DOCSEARCHbyAlgolia
Submit Docs

Grid.ai Documentation

website·Indexed 33 days ago

https://docs.grid.ai/

A platform for training ML models and running compute-optimized 'Runs' and interactive 'Sessions'. Features include a CLI for managing artifacts and datastores, Bring Your Own Cloud (BYOC) infrastructure support for AWS, JupyterLab integration, and scalable dataset management via Datastores.

Tokens
31.6K
Snippets
205
Records
327
Agent score
59%

What's inside Grid.ai

  1. Overview of High Performance Datastores (HPDs)

    High Performance Datastores (HPDs) are designed for Grid BYOC customers scaling large datasets to optimize latency and speed up data access. They are backed by AWS FSx for Lustre and offer several advantages over conventional S3-backed datastores:

    • Scalability: Supports up to 64TB of data, maintaining performance even with millions of small files.
    • Single-User Performance: Up to 6x faster than conventional datastores.
    • Concurrency: Better scalability for a larger number of concurrent experiments or sessions.
    • Automatic Sync: HPDs automatically update contents to track the underlying S3 bucket data.

    Constraints: HPDs must be provisioned using a pre-existing S3 bucket owned by the same AWS account as the BYOC cluster.

  2. Overview of Grid.ai Sessions

    Sessions provide a live machine environment for prototyping, debugging, and analyzing. Key features include the ability to change machine instance types, use spot instances, mount multiple GPUs, auto-mount datastores, and access the environment via pre-installed JupyterLab, SSH, or Visual Studio Code. Sessions can be paused to save work and reduce credit consumption, then resumed later with all previously installed packages intact.

  3. Overview of Grid.ai core concepts

    Grid is a platform for developing and training deep learning models at scale. Its workflow consists of three primary components:

    1. Datastores: High-performance, low-latency, versioned datasets.
    2. Sessions: Interactive cloud machines used for prototyping, debugging, and analyzing models.
    3. Runs: Execution of scripts on cloud machines, supporting large-scale training and hyperparameter sweeps.

  4. Overview of Resumable Experiments on Grid

    Resumable experiments allow Grid to automatically restart a training job from the last saved checkpoint when running on interruptible spot instances. When a spot instance is reclaimed, Grid sends a SIGTERM signal to the process. Once a new instance becomes available, Grid restores all artifacts (including checkpoints) and the experiment execution directory, then re-runs the original command. This ensures that training can continue without manual intervention despite hardware interruptions.

  5. Overview of Grid Runs

    Runs are "serverless" compute instances optimized for training ML models. Because they are serverless, users only pay for the actual time a script is running, which reduces costs in both Grid Cloud and custom clusters. Runs enable parallelized hyperparameter sweeps, allowing multiple experiments to run simultaneously in the cloud.

  6. Overview of Grid Datastores

    Datastores in Grid are high-performance, low-latency, and auto-versioned datasets. They preserve the original file format and directory structure of the uploaded data and can be attached to Runs or Sessions. Data storage for Datastores is provided at no additional charge.

  7. Overview of Bring Your Own Cloud (BYOC) infrastructure modes

    Bring Your Own Cloud (BYOC) allows Grid to create clusters inside your own cloud account, providing complete control over resources. This feature is available on the Teams plan. There are two modes for cluster creation:

    1. Grid-Managed Infrastructure Mode: Grid automatically manages and provisions AWS infrastructure. This requires higher permissions, including iam:FullAccess.
    2. Self-Managed Infrastructure Mode: You manage and provision AWS infrastructure using a Grid-provided Terraform module. This requires minimal permissions, limited to the created EKS cluster and S3 bucket.

  8. Monitor costs using the Grid Dashboard

    The Grid dashboard provides a graphical representation of costs associated with Runs and Sessions. It allows users to track spending and identify currently active resources. Note that all costs displayed on the dashboard are estimates until the respective Runs or Sessions are fully completed. For detailed cost breakdowns per experiment or session, refer to the specific Runs and Sessions pages.

  9. Install BYOC prerequisites on M1 macOS

    Install the required tools for BYOC on M1 macOS using Homebrew, pip3, and Go. This process includes a manual build of the terraform-provider-null plugin to ensure compatibility with the darwin_arm64 architecture.
    brew install git
brew tap hashicorp/tap
brew install hashicorp/tap/terraform
brew update
brew upgrade hashicorp/tap/terraform
brew install jq
pip3 install awscli --upgrade --user
export GOPATH=$(go env GOPATH)
mkdir -p $GOPATH/src/github.com/terraform-providers
cd $GOPATH/src/github.com/terraform-providers
git clone git@github.com:hashicorp/terraform-provider-null.git
cd terraform-providers-null
go vet
make fmt
make build
mkdir -p $HOME/.terraform.d/plugins/registry.terraform.io/hashicorp/null/2.1.2/darwin_arm64
cp $HOME/go/bin/terraform-provider-null $HOME/.terraform.d/plugins/registry.terraform.io/hashicorp/null/2.1.2/darwin_arm64/terraform-provider-null_2.1.2

  10. Deploy Grid via self-hosted or private cloud installations

    In addition to the cloud-hosted service, Grid supports private cloud and self-hosted installations, allowing users to orchestrate infrastructure on their own private cloud environments.

  11. Handle BYOC instance architecture constraints

    Bring Your Own Cloud (BYOC) only supports instance types with the x86_64 architecture. Instance families with ARM architecture are not supported, specifically: A1, T4g, M6g, C6g, G5g, and R6g.

  12. Grid Datastore capabilities overview

    Grid Datastores provide a scalable, immutable dataset resource with the following capabilities:

    1. Creation from S3 buckets (available in both Grid Cloud and BYOC environments).
    2. Ability to attach to Runs and Sessions.
    3. Creation from local machines, Sessions, or Clusters.