The llama.cpp repository has merged support for exact GELU activation functions specifically for ModernBERT encoders. This update improves inference accuracy for the ModernBERT architecture in the ggml-org runtime.
HOW THIS AFFECTS YOU
●
builderYou can now deploy ModernBERT encoders with higher fidelity in llama.cpp environments.
●
researcherThe implementation ensures that inference results more closely match the mathematical specifications of the encoder.