NetEase-FuXi / EETQ

Easy and Efficient Quantization for Transformers
Apache License 2.0
180 stars 14 forks source link

FIX: Use `matmul` instead of `mm` in `backward` #16

Closed younesbelkada closed 7 months ago

younesbelkada commented 7 months ago

Fixes the current PEFT integration with EETQ ! Using mm leads to surprising shape mismatch errors (as I think this does not support broadcasting), using matmul resolves it as it enables broadcasting

cc @dingjingzhen