quic / aimet

AIMET is a library that provides advanced quantization and compression techniques for trained neural network models.
https://quic.github.io/aimet-pages/index.html
Other
2.15k stars 383 forks source link

Preliminary refactoring for export versioning #3483

Closed quic-kyunggeu closed 3 days ago

quic-kyunggeu commented 4 days ago

This is a preliminary refactoring code to keep the SW maintenance cost of export versioning more manageable.

Main Changes

Defined some interface classes/methods so that aimet v2 EncodingBase will become the common middle layer between AIMET quantizers and different versions of QNN encoding formats. This will free us from the curse of maintaining exponential combinations -- {v1, v2} x {0.6.1, 1.0.0, ...} -- of encoding import/export logics

image