Cheaply Evaluating Inference Efficiency Metrics for Autoregressive Transformer Models

This repository contains the code used to produce the results in "Cheaply Evaluating Inference Efficiency Metrics for Autoregressive Transformer Models", accepted to NeurIPS 2023.

This repository contains the following:

• megatron_lm is a Git sub-module with Megatron-LM checked out to the commit hash used in our experiments.

  • scripts contains example scripts to profile models of interest on various prompt sizes and numbers of output tokens.
  • logs contains logfiles of profiling runs on V100 and A100 GPUs for various models.

• notebooks contains various IPython notebooks.

  • notebooks/fit_runtimes.ipynb is an IPython notebook that fits a linear regression model to the collected raw runtimes to determine per-model runtime of processing a prompt of a size ($\alpha_p$ in Equation 1) and the per-output-token runtimes ($\beta in Equation 2). These raw runtimes could be measured on a dedicated server deployment, or could be measured using black-box text generation APIs like OpenAI's davinci offering (in this repository, we use results from HELM's synthetic efficiency scenario that allows us to control both the prompt size and the number of output tokens). fit_runtimes.ipynb also dumps the learnt parameters into JSON files in processed_jsons/.
  • The other notebooks in this directory produce visualizations shown in the paper (e.g., end-to-end runtime versus number of output tokens for different models and prompt sizes, or prompt runtime versus prompt size for different models). These visualizations are collected in figures/.

Using Megatron-LM for autoregressive text generation

We use Megatron-LM in our experiments to generate text.

The main driver program is generate_text.py, which instantiates a model of the passed-in dimensions, and then generates the desired number of tokens given a prompt of a given size. The complete list of command line arguments supported by generate_text.py can be produced using the following command:

 PYTHONPATH=$PYTHONPATH:megatron_lm/ python generate_text.py -h
 ...
 text generation:
  --prompts-file PROMPTS_FILE
                        File with prompt (that can then be truncated)
  --all-num-input-tokens ALL_NUM_INPUT_TOKENS [ALL_NUM_INPUT_TOKENS ...]
                        Number of input tokens (must be > 0)
  --all-num-output-tokens ALL_NUM_OUTPUT_TOKENS [ALL_NUM_OUTPUT_TOKENS ...]
                        Number of tokens to generate
  --temperature TEMPERATURE
                        Sampling temperature.
  --top-p TOP_P         Top p sampling.
  --top-k TOP_K         Top k sampling.
  --out-seq-length OUT_SEQ_LENGTH
                        Size of the output generated text.

Citation

If you found this repository or our paper useful, feel free to cite our work:

@inproceedings{narayanan2023cheaply,
    title={{Cheaply Estimating Inference Efficiency Metrics for Autoregressive Transformer Models}},
    author={Deepak Narayanan and Keshav Santhanam and Peter Henderson and Rishi Bommasani and Tony Lee and Percy Liang},
    booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
    year={2023},
    url={https://openreview.net/forum?id=RJpAz15D0S}
}