Skip to content

Latest commit

 

History

History
117 lines (72 loc) · 4.18 KB

install.md

File metadata and controls

117 lines (72 loc) · 4.18 KB
Raw

MlAgility Installation Guide

The following describes how to install MlAgility.

Operating System Requirements

MLAgility is tested using ubuntu-latest, which currently defaults to Ubuntu 20.04. However, MLAgility is also expected to work on other recent versions of Ubuntu (e.g. 18.04, 22.04), and other flavors of Linux (e.g. CentOS).

Installing mlagility

We recommend that you install miniconda like this:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

Then create a virtual environment like this:

conda create -n mla python=3.8
conda activate mla

From here you have two options, installing from PyPI or cloning.

PyPI Install

The easiest way to get started:

pip install mlagility

Cloning Install

To clone and install mlagility:

git clone https://github.com/groq/mlagility.git
cd mlagility
pip install -e .

Additional Steps

If you are planning to use the mlagility tools with the MLAgility benchmark, Groq, or Slurm please see the corresponding sections below.

MLAgility Benchmark Requirements

The MLAgility benchmark's models are located at mlagility_install_path/models, which we refer to as models/ in most of the guides.

Note: The benchit models location command and mlagility.common.filesystem.MODELS_DIR are useful ways to locate the models directory. If you perform PyPI installation, we recommend that you take an additional step like this:

(mla) jfowers@jfowers:~/mlagility$ benchit models location

Info: The MLAgility models directory is: ~/mlagility/models
(mla) jfowers@jfowers:~/mlagility$ export models=~/mlagility/models

The mlagility package only requires the packages to run the MLAgility benchmarking tools. If you want to run the MLAgility benchmark itself, you will also have to install the benchmark's requirements.

In your miniconda environment:

pip install -r models/requirements.txt

Installing Groq support

When you install pip install mlagility[groq], pip will also install the groqflow pip package for you.

However, if you want to use mlagility with your Groq SDK and GroqChip processor hardware you must also follow the steps in the GroqFlow Install Guide, particularly:

  1. Prerequisites: install the GroqWare™ Suite.
  2. Add GroqWare Suite to Python Path

Installing Slurm support

Slurm is an open source workload manager for clusters. If you would like to use Slurm to build multiple models simultaneously, please follow the instructions below. Please note that MLAgility requires the Slurm nodes to have access to at least 128GB of RAM.

Setup your Slurm environment

Ensure that your MLAgility folder and your conda installation are both inside a shared volume that can be accessed by Slurm. Then, run the following command and wait for the Slurm job to finish:

sbatch --mem=128000 src/mlagility/cli/setup_venv.sh

Get an API token from Huggingface.co (optional)

Some models from Huggingface.co might require the use of an API token. You can find your api token under Settings from your Hugging Face account.

To allow slurm to use your api token, simply export your token as an environment variable as shown below:

export HUGGINGFACE_API_KEY=<YOUR_API_KEY>

Setup a shared download folder (optional)

Both Torch Hub and Hugging Face models save model content to a local cache. A good practice is to store that data with users that might use the same models using a shared folder. MLAgility allows you to setup a shared ML download cache folder when using Slurm by exporting an environment variable as shown below:

export SLURM_ML_CACHE=<PATH_TO_A_SHARED_FOLDER>

Test it

Go to the mlagility folder and build multiple models simultaneously using Slurm.

benchit models/selftest/*.py --use-slurm --build-only --cache-dir PATH_TO_A_CACHE_DIR