base_results.json

{
    "current_gpu_type": "Tesla T4",
    "current_gpu_total_memory": 15095.0625,
    "perplexity": 3.4586403369903564,
    "memory_inference_first": 810.0,
    "memory_inference": 808.0,
    "token_generation_latency_sync": 36.57732048034668,
    "token_generation_latency_async": 36.563923954963684,
    "token_generation_throughput_sync": 0.02733934544323193,
    "token_generation_throughput_async": 0.02734936220827159,
    "token_generation_CO2_emissions": 1.9446958828920205e-05,
    "token_generation_energy_consumption": 0.001798393284076898,
    "inference_latency_sync": 122.71987800598144,
    "inference_latency_async": 47.88346290588379,
    "inference_throughput_sync": 0.008148639130420741,
    "inference_throughput_async": 0.020884036769970592,
    "inference_CO2_emissions": 1.9399486097971613e-05,
    "inference_energy_consumption": 6.64519560542376e-05
}