{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 2.8998236656188965, "memory_inference_first": 358.0, "memory_inference": 358.0, "token_generation_latency_sync": 80.11734771728516, "token_generation_latency_async": 80.36270011216402, "token_generation_throughput_sync": 0.012481691275262373, "token_generation_throughput_async": 0.012443583884118847, "token_generation_CO2_emissions": 2.351045754117281e-05, "token_generation_energy_consumption": 0.005466914504574626, "inference_latency_sync": 79.62785682678222, "inference_latency_async": 59.72108840942383, "inference_throughput_sync": 0.012558419124293918, "inference_throughput_async": 0.016744503937108465, "inference_CO2_emissions": 2.3351902662235743e-05, "inference_energy_consumption": 8.117517466477338e-05 }