base_results.json

{
    "current_gpu_type": "Tesla T4",
    "current_gpu_total_memory": 15095.0625,
    "perplexity": 3.4586403369903564,
    "memory_inference_first": 808.0,
    "memory_inference": 808.0,
    "token_generation_latency_sync": 37.12965393066406,
    "token_generation_latency_async": 37.13082652539015,
    "token_generation_throughput_sync": 0.0269326507019807,
    "token_generation_throughput_async": 0.02693180016653272,
    "token_generation_CO2_emissions": 1.8527108646322842e-05,
    "token_generation_energy_consumption": 0.001813133805152576,
    "inference_latency_sync": 118.66042785644531,
    "inference_latency_async": 46.790528297424316,
    "inference_throughput_sync": 0.008427409356805911,
    "inference_throughput_async": 0.021371846747347947,
    "inference_CO2_emissions": 1.8868051579151425e-05,
    "inference_energy_consumption": 6.49589917025334e-05
}