base_results.json

{
    "current_gpu_type": "Tesla T4",
    "current_gpu_total_memory": 15095.0625,
    "perplexity": 2.2062017917633057,
    "memory_inference_first": 7296.0,
    "memory_inference": 7296.0,
    "token_generation_latency_sync": 36.99171943664551,
    "token_generation_latency_async": 36.34188920259476,
    "token_generation_throughput_sync": 0.027033077002886197,
    "token_generation_throughput_async": 0.027516456132076962,
    "token_generation_CO2_emissions": 0.00017776892139547917,
    "token_generation_energy_consumption": 0.002222163814551259,
    "inference_latency_sync": 1029.307325744629,
    "inference_latency_async": 132.0033311843872,
    "inference_throughput_sync": 0.0009715271377055175,
    "inference_throughput_async": 0.00757556639690526,
    "inference_CO2_emissions": 0.00017755345632530548,
    "inference_energy_consumption": 0.000612054947249437
}