#Getting started

Power has two inference entry points and one independent artifact surface. Pick the narrowest boundary your product needs.

#1. Embed the runtime

Add only the listener-free inference surface:

[dependencies]
a3s-power = { version = "1.0.0", default-features = false, features = ["embedded-inference"] }

Construct a runtime with an explicit device preference and resource limits:

use a3s_power::inference::{DevicePreference, EmbeddedRuntime, InferenceLimits};

fn main() -> Result<(), a3s_power::error::PowerError> {
    let runtime = EmbeddedRuntime::new(
        DevicePreference::Auto,
        InferenceLimits::default(),
    )?;

    println!("execution device: {}", runtime.device().name());
    Ok(())
}

Creating EmbeddedRuntime does not bind a socket, start a listener, download a model, or invoke another process. The model crate supplies its reviewed graph and retains semantic state; Power supplies the execution boundary.

Continue with Architecture to see the ownership contract.

#2. Run the hosted service

Install the default service profile and bind it to loopback:

cargo install a3s-power
a3s-power serve --host 127.0.0.1 --port 11434

In another terminal, pull and open a small GGUF model:

a3s-power models pull Qwen/Qwen2.5-0.5B-Instruct-GGUF:q4_k_m
a3s-power chat Qwen/Qwen2.5-0.5B-Instruct-GGUF:q4_k_m

Model manifests and content-addressed blobs live under ~/.a3s/power by default. Set A3S_POWER_HOME to choose another store.

#3. Send an OpenAI-compatible request

curl http://127.0.0.1:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "your-model",
    "messages": [{"role": "user", "content": "Explain capability-based security."}],
    "stream": true
  }'

Chat and completion responses include an attestation_receipt and its SHA-256 digest. Streaming responses emit the receipt before [DONE].

#4. Make policy explicit

The service reads A3S ACL from ~/.a3s/power/config.acl, or from the path passed to a3s-power serve --config.

host = "127.0.0.1"
port = 11434
max_loaded_models = 1
keep_alive = "5m"

flash_attention = true
num_parallel = 1

gpu {
  gpu_layers = -1
  main_gpu = 0
}

Invalid ACL, ranges, strategies, hashes, or unsupported explicit backends fail before inference. Production TEE policy adds verifier-owned measurements and model hashes; see Verification.

#Next steps

• Understand the execution and ownership boundary
• Choose optimization by graph, tensor, scheduling, weights, and evidence
• Inspect measured Qwen3.8 performance and quality evidence
• See how speculative decoding preserves exact target authority
• Choose a backend and production build profile