Strange test_linspace assertion errors in jax

[simonetgordon]

Hi there,

test_linspace is failing for the jax backend for ivy, and hypothesis picks up two distinct failures:

1. AssertionError: out[0]=ivy.array(nan), but should be 9.9792015476736e+291 [linspace(9.9792015476736e+291, -1.7976931348623157e+308, 2)]

• raised on line
• Falsifying example: linspace(9.9792015476736e+291, -1.7976931348623157e+308, num =2, dtype= 'float64', endpoint=False)
• The output for the same inputs in numpy is: ivy.array([[ nan, -inf]]) (i.e. out[0] = nan, the same as in jax)

2. AssertionError: out[0]=ivy.array(9.97920155e+291), should be ivy.array(nan) [linspace()]

• raised on line: ph.assert_array_elements("linspace", out, expected)
• Falsifying example: linspace(9.9792015476736e+291, -1.7976931348623157e+308, num =1, dtype= 'float64', endpoint=False)
• The output for the same inputs in numpy is: ivy.array([[ nan]]) (i.e. out[0] = nan, different from the jax output)

I went about fixing the second failure as jax was diverging from numpy. However, I then (locally) got:
AssertionError: out[0]=ivy.array(nan), but should be 9.9792015476736e+291 [linspace(9.9792015476736e+291, -1.7976931348623157e+308, 1)].

Summary: the expected value for linspace seems to vary from the numpy value when start=9.9792015476736e+291 and stop= -1.7976931348623157e+308 (close to float64 min value) with num=1/2, causing the test to fail for jax.

I can't reproduce these inputs for numpy test_linspace, so it is unclear if it would also fail for numpy or any other backend - jax is the only one that has raised this issue.

Reference stack trace from ivy CI (click Run Array API Tests'

[asmeurer]

This looks related to #141. Can you confirm whether you're using a version of the suite that includes that change? If so, perhaps the restriction there needs to be tightened.

[simonetgordon]

@asmeurer in fact I'm not, I will update the submodule and get back to you. Thanks!

[simonetgordon]

@asmeurer the update seems to have fixed the linspace issue - no failures in the first few commits. Thanks!