BUG: "constant / ScalarVariable" is unsupported + output_types_preference is missing in InRange

[t-ikegami]

Describe the issue:

While defining a custom Op, I have encountered two misimplementations in pytensor.scalar. A fix proposal is atached. Tell me if pull-request is more appropriate. (Tested on PyTensor 2.34.0/Python 3.11 and PyTensor 2.35.1/Python 3.14.)

1. __rtruediv__ is missing in ScalarVariable

>>> import pytensor.tensor as T
>>> x = T.dscalar("x")
>>> sx = T.scalar_from_tensor(x)
>>> 1.0 / sx
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: unsupported operand type(s) for /: 'float' and 'ScalarVariable'

2. Initialization of superclass is missing in InRange

>>> import pytensor
>>> import pytensor.tensor as T
>>> import pytensor.scalar as S
>>> func = T.elemwise.Elemwise(S.inopenrange, name = "func")
>>> x = T.dscalar("x")
>>> y = func(x, -1.0, 1.0)
>>> f = pytensor.function([x], y)    # f itself is successfully compiled.
ERROR (pytensor.graph.rewriting.basic): SequentialGraphRewriter apply <pytensor.graph.rewriting.basic.EquilibriumGraphRewriter object at 0x121d3c8d0>
ERROR (pytensor.graph.rewriting.basic): Traceback:
ERROR (pytensor.graph.rewriting.basic): Traceback (most recent call last):
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/basic.py", line 289, in apply
    sub_prof = rewriter.apply(fgraph)
               ^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/basic.py", line 2384, in apply
    for node_rewriter in self.node_tracker.get_trackers(node.op):
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/basic.py", line 1139, in get_trackers
    [
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/basic.py", line 1142, in <listcomp>
    if op_pattern.match_parameters(op)
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/unify.py", line 427, in match_parameters
    if param.parameters and not param.match_parameters(sub_op):
                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/unify.py", line 429, in match_parameters
    elif getattr(op, key) != param:
         ^^^^^^^^^^^^^^^^
AttributeError: 'InRange' object has no attribute 'output_types_preference'

Patch:

diff --git a/pytensor/scalar/basic.py b/pytensor/scalar/basic.py
index 5102d5f60..b0680e51b 100644
--- a/pytensor/scalar/basic.py
+++ b/pytensor/scalar/basic.py
@@ -924,6 +924,9 @@ class _scalar_py_operators:
     def __truediv__(self, other):
         return true_div(self, other)

+    def __rtruediv__(self, other):
+        return true_div(other, self)
+
     def __floordiv__(self, other):
         return int_div(self, other)

@@ -1654,6 +1657,7 @@ class InRange(LogicalComparison):
     nin = 3

     def __init__(self, openlow, openhi):
+        super().__init__()
         self.openlow = openlow
         self.openhi = openhi

Reproducable code example:

# 1.
import pytensor.tensor as T
x = T.dscalar("x")
sx = T.scalar_from_tensor(x)
1.0 / sx

# 2.
import pytensor
import pytensor.tensor as T
import pytensor.scalar as S
func = T.elemwise.Elemwise(S.inopenrange, name = "func")
x = T.dscalar("x")
y = func(x, -1.0, 1.0)
f = pytensor.function([x], y)

Error message:

1.
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: unsupported operand type(s) for /: 'float' and 'ScalarVariable'


2.
ERROR (pytensor.graph.rewriting.basic): SequentialGraphRewriter apply <pytensor.graph.rewriting.basic.EquilibriumGraphRewriter object at 0x121d3c8d0>
ERROR (pytensor.graph.rewriting.basic): Traceback:
ERROR (pytensor.graph.rewriting.basic): Traceback (most recent call last):
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/basic.py", line 289, in apply
    sub_prof = rewriter.apply(fgraph)
               ^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/basic.py", line 2384, in apply
    for node_rewriter in self.node_tracker.get_trackers(node.op):
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/basic.py", line 1139, in get_trackers
    [
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/basic.py", line 1142, in <listcomp>
    if op_pattern.match_parameters(op)
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/unify.py", line 427, in match_parameters
    if param.parameters and not param.match_parameters(sub_op):
                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/ikegami/venv3.11/lib/python3.11/site-packages/pytensor/graph/rewriting/unify.py", line 429, in match_parameters
    elif getattr(op, key) != param:
         ^^^^^^^^^^^^^^^^
AttributeError: 'InRange' object has no attribute 'output_types_preference'

PyTensor version information:

PyTensor 2.34.0:Python 3.11:MacOS 26.0.1 and
PyTensor 2.35.1:Python 3.14:MacOS 26.0.1

Context for the issue:

No response

[ricardoV94]

First is a duplicate of #1097

Re InRange: I don't think we ever use that, and it's not part of the Numpy API, so we should just remove it. I see it's not calling super().make_node() to initialize the Op with the output_types_preference, which is no longer optional. It also doesn't have proper eq / hash implementation and the base class doesn't handle the openlow/openhi arguments.

[ricardoV94]

Sorry first is not a duplicate of #1097, I thought you were trying to mix tensor and scalar, but it was just a python float

[t-ikegami]

I’m grad if InRange operator is kept maintained in the ScalarVariable space. In my personal project, the Op is not directly exported to the TensorVariable space, but is used to define grad method of custom functions. In more specific, it is used, in conjunction with switch Op, to dig a tiny hole around a singular point, to avoid NaN (due to 0/0) situation.

[ricardoV94]

I’m grad if InRange operator is kept maintained in the ScalarVariable space. In my personal project, the Op is not directly exported to the TensorVariable space, but is used to define grad method of custom functions. In more specific, it is used, in conjunction with switch Op, to dig a tiny hole around a singular point, to avoid NaN (due to 0/0) situation.

You can use a regular expression to achieve the same x > low & x < high. No need for a specific Op. PyTensor is good at fusing small scalar operations in a single kernel anyway, so you end up with equally fast code without needing such a specialized Op.

You can put it in a helper function to keep code readable as well

[ricardoV94]

Performance is similar on the following test:

import numpy as np

import pytensor
import pytensor.tensor as pt
from pytensor.scalar import inopenrange
from pytensor.tensor.elemwise import Elemwise


inrange = Elemwise(inopenrange)

x = pt.vector("x")
low = pt.vector("low")
high = pt.vector("high")

out1 = inrange(x, low, high)
out2 = (x > low) & (x < high)
fn1 = pytensor.function([x, low, high], out1, trust_input=True)
fn2 = pytensor.function([x, low, high], out2, trust_input=True)

x_test = np.random.normal(size=(1000,))
low_test = np.random.normal(size=(1000,))
high_test = np.random.normal(size=(1000,)) + low_test

np.testing.assert_allclose(
    fn1(x_test, low_test, high_test),
    fn2(x_test, low_test, high_test),
)

%timeit fn1(x_test, low_test, high_test)  # 4.83 μs ± 534 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
%timeit fn2(x_test, low_test, high_test)  # 4.82 μs ± 762 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

I patched the __init__ method of InRange to run that snippet.

    def __init__(self, openlow, openhi):
        self.openlow = openlow
        self.openhi = openhi
        self.output_types_preference = None

[ricardoV94]

@t-ikegami Do you want to open a PR for the division dunder methods? We should add some simple test in tests/scalar/test_basic.py as well.

If not, someone else can probably take this issue.

[ricardoV94]

Unrelated @t-ikegami do you have a Dvorak keyboard. I just did the same type grad instead of glad you did :O

[t-ikegami]

Oops. I'm using Qwerty, but my fingers remember more grad than glad. And thank you for the benchmark code. I've confirmed it works in my project.

I'm glad if someone makes a PR for __rtruediv__ case. I just take a look at test_basic.py, and it is beyond my ability to set up the test case.