BTT InferGrid Documentation

This documentation describes the miner-side components in the BTT InferGrid compute network, for developers and operators who need to deploy and manage inference services on NVIDIA GPU hosts.

Prerequisites

Familiarity with Linux command line, Docker containers, and GPU inference concepts is recommended before reading.

Core Projects

The current repository set consists of two core components:

Project	Purpose	Entry Command
miner-cli	Miner service command-line deployment tool	miner-cli
miner-agent	Node control-plane agent	miner-agent

miner-cli

Docker-based LLM deployment helper for single Linux hosts, responsible for:

• Checking Linux GPU host environment
• Rendering Docker Compose configurations
• Starting inference runtime and managing lifecycle

miner-agent

FastAPI sidecar running in the inference service network, responsible for:

• Node registration and heartbeat
• Challenge verification
• Local diagnostics API

Typical Deployment Topology

The default operating model is a three-container topology on a single NVIDIA GPU miner host:

Container	Responsibility
LLM runtime	Runs vllm or sglang, exposes OpenAI-compatible /v1 API
dcgm-exporter	Exposes NVIDIA GPU metrics on /metrics
miner-agent	Registers node, signs control-plane messages, sends heartbeats, handles challenges, exposes local health API

Quick Start Flow

1. Install Python 3.10+ and use uv or pip to install the project
2. Use miner-cli init to generate a deployment YAML
3. Run miner-cli doctor to check the host
4. Run miner-cli toolkit verify to validate GPU container support
5. Run miner-cli runtime prepare to prepare the runtime
6. Use miner-cli up to start the model runtime
7. Check liveness, readiness, identity, and control-plane state via the agent local API

Related Documentation

• Detailed command reference: miner-cli Commands
• Configuration: miner-cli Configuration
• Troubleshooting: Operations Guide