{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 2.7251362800598145, "memory_inference_first": 1690.0, "memory_inference": 1690.0, "token_generation_latency_sync": 155.3000717163086, "token_generation_latency_async": 155.34533187747002, "token_generation_throughput_sync": 0.006439147058648696, "token_generation_throughput_async": 0.0064372710007711, "token_generation_CO2_emissions": 0.00020285744923935242, "token_generation_energy_consumption": 0.009233308010788546, "inference_latency_sync": 1166.3823852539062, "inference_latency_async": 429.13405895233154, "inference_throughput_sync": 0.0008573517678615431, "inference_throughput_async": 0.0023302741396041943, "inference_CO2_emissions": 0.000202412501568058, "inference_energy_consumption": 0.0006962767357209158 }