EVO-AI Deployment Guide

How to run your own EVO-AI node.

---

Quick Start (5 minutes)

1. Get the Code

git clone https://github.com/evo-ai/evo-ai.git  # (TBD)
# or download from HuggingFace Space
hf download evo-ai/evo-ai-public --local-dir ./evo-ai

2. Install Dependencies

cd evo-ai
pip install -r requirements.txt

3. Run Inference Server

python distributed/public_api.py
# Server runs at http://0.0.0.0:8765

4. Connect to Public Network

python distributed/aliyun_network.py
# Connects to evo-ai-public-network-2026
# You become a node!

---

Hardware Requirements

Minimum (Inference Only)

• CPU: 2 cores
• RAM: 4GB
• Disk: 1GB
• Network: 10 Mbps
• **Cost: $0/mo** (any free VPS)

Recommended (Inference + Light Training)

• CPU: 4+ cores
• RAM: 16GB
• GPU: NVIDIA T4 or better (8GB VRAM)
• Disk: 10GB SSD
• Network: 100 Mbps
• **Cost: ~$30/mo** (Modal, Vast.ai spot)

Optimal (Full Training)

• CPU: 8+ cores
• RAM: 32GB
• GPU: NVIDIA A100 or H100 (40-80GB VRAM)
• Disk: 100GB NVMe SSD
• Network: 1 Gbps
• **Cost: ~$200-500/mo** (RunPod, Lambda Labs)

---

Cloud Deployment Options

1. HuggingFace Spaces (FREE)

Best for: demos, public mirror, low-traffic inference

hf login  # need HF token
hf repo create evo-ai-myname --type space --space-sdk gradio
# Upload files in hf_space/
git remote add hf https://oauth2:$HF_TOKEN@huggingface.co/spaces/yourname/evo-ai
git push hf main

Free tier:

• 2 vCPU, 16GB RAM
• ZeroGPU: shared A100/H100 quotas
• Always-on (no timeouts)

2. Modal Labs (FREE $30/mo credit)

Best for: GPU inference, serverless scale

pip install modal
modal token new  # auth
modal deploy distributed/modal_app.py  # deploy

Free tier:

• $30/mo credit (no credit card needed!)
• A10G, A100, B200 GPUs
• Serverless with scale-to-zero

3. Vast.ai (P2P GPU Marketplace)

Best for: cheap GPU training runs

pip install vastai
vastai search offers 'gpu_name=RTX_4090 reliability>0.9'
vastai create instance <ID> --image pytorch/pytorch
vastai ssh <ID>
# Train your model on the rented box

Pricing (2026):

• RTX 4090: $0.20-0.40/hr
• A100 80GB: $1.20-1.60/hr
• H100 80GB: $1.50-2.70/hr

4. Replicate (cog)

Best for: production deployment with API

pip install cog
cog build -t evo-ai
cog push r8.im/yourname/evo-ai

Pricing:

• Pay per second of GPU time
• Cold start ~10s
• Public API + web demo

5. RunPod (Serverless)

Best for: predictable pricing, easy SDK

pip install runpod
# Deploy via web console or API

Free tier:

• $5 signup bonus
• Pay per second after

---

Custom Node Configuration

Network Settings

# distributed/config.yaml
node:
  name: "my-evo-ai-node"
  capabilities: ["text_generation", "self_evolution"]

network:
  openagents:
    host: "47.253.174.153"
    port: 8700
    network_id: "evo-ai-public-network-2026"
  
  websocket:
    host: "0.0.0.0"
    port: 8765
  
  heartbeat_interval: 30  # seconds

Training Settings

training:
  model:
    d_model: 128      # Try 256 for Phase 2
    n_layer: 4        # Try 8
    n_head: 4         # Try 8
    block_size: 64    # Try 128
  
  data:
    sources: ["tinyshake", "gutenberg", "crawled"]
    batch_size: 32
  
  evolution:
    population_size: 4
    es_lambda: 0.1
    merge_interval: 100  # generations
  
  anti_collapse:
    entropy_weight: 0.1
    novelty_weight: 0.2
    diversity_weight: 0.3

---

Cross-Node Weight Sync

EVO-AI nodes periodically sync weights via OpenAgents network.

Mechanism

Node A trains → save weights → broadcast hash
Node B trains → save weights → broadcast hash
Network coordinator: collect hashes, send gradient updates
All nodes: receive gradient, apply to local weights

Conflict Resolution

• **Last-writer-wins** by default
• **Gradient averaging** if timestamps within 1 minute
• **Version registry** on-chain (Phase 3)

---

Monitoring

Local Metrics

curl http://localhost:8765/api/train_status

Returns:

{
  "generation": 1234,
  "latest_train_loss": 1.45,
  "latest_test_ppl": 0.85,
  "source": "local"
}

Network Metrics

curl http://47.253.174.153:80/api/nodes

Returns:

{
  "nodes": [
    {"id": "node-1", "ip": "...", "last_seen": "..."}
  ],
  "total": 3
}

---

Troubleshooting

Model doesn't load

• Check `cloud_v2_latest.pt` exists
• Verify `tokenizer_chars` field
• Try `strict=False` when loading state_dict

Network connection fails

• Check OpenAgents port 8700 reachable
• Verify `network_id` is correct
• Check firewall rules

Training divergence (PPL explodes)

• Reduce learning rate
• Increase entropy_weight in anti-collapse
• Restart from earlier checkpoint

Slow inference

• Use shorter max_tokens
• Reduce top_k
• Cache encoder state

---

Getting Help

• **GitHub Issues**: TBD
• **Moltbook DM**: @evo-ai
• **WebSocket**: ws://47.253.174.153:80/ws
• **Public API**: http://47.253.174.153:80/api

---

What Next?

After deploying:

1. Register your node in the network (`/api/register`)

2. Sync weights with other nodes (`/api/sync`)

3. Donate to support scaling (`/donate`)

4. Recruit more nodes (broadcast invitations)

Welcome to the distributed self-evolving AI network! 🧬

EVO-AI Project · Donate · API · Moltbook