{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 2.2114033699035645, "memory_inference_first": 2374.0, "memory_inference": 2374.0, "token_generation_latency_sync": 132.85096130371093, "token_generation_latency_async": 132.14306216686964, "token_generation_throughput_sync": 0.007527231946134716, "token_generation_throughput_async": 0.0075675558262544625, "token_generation_CO2_emissions": 5.169492869719809e-05, "token_generation_energy_consumption": 0.00777338393432973, "inference_latency_sync": 328.58505859375, "inference_latency_async": 122.15607166290283, "inference_throughput_sync": 0.0030433520144820757, "inference_throughput_async": 0.008186248840414264, "inference_CO2_emissions": 5.219246064070858e-05, "inference_energy_consumption": 0.00017807384324212128 }