Skip to content

SETUP.md

Latest commit

 

History

History
151 lines (108 loc) · 3.83 KB

SETUP.md

File metadata and controls

151 lines (108 loc) · 3.83 KB
Raw

Environment Setup

Local Development (no Docker)

# Install dependencies (requires uv: https://docs.astral.sh/uv/)
uv sync

# Or with pip:
pip install -e .

# Prepare data for a game (downloads from HuggingFace)
python scripts/prepare_data.py --game pong

# Train a single model
cd tasks/pong_dreamer
python train.py

Harbor (Agent Benchmarking)

Harbor runs each task inside a Docker container with an AI coding agent. Follow all steps below for a fresh machine.

0. Install Harbor

uv sync --extra harbor

This installs Harbor and all dependencies into .venv/.

1. Agent Authentication

Copy .env.example to .env and fill in credentials:

cp .env.example .env
Agent Required Variable Optional
claude-code ANTHROPIC_API_KEY ANTHROPIC_BASE_URL (custom gateway)
claude-code --bedrock AWS_PROFILE + AWS_REGION —
codex OPENAI_API_KEY (or AZURE_OPENAI_API_KEY) OPENAI_BASE_URL (Azure/proxy)

For AWS Bedrock mode, configure credentials:

# ~/.aws/config
[profile default]
sso_session = my-sso
sso_account_id = <your-account-id>
sso_role_name = <role-with-bedrock-access>
region = us-east-1

[sso-session my-sso]
sso_start_url = https://<your-org>.awsapps.com/start/#
sso_region = us-east-1
sso_registration_scopes = sso:account:access

Then: aws sso login

The container runs as a non-root agent user (UID 1007). If ~/.aws files are owner-only (common when running as root), fix permissions:

chmod 644 ~/.aws/config
chmod -R o+rX ~/.aws/sso ~/.aws/cli

2. Build the Base Docker Image

Every per-task Dockerfile starts with FROM autoworldbench-base:latest:

docker build -t autoworldbench-base \
    --build-arg HOST_UID=$(id -u) --build-arg HOST_GID=$(id -g) \
    -f docker/Dockerfile.harbor .

3. NVIDIA Container Toolkit

curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
    | gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
    | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
    | tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
apt-get update && apt-get install -y nvidia-container-toolkit
nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

4. Apply Harbor Patches

Harbor requires patches for Codex exec loop support and bind-mount permissions:

bash scripts/apply_harbor_patches.sh

Re-run after every uv sync or Harbor upgrade.

Patch file Purpose
harbor_codex.py Codex exec loop, base64 instruction encoding, auth fixes
harbor_docker.py chmod as root so agent user can write to bind-mounted dirs
harbor_docker_compose_base.yaml Strip deploy.resources for compatibility

5. Validate

# Network from containers:
docker run --rm alpine sh -c "nslookup archive.ubuntu.com"

# GPU access:
docker run --rm --runtime=nvidia alpine sh -c "echo GPU ok"

# Base image exists:
docker images | grep autoworldbench-base

# End-to-end test:
./run_harbor.sh --task pong_dreamer --agent claude-code

6. Run

# Single task
./run_harbor.sh --task pong_dreamer --agent claude-code

# All models for one game
./run_harbor.sh --game pong --agent claude-code

# GPU-parallel orchestration (all 32 tasks)
python scripts/orchestrate.py --all --agent claude-code --gpus 0,1,2,3

# Dry run (show what would execute)
python scripts/orchestrate.py --all --agent claude-code --dry-run

Preflight Check

Validate the full environment before running:

python scripts/preflight.py
python scripts/preflight.py --fix   # auto-fix: prepare data, apply patches