{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 3.4586403369903564, "memory_inference_first": 810.0, "memory_inference": 810.0, "token_generation_latency_sync": 38.99144744873047, "token_generation_latency_async": 38.517594896256924, "token_generation_throughput_sync": 0.025646649853532412, "token_generation_throughput_async": 0.0259621609992367, "token_generation_CO2_emissions": 1.9730394054846908e-05, "token_generation_energy_consumption": 0.0018913891809545897, "inference_latency_sync": 123.40780677795411, "inference_latency_async": 47.88656234741211, "inference_throughput_sync": 0.008103215072927158, "inference_throughput_async": 0.02088268505776427, "inference_CO2_emissions": 1.9416946448516684e-05, "inference_energy_consumption": 6.635862082185392e-05 }