| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": Infinity, | |
| "base_token_generation_latency_sync": 34.59449996948242, | |
| "base_token_generation_latency_async": 34.055160731077194, | |
| "base_token_generation_throughput_sync": 0.02890632906624322, | |
| "base_token_generation_throughput_async": 0.029364125099766315, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 127.86851806640625, | |
| "base_inference_latency_async": 31.493091583251953, | |
| "base_inference_throughput_sync": 0.007820533272159045, | |
| "base_inference_throughput_async": 0.03175299564847424, | |
| "base_inference_CO2_emissions": null, | |
| "base_inference_energy_consumption": null, | |
| "smashed_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "smashed_current_gpu_total_memory": 40339.3125, | |
| "smashed_perplexity": Infinity, | |
| "smashed_token_generation_latency_sync": 40.42353210449219, | |
| "smashed_token_generation_latency_async": 41.0306828096509, | |
| "smashed_token_generation_throughput_sync": 0.024738065872498854, | |
| "smashed_token_generation_throughput_async": 0.024372004839382986, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 181.8393585205078, | |
| "smashed_inference_latency_async": 64.73753452301025, | |
| "smashed_inference_throughput_sync": 0.005499359479357271, | |
| "smashed_inference_throughput_async": 0.015446989252341095, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |