Kinfer Runtime

K-Scale kinfer Skill

"The K-Scale model export and inference tool"

Trigger Conditions

• User asks about deploying RL policies to real robots
• Questions about ONNX model inference, Rust ML runtime
• Policy execution on embedded systems
• Real-time neural network inference

Overview

kinfer is K-Scale's model inference engine for deploying trained policies:

1. Model Loading: ONNX format support via ort (ONNX Runtime)
2. Real-time Execution: Rust implementation for low latency
3. Logging: NDJSON telemetry for debugging
4. Integration: Seamless connection with KOS firmware

Architecture

┌─────────────────────────────────────────────────────────────────────────┐
│  kinfer Inference Pipeline                                               │
│                                                                          │
│  ┌──────────────┐      load      ┌──────────────┐                       │
│  │  ONNX Model  │───────────────▶│   Runtime    │                       │
│  │  (.onnx)     │                │  (ort-sys)   │                       │
│  └──────────────┘                └──────┬───────┘                       │
│                                         │                                │
│  ┌──────────────┐      step      ┌──────┴───────┐      output           │
│  │ Observation  │───────────────▶│   Inference  │───────────────▶Action │
│  │  (sensors)   │                │    Engine    │                       │
│  └──────────────┘                └──────────────┘                       │
│                                         │                                │
│                                         ▼                                │
│                                  ┌──────────────┐                       │
│                                  │   Logger     │                       │
│                                  │  (NDJSON)    │                       │
│                                  └──────────────┘                       │
└─────────────────────────────────────────────────────────────────────────┘

Key Features

1. Single Tokio Runtime

// Efficient async execution with GIL management
lazy_static! {
    static ref RUNTIME: Runtime = Runtime::new().unwrap();
}

2. Pre-fetch Inputs

// Minimize latency by preparing inputs ahead of time
fn step_and_take_action(&mut self, observation: &[f32]) -> Vec<f32> {
    // Pre-fetch next input while processing current
    ...
}

3. NDJSON Logging

// Async logging thread for telemetry
struct Logger {
    file: File,
    tx: Sender<LogEntry>,
}

Language & Stack

• Primary: Rust (performance-critical)
• ML Runtime: ONNX Runtime (ort, ort-sys)
• Async: Tokio for non-blocking I/O
• Bindings: Python via PyO3

GF(3) Trit Assignment

Trit: -1 (MINUS)
Role: Verification/Validation (inference must be correct)
Color: #6E5FE4
URI: skill://kscale-kinfer#6E5FE4

Balanced Triads

kscale-kinfer (-1) ⊗ kscale-ksim (0) ⊗ onnx-export (+1) = 0 ✓
kscale-kinfer (-1) ⊗ rust-ml (0) ⊗ policy-training (+1) = 0 ✓

Key Contributors

Example Usage

import kinfer

# Load model
model = kinfer.load_model("walking_policy.onnx")

# Get observation from sensors
obs = get_sensor_data()

# Run inference
action = model.step(obs)

# Apply to actuators
apply_action(action)

Rust API

use kinfer::InferenceEngine;

let mut engine = InferenceEngine::load("policy.onnx")?;

loop {
    let obs = get_observation();
    let action = engine.step_and_take_action(&obs);
    send_to_actuators(&action);
}

References

• kscalelabs/kinfer - Main repository (17 stars)
• kscalelabs/kinfer-sim - Simulation visualization
• ONNX Runtime - Inference backend