search

neuralmagic / sparseml

Libraries for applying sparsification recipes to neural networks with a few lines of code, enabling faster and smaller models
Apache License 2.0
2.07k stars 148 forks source link

[GTPQ] fix slice of scale/zp for group_size #2290

Closed bfineran closed 6 months ago

bfineran commented 6 months ago

fixes the input dim value to be relative to the entire weight, not just the W1 slice which is incorrect (resets every blocksize slice)