| { | |
| "current_gpu_type": "Tesla T4", | |
| "current_gpu_total_memory": 15095.0625, | |
| "perplexity": 2.2114033699035645, | |
| "memory_inference_first": 2354.0, | |
| "memory_inference": 2354.0, | |
| "token_generation_latency_sync": 132.0942596435547, | |
| "token_generation_latency_async": 132.1555681526661, | |
| "token_generation_throughput_sync": 0.0075703516768890365, | |
| "token_generation_throughput_async": 0.0075668397024694425, | |
| "token_generation_CO2_emissions": 5.2118064791246415e-05, | |
| "token_generation_energy_consumption": 0.007773342725803954, | |
| "inference_latency_sync": 328.54746704101564, | |
| "inference_latency_async": 121.8498706817627, | |
| "inference_throughput_sync": 0.0030437002269603883, | |
| "inference_throughput_async": 0.008206820363492354, | |
| "inference_CO2_emissions": 5.156530753032632e-05, | |
| "inference_energy_consumption": 0.0001779811755834419 | |
| } |