Fix non-deterministic NUTS initialization when using ADVI

pymc/sampling.py

@@ -2531,7 +2531,9 @@ def init_nuts(
             progressbar=progressbar,
             obj_optimizer=pm.adagrad_window,
         )
-        approx_sample = approx.sample(draws=chains, return_inferencedata=False)
+        approx_sample = approx.sample(
+            draws=chains, random_seed=random_seed_list[0], return_inferencedata=False
+        )
         initial_points = [approx_sample[i] for i in range(chains)]
         std_apoint = approx.std.eval()
         cov = std_apoint**2
@@ -2549,7 +2551,9 @@ def init_nuts(
             progressbar=progressbar,
             obj_optimizer=pm.adagrad_window,
         )
-        approx_sample = approx.sample(draws=chains, return_inferencedata=False)
+        approx_sample = approx.sample(
+            draws=chains, random_seed=random_seed_list[0], return_inferencedata=False
+        )
         initial_points = [approx_sample[i] for i in range(chains)]
         cov = approx.std.eval() ** 2
         potential = quadpotential.QuadPotentialDiag(cov)
@@ -2564,7 +2568,9 @@ def init_nuts(
             progressbar=progressbar,
             obj_optimizer=pm.adagrad_window,
         )
-        approx_sample = approx.sample(draws=chains, return_inferencedata=False)
+        approx_sample = approx.sample(
+            draws=chains, random_seed=random_seed_list[0], return_inferencedata=False
+        )
         initial_points = [approx_sample[i] for i in range(chains)]
         cov = approx.std.eval() ** 2
         potential = quadpotential.QuadPotentialDiag(cov)

pymc/tests/test_sampling.py

@@ -100,11 +100,6 @@ def test_random_seed(self, chains, seeds, cores, init):
         allequal = np.all(tr1["x"] == tr2["x"])
         if seeds is None:
             assert not allequal
-        # TODO: ADVI init methods are not correctly seeded, as they rely on the state of
-        #  the model RandomState/Generators which is updated in place when the function
-        #  is compiled and evaluated. This elif branch must be removed once this is fixed
-        elif init == "advi":
-            assert not allequal
         else:
             assert allequal
 

pymc/variational/opvi.py

@@ -57,11 +57,20 @@
 
 import pymc as pm
 
-from pymc.aesaraf import at_rng, compile_pymc, identity, rvs_to_value_vars
+from pymc.aesaraf import (
+    SeedSequenceSeed,
+    at_rng,
+    compile_pymc,
+    find_rng_nodes,
+    identity,
+    reseed_rngs,
+    rvs_to_value_vars,
+)
 from pymc.backends import NDArray
 from pymc.blocking import DictToArrayBijection
 from pymc.initial_point import make_initial_point_fn
 from pymc.model import modelcontext
+from pymc.sampling import RandomState, _get_seeds_per_chain
 from pymc.util import WithMemoization, locally_cachedmethod
 from pymc.variational.updates import adagrad_window
 from pymc.vartypes import discrete_types
@@ -1641,22 +1650,30 @@ def sample_dict_fn(self):
         sampled = [self.rslice(name) for name in names]
         sampled = self.set_size_and_deterministic(sampled, s, 0)
         sample_fn = compile_pymc([s], sampled)
+        rng_nodes = find_rng_nodes(sampled)
 
-        def inner(draws=100):
+        def inner(draws=100, *, random_seed: SeedSequenceSeed = None):
+            if random_seed is not None:
+                reseed_rngs(rng_nodes, random_seed)
             _samples = sample_fn(draws)
+
             return {v_: s_ for v_, s_ in zip(names, _samples)}
 
         return inner
 
-    def sample(self, draws=500, return_inferencedata=True, **kwargs):
+    def sample(
+        self, draws=500, *, random_seed: RandomState = None, return_inferencedata=True, **kwargs
+    ):
         """Draw samples from variational posterior.
 
         Parameters
         ----------
-        draws: `int`
+        draws : int
             Number of random samples.
-        return_inferencedata: `bool`
-            Return trace in Arviz format
+        random_seed : int, RandomState or Generator, optional
+            Seed for the random number generator.
+        return_inferencedata : bool
+            Return trace in Arviz format.
 
         Returns
         -------
@@ -1666,7 +1683,9 @@ def sample(self, draws=500, return_inferencedata=True, **kwargs):
         # TODO: add tests for include_transformed case
         kwargs["log_likelihood"] = False
 
-        samples = self.sample_dict_fn(draws)  # type: dict
+        if random_seed is not None:
+            (random_seed,) = _get_seeds_per_chain(random_seed, 1)
+        samples = self.sample_dict_fn(draws, random_seed=random_seed)  # type: dict
         points = ({name: records[i] for name, records in samples.items()} for i in range(draws))
 
         trace = NDArray(