{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 3.4586403369903564, "memory_inference_first": 808.0, "memory_inference": 808.0, "token_generation_latency_sync": 38.27063522338867, "token_generation_latency_async": 37.724124267697334, "token_generation_throughput_sync": 0.026129694324719784, "token_generation_throughput_async": 0.026508236292082377, "token_generation_CO2_emissions": 1.899763620684379e-05, "token_generation_energy_consumption": 0.0018430467100968824, "inference_latency_sync": 118.64526290893555, "inference_latency_async": 47.777557373046875, "inference_throughput_sync": 0.008428486527671445, "inference_throughput_async": 0.02093032911230698, "inference_CO2_emissions": 1.8905640125185246e-05, "inference_energy_consumption": 6.517715960578611e-05 }