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": 36.65739364624024,
    "token_generation_latency_async": 36.3558666780591,
    "token_generation_throughput_sync": 0.02727962630541697,
    "token_generation_throughput_async": 0.027505877080452153,
    "token_generation_CO2_emissions": 0.00018283805669861312,
    "token_generation_energy_consumption": 0.002242028576060243,
    "inference_latency_sync": 1057.4123092651366,
    "inference_latency_async": 136.27676963806152,
    "inference_throughput_sync": 0.000945704897926679,
    "inference_throughput_async": 0.007338007810545461,
    "inference_CO2_emissions": 0.00018305297379900547,
    "inference_energy_consumption": 0.0006297497502826921
}