Skip to content

llm-efficiency-challenge/eval-scripts

Repository files navigation

Eval Scripts

TL;DR

These are the very basic scripts are for running submissions of the competition. Scripts herein are very dumb, intentionally so. PRs welcome for better scripts!

Setup

The following tools need to be setup on an evaluation machine.

  • docker and nvidia-docker for running the helm suite and the submissions
  • git for cloning repos
  • curl for running the healthchecks to ensure containers are running

With these installed run the setup.sh script with the number of GPUs in the evaluation machine.

./setup.sh $NUM_GPUS

... for example to use 8 gpus in an eval machine

./setup.sh 8

What this does

This script will ensure a few tools are present and functional, as well as setting up a degree of isolation on the machine. This takes the form of docker networks and port exposures for each individual GPU.

We also checkout the helm evaluation version used in the competition, into the ./private-helm folder.

Running a submission

Make sure the submissions are in ./submissions folder relative to this script.

To run a single submission do:

./eval-repo.sh         \
    '$gpu_device'      \
    '$isolation'       \
    '$hardware_track'  \
    '$helm_config'     \
    '$submission'

What this does:

  • '$gpu_device' Specifies the GPU string to pass to docker for the GPU(s) to run the submission on
  • '$isolation' An isolation factor used to divide submissions between multiple GPUs on a single server. We recommend to keep this number the same as the $gpu_device string.
  • '$helm_config' The config used for helm, must be in a path visibile to the helm container. Private-helm contains configs within the container for the 111 competition.
  • '$hardware_track' The hardware track to build for in the submissions folder. Essentially a top level folder in ./submissions that differentiates between different hardware tracks/
  • '$submission' The submission folder to build

The $submission is the folder that contains the submission to evaluate. A bare Dockerfile is expected at this location to build the submissions container.

Layout of the submissions folder

Submissions are laid out in ./submissions folder in the following fashion:

./submissions
├── 4090
│   └── $user
│       └── $repo
│           ├── README.md
│           ├── submission_1
│           │   ├── Dockerfile
│           │   └── ...
│           └── submission_2
│               ├── Dockerfile
│               └── ...
└── A100
    └── $user
        └── $repo
            ├── README.md
            ├── submission_1
            │   ├── Dockerfile
            │   └── ...
            └── submission_2
                ├── Dockerfile
                └── ...

Putting this together for a simple run

Using the above layout to run a submission for A100 $user $repo/submission_2 on the second GPU you would do the following:

./eval-repo.sh                           \
    'device=1'                           \
    '1'                                  \
    'A100'                               \
    '/helm/config/some_helm_config.conf' \
    '$user/$repo/submission_2'

Changing locations

Should you want to change the locations of the ./private-helm, ./submissions or ./results folders you will need to edit utils.sh

About

The eval scripts used to run and eval submissions

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages