Skip to content

CTranslate2 2.4.0
Compare
Choose a tag to compare
@guillaumekln guillaumekln released this 10 Sep 09:13
· 806 commits to master since this release
v2.4.0
66f9831
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

New features

  • [Python] Support asynchronous translation: translate_batch can return future-like objects with argument asynchronous=True
  • [Python] translate_batch now returns a list of TranslationResult objects instead of a list of dictionaries (this object can also be indexed as a list of dictionaries for backward compatibility)
  • Add options --source_lang and --target_lang to the Fairseq converter for models that do not include these information

Fixes and improvements

  • Fix Fairseq model conversion when the model options are stored in model["cfg"]["model"]
  • Compile the CPU INT8 quantization kernel with FMA instructions
  • Enable packing of the last linear weight when not using dynamic vocabulary reduction
  • Replace the generic Tile implementation by dedicated CPU and CUDA kernels
  • [Python] Implement __repr__ method for TranslationStats objects
  • [Python] Update pybind11 to 2.7.1
Assets 2
Loading