{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 2.7025578022003174, "memory_inference_first": 1710.0, "memory_inference": 1710.0, "token_generation_latency_sync": 43.00895957946777, "token_generation_latency_async": 42.318619787693024, "token_generation_throughput_sync": 0.02325096932773501, "token_generation_throughput_async": 0.023630260273536072, "token_generation_CO2_emissions": 4.656282724561804e-05, "token_generation_energy_consumption": 0.002189405360918023, "inference_latency_sync": 278.88348388671875, "inference_latency_async": 101.10292434692383, "inference_throughput_sync": 0.003585726863646739, "inference_throughput_async": 0.009890910737345316, "inference_CO2_emissions": 4.665015602960695e-05, "inference_energy_consumption": 0.00016071153594523843 }