Skip to content

Conversation

@iksnagreb
Copy link
Contributor

Note: Keep the name indicating only handling signed-bias but adjust the datatype inference after absorbing the bias to handle signed bias as well as outputs.

Draft PR for now, related to work on streamlining for more generalized activation function support via multi-thresholds, but this could be of interest in any situation where it is easier to move some bias (not just the exported sign-bias but any scalar Add) upwards instead of downwards through the graph, e.g., if there is some joining Mul or MatMul operation blocking the usual flow of streamlining transformation.

Note: Keep the name indicating only handling signed-bias but adjust the
datatype inference after absorbing the bias to handle signed bias as
well as outputs.
Adds more (proably too much) code documentation and more usefule
warnings instead of a single hard assertion failure...
@iksnagreb iksnagreb marked this pull request as ready for review January 31, 2025 11:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant