wren runs its model on one BLAS thread

threadpoolctl scopes the limit to wren.act, so other numpy calls in the
engine keep their threads.

backend/fluksio/flow/nodes/wren.py

@@ -14,6 +14,7 @@ from collections.abc import Iterable
 from typing import Any
 
 from pydantic import BaseModel, ConfigDict, Field
+from threadpoolctl import threadpool_limits
 
 from fluksio.flow.messages import DType, MessageSpec
 from fluksio.flow.nodes.base import Node
@@ -306,7 +307,11 @@ class WrenNode(Node):
         wren = _wren()
         specs = {spec.port: spec for spec in self.output_ports}
         house = self._house(wren, specs, presence, outdoor, dark)
-        calls, confidence = wren.act(prompt, house)
+        # One BLAS thread answers in about half the time for a model this
+        # small. Scoped to the call, so numpy elsewhere in the process keeps
+        # its threads.
+        with threadpool_limits(1):
+            calls, confidence = wren.act(prompt, house)
         confidence = round(float(confidence), 3)
         write = self.cfg.apply and confidence >= self.cfg.min_confidence
         writes: dict[str, Any] = {}

backend/pyproject.toml

@@ -58,6 +58,10 @@ dependencies = [
     # The `mlp` node, which is registered at boot. Kept out of the extras
     # because a data-science environment has it already.
     "numpy>=2.2.6",
+    # The `wren` node runs its model on one BLAS thread, which answers in about
+    # half the time. Scoped to the call: OPENBLAS_NUM_THREADS would bind every
+    # other numpy call in the engine too, and only counts before numpy loads.
+    "threadpoolctl>=3.5",
     # The `inject` node's cron. Imported where it is used, behind a message
     # that cannot fire while this line is here.
     "croniter>=1.3.0",
@@ -164,6 +168,11 @@ ignore_missing_imports = true
 module = ["wren"]
 ignore_missing_imports = true
 
+# What scopes the `wren` node's BLAS threads to its call; it ships no types.
+[[tool.mypy.overrides]]
+module = ["threadpoolctl"]
+ignore_missing_imports = true
+
 
 [tool.ruff]
 target-version = "py312"

backend/tests/flow/test_wren.py

@@ -143,3 +143,25 @@ def test_nothing_is_written_unless_applied_and_sure(fake_wren, params, confidenc
             }
         ]
     }
+
+
+def test_the_model_runs_on_one_blas_thread(fake_wren, monkeypatch):
+    """About twice as fast for a model this small; numpy is loaded by `mlp`."""
+    from threadpoolctl import threadpool_info
+
+    threads: list[int] = []
+    act = fake_wren.act
+
+    def counted(command: str, house: Any) -> tuple[list[Any], float]:
+        threads.extend(
+            pool["num_threads"]
+            for pool in threadpool_info()
+            if pool["user_api"] == "blas"
+        )
+        return act(command, house)
+
+    monkeypatch.setattr(fake_wren, "act", counted)
+    _node(("kitchen_lamp", DType.BOOL)).execute({"house.command": "lights on"})
+
+    assert threads
+    assert set(threads) == {1}