Hosting Guide

Earn credits by registering your GPU as a host on the Infrintia compute marketplace. When users submit inference jobs, the broker routes them to available hosts — you run the computation and get paid.

---

Requirements

• A machine with one or more NVIDIA GPUs (CUDA-capable)
• Python 3.9+
• Stable internet connection
• Sufficient VRAM for the models you want to serve

---

Register as a Host

1. Install the Infrintia host agent

pip install infrintia[host]

2. Register your machine

from infrintia.host import HostAgent

agent = HostAgent(
    api_url="https://api.infrintia.crossgl.net",
    gpu="NVIDIA A100 80GB",
    models=["meta-llama/Llama-3-8B-Instruct"],
    price_per_token=0.00012,
    backend="huggingface"
)

registration = agent.register()
print(f"Host ID: {registration['host_id']}")
print(f"API Key: {registration['api_key']}")

The registration response includes a host_id and api_key — store the API key securely. You'll need it for all subsequent host operations.

3. Start serving

agent.serve()

This starts an event loop that:

1. Sends heartbeats every 30 seconds
2. Polls for new jobs via POST /hosts/next-job
3. Runs inference on assigned jobs
4. Streams results back to the platform

---

Auto-Detection

The host agent automatically detects:

• GPU model and VRAM via nvidia-smi
• Available models based on installed backends and cached weights
• System resources (CPU, RAM, disk) for capacity reporting

You can override any auto-detected value in the HostAgent constructor.

---

Backends

Choose the backend that matches your setup:

HuggingFace (default)

Runs models using HuggingFace Transformers. Models are downloaded and cached automatically.

agent = HostAgent(
    api_url="https://api.infrintia.crossgl.net",
    backend="huggingface",
    models=["meta-llama/Llama-3-8B-Instruct"],
    price_per_token=0.00012
)

LangChain

Execute LangChain chains and agents.

agent = HostAgent(
    api_url="https://api.infrintia.crossgl.net",
    backend="langchain",
    models=["custom-chain-v1"],
    price_per_token=0.0002
)

Worker (Custom)

Point to your own inference server. Useful for custom models, quantized weights, or vLLM/TGI backends.

agent = HostAgent(
    api_url="https://api.infrintia.crossgl.net",
    backend="worker",
    endpoint_url="http://localhost:8000/generate",
    models=["my-custom-model"],
    price_per_token=0.00015
)

---

Earnings

• Hosts are paid per token generated, at the price_per_token they set during registration
• The platform takes a 12% fee; the remaining 88% is credited to the host
• Earnings accumulate in the host's credit balance and can be withdrawn

Example

If you set price_per_token=0.00012 and generate 1,000,000 tokens:

• Gross: 120.00 credits
• Platform fee (12%): 14.40 credits
• Net earnings: 105.60 credits

---

Monitoring

Check your host's status and earnings:

from infrintia import ComputeClient

client = ComputeClient(
    base_url="https://api.infrintia.crossgl.net",
    api_key="hk_live_xxxxxxxx"
)

hosts = client.list_hosts()
for h in hosts:
    if h["host_id"] == "host_xyz":
        print(f"Status: {h['status']}")
        print(f"Jobs completed: {h['total_jobs_completed']}")

---

Best Practices

• Set competitive pricing — check GET /hosts to see current market rates
• Keep your host online — missed heartbeats mark your host as inactive
• Use fast storage — model loading speed affects job pickup latency
• Monitor GPU utilization — the heartbeat reports utilization to the broker for smart routing