Repository navigation
feat: convert TFLite int8 CUMSUM at any input zero point - #126
Merged
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
TFLite's int8 CUMSUM seeds its sum with the scaled input zero point, while ONNX's CumSum sums dequantized values, so a non-zero zero point was refused. The frontend now emits the compiler's own CumSum for int8, which normalizes to CumSum with TFLite's semantics. The runtime already computes those semantics, so the plan format and runtime are unchanged. ONNX's own CumSum keeps its zero-point-0 rule, now covered by a test.
New int8 fixtures at input zero point -128, plain and exclusive-reverse, bit-exact against TFLite Micro.