{
    "base_current_gpu_type": "NVIDIA A100-PCIE-40GB",
    "base_current_gpu_total_memory": 40339.3125,
    "base_perplexity": 8.891793251037598,
    "base_token_generation_latency_sync": 39.7899730682373,
    "base_token_generation_latency_async": 39.293849654495716,
    "base_token_generation_throughput_sync": 0.025131959709675168,
    "base_token_generation_throughput_async": 0.02544927536479204,
    "base_token_generation_CO2_emissions": null,
    "base_token_generation_energy_consumption": null,
    "base_inference_latency_sync": 117.77751007080079,
    "base_inference_latency_async": 38.930606842041016,
    "base_inference_throughput_sync": 0.008490585336698492,
    "base_inference_throughput_async": 0.025686730341950483,
    "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_perplexity": 9.283415794372559,
    "smashed_token_generation_latency_sync": 42.9160888671875,
    "smashed_token_generation_latency_async": 42.69602820277214,
    "smashed_token_generation_throughput_sync": 0.02330128458571101,
    "smashed_token_generation_throughput_async": 0.023421382318064718,
    "smashed_token_generation_CO2_emissions": null,
    "smashed_token_generation_energy_consumption": null,
    "smashed_inference_latency_sync": 160.85330200195312,
    "smashed_inference_latency_async": 69.06430721282959,
    "smashed_inference_throughput_sync": 0.0062168447122574935,
    "smashed_inference_throughput_async": 0.014479259118872868,
    "smashed_inference_CO2_emissions": null,
    "smashed_inference_energy_consumption": null
}