{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 4.264157295227051, "memory_inference_first": 304.0, "memory_inference": 304.0, "token_generation_latency_sync": 100.45579986572265, "token_generation_latency_async": 97.21409808844328, "token_generation_throughput_sync": 0.009954626824301642, "token_generation_throughput_async": 0.010286573857736372, "token_generation_CO2_emissions": 1.275328979599919e-05, "token_generation_energy_consumption": 0.0046417158591284475, "inference_latency_sync": 98.32678298950195, "inference_latency_async": 86.0386610031128, "inference_throughput_sync": 0.010170168997665334, "inference_throughput_async": 0.011622682040156587, "inference_CO2_emissions": 1.2610012739805677e-05, "inference_energy_consumption": 4.2289036494555656e-05 }