Skip to content

CTranslate2 2.14.0
Compare
Choose a tag to compare
@guillaumekln guillaumekln released this 16 Mar 10:29
· 621 commits to master since this release
v2.14.0
3e2147b

New features

  • Include BART and MBART in the list of supported Fairseq architectures
  • Add Fairseq converter option --no_default_special_tokens to require all special tokens to be set by the user during inference, including the decoder start tokens (for example, this is required by MBART-25 to properly set the language tokens)

Fixes and improvements

  • Fix conversion of Post-Norm Transformers trained with OpenNMT-tf
  • Fix scoring with Fairseq models that used an incorrect decoder start token (Fairseq uses </s> as the decoder start token, not <s>)
  • Fix scoring result to include the end of sentence token
  • Ignore OpenNMT-py options --alignment_layer and --alignment_heads for models that are not trained with alignments
  • Enable batch encoding in return_alternatives translation mode (the decoding still runs sequentially)
  • Make enumerations ctranslate2.specs.Activation and ctranslate2.specs.EmbeddingsMerge public since they could be used to configure the Transformer specification
  • Update oneDNN to 2.5.3
  • Update cpu_features to 0.7.0
  • Update cxxopts to 3.0.0
  • Update spdlog to 1.9.2
Assets 2
Loading