Skip to content

Undercurrent Undercurrent

Wrynx's activation-probing platform. See the undercurrent before it surfaces.

Undercurrent is production-grade activation probing for live inference: it reads a model's internal activations while it generates, runs your probes on them, and can stop a generation the moment a probe objects.

  • Declarative specs. A short YAML spec says which tensor to capture, at which layers and token positions, and which probe receives it. No model surgery, no forked serving code.
  • Inline intervention. Probes run on the generation path and can abort a generation mid-stream, before the next token is produced.
  • Production-ready async execution on vLLM. Observe-only probes run off the hot path on bounded worker queues with overflow policies, and report to file, webhook and metrics sinks.
from undercurrent.core import probe
from undercurrent.model import ProbedModel


@probe("norm_gate", threshold=500.0)
def norm_gate(record):
    return float(record.tensor.norm())


with ProbedModel.from_pretrained("openai-community/gpt2", spec=SPEC, probes={"norm_gate": norm_gate}) as model:
    out = model.generate("The quick brown fox", max_new_tokens=20, temperature=0)
print(out.text, out.probe_results)

SPEC is a few lines of YAML; the Quickstart walks through it end to end on a laptop CPU.