Improve TriangularMultiplicativeUpdate stability in fp16 mode #295
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Outdated suggestions cannot be applied.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 solution to (observed) fp16 overflow in TriangularMultiplicativeUpdate torch.matmul caused by hugh std of a*b elements. It produces the same output as it's followed by layer norm which performs equivalent std-normalization.
autocast=False
mode doesn't seem to solve this in pure fp16 mode unfortunately.by @adamlerer