New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We鈥檒l occasionally send you account related emails.
Already on GitHub? Sign in to your account
[chalf] reference_testing: low quality test for fast growing ops #78170
Comments
We should have these problems already with |
Reference test against NumPy doesn't run on Lines 331 to 332 in 141238a
and for UnaryUfuncs we use values based on dtype pytorch/test/test_unary_ufuncs.py Lines 83 to 89 in 141238a
However in this case, since we are testing all types of functions. Also, te value range of Tensors generated by make_tensor is conservative. So this an issue only for selected functions like |
Does this issue arise because we go back and forth between cfloat and chalf? If yes, then BFloat16 should also have a similar issue right?
It seems like the permissible range should also depend on the type on function right? |
It happens because of the difference in the range of values supported by
Yes, the permissible range is dependent on function. I think only few functions |
馃悰 Describe the bug
In PR #77640:
Since range of chalf is much less compared to cfloat, we get
inf
s easily (eg. withpow
,exp
), so we castcfloat
back tochalf
.However, this is might mask an actual issue as we don't control the percent of input that will be valid. The correct approach would be to sample input which are valid given the range of
chalf
.One of the approach would be to add extra meta-data to OpInfo.
cc: @ngimel @mruberry @anjali411
Versions
master
cc @ezyang @anjali411 @dylanbespalko @mruberry @lezcano @nikitaved
The text was updated successfully, but these errors were encountered: