Autofoundry

Run any ML experiment script across GPUs on multiple cloud providers with a single command.

Autofoundry is a CLI companion to Karpathy's autoresearch. Point it at a shell script, pick your GPU configuration, and it handles the rest: provisioning instances, distributing experiment runs, streaming results live, and producing a final metrics report.

Supported Providers

• RunPod — Secure and Community cloud
• Vast.ai — Global GPU marketplace
• PRIME Intellect — Decentralized GPU network
• Lambda Labs — On-demand cloud GPUs

Quickstart

git clone https://github.com/autofoundry/autofoundry.git
cd autofoundry
uv tool install autofoundry

Then run:

autofoundry run

On first run, Autofoundry walks you through configuring provider API keys, SSH key path, minimum download bandwidth (default 5000 Mbps — filters out slow Vast.ai hosts), and HuggingFace token. Config is saved to ~/.config/autofoundry/config.toml.

Examples

Run experiments

# Interactive mode — walks you through everything
autofoundry run

# Run a specific script
autofoundry run scripts/run_autoresearch.sh

# Specific GPU with multiple experiment runs
autofoundry run train.sh --gpu H100 --num 4

# Auto-select cheapest datacenter GPU with 80GB+ VRAM
autofoundry run train.sh --segment datacenter --min-vram 80 --auto

# Target a specific provider
autofoundry run train.sh --segment datacenter --min-vram 80 --provider runpod --auto
autofoundry run train.sh --segment datacenter --provider lambdalabs --auto
autofoundry run train.sh --segment workstation --provider vastai --auto
autofoundry run train.sh --segment datacenter --provider primeintellect --auto

# Attach a network volume (RunPod, Lambda Labs)
autofoundry run train.sh --volume my-data --provider runpod

# Resume a previous session
autofoundry run --resume <session-id>

Browse GPU inventory

# Browse all available GPUs across providers
autofoundry inventory

# Filter by segment, VRAM, or GPU name
autofoundry inventory --segment datacenter --min-vram 80
autofoundry inventory --gpu A100

Configure

# Interactive setup for API keys, SSH key, and defaults
autofoundry config

Manage volumes

# List volumes across providers
autofoundry volumes list

# Create a new volume
autofoundry volumes create --name my-data --provider runpod

Monitor and manage sessions

# Show all sessions
autofoundry status

# Show a specific session
autofoundry status <session-id>

# View metrics from most recent run
autofoundry results

# Terminate instances for a session
autofoundry teardown <session-id>

See the full documentation for writing experiment scripts, network volumes, resuming sessions, custom images, CLI reference, and architecture details.

Requirements

• Python 3.11+
• SSH key pair (ed25519 or RSA)
• At least one provider API key (RunPod, Vast.ai, PRIME Intellect, or Lambda Labs)

License

MIT