{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 343688032.0, "memory_inference_first": 344.0, "memory_inference": 344.0, "token_generation_latency_sync": 94.73909301757813, "token_generation_latency_async": 94.9408158659935, "token_generation_throughput_sync": 0.010555304765419883, "token_generation_throughput_async": 0.010532877676251214, "token_generation_CO2_emissions": 2.0721120007850446e-05, "token_generation_energy_consumption": 0.004636361917186246, "inference_latency_sync": 147.19097137451172, "inference_latency_async": 94.34385299682617, "inference_throughput_sync": 0.006793894969655487, "inference_throughput_async": 0.010599524698589966, "inference_CO2_emissions": 2.0695969686373567e-05, "inference_energy_consumption": 7.278868671603012e-05 }