base_results.json

{
    "current_gpu_type": "Tesla T4",
    "current_gpu_total_memory": 15095.0625,
    "perplexity": 2.2062017917633057,
    "memory_inference_first": 7278.0,
    "memory_inference": 7278.0,
    "token_generation_latency_sync": 35.36079788208008,
    "token_generation_latency_async": 34.97263416647911,
    "token_generation_throughput_sync": 0.0282799048634243,
    "token_generation_throughput_async": 0.028593785507826836,
    "token_generation_CO2_emissions": 0.00018337661498006707,
    "token_generation_energy_consumption": 0.002160617961299883,
    "inference_latency_sync": 1060.47254486084,
    "inference_latency_async": 136.61155700683594,
    "inference_throughput_sync": 0.000942975850573505,
    "inference_throughput_async": 0.007320024907921669,
    "inference_CO2_emissions": 0.0001840978910408864,
    "inference_energy_consumption": 0.0006337665377060597
}