| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 47.17208595275879, | |
| "base_token_generation_latency_async": 47.1084987744689, | |
| "base_token_generation_throughput_sync": 0.02119897773868778, | |
| "base_token_generation_throughput_async": 0.021227592175829722, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 45.285990142822264, | |
| "base_inference_latency_async": 43.656325340270996, | |
| "base_inference_throughput_sync": 0.022081884416046006, | |
| "base_inference_throughput_async": 0.022906188100021898, | |
| "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": 178.66405029296874, | |
| "smashed_token_generation_latency_async": 179.02547232806683, | |
| "smashed_token_generation_throughput_sync": 0.005597096888603082, | |
| "smashed_token_generation_throughput_async": 0.005585797300215945, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 186.31659545898438, | |
| "smashed_inference_latency_async": 133.17112922668457, | |
| "smashed_inference_throughput_sync": 0.005367208420358558, | |
| "smashed_inference_throughput_async": 0.007509135094122352, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |