Steering backends & unified runner¶
Generate from a steered model at scale. The steer.run() entry point dispatches to the correct backend.
| Backend | Speed | Use case |
|---|---|---|
| Unified runner | — | Auto-dispatches based on wrapper type |
| VLLMHookSteer | ~6× HF | Activation-steering at scale |
| VLLMDecodeSteer | vLLM | Decoding-steering at scale |
| SteerSpec | — | Serializable steering spec |
| DecodeSteerSpec | — | Serializable decoding spec |