{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 5.228277683258057, "memory_inference_first": 252.0, "memory_inference": 252.0, "token_generation_latency_sync": 64.88516998291016, "token_generation_latency_async": 64.88899011164904, "token_generation_throughput_sync": 0.015411842186178845, "token_generation_throughput_async": 0.015410934863978988, "token_generation_CO2_emissions": 1.3202058888470628e-05, "token_generation_energy_consumption": 0.003116907546278766, "inference_latency_sync": 94.321435546875, "inference_latency_async": 77.33259201049805, "inference_throughput_sync": 0.01060204389598194, "inference_throughput_async": 0.012931158441763443, "inference_CO2_emissions": 1.3341487223188539e-05, "inference_energy_consumption": 4.514158086780525e-05 }