Vllm support

[IRENEKO]

AISteer360 steers models by registering PyTorch forward hooks (pre/forward/backward) on specific layers. This works great when you own the model object. But vLLM spawns worker subprocesses that each load their own copy of the model, and you can't just pass a hooked model across process boundaries.

Main components:

1. backend.py

• Serializes the StateControl into a JSON "recipe" (via recipe.py)
• Writes it to a temp file, sets VLLM_AISTEER360_CONFIG env var
• Registers AISteer360Worker in vLLM-Hook's plugin registry
• Boots a HookLLM engine that will use that worker

2. recipe.py

• serialize_state_control() : Extracts everything needed to rebuild the StateControl, including

  • The class path (e.g. aisteer360.algorithms.state_control.caa.control.CAAControl)
  • Constructor args from the dataclass
  • Pre-computed steering vectors saved as .svec files
  • Resolved layer IDs/names (so the worker doesn't re-compute them)

• reconstruct_state_control(): Rebuilds it by dynamically importing the class, loading vectors from disk, and swapping prompt args with pre-computed vectors to skip extraction.

3. worker.py

AISteer360Worker extends vLLM's V1Worker. After load_model():

• Reads the recipe JSON from the env var path
• Reconstructs the StateControl
• Calls state_control.steer(model) on the real model
• Gets hooks and wraps them with _gated_hook() which does two things:
  • Flag-file gating: HookLLM can toggle steering on/off by creating/removing a flag file
  • Tensor shape normalization: AISteer360 hooks expect [B, T, H] (3-D) but vLLM passes [N, H] (2-D), so it unsqueezes before the hook and squeezes after

Additional revisions:

1. Minor helpers for vllm I/O
2. Backend reroute and support in aisteer360/algorithms/core/steering_pipeline.py for backend="vllm"
3. Tiny dummy example test_vllm_state.py showing how to use

@emiehling Please advise

[emiehling]

Thanks @IRENEKO I'll dig into this shortly