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

Grouped GPTQ Bug Fix #2308

Closed Satrat closed 6 months ago

Satrat commented 6 months ago

The dimmensions of grouped scale/zp was changed from 3-dim to 2, but this change was never made in SparseML. Updating the GPTQ wrapper to expect grouped scale/zp as 2 dim instead of 3