Bases: Worker
Custom vLLM GPU worker that uses :class:NNsightGPUModelRunner.
Monkey-patches the default GPUModelRunner class before
initialization so vLLM creates NNsight-aware model runners
that can execute intervention code during model forward passes.
collect_nnsight
collect_nnsight(req_ids: list[str], finished_req_ids: list[str] | None = None)