Precompile fn in ZarrChain and borrow inputs

pymc/backends/zarr.py

@@ -11,7 +11,7 @@
 #   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 #   See the License for the specific language governing permissions and
 #   limitations under the License.
-from collections.abc import Mapping, MutableMapping, Sequence
+from collections.abc import Callable, Mapping, MutableMapping, Sequence
 from typing import Any
 
 import arviz as az
@@ -85,8 +85,9 @@ def __init__(
         vars: Sequence[TensorVariable] | None = None,
         test_point: dict[str, np.ndarray] | None = None,
         draws_per_chunk: int = 1,
+        fn: Callable | None = None,
     ):
-        super().__init__(name="zarr", model=model, vars=vars, test_point=test_point)
+        super().__init__(name="zarr", model=model, vars=vars, test_point=test_point, fn=fn)
         self._step_method: BlockedStep | CompoundStep | None = None
         self.unconstrained_variables = {
             var.name for var in self.vars if is_transformed_name(var.name)
@@ -442,7 +443,12 @@ def init_trace(
         )
         self.vars = [var for var in vars if var.name in self.varnames]
 
-        self.fn = model.compile_fn(self.vars, inputs=model.value_vars, on_unused_input="ignore")
+        self.fn = model.compile_fn(
+            self.vars,
+            inputs=model.value_vars,
+            on_unused_input="ignore",
+            borrow_vars=True,
+        )
 
         # Get variable shapes. Most backends will need this
         # information.
@@ -518,6 +524,7 @@ def init_trace(
                 test_point=test_point,
                 stats_bijection=StatsBijection(step.stats_dtypes),
                 draws_per_chunk=self.draws_per_chunk,
+                fn=self.fn,
             )
             for _ in range(chains)
         ]

pymc/model/core.py

@@ -1611,6 +1611,7 @@ def compile_fn(
         inputs: Sequence[Variable] | None = None,
         mode=None,
         point_fn: bool = True,
+        borrow_vars: bool = False,
         **kwargs,
     ) -> PointFunc | Function:
         """Compiles a PyTensor function.
@@ -1636,6 +1637,10 @@ def compile_fn(
         if inputs is None:
             inputs = inputvars(outs)
 
+        if borrow_vars:
+            inputs = [pytensor.In(v, borrow=True) for v in inputs]
+            outs = [pytensor.Out(v, borrow=True) for v in outs]
+
         with self:
             fn = compile(
                 inputs,
@@ -1645,6 +1650,8 @@ def compile_fn(
                 mode=mode,
                 **kwargs,
             )
+        if borrow_vars:
+            fn.trust_input = True
 
         if point_fn:
             return PointFunc(fn)

pymc/sampling/parallel.py

@@ -110,8 +110,10 @@ def __init__(
         zarr_chains: list[ZarrChain] | bytes | None = None,
         zarr_chains_is_pickled: bool = False,
     ):
-        # For some strange reason, spawn multiprocessing doesn't copy the rng
-        # seed sequence, so we have to rebuild it from scratch
+        # Because of https://github.com/numpy/numpy/issues/27727, we can't send
+        # the rng instance to the child process because pickling (copying) looses
+        # the seed sequence state information. For this reason, we send a
+        # RandomGeneratorState instead.
         rng = random_generator_from_state(rng_state)
         self._msg_pipe = msg_pipe
         self._step_method = step_method