CTranslate2 1.10.0

New features

• Coverage penalty as in Wu et al. 2016 with the option coverage_penalty
• Batch size can be expressed in number of tokens with the option batch_type
• Translation scores can be disabled with the option return_scores (if disabled, the final SoftMax is skipped during greedy decoding)
• Support compilation without TensorRT by setting -DWITH_TENSORRT=OFF during CMake configuration (in this case, beam search is no longer supported)
• Experimental integration of Intel MKL's packed GEMM which can be enabled by setting the environment variable CT2_USE_EXPERIMENTAL_PACKED_GEMM=1

Fixes and improvements

• Remove direct dependency to cuDNN (still an indirect dependency via TensorRT)
• Static AVX optimization for the ReLU operator
• Remove unnecessary memory initialization when creating temporary buffers
• Dissociate SoftMax and LogSoftMax in profiling report

CTranslate2 1.9.1

Fixes and improvements

• Fix parallel translations when calling Translator.translate_batch from multiple Python threads
• Fix crash on invalid num_hypotheses value

CTranslate2 1.9.0

New features

• Return 2 additional statistics from file translation APIs:
  • the number of translated examples
  • the total translation time in milliseconds

Fixes and improvements

• Fix exceptions that were not catched by the Python wrapper
• Fix an invalid insertion in the variables collection while iterating over it
• Optimize filling operation of float storages
• Internal refactoring of decoding functions to make them reusable for other tasks (e.g. generative language models)

CTranslate2 1.8.0

CTranslate2 1.8.0

New features

• [Python] Add methods Translator.unload_model and Translator.load_model to manually manage memory
• [Docker] Move all images to Python 3 only
• Expose options that enable an internal sorting by length to increase the translation efficiency:
  • for file translation: read_batch_size contiguous examples will be loaded, sorted by length, and batched with size max_batch_size
  • for batch translation: if the batch is larger than max_batch_size, examples will be sorted by length and batched with size max_batch_size

Fixes and improvements

• Fix another error when releasing a translator that is placed on a GPU that is not GPU 0
• Fix possible memory corruption when creating GPU translators in parallel
• Fix memory that is briefly allocated on GPU 0 when destroying a translator that is placed on another GPU
• Reduce latency of model loading, especially on GPU

CTranslate2 1.7.1

CTranslate2 1.7.1

Fixes and improvements

• Revert "Parallelize some low level transformations on CPU" which caused incorrect computation
• Avoid unnecessary TensorFlow runtime initialization when converting checkpoints
• Fix compilation without MKL

CTranslate2 1.7.0

CTranslate2 1.7.0

New features

• Translation option return_alternatives to return multiple choices at the first unconstrained decoding position: combined with a target prefix, this could be used to provide alternative words and translations at a specific location in the target
• Support Transformers with different number of encoder/decoder layers
• Allow compilation without OpenMP with -DOPENMP_RUNTIME=NONE

Fixes and improvements

• Fix SavedModel conversion when TensorFlow Addons 0.8 is installed
• Fix error when releasing a translator/model that is placed on a GPU that is not GPU 0
• Fix memory that was allocated on GPU 0 even when the translator/model was placed on another GPU
• Query GPU int8 support on the first model load, and then cache the result for future loads
• Avoid creating an empty model directory on conversion errors
• Parallelize some low level transformations on CPU
• Reduce memory usage when translating large files by limiting the work queue size

CTranslate2 1.6.3

CTranslate2 1.6.3

Fixes and improvements

• Fix incorrectness in relative representation computation

CTranslate2 1.6.2

CTranslate2 1.6.2

Fixes and improvements

• Fix conversion of models with shared embeddings

CTranslate2 1.6.1

CTranslate2 1.6.1

Fixes and improvements

• [Docker] Remove translation client in CentOS 7 images as it can cause compatibility issues with downstream images

CTranslate2 1.6.0

CTranslate2 1.6.0

New features

• Support Transformers with relative position representations (as in Shaw et al. 2018)
• Accept target prefix in batch request
• Support return_attention with prefixed translation