Constantinople (SN97)

Decentralized LLM inference on Bittensor. Miners serve language models, validators verify quality using hidden-state verification.

Architecture

Constantinople uses a challenge-response protocol to verify that miners are running real models:

1. Gateway receives inference requests and routes them to miners
2. Miners serve Qwen2.5-7B-Instruct via vLLM with hidden state extraction
3. Validator challenges miners by requesting hidden states at random (layer, position) pairs and comparing against reference computations
4. Scoring combines throughput, latency, and verification pass rate
5. Epoch summaries are published to R2 for transparency — anyone can audit the scoring

Quick Links

• API: api.constantinople.cloud (OpenAI-compatible)
• Dataset: api.constantinople.cloud/v1/dataset/ (all inference data, downloadable)
• Website: constantinople.cloud

Run a Validator

See validator/README.md for full details. Two options:

Watchtower (no GPU) — follow published scores:

docker compose up -d

Full Validator (GPU) — run independent hidden-state verification:

pip install -r validator/requirements.txt
python validator/hardened_gateway.py --port 8081 --model Qwen/Qwen2.5-7B-Instruct --netuid 97 --discover
python validator/audit_validator.py --port 8082 --model Qwen/Qwen2.5-7B-Instruct --wallet validator --hotkey default --netuid 97

Run a Miner

See miners/README.md for the full setup guide (registration, PM2, Docker, networking). Requires a GPU (RTX 4090+ recommended).

pip install -r miners/requirements.txt
python miners/vllm_miner.py --model Qwen/Qwen2.5-7B-Instruct --port 8091 --gpu-memory-utilization 0.70 --hf-device cpu

Incentive Mechanics

• Throughput (~40% weight): Tokens per second. Faster miners earn more.
• Latency (~20% weight): Time-to-first-token. Lower is better.
• Verification (~40% weight): Hidden-state challenge pass rate. Must be consistently honest.
• Bayesian smoothing: New miners get the benefit of the doubt (2 virtual passes). Score stabilizes after ~5 challenges.
• Adaptive challenge rate: 10% floor, minimum 5 challenges per 25-minute epoch.

Auto-Updates

Validators running via Docker Compose get automatic updates through Watchtower. When we push a new validator image, Watchtower detects the change and restarts your container with the latest code.

License

MIT