base_results.json

{
    "current_gpu_type": "Tesla T4",
    "current_gpu_total_memory": 15095.0625,
    "perplexity": 2.2062017917633057,
    "memory_inference_first": 7276.0,
    "memory_inference": 7276.0,
    "token_generation_latency_sync": 37.20788230895996,
    "token_generation_latency_async": 36.32834870368242,
    "token_generation_throughput_sync": 0.026876025668335118,
    "token_generation_throughput_async": 0.027526712214657724,
    "token_generation_CO2_emissions": 0.00018383593221655306,
    "token_generation_energy_consumption": 0.0022573257956992913,
    "inference_latency_sync": 1062.5710464477538,
    "inference_latency_async": 137.16118335723877,
    "inference_throughput_sync": 0.0009411135409185738,
    "inference_throughput_async": 0.007290692421306122,
    "inference_CO2_emissions": 0.00018406364335004676,
    "inference_energy_consumption": 0.0006330725272366455
}