{ "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", "base_current_gpu_total_memory": 40339.3125, "base_memory_inference_first": 690.0, "base_memory_inference": 570.0, "base_token_generation_latency_sync": 25.69969253540039, "smashed_current_gpu_type": "NVIDIA A100-PCIE-40GB", "smashed_current_gpu_total_memory": 40339.3125, "smashed_memory_inference_first": 186.0, "smashed_memory_inference": 206.0, "smashed_token_generation_latency_sync": 20.34065399169922 }