base_results.json

{
    "current_gpu_type": "Tesla T4",
    "current_gpu_total_memory": 15095.0625,
    "perplexity": 3.4586403369903564,
    "memory_inference_first": 808.0,
    "memory_inference": 808.0,
    "token_generation_latency_sync": 38.3092342376709,
    "token_generation_latency_async": 38.34086339920759,
    "token_generation_throughput_sync": 0.02610336697925073,
    "token_generation_throughput_async": 0.02608183309770399,
    "token_generation_CO2_emissions": 1.9478876919620074e-05,
    "token_generation_energy_consumption": 0.0018778012307320703,
    "inference_latency_sync": 122.27628936767579,
    "inference_latency_async": 48.573970794677734,
    "inference_throughput_sync": 0.008178200411308473,
    "inference_throughput_async": 0.02058715776453611,
    "inference_CO2_emissions": 1.9799463696101588e-05,
    "inference_energy_consumption": 6.69570842513499e-05
}