BUG: Raise TypeError for non-integer float observed data on discrete distributions

pymc/model/core.py

@@ -1320,7 +1320,36 @@ def make_obs_var(
         TensorVariable
         """
         name = rv_var.name
-        data = convert_observed_data(data).astype(rv_var.dtype)
+        data = convert_observed_data(data)
+
+        # Check if float observed data passed to an integer-dtype distribution
+        # would lose information when cast (e.g. Binomial, Poisson).
+        # Only applies to concrete numpy arrays with float dtype going to integer dtype.
+        # Float arrays containing exact integer values (e.g. [0.0, 1.0]) are allowed.
+        rv_dtype = np.dtype(rv_var.dtype)
+        data_dtype = getattr(data, "dtype", None)
+        if (
+            data_dtype is not None
+            and not isinstance(data, Variable)
+            and np.issubdtype(rv_dtype, np.integer)
+            and np.issubdtype(data_dtype, np.floating)
+        ):
+            if isinstance(data, np.ma.MaskedArray):
+                check_data = data.compressed()
+            elif sparse.issparse(data):
+                check_data = data.data
+            else:
+                check_data = np.asarray(data)
+            cast_check = check_data.astype(rv_dtype)
+            if not np.array_equal(check_data, cast_check):
+                raise TypeError(
+                    f"Observed data for '{name}' has dtype {data_dtype} with non-integer values "
+                    f"that cannot be safely cast to the expected dtype {rv_dtype} of the "
+                    f"{rv_var.owner.op} distribution. If the cast is intentional, convert the "
+                    f"data explicitly before passing it as observed."
+                )
+
+        data = data.astype(rv_var.dtype)
 
         if data.ndim != rv_var.ndim:
             raise ShapeError(

tests/model/test_core.py

@@ -687,6 +687,45 @@ def test_make_obs_var():
     del fake_model.named_vars[fake_distribution.name]
 
 
+@pytest.mark.parametrize(
+    "dist_cls,dist_kwargs",
+    [
+        (pm.Binomial, {"n": 5, "p": 0.5}),
+        (pm.Poisson, {"mu": 3.0}),
+    ],
+)
+def test_discrete_float_observed_raises(dist_cls, dist_kwargs):
+    """Non-integer float observed data passed to discrete distributions should raise.
+
+    Regression test for https://github.com/pymc-devs/pymc/issues/8282.
+    """
+    with pm.Model():
+        with pytest.raises(TypeError, match="non-integer values"):
+            dist_cls("bad", observed=np.array([0.11, 1.5, 2.89], dtype=np.float32), **dist_kwargs)
+
+
+def test_discrete_float_observed_exact_ints_allowed():
+    """Float arrays containing exact integer values should be accepted by discrete dists.
+
+    Regression test for https://github.com/pymc-devs/pymc/issues/8282.
+    """
+    with pm.Model():
+        # Common pattern: floatX zeros/ones passed to Bernoulli
+        pm.Bernoulli("b", p=0.5, observed=np.array([0.0, 1.0, 0.0], dtype=np.float64))
+        # Integer-valued data in native int dtype
+        pm.Binomial("x", n=5, p=0.5, observed=np.array([0, 1, 2], dtype=np.int64))
+
+
+def test_discrete_float_observed_scalar():
+    """Scalar non-integer float to discrete should raise.
+
+    Regression test for https://github.com/pymc-devs/pymc/issues/8282.
+    """
+    with pm.Model():
+        with pytest.raises(TypeError, match="non-integer values"):
+            pm.Poisson("bad", mu=3.0, observed=np.float64(1.5))
+
+
 def test_initial_point():
     with pm.Model() as model:
         a = pm.Uniform("a")