{ | |
"current_gpu_type": "Tesla T4", | |
"current_gpu_total_memory": 15095.0625, | |
"perplexity": 3.4594500064849854, | |
"memory_inference_first": 358.0, | |
"memory_inference": 358.0, | |
"token_generation_latency_sync": 73.98157806396485, | |
"token_generation_latency_async": 74.57008305937052, | |
"token_generation_throughput_sync": 0.01351687847392759, | |
"token_generation_throughput_async": 0.013410203649683872, | |
"token_generation_CO2_emissions": 9.466104547144268e-06, | |
"token_generation_energy_consumption": 0.003595659407598109, | |
"inference_latency_sync": 85.0370491027832, | |
"inference_latency_async": 66.51849746704102, | |
"inference_throughput_sync": 0.011759580212988254, | |
"inference_throughput_async": 0.015033412330088874, | |
"inference_CO2_emissions": 9.849718661233337e-06, | |
"inference_energy_consumption": 3.3456618429771876e-05 | |
} |