README.md

metadata

dataset_info:
  features:
    - name: task
      dtype: string
    - name: org
      dtype: string
    - name: model
      dtype: string
    - name: hardware
      dtype: string
    - name: date
      dtype: string
    - name: prefill
      struct:
        - name: efficency
          struct:
            - name: unit
              dtype: string
            - name: value
              dtype: float64
        - name: energy
          struct:
            - name: cpu
              dtype: float64
            - name: gpu
              dtype: float64
            - name: ram
              dtype: float64
            - name: total
              dtype: float64
            - name: unit
              dtype: string
    - name: decode
      struct:
        - name: efficiency
          struct:
            - name: unit
              dtype: string
            - name: value
              dtype: float64
        - name: energy
          struct:
            - name: cpu
              dtype: float64
            - name: gpu
              dtype: float64
            - name: ram
              dtype: float64
            - name: total
              dtype: float64
            - name: unit
              dtype: string
    - name: preprocess
      struct:
        - name: efficiency
          struct:
            - name: unit
              dtype: string
            - name: value
              dtype: float64
        - name: energy
          struct:
            - name: cpu
              dtype: float64
            - name: gpu
              dtype: float64
            - name: ram
              dtype: float64
            - name: total
              dtype: float64
            - name: unit
              dtype: string
  splits:
    - name: benchmark_results
      num_bytes: 1886
      num_examples: 7
    - name: train
      num_bytes: 2446
      num_examples: 9
  download_size: 75548
  dataset_size: 4332
configs:
  - config_name: default
    data_files:
      - split: benchmark_results
        path: data/train-*
      - split: train
        path: data/train-*

Analysis of energy usage for HUGS models

Based on the energy_star branch of optimum-benchmark, and using codecarbon.

Fields

• task: Task the model was benchmarked on.
• org: Organization hosting the model.
• model: The specific model. Model names at HF are usually constructed with {org}/{model}.
• date: The date that the benchmark was run.
• prefill: The esimated energy and efficiency for prefilling.
• decode: The estimated energy and efficiency for decoding.
• preprocess: The estimated energy and efficiency for preprocessing.

Code to Reproduce

As I'm devving, I'm hopping between https://huggingface.co/spaces/AIEnergyScore/benchmark-hugs-models and https://huggingface.co/spaces/meg/CalculateCarbon

From there, python code/make_pretty_dataset.py (included in this repository) takes the raw results and uploads them to the dataset here.