Model-internal monitoring needs access to the model runtime. WoolyAI runs inside the address space of the inference environment serving the open-source model and reads internal signals while the model is generating.
You do not control the model’s inference runtime.
The request format does not matter. Where the model runs does.
Yes. vLLM serving Qwen at /v1/chat/completions can look like the OpenAI API to your application. WoolyAI still has access to the model runtime underneath.
Different case. You do not control that model execution, so you cannot read its internal state.
Internal neuron signals are only visible within the model serving. Input and output monitoring only see the edges and the response is generated.