automl / RNAformer

Scalable Deep Learning for RNA Secondary Structure Prediction
Apache License 2.0
27 stars 9 forks source link

Fix flash attention bug and add cpu-only inference and evaluation to get the same results. #5

Closed ivelet closed 5 months ago

ivelet commented 5 months ago

Fixed attention implementation to get the same performance with flash-attn=False/True.

Automatically checks for flash-attn package and also device to be able to run inference and evaluation without a GPU or flash attention installed.