| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 58.74873008728027, | |
| "base_token_generation_latency_async": 56.17937967181206, | |
| "base_token_generation_throughput_sync": 0.017021644527708876, | |
| "base_token_generation_throughput_async": 0.017800125345665732, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 55.618355178833006, | |
| "base_inference_latency_async": 54.12294864654541, | |
| "base_inference_throughput_sync": 0.017979675896287125, | |
| "base_inference_throughput_async": 0.01847645084030041, | |
| "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": 163.7388656616211, | |
| "smashed_token_generation_latency_async": 164.2250455915928, | |
| "smashed_token_generation_throughput_sync": 0.006107285499745531, | |
| "smashed_token_generation_throughput_async": 0.006089205190339087, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 170.35724792480468, | |
| "smashed_inference_latency_async": 127.94063091278076, | |
| "smashed_inference_throughput_sync": 0.005870017344031043, | |
| "smashed_inference_throughput_async": 0.00781612528299721, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |