Change8

b7605

📦 llama-cppView on GitHub →
5 features🐛 1 fixes🔧 4 symbols

Summary

This release adds support for JinaBertModel non-gated FFN variants and the German_Semantic_V3 tokenizer. It also improves model metadata handling by automatically detecting FFN variants from tensor dimensions.

Migration Steps

  1. Re-convert Jina BERT models using the updated conversion scripts to include new feed_forward_type metadata if automatic detection is not used.

✨ New Features

  • Added support for JinaBertModel with non-gated Feed-Forward Network (FFN) variants.
  • Added jina-v2-de tokenizer variant support for German_Semantic_V3.
  • Introduced feed_forward_type metadata for jina-bert-v2 models.
  • Implemented automatic FFN variant detection based on tensor dimensions.
  • Added support for standard GELU FFN variant in jina-bert-v2.

🐛 Bug Fixes

  • Fixed token collision issues in BERT phantom vocabulary conversion.

🔧 Affected Symbols

JinaBertModelllama-model.cppsrc/models/bert.cppjina-v2-de