avoids setting jax tracer as lazy property attribute (#1843)

* remove tracer as attribute of truncated dist

* streamline test

* fix CI test run failure

* Update test name

Co-authored-by: Dylan H. Morris <[email protected]>

* Move test from test_distributions.py to test_distributions_util.py

* Direct tests of tracer leaks

* pre-commit changes

---------

Co-authored-by: Dylan H. Morris <[email protected]>
Co-authored-by: Dylan H. Morris <[email protected]>

numpyro/distributions/util.py

@@ -14,6 +14,8 @@
 from jax.scipy.linalg import solve_triangular
 from jax.scipy.special import digamma
 
+from numpyro.util import not_jax_tracer
+
 # Parameters for Transformed Rejection with Squeeze (TRS) algorithm - page 3.
 _tr_params = namedtuple(
     "tr_params", ["c", "b", "a", "alpha", "u_r", "v_r", "m", "log_p", "log1_p", "log_h"]
@@ -692,7 +694,8 @@ def __get__(self, instance, obj_type=None):
         if instance is None:
             return self
         value = self.wrapped(instance)
-        setattr(instance, self.wrapped.__name__, value)
+        if not_jax_tracer(value):
+            setattr(instance, self.wrapped.__name__, value)
         return value
 
 

test/test_distributions_util.py

@@ -8,10 +8,12 @@
 import pytest
 import scipy
 
+import jax
 from jax import lax, random, vmap
 import jax.numpy as jnp
 from jax.scipy.special import expit, xlog1py, xlogy
 
+import numpyro.distributions as dist
 from numpyro.distributions.util import (
     add_diag,
     binary_cross_entropy_with_logits,
@@ -182,3 +184,45 @@ def test_add_diag(matrix_shape: tuple, diag_shape: tuple) -> None:
     expected = matrix + diag[..., None] * jnp.eye(matrix.shape[-1])
     actual = add_diag(matrix, diag)
     np.testing.assert_allclose(actual, expected)
+
+
+@pytest.mark.parametrize(
+    "my_dist",
+    [
+        dist.TruncatedNormal(low=-1.0, high=2.0),
+        dist.TruncatedCauchy(low=-5, high=10),
+        dist.TruncatedDistribution(dist.StudentT(3), low=1.5),
+    ],
+)
+def test_no_tracer_leak_at_lazy_property_log_prob(my_dist):
+    """
+    Tests that truncated distributions, which use @lazy_property
+    values in their log_prob() methods, do not
+    have tracer leakage when log_prob() is called.
+    Reference: https://github.com/pyro-ppl/numpyro/issues/1836, and
+    https://github.com/CDCgov/multisignal-epi-inference/issues/282
+    """
+    jit_lp = jax.jit(my_dist.log_prob)
+    with jax.check_tracer_leaks():
+        jit_lp(1.0)
+
+
+@pytest.mark.parametrize(
+    "my_dist",
+    [
+        dist.TruncatedNormal(low=-1.0, high=2.0),
+        dist.TruncatedCauchy(low=-5, high=10),
+        dist.TruncatedDistribution(dist.StudentT(3), low=1.5),
+    ],
+)
+def test_no_tracer_leak_at_lazy_property_sample(my_dist):
+    """
+    Tests that truncated distributions, which use @lazy_property
+    values in their sample() methods, do not
+    have tracer leakage when sample() is called.
+    Reference: https://github.com/pyro-ppl/numpyro/issues/1836, and
+    https://github.com/CDCgov/multisignal-epi-inference/issues/282
+    """
+    jit_sample = jax.jit(my_dist.sample)
+    with jax.check_tracer_leaks():
+        jit_sample(jax.random.key(5))