Fix integer overflow in quantization #129127
Open
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The
static_cast<int64_t>
can overflow for large float values and/or a small scale (e.g. 9.2e14 & 1e-4)Fix a similar issue in the mask calculation where
std::lrint
is used which may convert to a 32 bit float returning an implementation defined value on overflow.Stay in float mode using
std::round
andfmin/fmax
to avoid this.Fixes #111471
I actually fixed the CUDA code first and copied that.
I didn't touch the code duplication which can likely be removed by using a fitting
AT_DISPATCH
but for some reason the zero_point isint32_t
while the limits areint64_t
which to me doesn't make much sense and the actual type could always be used.After #113861 failed for some reason
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10