{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 108340.9609375, "memory_inference_first": 262.0, "memory_inference": 262.0, "token_generation_latency_sync": 110.18281021118165, "token_generation_latency_async": 110.60973107814789, "token_generation_throughput_sync": 0.009075825875954263, "token_generation_throughput_async": 0.009040795870785373, "token_generation_CO2_emissions": 1.4485038513711413e-05, "token_generation_energy_consumption": 0.005279731847863916, "inference_latency_sync": 107.71225280761719, "inference_latency_async": 99.97386932373047, "inference_throughput_sync": 0.009283994846770878, "inference_throughput_async": 0.010002613750617666, "inference_CO2_emissions": 1.4262623257981447e-05, "inference_energy_consumption": 4.912709554389577e-05 }