{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 3.4586403369903564, "memory_inference_first": 790.0, "memory_inference": 790.0, "token_generation_latency_sync": 37.40723571777344, "token_generation_latency_async": 37.517561949789524, "token_generation_throughput_sync": 0.026732795963452235, "token_generation_throughput_async": 0.026654184014897324, "token_generation_CO2_emissions": 1.942063140736845e-05, "token_generation_energy_consumption": 0.0018457401360101422, "inference_latency_sync": 122.10714988708496, "inference_latency_async": 47.79155254364014, "inference_throughput_sync": 0.008189528630589782, "inference_throughput_async": 0.020924199921876677, "inference_CO2_emissions": 1.9370730362422426e-05, "inference_energy_consumption": 6.823837478184117e-05 }