base_results.json

{
    "current_gpu_type": "NVIDIA L40S",
    "current_gpu_total_memory": 45372.6875,
    "memory_inference_first": 25590.0,
    "memory_inference": 25590.0,
    "token_generation_latency_sync": 30.128338623046876,
    "token_generation_latency_async": 30.288023874163628,
    "token_generation_throughput_sync": 0.033191342294428515,
    "token_generation_throughput_async": 0.03301635009780294,
    "token_generation_CO2_emissions": 3.236854331824288e-06,
    "token_generation_energy_consumption": 0.001333430427257508,
    "inference_latency_sync": 23.984115219116212,
    "inference_latency_async": 22.998762130737305,
    "inference_throughput_sync": 0.04169426267611338,
    "inference_throughput_async": 0.0434806010130225,
    "inference_CO2_emissions": 2.2244397934888317e-06,
    "inference_energy_consumption": 7.886087285242966e-06
}