nod-ai / sharktank

SHARK Inference Modeling and Serving
Apache License 2.0
7 stars 9 forks source link

[punet] Switch weight quantization to signed. #61

Closed stellaraccident closed 4 weeks ago

stellaraccident commented 4 weeks ago

This eliminates the use of mixed signedness mm/conv by shifting the weight zero point to use signed i8 values. Note that this only requires changing the brevitas import script as the rest of the tooling just responds to the type change.