| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 55.83882064819336, | |
| "base_token_generation_latency_async": 54.52581588178873, | |
| "base_token_generation_throughput_sync": 0.017908687690601405, | |
| "base_token_generation_throughput_async": 0.018339936483811395, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.76962547302246, | |
| "base_inference_latency_async": 52.814698219299316, | |
| "base_inference_throughput_sync": 0.01859786061745445, | |
| "base_inference_throughput_async": 0.018934123145942436, | |
| "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": 172.9581268310547, | |
| "smashed_token_generation_latency_async": 172.84861095249653, | |
| "smashed_token_generation_throughput_sync": 0.005781746242989778, | |
| "smashed_token_generation_throughput_async": 0.005785409523914699, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 179.0017532348633, | |
| "smashed_inference_latency_async": 156.34374618530273, | |
| "smashed_inference_throughput_sync": 0.005586537460825467, | |
| "smashed_inference_throughput_async": 0.006396162458681101, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |