{ | |
"current_gpu_type": "Tesla T4", | |
"current_gpu_total_memory": 15095.0625, | |
"perplexity": 2.548238754272461, | |
"memory_inference_first": 1458.0, | |
"memory_inference": 1458.0, | |
"token_generation_latency_sync": 57.3639274597168, | |
"token_generation_latency_async": 57.42117892950773, | |
"token_generation_throughput_sync": 0.01743255812988466, | |
"token_generation_throughput_async": 0.017415177093936633, | |
"token_generation_CO2_emissions": 0.00017302643152991706, | |
"token_generation_energy_consumption": 0.003290108461267099, | |
"inference_latency_sync": 964.0653579711914, | |
"inference_latency_async": 333.02977085113525, | |
"inference_throughput_sync": 0.0010372740724803456, | |
"inference_throughput_async": 0.0030027345526625645, | |
"inference_CO2_emissions": 0.0001732807294008593, | |
"inference_energy_consumption": 0.0005966248434902247 | |
} |