| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 47.90380516052246, | |
| "base_token_generation_latency_async": 48.11076000332832, | |
| "base_token_generation_throughput_sync": 0.020875168405705278, | |
| "base_token_generation_throughput_async": 0.02078537108810627, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 48.23736343383789, | |
| "base_inference_latency_async": 45.93331813812256, | |
| "base_inference_throughput_sync": 0.02073081795549615, | |
| "base_inference_throughput_async": 0.021770689350004646, | |
| "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_token_generation_latency_sync": 35.530513763427734, | |
| "smashed_token_generation_latency_async": 35.968757420778275, | |
| "smashed_token_generation_throughput_sync": 0.02814482240978232, | |
| "smashed_token_generation_throughput_async": 0.02780190564554572, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 60.0002555847168, | |
| "smashed_inference_latency_async": 34.74392890930176, | |
| "smashed_inference_throughput_sync": 0.016666595671214424, | |
| "smashed_inference_throughput_async": 0.028782006853930578, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |