5204 lines
150 KiB
JSON
5204 lines
150 KiB
JSON
{
|
|
"runs": [
|
|
{
|
|
"run_name": "run_1763326618_non_optimized",
|
|
"timestamp": 1763326618.9855335,
|
|
"optimized": false,
|
|
"prompt_length": 53,
|
|
"generated_length": 5,
|
|
"total_time": 0.4859466552734375,
|
|
"tokens_per_second": 10.289195214619902,
|
|
"time_per_token": 97.1893310546875,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326618_optimized",
|
|
"timestamp": 1763326619.026237,
|
|
"optimized": true,
|
|
"prompt_length": 53,
|
|
"generated_length": 14,
|
|
"total_time": 0.04017782211303711,
|
|
"tokens_per_second": 348.4509429259782,
|
|
"time_per_token": 2.869844436645508,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326622_non_optimized",
|
|
"timestamp": 1763326622.7140436,
|
|
"optimized": false,
|
|
"prompt_length": 66,
|
|
"generated_length": 0,
|
|
"total_time": 0.3553469181060791,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326622_optimized",
|
|
"timestamp": 1763326622.717709,
|
|
"optimized": true,
|
|
"prompt_length": 66,
|
|
"generated_length": 0,
|
|
"total_time": 0.003069162368774414,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326626_non_optimized",
|
|
"timestamp": 1763326626.4152546,
|
|
"optimized": false,
|
|
"prompt_length": 60,
|
|
"generated_length": 14,
|
|
"total_time": 0.3536839485168457,
|
|
"tokens_per_second": 39.58336265671155,
|
|
"time_per_token": 25.263139179774694,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326626_optimized",
|
|
"timestamp": 1763326626.4510481,
|
|
"optimized": true,
|
|
"prompt_length": 60,
|
|
"generated_length": 12,
|
|
"total_time": 0.03518319129943848,
|
|
"tokens_per_second": 341.0719595579017,
|
|
"time_per_token": 2.9319326082865396,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326629_non_optimized",
|
|
"timestamp": 1763326630.1408024,
|
|
"optimized": false,
|
|
"prompt_length": 61,
|
|
"generated_length": 16,
|
|
"total_time": 0.35126209259033203,
|
|
"tokens_per_second": 45.55003325867101,
|
|
"time_per_token": 21.953880786895752,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326629_optimized",
|
|
"timestamp": 1763326630.1444962,
|
|
"optimized": true,
|
|
"prompt_length": 61,
|
|
"generated_length": 0,
|
|
"total_time": 0.003066539764404297,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326633_non_optimized",
|
|
"timestamp": 1763326633.8381388,
|
|
"optimized": false,
|
|
"prompt_length": 60,
|
|
"generated_length": 0,
|
|
"total_time": 0.35619688034057617,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326633_optimized",
|
|
"timestamp": 1763326633.8418753,
|
|
"optimized": true,
|
|
"prompt_length": 60,
|
|
"generated_length": 0,
|
|
"total_time": 0.003074169158935547,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326637_non_optimized",
|
|
"timestamp": 1763326637.540628,
|
|
"optimized": false,
|
|
"prompt_length": 64,
|
|
"generated_length": 8,
|
|
"total_time": 0.3398597240447998,
|
|
"tokens_per_second": 23.539123450078044,
|
|
"time_per_token": 42.482465505599976,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326637_optimized",
|
|
"timestamp": 1763326637.5471995,
|
|
"optimized": true,
|
|
"prompt_length": 64,
|
|
"generated_length": 1,
|
|
"total_time": 0.005898714065551758,
|
|
"tokens_per_second": 169.52847500101046,
|
|
"time_per_token": 5.898714065551758,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326640_non_optimized",
|
|
"timestamp": 1763326641.2843578,
|
|
"optimized": false,
|
|
"prompt_length": 54,
|
|
"generated_length": 30,
|
|
"total_time": 0.37198519706726074,
|
|
"tokens_per_second": 80.64837051721585,
|
|
"time_per_token": 12.399506568908691,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326640_optimized",
|
|
"timestamp": 1763326641.3555984,
|
|
"optimized": true,
|
|
"prompt_length": 54,
|
|
"generated_length": 25,
|
|
"total_time": 0.0705265998840332,
|
|
"tokens_per_second": 354.4761840370508,
|
|
"time_per_token": 2.821063995361328,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326644_non_optimized",
|
|
"timestamp": 1763326645.0330086,
|
|
"optimized": false,
|
|
"prompt_length": 67,
|
|
"generated_length": 0,
|
|
"total_time": 0.333660364151001,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326644_optimized",
|
|
"timestamp": 1763326645.0367913,
|
|
"optimized": true,
|
|
"prompt_length": 67,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030765533447265625,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326648_non_optimized",
|
|
"timestamp": 1763326648.755835,
|
|
"optimized": false,
|
|
"prompt_length": 63,
|
|
"generated_length": 17,
|
|
"total_time": 0.34771132469177246,
|
|
"tokens_per_second": 48.891131213714694,
|
|
"time_per_token": 20.453607334810144,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326648_optimized",
|
|
"timestamp": 1763326648.7840471,
|
|
"optimized": true,
|
|
"prompt_length": 63,
|
|
"generated_length": 9,
|
|
"total_time": 0.027483701705932617,
|
|
"tokens_per_second": 327.46680546519195,
|
|
"time_per_token": 3.053744633992513,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326652_non_optimized",
|
|
"timestamp": 1763326652.4592273,
|
|
"optimized": false,
|
|
"prompt_length": 73,
|
|
"generated_length": 19,
|
|
"total_time": 0.31671810150146484,
|
|
"tokens_per_second": 59.990256035025276,
|
|
"time_per_token": 16.66937376323499,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326652_optimized",
|
|
"timestamp": 1763326652.4657354,
|
|
"optimized": true,
|
|
"prompt_length": 73,
|
|
"generated_length": 1,
|
|
"total_time": 0.005753517150878906,
|
|
"tokens_per_second": 173.80672965357203,
|
|
"time_per_token": 5.753517150878906,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326655_non_optimized",
|
|
"timestamp": 1763326656.2020977,
|
|
"optimized": false,
|
|
"prompt_length": 46,
|
|
"generated_length": 0,
|
|
"total_time": 0.392134428024292,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326655_optimized",
|
|
"timestamp": 1763326656.2059286,
|
|
"optimized": true,
|
|
"prompt_length": 46,
|
|
"generated_length": 0,
|
|
"total_time": 0.003073453903198242,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326659_non_optimized",
|
|
"timestamp": 1763326659.8722565,
|
|
"optimized": false,
|
|
"prompt_length": 70,
|
|
"generated_length": 17,
|
|
"total_time": 0.33252859115600586,
|
|
"tokens_per_second": 51.12342352548099,
|
|
"time_per_token": 19.56050536211799,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326659_optimized",
|
|
"timestamp": 1763326659.913114,
|
|
"optimized": true,
|
|
"prompt_length": 70,
|
|
"generated_length": 14,
|
|
"total_time": 0.040065765380859375,
|
|
"tokens_per_second": 349.42549747691135,
|
|
"time_per_token": 2.861840384347098,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326663_non_optimized",
|
|
"timestamp": 1763326663.6047928,
|
|
"optimized": false,
|
|
"prompt_length": 66,
|
|
"generated_length": 14,
|
|
"total_time": 0.3415799140930176,
|
|
"tokens_per_second": 40.986016514389014,
|
|
"time_per_token": 24.3985652923584,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326663_optimized",
|
|
"timestamp": 1763326663.6473217,
|
|
"optimized": true,
|
|
"prompt_length": 66,
|
|
"generated_length": 14,
|
|
"total_time": 0.041661977767944336,
|
|
"tokens_per_second": 336.037815534814,
|
|
"time_per_token": 2.9758555548531667,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326666_non_optimized",
|
|
"timestamp": 1763326667.383774,
|
|
"optimized": false,
|
|
"prompt_length": 50,
|
|
"generated_length": 0,
|
|
"total_time": 0.38038086891174316,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326666_optimized",
|
|
"timestamp": 1763326667.4496503,
|
|
"optimized": true,
|
|
"prompt_length": 50,
|
|
"generated_length": 23,
|
|
"total_time": 0.06507229804992676,
|
|
"tokens_per_second": 353.45301594164135,
|
|
"time_per_token": 2.829230349996816,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326670_non_optimized",
|
|
"timestamp": 1763326671.1408262,
|
|
"optimized": false,
|
|
"prompt_length": 66,
|
|
"generated_length": 0,
|
|
"total_time": 0.34095191955566406,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326670_optimized",
|
|
"timestamp": 1763326671.1447349,
|
|
"optimized": true,
|
|
"prompt_length": 66,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030736923217773438,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326674_non_optimized",
|
|
"timestamp": 1763326674.9176548,
|
|
"optimized": false,
|
|
"prompt_length": 43,
|
|
"generated_length": 43,
|
|
"total_time": 0.4153327941894531,
|
|
"tokens_per_second": 103.53143455458913,
|
|
"time_per_token": 9.6589021904524,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326674_optimized",
|
|
"timestamp": 1763326675.01074,
|
|
"optimized": true,
|
|
"prompt_length": 43,
|
|
"generated_length": 32,
|
|
"total_time": 0.09222531318664551,
|
|
"tokens_per_second": 346.97632238167006,
|
|
"time_per_token": 2.882041037082672,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326678_non_optimized",
|
|
"timestamp": 1763326678.7116284,
|
|
"optimized": false,
|
|
"prompt_length": 61,
|
|
"generated_length": 0,
|
|
"total_time": 0.34818243980407715,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326678_optimized",
|
|
"timestamp": 1763326678.7930346,
|
|
"optimized": true,
|
|
"prompt_length": 61,
|
|
"generated_length": 29,
|
|
"total_time": 0.08052730560302734,
|
|
"tokens_per_second": 360.1262923530596,
|
|
"time_per_token": 2.776803641483702,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326682_non_optimized",
|
|
"timestamp": 1763326682.5197363,
|
|
"optimized": false,
|
|
"prompt_length": 58,
|
|
"generated_length": 0,
|
|
"total_time": 0.3616158962249756,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326682_optimized",
|
|
"timestamp": 1763326682.579515,
|
|
"optimized": true,
|
|
"prompt_length": 58,
|
|
"generated_length": 19,
|
|
"total_time": 0.05885767936706543,
|
|
"tokens_per_second": 322.8125913953667,
|
|
"time_per_token": 3.0977725982666016,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326685_non_optimized",
|
|
"timestamp": 1763326686.2714303,
|
|
"optimized": false,
|
|
"prompt_length": 66,
|
|
"generated_length": 0,
|
|
"total_time": 0.3391129970550537,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326685_optimized",
|
|
"timestamp": 1763326686.275417,
|
|
"optimized": true,
|
|
"prompt_length": 66,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030994415283203125,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326689_non_optimized",
|
|
"timestamp": 1763326689.9624639,
|
|
"optimized": false,
|
|
"prompt_length": 67,
|
|
"generated_length": 28,
|
|
"total_time": 0.3353545665740967,
|
|
"tokens_per_second": 83.4937191583267,
|
|
"time_per_token": 11.976948806217738,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326689_optimized",
|
|
"timestamp": 1763326689.9768,
|
|
"optimized": true,
|
|
"prompt_length": 67,
|
|
"generated_length": 3,
|
|
"total_time": 0.011225700378417969,
|
|
"tokens_per_second": 267.24390451108655,
|
|
"time_per_token": 3.7419001261393228,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326693_non_optimized",
|
|
"timestamp": 1763326693.711334,
|
|
"optimized": false,
|
|
"prompt_length": 59,
|
|
"generated_length": 24,
|
|
"total_time": 0.3620493412017822,
|
|
"tokens_per_second": 66.28930719866715,
|
|
"time_per_token": 15.085389216740927,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326693_optimized",
|
|
"timestamp": 1763326693.7152996,
|
|
"optimized": true,
|
|
"prompt_length": 59,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030515193939208984,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326697_non_optimized",
|
|
"timestamp": 1763326697.4116411,
|
|
"optimized": false,
|
|
"prompt_length": 63,
|
|
"generated_length": 0,
|
|
"total_time": 0.3452737331390381,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326697_optimized",
|
|
"timestamp": 1763326697.4156432,
|
|
"optimized": true,
|
|
"prompt_length": 63,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030455589294433594,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326700_non_optimized",
|
|
"timestamp": 1763326701.1587586,
|
|
"optimized": false,
|
|
"prompt_length": 53,
|
|
"generated_length": 19,
|
|
"total_time": 0.3794400691986084,
|
|
"tokens_per_second": 50.07378382606958,
|
|
"time_per_token": 19.970529957821494,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326700_optimized",
|
|
"timestamp": 1763326701.2194357,
|
|
"optimized": true,
|
|
"prompt_length": 53,
|
|
"generated_length": 21,
|
|
"total_time": 0.05968618392944336,
|
|
"tokens_per_second": 351.84021858098123,
|
|
"time_per_token": 2.8421992347353977,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326704_non_optimized",
|
|
"timestamp": 1763326704.906457,
|
|
"optimized": false,
|
|
"prompt_length": 67,
|
|
"generated_length": 0,
|
|
"total_time": 0.3325045108795166,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326704_optimized",
|
|
"timestamp": 1763326704.9105146,
|
|
"optimized": true,
|
|
"prompt_length": 67,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030672550201416016,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326708_non_optimized",
|
|
"timestamp": 1763326708.6155152,
|
|
"optimized": false,
|
|
"prompt_length": 64,
|
|
"generated_length": 4,
|
|
"total_time": 0.3434574604034424,
|
|
"tokens_per_second": 11.646274899084734,
|
|
"time_per_token": 85.8643651008606,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326708_optimized",
|
|
"timestamp": 1763326708.6492803,
|
|
"optimized": true,
|
|
"prompt_length": 64,
|
|
"generated_length": 11,
|
|
"total_time": 0.03272819519042969,
|
|
"tokens_per_second": 336.1016376245702,
|
|
"time_per_token": 2.9752904718572446,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326712_non_optimized",
|
|
"timestamp": 1763326712.3875144,
|
|
"optimized": false,
|
|
"prompt_length": 54,
|
|
"generated_length": 5,
|
|
"total_time": 0.3697817325592041,
|
|
"tokens_per_second": 13.521490002753103,
|
|
"time_per_token": 73.95634651184082,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326712_optimized",
|
|
"timestamp": 1763326712.3916335,
|
|
"optimized": true,
|
|
"prompt_length": 54,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030570030212402344,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326715_non_optimized",
|
|
"timestamp": 1763326716.129369,
|
|
"optimized": false,
|
|
"prompt_length": 51,
|
|
"generated_length": 0,
|
|
"total_time": 0.38543081283569336,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326715_optimized",
|
|
"timestamp": 1763326716.133514,
|
|
"optimized": true,
|
|
"prompt_length": 51,
|
|
"generated_length": 0,
|
|
"total_time": 0.003108978271484375,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326719_non_optimized",
|
|
"timestamp": 1763326719.845239,
|
|
"optimized": false,
|
|
"prompt_length": 60,
|
|
"generated_length": 10,
|
|
"total_time": 0.35239410400390625,
|
|
"tokens_per_second": 28.377319275152093,
|
|
"time_per_token": 35.239410400390625,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326719_optimized",
|
|
"timestamp": 1763326719.8974886,
|
|
"optimized": true,
|
|
"prompt_length": 60,
|
|
"generated_length": 18,
|
|
"total_time": 0.051172733306884766,
|
|
"tokens_per_second": 351.7498252839718,
|
|
"time_per_token": 2.842929628160265,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326723_non_optimized",
|
|
"timestamp": 1763326723.6197646,
|
|
"optimized": false,
|
|
"prompt_length": 56,
|
|
"generated_length": 0,
|
|
"total_time": 0.3644559383392334,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326723_optimized",
|
|
"timestamp": 1763326723.6239054,
|
|
"optimized": true,
|
|
"prompt_length": 56,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030570030212402344,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326726_non_optimized",
|
|
"timestamp": 1763326727.3191404,
|
|
"optimized": false,
|
|
"prompt_length": 65,
|
|
"generated_length": 0,
|
|
"total_time": 0.34804534912109375,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326726_optimized",
|
|
"timestamp": 1763326727.3233125,
|
|
"optimized": true,
|
|
"prompt_length": 65,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030825138092041016,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326730_non_optimized",
|
|
"timestamp": 1763326731.044284,
|
|
"optimized": false,
|
|
"prompt_length": 59,
|
|
"generated_length": 16,
|
|
"total_time": 0.35839104652404785,
|
|
"tokens_per_second": 44.64397242950211,
|
|
"time_per_token": 22.39944040775299,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326730_optimized",
|
|
"timestamp": 1763326731.048486,
|
|
"optimized": true,
|
|
"prompt_length": 59,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030989646911621094,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326734_non_optimized",
|
|
"timestamp": 1763326734.8039234,
|
|
"optimized": false,
|
|
"prompt_length": 54,
|
|
"generated_length": 0,
|
|
"total_time": 0.3762962818145752,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326734_optimized",
|
|
"timestamp": 1763326734.84245,
|
|
"optimized": true,
|
|
"prompt_length": 54,
|
|
"generated_length": 13,
|
|
"total_time": 0.03740096092224121,
|
|
"tokens_per_second": 347.58465235766965,
|
|
"time_per_token": 2.8769969940185547,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326738_non_optimized",
|
|
"timestamp": 1763326738.6017425,
|
|
"optimized": false,
|
|
"prompt_length": 54,
|
|
"generated_length": 41,
|
|
"total_time": 0.370577335357666,
|
|
"tokens_per_second": 110.63817478321626,
|
|
"time_per_token": 9.038471594089415,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326738_optimized",
|
|
"timestamp": 1763326738.6901221,
|
|
"optimized": true,
|
|
"prompt_length": 54,
|
|
"generated_length": 31,
|
|
"total_time": 0.08722376823425293,
|
|
"tokens_per_second": 355.4077131447096,
|
|
"time_per_token": 2.813669943040417,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326742_non_optimized",
|
|
"timestamp": 1763326742.419483,
|
|
"optimized": false,
|
|
"prompt_length": 59,
|
|
"generated_length": 12,
|
|
"total_time": 0.3536853790283203,
|
|
"tokens_per_second": 33.928459335716944,
|
|
"time_per_token": 29.47378158569336,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326742_optimized",
|
|
"timestamp": 1763326742.4236658,
|
|
"optimized": true,
|
|
"prompt_length": 59,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030400753021240234,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326745_non_optimized",
|
|
"timestamp": 1763326746.1228704,
|
|
"optimized": false,
|
|
"prompt_length": 65,
|
|
"generated_length": 19,
|
|
"total_time": 0.3396615982055664,
|
|
"tokens_per_second": 55.93802802665087,
|
|
"time_per_token": 17.8769262213456,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326745_optimized",
|
|
"timestamp": 1763326746.1515133,
|
|
"optimized": true,
|
|
"prompt_length": 65,
|
|
"generated_length": 9,
|
|
"total_time": 0.02747321128845215,
|
|
"tokens_per_second": 327.59184594423374,
|
|
"time_per_token": 3.052579032050239,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326749_non_optimized",
|
|
"timestamp": 1763326749.8801262,
|
|
"optimized": false,
|
|
"prompt_length": 60,
|
|
"generated_length": 0,
|
|
"total_time": 0.35973238945007324,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326749_optimized",
|
|
"timestamp": 1763326749.9388814,
|
|
"optimized": true,
|
|
"prompt_length": 60,
|
|
"generated_length": 20,
|
|
"total_time": 0.05757498741149902,
|
|
"tokens_per_second": 347.37306770136695,
|
|
"time_per_token": 2.878749370574951,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326753_non_optimized",
|
|
"timestamp": 1763326753.6588142,
|
|
"optimized": false,
|
|
"prompt_length": 56,
|
|
"generated_length": 22,
|
|
"total_time": 0.3657264709472656,
|
|
"tokens_per_second": 60.154245720901606,
|
|
"time_per_token": 16.62393049760298,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326753_optimized",
|
|
"timestamp": 1763326753.6631608,
|
|
"optimized": true,
|
|
"prompt_length": 56,
|
|
"generated_length": 0,
|
|
"total_time": 0.0031464099884033203,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326757_non_optimized",
|
|
"timestamp": 1763326757.3905053,
|
|
"optimized": false,
|
|
"prompt_length": 59,
|
|
"generated_length": 33,
|
|
"total_time": 0.36070775985717773,
|
|
"tokens_per_second": 91.4868036469921,
|
|
"time_per_token": 10.930538177490234,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326757_optimized",
|
|
"timestamp": 1763326757.4786851,
|
|
"optimized": true,
|
|
"prompt_length": 59,
|
|
"generated_length": 31,
|
|
"total_time": 0.08695292472839355,
|
|
"tokens_per_second": 356.5147474548062,
|
|
"time_per_token": 2.804933055754631,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326760_non_optimized",
|
|
"timestamp": 1763326761.2264946,
|
|
"optimized": false,
|
|
"prompt_length": 49,
|
|
"generated_length": 17,
|
|
"total_time": 0.38045310974121094,
|
|
"tokens_per_second": 44.683561691908935,
|
|
"time_per_token": 22.379594690659466,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326760_optimized",
|
|
"timestamp": 1763326761.3133712,
|
|
"optimized": true,
|
|
"prompt_length": 49,
|
|
"generated_length": 31,
|
|
"total_time": 0.08566641807556152,
|
|
"tokens_per_second": 361.8687543659949,
|
|
"time_per_token": 2.763432841147146,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326764_non_optimized",
|
|
"timestamp": 1763326765.0356681,
|
|
"optimized": false,
|
|
"prompt_length": 57,
|
|
"generated_length": 37,
|
|
"total_time": 0.3631327152252197,
|
|
"tokens_per_second": 101.89112258049266,
|
|
"time_per_token": 9.814397708789722,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326764_optimized",
|
|
"timestamp": 1763326765.0750856,
|
|
"optimized": true,
|
|
"prompt_length": 57,
|
|
"generated_length": 13,
|
|
"total_time": 0.03816509246826172,
|
|
"tokens_per_second": 340.62540293360655,
|
|
"time_per_token": 2.9357763437124396,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326768_non_optimized",
|
|
"timestamp": 1763326768.767545,
|
|
"optimized": false,
|
|
"prompt_length": 68,
|
|
"generated_length": 19,
|
|
"total_time": 0.331697940826416,
|
|
"tokens_per_second": 57.281030906197486,
|
|
"time_per_token": 17.457786359285052,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326768_optimized",
|
|
"timestamp": 1763326768.7959456,
|
|
"optimized": true,
|
|
"prompt_length": 68,
|
|
"generated_length": 9,
|
|
"total_time": 0.027106285095214844,
|
|
"tokens_per_second": 332.0263167153362,
|
|
"time_per_token": 3.011809455023872,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326772_non_optimized",
|
|
"timestamp": 1763326772.5260317,
|
|
"optimized": false,
|
|
"prompt_length": 57,
|
|
"generated_length": 10,
|
|
"total_time": 0.36326122283935547,
|
|
"tokens_per_second": 27.52839932056906,
|
|
"time_per_token": 36.32612228393555,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326772_optimized",
|
|
"timestamp": 1763326772.5303962,
|
|
"optimized": true,
|
|
"prompt_length": 57,
|
|
"generated_length": 0,
|
|
"total_time": 0.003071308135986328,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326775_non_optimized",
|
|
"timestamp": 1763326776.254764,
|
|
"optimized": false,
|
|
"prompt_length": 57,
|
|
"generated_length": 25,
|
|
"total_time": 0.3601036071777344,
|
|
"tokens_per_second": 69.42446424220596,
|
|
"time_per_token": 14.404144287109375,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326775_optimized",
|
|
"timestamp": 1763326776.2603817,
|
|
"optimized": true,
|
|
"prompt_length": 57,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030698776245117188,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326779_non_optimized",
|
|
"timestamp": 1763326779.9732888,
|
|
"optimized": false,
|
|
"prompt_length": 62,
|
|
"generated_length": 0,
|
|
"total_time": 0.3485872745513916,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326779_optimized",
|
|
"timestamp": 1763326780.019289,
|
|
"optimized": true,
|
|
"prompt_length": 62,
|
|
"generated_length": 15,
|
|
"total_time": 0.0446467399597168,
|
|
"tokens_per_second": 335.9707789086948,
|
|
"time_per_token": 2.9764493306477866,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326783_non_optimized",
|
|
"timestamp": 1763326783.7825809,
|
|
"optimized": false,
|
|
"prompt_length": 48,
|
|
"generated_length": 37,
|
|
"total_time": 0.3960244655609131,
|
|
"tokens_per_second": 93.42857125656288,
|
|
"time_per_token": 10.703363934078732,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326783_optimized",
|
|
"timestamp": 1763326783.8163621,
|
|
"optimized": true,
|
|
"prompt_length": 48,
|
|
"generated_length": 11,
|
|
"total_time": 0.032434701919555664,
|
|
"tokens_per_second": 339.142934850523,
|
|
"time_per_token": 2.9486092654141514,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326787_non_optimized",
|
|
"timestamp": 1763326787.5348887,
|
|
"optimized": false,
|
|
"prompt_length": 65,
|
|
"generated_length": 0,
|
|
"total_time": 0.33835744857788086,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326787_optimized",
|
|
"timestamp": 1763326787.539357,
|
|
"optimized": true,
|
|
"prompt_length": 65,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030851364135742188,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326790_non_optimized",
|
|
"timestamp": 1763326791.2828848,
|
|
"optimized": false,
|
|
"prompt_length": 54,
|
|
"generated_length": 0,
|
|
"total_time": 0.37782812118530273,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326790_optimized",
|
|
"timestamp": 1763326791.2873359,
|
|
"optimized": true,
|
|
"prompt_length": 54,
|
|
"generated_length": 0,
|
|
"total_time": 0.003094911575317383,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326794_non_optimized",
|
|
"timestamp": 1763326794.9969409,
|
|
"optimized": false,
|
|
"prompt_length": 60,
|
|
"generated_length": 15,
|
|
"total_time": 0.3552377223968506,
|
|
"tokens_per_second": 42.225245390023325,
|
|
"time_per_token": 23.682514826456707,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326794_optimized",
|
|
"timestamp": 1763326795.0417602,
|
|
"optimized": true,
|
|
"prompt_length": 60,
|
|
"generated_length": 15,
|
|
"total_time": 0.043425798416137695,
|
|
"tokens_per_second": 345.41679248494296,
|
|
"time_per_token": 2.895053227742513,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326798_non_optimized",
|
|
"timestamp": 1763326798.768397,
|
|
"optimized": false,
|
|
"prompt_length": 56,
|
|
"generated_length": 0,
|
|
"total_time": 0.3616046905517578,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326798_optimized",
|
|
"timestamp": 1763326798.7728815,
|
|
"optimized": true,
|
|
"prompt_length": 56,
|
|
"generated_length": 0,
|
|
"total_time": 0.0030786991119384766,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326802_non_optimized",
|
|
"timestamp": 1763326802.5161467,
|
|
"optimized": false,
|
|
"prompt_length": 55,
|
|
"generated_length": 8,
|
|
"total_time": 0.367128849029541,
|
|
"tokens_per_second": 21.790714680001297,
|
|
"time_per_token": 45.89110612869263,
|
|
"memory_used_mb": 8.12548828125,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326802_optimized",
|
|
"timestamp": 1763326802.5774758,
|
|
"optimized": true,
|
|
"prompt_length": 55,
|
|
"generated_length": 21,
|
|
"total_time": 0.05986332893371582,
|
|
"tokens_per_second": 350.79906804468607,
|
|
"time_per_token": 2.8506347111293246,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326805_non_optimized",
|
|
"timestamp": 1763326806.225911,
|
|
"optimized": false,
|
|
"prompt_length": 86,
|
|
"generated_length": 14,
|
|
"total_time": 0.2903730869293213,
|
|
"tokens_per_second": 48.21383464992984,
|
|
"time_per_token": 20.740934780665807,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326805_optimized",
|
|
"timestamp": 1763326806.23062,
|
|
"optimized": true,
|
|
"prompt_length": 86,
|
|
"generated_length": 0,
|
|
"total_time": 0.003251314163208008,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326809_non_optimized",
|
|
"timestamp": 1763326809.8743262,
|
|
"optimized": false,
|
|
"prompt_length": 87,
|
|
"generated_length": 0,
|
|
"total_time": 0.2792494297027588,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326809_optimized",
|
|
"timestamp": 1763326809.8790207,
|
|
"optimized": true,
|
|
"prompt_length": 87,
|
|
"generated_length": 0,
|
|
"total_time": 0.0032303333282470703,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326813_non_optimized",
|
|
"timestamp": 1763326813.5236838,
|
|
"optimized": false,
|
|
"prompt_length": 92,
|
|
"generated_length": 8,
|
|
"total_time": 0.2737746238708496,
|
|
"tokens_per_second": 29.22111584663858,
|
|
"time_per_token": 34.2218279838562,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326813_optimized",
|
|
"timestamp": 1763326813.5475192,
|
|
"optimized": true,
|
|
"prompt_length": 92,
|
|
"generated_length": 8,
|
|
"total_time": 0.022355318069458008,
|
|
"tokens_per_second": 357.8566842638511,
|
|
"time_per_token": 2.794414758682251,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326816_non_optimized",
|
|
"timestamp": 1763326817.1884215,
|
|
"optimized": false,
|
|
"prompt_length": 91,
|
|
"generated_length": 9,
|
|
"total_time": 0.2681541442871094,
|
|
"tokens_per_second": 33.56278540436731,
|
|
"time_per_token": 29.794904920789932,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326816_optimized",
|
|
"timestamp": 1763326817.193147,
|
|
"optimized": true,
|
|
"prompt_length": 91,
|
|
"generated_length": 0,
|
|
"total_time": 0.0032160282135009766,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326820_non_optimized",
|
|
"timestamp": 1763326820.8178222,
|
|
"optimized": false,
|
|
"prompt_length": 96,
|
|
"generated_length": 0,
|
|
"total_time": 0.25312256813049316,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326820_optimized",
|
|
"timestamp": 1763326820.8225522,
|
|
"optimized": true,
|
|
"prompt_length": 96,
|
|
"generated_length": 0,
|
|
"total_time": 0.003218412399291992,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326824_non_optimized",
|
|
"timestamp": 1763326824.1973226,
|
|
"optimized": false,
|
|
"prompt_length": 100,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023245811462402344,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326824_optimized",
|
|
"timestamp": 1763326824.198983,
|
|
"optimized": true,
|
|
"prompt_length": 100,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003190040588378906,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326827_non_optimized",
|
|
"timestamp": 1763326827.7810934,
|
|
"optimized": false,
|
|
"prompt_length": 98,
|
|
"generated_length": 2,
|
|
"total_time": 0.24810361862182617,
|
|
"tokens_per_second": 8.061148044150517,
|
|
"time_per_token": 124.05180931091309,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326827_optimized",
|
|
"timestamp": 1763326827.7893782,
|
|
"optimized": true,
|
|
"prompt_length": 98,
|
|
"generated_length": 2,
|
|
"total_time": 0.006571054458618164,
|
|
"tokens_per_second": 304.3651536591561,
|
|
"time_per_token": 3.285527229309082,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326831_non_optimized",
|
|
"timestamp": 1763326831.4181712,
|
|
"optimized": false,
|
|
"prompt_length": 95,
|
|
"generated_length": 5,
|
|
"total_time": 0.26477837562561035,
|
|
"tokens_per_second": 18.883717328449315,
|
|
"time_per_token": 52.95567512512207,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326831_optimized",
|
|
"timestamp": 1763326831.425718,
|
|
"optimized": true,
|
|
"prompt_length": 95,
|
|
"generated_length": 1,
|
|
"total_time": 0.00593876838684082,
|
|
"tokens_per_second": 168.38508169737847,
|
|
"time_per_token": 5.93876838684082,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326834_non_optimized",
|
|
"timestamp": 1763326835.0461798,
|
|
"optimized": false,
|
|
"prompt_length": 99,
|
|
"generated_length": 1,
|
|
"total_time": 0.25382065773010254,
|
|
"tokens_per_second": 3.9397896488919324,
|
|
"time_per_token": 253.82065773010254,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326834_optimized",
|
|
"timestamp": 1763326835.0514338,
|
|
"optimized": true,
|
|
"prompt_length": 99,
|
|
"generated_length": 1,
|
|
"total_time": 0.0036373138427734375,
|
|
"tokens_per_second": 274.9281594126901,
|
|
"time_per_token": 3.6373138427734375,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326838_non_optimized",
|
|
"timestamp": 1763326838.404944,
|
|
"optimized": false,
|
|
"prompt_length": 107,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002460479736328125,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326838_optimized",
|
|
"timestamp": 1763326838.4067247,
|
|
"optimized": true,
|
|
"prompt_length": 107,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032591819763183594,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326841_non_optimized",
|
|
"timestamp": 1763326841.7476485,
|
|
"optimized": false,
|
|
"prompt_length": 111,
|
|
"generated_length": 0,
|
|
"total_time": 0.00024366378784179688,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326841_optimized",
|
|
"timestamp": 1763326841.7494369,
|
|
"optimized": true,
|
|
"prompt_length": 111,
|
|
"generated_length": 0,
|
|
"total_time": 0.000324249267578125,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326845_non_optimized",
|
|
"timestamp": 1763326845.076511,
|
|
"optimized": false,
|
|
"prompt_length": 104,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002353191375732422,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326845_optimized",
|
|
"timestamp": 1763326845.078336,
|
|
"optimized": true,
|
|
"prompt_length": 104,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003311634063720703,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326848_non_optimized",
|
|
"timestamp": 1763326848.650482,
|
|
"optimized": false,
|
|
"prompt_length": 97,
|
|
"generated_length": 3,
|
|
"total_time": 0.2599804401397705,
|
|
"tokens_per_second": 11.539329644903832,
|
|
"time_per_token": 86.66014671325684,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326848_optimized",
|
|
"timestamp": 1763326848.655401,
|
|
"optimized": true,
|
|
"prompt_length": 97,
|
|
"generated_length": 0,
|
|
"total_time": 0.0032651424407958984,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326851_non_optimized",
|
|
"timestamp": 1763326852.0037806,
|
|
"optimized": false,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00024008750915527344,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326851_optimized",
|
|
"timestamp": 1763326852.0057652,
|
|
"optimized": true,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032830238342285156,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326855_non_optimized",
|
|
"timestamp": 1763326855.3215888,
|
|
"optimized": false,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002357959747314453,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326855_optimized",
|
|
"timestamp": 1763326855.3233974,
|
|
"optimized": true,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003199577331542969,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326858_non_optimized",
|
|
"timestamp": 1763326858.6427326,
|
|
"optimized": false,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002455711364746094,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326858_optimized",
|
|
"timestamp": 1763326858.6446369,
|
|
"optimized": true,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003333091735839844,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326861_non_optimized",
|
|
"timestamp": 1763326862.2181683,
|
|
"optimized": false,
|
|
"prompt_length": 94,
|
|
"generated_length": 0,
|
|
"total_time": 0.26184773445129395,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326861_optimized",
|
|
"timestamp": 1763326862.2230783,
|
|
"optimized": true,
|
|
"prompt_length": 94,
|
|
"generated_length": 0,
|
|
"total_time": 0.0032091140747070312,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326865_non_optimized",
|
|
"timestamp": 1763326865.582198,
|
|
"optimized": false,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.000240325927734375,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326865_optimized",
|
|
"timestamp": 1763326865.5841131,
|
|
"optimized": true,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003209114074707031,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326868_non_optimized",
|
|
"timestamp": 1763326868.9020677,
|
|
"optimized": false,
|
|
"prompt_length": 103,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002429485321044922,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326868_optimized",
|
|
"timestamp": 1763326868.9039805,
|
|
"optimized": true,
|
|
"prompt_length": 103,
|
|
"generated_length": 0,
|
|
"total_time": 0.00033211708068847656,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326872_non_optimized",
|
|
"timestamp": 1763326872.2377539,
|
|
"optimized": false,
|
|
"prompt_length": 107,
|
|
"generated_length": 0,
|
|
"total_time": 0.00024199485778808594,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326872_optimized",
|
|
"timestamp": 1763326872.2396657,
|
|
"optimized": true,
|
|
"prompt_length": 107,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032520294189453125,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326875_non_optimized",
|
|
"timestamp": 1763326875.572948,
|
|
"optimized": false,
|
|
"prompt_length": 110,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002472400665283203,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326875_optimized",
|
|
"timestamp": 1763326875.5749235,
|
|
"optimized": true,
|
|
"prompt_length": 110,
|
|
"generated_length": 0,
|
|
"total_time": 0.000339508056640625,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326878_non_optimized",
|
|
"timestamp": 1763326878.917066,
|
|
"optimized": false,
|
|
"prompt_length": 106,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023293495178222656,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326878_optimized",
|
|
"timestamp": 1763326878.9190118,
|
|
"optimized": true,
|
|
"prompt_length": 106,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032639503479003906,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326882_non_optimized",
|
|
"timestamp": 1763326882.2492902,
|
|
"optimized": false,
|
|
"prompt_length": 109,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023746490478515625,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326882_optimized",
|
|
"timestamp": 1763326882.251233,
|
|
"optimized": true,
|
|
"prompt_length": 109,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003230571746826172,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326885_non_optimized",
|
|
"timestamp": 1763326885.5688546,
|
|
"optimized": false,
|
|
"prompt_length": 110,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002415180206298828,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326885_optimized",
|
|
"timestamp": 1763326885.5708404,
|
|
"optimized": true,
|
|
"prompt_length": 110,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032401084899902344,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326888_non_optimized",
|
|
"timestamp": 1763326888.8873081,
|
|
"optimized": false,
|
|
"prompt_length": 114,
|
|
"generated_length": 0,
|
|
"total_time": 0.00024056434631347656,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326888_optimized",
|
|
"timestamp": 1763326888.8893402,
|
|
"optimized": true,
|
|
"prompt_length": 114,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003287792205810547,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326892_non_optimized",
|
|
"timestamp": 1763326892.2040594,
|
|
"optimized": false,
|
|
"prompt_length": 113,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002467632293701172,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326892_optimized",
|
|
"timestamp": 1763326892.2060897,
|
|
"optimized": true,
|
|
"prompt_length": 113,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032711029052734375,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326895_non_optimized",
|
|
"timestamp": 1763326895.5235405,
|
|
"optimized": false,
|
|
"prompt_length": 113,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023317337036132812,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326895_optimized",
|
|
"timestamp": 1763326895.5255647,
|
|
"optimized": true,
|
|
"prompt_length": 113,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003211498260498047,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326898_non_optimized",
|
|
"timestamp": 1763326898.837703,
|
|
"optimized": false,
|
|
"prompt_length": 112,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002307891845703125,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326898_optimized",
|
|
"timestamp": 1763326898.839746,
|
|
"optimized": true,
|
|
"prompt_length": 112,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032401084899902344,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326902_non_optimized",
|
|
"timestamp": 1763326902.1575205,
|
|
"optimized": false,
|
|
"prompt_length": 109,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002338886260986328,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326902_optimized",
|
|
"timestamp": 1763326902.1595387,
|
|
"optimized": true,
|
|
"prompt_length": 109,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003142356872558594,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326905_non_optimized",
|
|
"timestamp": 1763326905.4874718,
|
|
"optimized": false,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023674964904785156,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326905_optimized",
|
|
"timestamp": 1763326905.4895573,
|
|
"optimized": true,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00033164024353027344,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326908_non_optimized",
|
|
"timestamp": 1763326908.803102,
|
|
"optimized": false,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.00025081634521484375,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326908_optimized",
|
|
"timestamp": 1763326908.8052208,
|
|
"optimized": true,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.00031828880310058594,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326912_non_optimized",
|
|
"timestamp": 1763326912.1360228,
|
|
"optimized": false,
|
|
"prompt_length": 109,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023627281188964844,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326912_optimized",
|
|
"timestamp": 1763326912.1381037,
|
|
"optimized": true,
|
|
"prompt_length": 109,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003142356872558594,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326915_non_optimized",
|
|
"timestamp": 1763326915.4476118,
|
|
"optimized": false,
|
|
"prompt_length": 103,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002300739288330078,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326915_optimized",
|
|
"timestamp": 1763326915.449725,
|
|
"optimized": true,
|
|
"prompt_length": 103,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003192424774169922,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326918_non_optimized",
|
|
"timestamp": 1763326918.7690768,
|
|
"optimized": false,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023889541625976562,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326918_optimized",
|
|
"timestamp": 1763326918.7712831,
|
|
"optimized": true,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032782554626464844,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326922_non_optimized",
|
|
"timestamp": 1763326922.0759387,
|
|
"optimized": false,
|
|
"prompt_length": 107,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002472400665283203,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326922_optimized",
|
|
"timestamp": 1763326922.078079,
|
|
"optimized": true,
|
|
"prompt_length": 107,
|
|
"generated_length": 0,
|
|
"total_time": 0.000324249267578125,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326925_non_optimized",
|
|
"timestamp": 1763326925.3886678,
|
|
"optimized": false,
|
|
"prompt_length": 112,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023698806762695312,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326925_optimized",
|
|
"timestamp": 1763326925.3908367,
|
|
"optimized": true,
|
|
"prompt_length": 112,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032520294189453125,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326928_non_optimized",
|
|
"timestamp": 1763326928.6967072,
|
|
"optimized": false,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002300739288330078,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326928_optimized",
|
|
"timestamp": 1763326928.6988983,
|
|
"optimized": true,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003209114074707031,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326932_non_optimized",
|
|
"timestamp": 1763326932.2863104,
|
|
"optimized": false,
|
|
"prompt_length": 97,
|
|
"generated_length": 3,
|
|
"total_time": 0.25337719917297363,
|
|
"tokens_per_second": 11.840055102795507,
|
|
"time_per_token": 84.45906639099121,
|
|
"memory_used_mb": 8.1259765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326932_optimized",
|
|
"timestamp": 1763326932.2974646,
|
|
"optimized": true,
|
|
"prompt_length": 97,
|
|
"generated_length": 3,
|
|
"total_time": 0.008779525756835938,
|
|
"tokens_per_second": 341.70410601781447,
|
|
"time_per_token": 2.926508585611979,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326935_non_optimized",
|
|
"timestamp": 1763326935.6391814,
|
|
"optimized": false,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023484230041503906,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326935_optimized",
|
|
"timestamp": 1763326935.6413832,
|
|
"optimized": true,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00031876564025878906,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326938_non_optimized",
|
|
"timestamp": 1763326938.9412687,
|
|
"optimized": false,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023102760314941406,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326938_optimized",
|
|
"timestamp": 1763326938.9434795,
|
|
"optimized": true,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003178119659423828,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326942_non_optimized",
|
|
"timestamp": 1763326942.2639933,
|
|
"optimized": false,
|
|
"prompt_length": 109,
|
|
"generated_length": 0,
|
|
"total_time": 0.00024819374084472656,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326942_optimized",
|
|
"timestamp": 1763326942.266253,
|
|
"optimized": true,
|
|
"prompt_length": 109,
|
|
"generated_length": 0,
|
|
"total_time": 0.00033283233642578125,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326945_non_optimized",
|
|
"timestamp": 1763326945.5876389,
|
|
"optimized": false,
|
|
"prompt_length": 112,
|
|
"generated_length": 0,
|
|
"total_time": 0.00024247169494628906,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326945_optimized",
|
|
"timestamp": 1763326945.5899763,
|
|
"optimized": true,
|
|
"prompt_length": 112,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003368854522705078,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326948_non_optimized",
|
|
"timestamp": 1763326948.9048371,
|
|
"optimized": false,
|
|
"prompt_length": 107,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002422332763671875,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326948_optimized",
|
|
"timestamp": 1763326948.9071276,
|
|
"optimized": true,
|
|
"prompt_length": 107,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003273487091064453,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326952_non_optimized",
|
|
"timestamp": 1763326952.2211714,
|
|
"optimized": false,
|
|
"prompt_length": 106,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023603439331054688,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326952_optimized",
|
|
"timestamp": 1763326952.2234442,
|
|
"optimized": true,
|
|
"prompt_length": 106,
|
|
"generated_length": 0,
|
|
"total_time": 0.00032329559326171875,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326955_non_optimized",
|
|
"timestamp": 1763326955.5458589,
|
|
"optimized": false,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002446174621582031,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326955_optimized",
|
|
"timestamp": 1763326955.5481744,
|
|
"optimized": true,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.00031685829162597656,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326958_non_optimized",
|
|
"timestamp": 1763326958.8506002,
|
|
"optimized": false,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.00022912025451660156,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326958_optimized",
|
|
"timestamp": 1763326958.85293,
|
|
"optimized": true,
|
|
"prompt_length": 102,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003216266632080078,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326962_non_optimized",
|
|
"timestamp": 1763326962.1738048,
|
|
"optimized": false,
|
|
"prompt_length": 106,
|
|
"generated_length": 0,
|
|
"total_time": 0.00024080276489257812,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326962_optimized",
|
|
"timestamp": 1763326962.1762178,
|
|
"optimized": true,
|
|
"prompt_length": 106,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003228187561035156,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326965_non_optimized",
|
|
"timestamp": 1763326965.4970171,
|
|
"optimized": false,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.00023508071899414062,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326965_optimized",
|
|
"timestamp": 1763326965.4993877,
|
|
"optimized": true,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003216266632080078,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326968_non_optimized",
|
|
"timestamp": 1763326968.8236582,
|
|
"optimized": false,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.0002491474151611328,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326968_optimized",
|
|
"timestamp": 1763326968.826099,
|
|
"optimized": true,
|
|
"prompt_length": 105,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003414154052734375,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326972_non_optimized",
|
|
"timestamp": 1763326972.1689265,
|
|
"optimized": false,
|
|
"prompt_length": 110,
|
|
"generated_length": 0,
|
|
"total_time": 0.00025177001953125,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"memory_used_mb": 0.0009765625,
|
|
"gpu_utilization": null,
|
|
"device": "cuda"
|
|
},
|
|
{
|
|
"run_name": "run_1763326972_optimized",
|
|
"timestamp": 1763326972.171371,
|
|
"optimized": true,
|
|
"prompt_length": 110,
|
|
"generated_length": 0,
|
|
"total_time": 0.0003230571746826172,
|
|
"tokens_per_second": 0,
|
|
"time_per_token": 0,
|
|
"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"
|
|
}
|
|
]
|
|
} |