{ "current_gpu_type": "Tesla T4", "current_gpu_total_memory": 15095.0625, "perplexity": 2.2062017917633057, "memory_inference_first": 7264.0, "memory_inference": 7264.0, "token_generation_latency_sync": 37.2805679321289, "token_generation_latency_async": 36.650632694363594, "token_generation_throughput_sync": 0.026823625697455812, "token_generation_throughput_async": 0.027284658585274228, "token_generation_CO2_emissions": 0.00017765237369534836, "token_generation_energy_consumption": 0.0022326934668068926, "inference_latency_sync": 1031.3929580688477, "inference_latency_async": 131.7535161972046, "inference_throughput_sync": 0.0009695625631111279, "inference_throughput_async": 0.007589930264200547, "inference_CO2_emissions": 0.00017810886568024814, "inference_energy_consumption": 0.0006130412240773345 }