Fix gradient bug in models with max operation

pymc/logprob/rewriting.py

@@ -365,8 +365,6 @@ def incsubtensor_rv_replace(fgraph, node):
     "local_exp_over_1_plus_exp", out2in(local_exp_over_1_plus_exp), "basic"
 )
 logprob_rewrites_db.register("pre-canonicalize", optdb.query("+canonicalize"), "basic")
-# Split max_and_argmax
-logprob_rewrites_db.register("local_max_and_argmax", out2in(local_max_and_argmax), "basic")
 
 # These rewrites convert un-measurable variables into their measurable forms,
 # but they need to be reapplied, because some of the measurable forms require
@@ -376,6 +374,12 @@ def incsubtensor_rv_replace(fgraph, node):
 
 logprob_rewrites_db.register("measurable_ir_rewrites", measurable_ir_rewrites_db, "basic")
 
+# Split max_and_argmax
+# We only register this in the measurable IR db because max does not have a grad implemented
+# And running this on any MaxAndArgmax would lead to issues: https://github.com/pymc-devs/pymc/issues/7251
+# This special registering can be removed after https://github.com/pymc-devs/pytensor/issues/334 is fixed
+measurable_ir_rewrites_db.register("local_max_and_argmax", local_max_and_argmax, "basic")
+
 # These rewrites push random/measurable variables "down", making them closer to
 # (or eventually) the graph outputs.  Often this is done by lifting other `Op`s
 # "up" through the random/measurable variables and into their inputs.

tests/logprob/test_order.py

@@ -45,6 +45,7 @@
 import pymc as pm
 
 from pymc import logp
+from pymc.logprob import conditional_logp
 from pymc.testing import assert_no_rvs
 
 
@@ -293,3 +294,18 @@ def test_min_max_bernoulli():
     min_logp_fn = pytensor.function([value], pm.logp(pt.min(x), value))
     np.testing.assert_allclose(min_logp_fn(1), np.log(p**n))
     np.testing.assert_allclose(min_logp_fn(0), np.log(1 - p**n))
+
+
+def test_non_measurable_max_grad():
+    # Regression test for https://github.com/pymc-devs/pytensor/issues/711
+    x = pt.random.normal(0, 1, size=(3,))
+    max_x = x.max()
+    y = pt.random.normal(max_x, 1)
+
+    x_vv = x.type()
+    y_vv = y.type()
+    logp_terms = conditional_logp({x: x_vv, y: y_vv}).values()
+    joint_logp = pt.sum([term.sum() for term in logp_terms])
+
+    # Test that calling gradient does not raise a NotImplementedError
+    assert pt.grad(joint_logp, x_vv)