[Fix] l2_exp random fail in half-float32 mixed precision on self-neighboring #596
Conversation
rhdong
added
bug
| @@ -21,6 +21,7 @@ | |||
| from cuvs.distance import pairwise_distance | |||
|
|
|||
|
|
|||
| @pytest.mark.parametrize("times", range(20)) | |||
There was a problem hiding this comment.
Whats this times parameter used for? I don't see it used in the test it self -
Are you just trying to run this test multiple times here to stress test it?
There was a problem hiding this comment.
Yeah, it's just for testing multiple times and to guarantee the reproducing on one going because the possibility is close to ~10% empirically.
| @@ -79,7 +80,5 @@ def test_distance(n_rows, n_cols, inplace, order, metric, dtype): | |||
| actual = output_device.copy_to_host() | |||
|
|
|||
| tol = 1e-3 | |||
| if np.issubdtype(dtype, np.float16): | |||
| tol = 1e-1 | |||
There was a problem hiding this comment.
I think I added this reduced tolerance because I was seeing failures - is this no longer needed?
There was a problem hiding this comment.
Yeah, I tried it successfully at local machines. I think this change can help us block potential actual failures in the future, so I made it.
|
Looks like Python tests are now passing. We are waiting on one more C++ test job. I will go ahead and trigger a merge once CI finishes to unblock CI, since several PRs depend on this. |
|
/merge |
No description provided.