Hey @jaiswalshash, this issue is about adding four numerical constants to the torch namespace, since this is required by the array API. I'll explain a little, how I would go about it (that is, you can go about this totally different as long as the result is achieved).

@gxyd wanted to work on this so I assigned him. Maybe he can give us a status update?

First, we can write a test to make sure that whatever we implement behaves actually like we want it. For example we might add a test case to test/test_torch.py:

import math
import torch

class TestConstants(TestCase):
    def test_e(self):
        self.assertIsInstance(torch.e, float)
        torch.testing.assert_close(torch.e, math.e)

    def test_pi(self):
         ...

Note that these tests will fail until we actually implement the constants.

To actually have the constants under the torch namespace, we have two options:

Implement them directly in torch/__init__.py or
implement them somewhere else and import them in torch/__init__.py.

I would prefer option 2, but since there is already a lot of functionality in the __init__.py, option 1 might also be viable. For my example I'm going to use option 2.

Since #58739 (comment) clarified that the constants should be python float's, the easiest solution is to simply import them from the builtin math library. For example we might add a torch/_constants.py module with

from math import e, inf, nan, pi

__all__ = ["e", "inf", "nan", "pi"]

and import the constants in torch/__init__.py with

from ._constants import *

After this our tests should pass and you are ready to send your PR. Although it is quite a read, I recommend https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md if you don't know how to do that.

Python Array API: support constants #58739

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions