Bittensor GPT Subnet Workflow Guide

This project enables a decentralized network for training GPT models via Bittensor. Miners submit model implementations, and validators evaluate them to determine network weights.

Prerequisites

• Python 3.8+
• Docker + NVIDIA Container Toolkit (for GPU validation)
• Bittensor wallet (btcli w new_coldkey, btcli w new_hotkey)
• GitHub Personal Access Token (for validator)
• Clone this repository

git clone [REPO_URL] && cd [REPO_NAME]
pip install -r requirements.txt

1. Configuration

Edit config.py:

# For testnet
network = "test"
netuid = 100  # Subnet UID
wallet_name = "your_wallet"
wallet_hotkey = "your_hotkey"
weight_server_url = "http://your-server:8000"

2. Miner Submission

1. Implement Components in miner_submission.py:

   • Custom AttentionBlock
   • GPTModel architecture
   • Training Optimizer
   • LossFunction

2. Push to GitHub repository containing:

   • miner_submission.py
   • hotkey.json (contains your Bittensor hotkey SS58 address)

3. Submit to Chain:

python submit_to_chain.py \
  --repo yourgithub/repo \
  --wallet-name your_wallet \
  --wallet-hotkey your_hotkey \
  --network test for testnet finney for mainnet \
  --subnet-uid 100 for testnet 49 for mainnet

3. Validator Operation

1. Start Weight Server (separate terminal):

python weight_server.py

2. Run Validator:

export GITHUB_TOKEN=your_github_token
python main.py

Validation Process:

1. Syncs with Bittensor chain every 10 minutes
2. Pulls new submissions from GitHub
3. Runs Docker containers to:
   • Generate model fingerprints
   • Check for code duplicates
   • Evaluate training performance
4. Scores models based on validation loss
5. Updates network weights via weight server

4. Additional Components

Weight Updates

• Validators send scores to weight_server.py
• Server verifies signatures and updates on-chain weights

Security Features

• Container isolation with read-only filesystems
• Signature verification for weight updates
• Duplicate detection through code hashing

Monitoring

Check databases for validation results:

• submissions.db - Current submissions
• model_records.db - Historical validation data

Troubleshooting

1. Docker Permissions: Ensure user in docker group
2. GPU Access: Verify nvidia-docker installation
3. Chain Sync: Check btt_connector.py logs for sync errors
4. Rate Limits: Validator automatically handles GitHub API limits