{
    "base_current_gpu_type": "NVIDIA A100-PCIE-40GB",
    "base_current_gpu_total_memory": 40339.3125,
    "base_perplexity": 6.637787342071533,
    "base_token_generation_latency_sync": 39.003237915039065,
    "base_token_generation_latency_async": 38.958216086030006,
    "base_token_generation_throughput_sync": 0.025638897011020078,
    "base_token_generation_throughput_async": 0.025668526448740275,
    "base_token_generation_CO2_emissions": null,
    "base_token_generation_energy_consumption": null,
    "base_inference_latency_sync": 120.74454956054687,
    "base_inference_latency_async": 39.36746120452881,
    "base_inference_throughput_sync": 0.008281947331283504,
    "base_inference_throughput_async": 0.02540168884156951,
    "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": 7.51709508895874,
    "smashed_token_generation_latency_sync": 64.69257469177246,
    "smashed_token_generation_latency_async": 63.47550693899393,
    "smashed_token_generation_throughput_sync": 0.015457724549757006,
    "smashed_token_generation_throughput_async": 0.015754108131205573,
    "smashed_token_generation_CO2_emissions": null,
    "smashed_token_generation_energy_consumption": null,
    "smashed_inference_latency_sync": 197.72528839111328,
    "smashed_inference_latency_async": 111.56883239746094,
    "smashed_inference_throughput_sync": 0.0050575220202584105,
    "smashed_inference_throughput_async": 0.008963076681106845,
    "smashed_inference_CO2_emissions": null,
    "smashed_inference_energy_consumption": null
}