{ | |
"current_gpu_type": "Tesla T4", | |
"current_gpu_total_memory": 15095.0625, | |
"perplexity": 2.3863821029663086, | |
"memory_inference_first": 1598.0, | |
"memory_inference": 1598.0, | |
"token_generation_latency_sync": 153.33177795410157, | |
"token_generation_latency_async": 152.51875221729279, | |
"token_generation_throughput_sync": 0.0065218052861771465, | |
"token_generation_throughput_async": 0.006556570818093925, | |
"token_generation_CO2_emissions": 5.481785809408624e-05, | |
"token_generation_energy_consumption": 0.008956993263524471, | |
"inference_latency_sync": 345.3065460205078, | |
"inference_latency_async": 190.85931777954102, | |
"inference_throughput_sync": 0.002895977535104735, | |
"inference_throughput_async": 0.005239461251533375, | |
"inference_CO2_emissions": 5.5271057652387143e-05, | |
"inference_energy_consumption": 0.00018832182125658322 | |
} |