806762695312, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328417_optimized", + "timestamp": 1763328418.000359, + "optimized": true, + "prompt_length": 114, + "generated_length": 0, + "total_time": 0.00033736228942871094, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328421_non_optimized", + "timestamp": 1763328421.3330092, + "optimized": false, + "prompt_length": 113, + "generated_length": 0, + "total_time": 0.0002493858337402344, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328421_optimized", + "timestamp": 1763328421.3366413, + "optimized": true, + "prompt_length": 113, + "generated_length": 0, + "total_time": 0.0003311634063720703, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328424_non_optimized", + "timestamp": 1763328424.6846545, + "optimized": false, + "prompt_length": 113, + "generated_length": 0, + "total_time": 0.0002396106719970703, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328424_optimized", + "timestamp": 1763328424.6882706, + "optimized": true, + "prompt_length": 113, + "generated_length": 0, + "total_time": 0.0003235340118408203, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328428_non_optimized", + "timestamp": 1763328428.0342734, + "optimized": false, + "prompt_length": 112, + "generated_length": 0, + "total_time": 0.00023412704467773438, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328428_optimized", + "timestamp": 1763328428.0378458, + "optimized": true, + "prompt_length": 112, + "generated_length": 0, + "total_time": 0.0003237724304199219, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328431_non_optimized", + "timestamp": 1763328431.3756785, + "optimized": false, + "prompt_length": 109, + "generated_length": 0, + "total_time": 0.00023365020751953125, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328431_optimized", + "timestamp": 1763328431.3793805, + "optimized": true, + "prompt_length": 109, + "generated_length": 0, + "total_time": 0.0003218650817871094, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328434_non_optimized", + "timestamp": 1763328434.742513, + "optimized": false, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.00023174285888671875, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328434_optimized", + "timestamp": 1763328434.7461731, + "optimized": true, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.00031876564025878906, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328438_non_optimized", + "timestamp": 1763328438.0807612, + "optimized": false, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.0002357959747314453, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328438_optimized", + "timestamp": 1763328438.0844612, + "optimized": true, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.0003268718719482422, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328441_non_optimized", + "timestamp": 1763328441.415201, + "optimized": false, + "prompt_length": 109, + "generated_length": 0, + "total_time": 0.00023245811462402344, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328441_optimized", + "timestamp": 1763328441.4189343, + "optimized": true, + "prompt_length": 109, + "generated_length": 0, + "total_time": 0.00034737586975097656, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328444_non_optimized", + "timestamp": 1763328444.7733834, + "optimized": false, + "prompt_length": 103, + "generated_length": 0, + "total_time": 0.00023651123046875, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328444_optimized", + "timestamp": 1763328444.7771332, + "optimized": true, + "prompt_length": 103, + "generated_length": 0, + "total_time": 0.0003209114074707031, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328448_non_optimized", + "timestamp": 1763328448.1191192, + "optimized": false, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.00023126602172851562, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328448_optimized", + "timestamp": 1763328448.1228065, + "optimized": true, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.0003218650817871094, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328451_non_optimized", + "timestamp": 1763328451.461982, + "optimized": false, + "prompt_length": 107, + "generated_length": 0, + "total_time": 0.00024175643920898438, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328451_optimized", + "timestamp": 1763328451.4657404, + "optimized": true, + "prompt_length": 107, + "generated_length": 0, + "total_time": 0.00033092498779296875, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328454_non_optimized", + "timestamp": 1763328454.8150704, + "optimized": false, + "prompt_length": 112, + "generated_length": 0, + "total_time": 0.00024199485778808594, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328454_optimized", + "timestamp": 1763328454.818905, + "optimized": true, + "prompt_length": 112, + "generated_length": 0, + "total_time": 0.0003361701965332031, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328458_non_optimized", + "timestamp": 1763328458.150419, + "optimized": false, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.0002446174621582031, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328458_optimized", + "timestamp": 1763328458.154374, + "optimized": true, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.0003304481506347656, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328461_non_optimized", + "timestamp": 1763328461.7283993, + "optimized": false, + "prompt_length": 97, + "generated_length": 3, + "total_time": 0.2511739730834961, + "tokens_per_second": 11.943912672045563, + "time_per_token": 83.7246576944987, + "memory_used_mb": 84.06982421875, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328461_optimized", + "timestamp": 1763328461.7408895, + "optimized": true, + "prompt_length": 97, + "generated_length": 3, + "total_time": 0.008848905563354492, + "tokens_per_second": 339.0249764246262, + "time_per_token": 2.9496351877848306, + "memory_used_mb": 84.07080078125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328465_non_optimized", + "timestamp": 1763328465.1261904, + "optimized": false, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.0002315044403076172, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328465_optimized", + "timestamp": 1763328465.1299899, + "optimized": true, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.00033164024353027344, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328468_non_optimized", + "timestamp": 1763328468.4668736, + "optimized": false, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.0002491474151611328, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328468_optimized", + "timestamp": 1763328468.4707043, + "optimized": true, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.00032591819763183594, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328471_non_optimized", + "timestamp": 1763328471.8006985, + "optimized": false, + "prompt_length": 109, + "generated_length": 0, + "total_time": 0.00022912025451660156, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328471_optimized", + "timestamp": 1763328471.8047862, + "optimized": true, + "prompt_length": 109, + "generated_length": 0, + "total_time": 0.0003237724304199219, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328475_non_optimized", + "timestamp": 1763328475.1509132, + "optimized": false, + "prompt_length": 112, + "generated_length": 0, + "total_time": 0.0002551078796386719, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328475_optimized", + "timestamp": 1763328475.1548111, + "optimized": true, + "prompt_length": 112, + "generated_length": 0, + "total_time": 0.0003528594970703125, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328478_non_optimized", + "timestamp": 1763328478.5046594, + "optimized": false, + "prompt_length": 107, + "generated_length": 0, + "total_time": 0.00023484230041503906, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328478_optimized", + "timestamp": 1763328478.5085387, + "optimized": true, + "prompt_length": 107, + "generated_length": 0, + "total_time": 0.0003216266632080078, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328481_non_optimized", + "timestamp": 1763328481.850465, + "optimized": false, + "prompt_length": 106, + "generated_length": 0, + "total_time": 0.000232696533203125, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328481_optimized", + "timestamp": 1763328481.854363, + "optimized": true, + "prompt_length": 106, + "generated_length": 0, + "total_time": 0.0003254413604736328, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328485_non_optimized", + "timestamp": 1763328485.2010415, + "optimized": false, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.0002512931823730469, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328485_optimized", + "timestamp": 1763328485.2049637, + "optimized": true, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.0003311634063720703, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328488_non_optimized", + "timestamp": 1763328488.538209, + "optimized": false, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.00025153160095214844, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328488_optimized", + "timestamp": 1763328488.5422387, + "optimized": true, + "prompt_length": 102, + "generated_length": 0, + "total_time": 0.0003306865692138672, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328491_non_optimized", + "timestamp": 1763328491.8936143, + "optimized": false, + "prompt_length": 106, + "generated_length": 0, + "total_time": 0.0002357959747314453, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328491_optimized", + "timestamp": 1763328491.8975422, + "optimized": true, + "prompt_length": 106, + "generated_length": 0, + "total_time": 0.0003228187561035156, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328495_non_optimized", + "timestamp": 1763328495.2653213, + "optimized": false, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.00024080276489257812, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328495_optimized", + "timestamp": 1763328495.2692711, + "optimized": true, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.0003209114074707031, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328498_non_optimized", + "timestamp": 1763328498.633191, + "optimized": false, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.0002338886260986328, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328498_optimized", + "timestamp": 1763328498.6372032, + "optimized": true, + "prompt_length": 105, + "generated_length": 0, + "total_time": 0.0003268718719482422, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328501_non_optimized", + "timestamp": 1763328501.9915178, + "optimized": false, + "prompt_length": 110, + "generated_length": 0, + "total_time": 0.00023627281188964844, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.24609375, + "gpu_utilization": null, + "device": "cuda" + }, + { + "run_name": "run_1763328501_optimized", + "timestamp": 1763328501.9954562, + "optimized": true, + "prompt_length": 110, + "generated_length": 0, + "total_time": 0.0003197193145751953, + "tokens_per_second": 0, + "time_per_token": 0, + "memory_used_mb": 73.2470703125, + "gpu_utilization": null, + "device": "cuda" } ] } \ No newline at end of file