v1.2.2 (2019-11-25)
- Fix
PositionEncoderinternal state that was shared with other instances on the same thread - Replace Boost.Python by pybind11
- Include a Python source distribution in the Docker images
v1.2.1 (2019-11-06)
- Avoid copying decoder states when possible to improve decoding performance (10% to 20% faster)
- Fix execution profiling on GPU (device was not synchronized before measuring the time)
- Include
Muloperation in profiling report - Add a Python 3 wheel in Ubuntu Docker images
v1.2.0 (2019-10-28)
- Accept Transformer models with custom number of layers and heads
--log-profilingclient option to profile ops execution
- Fix conversion error for models having 2 different weights with the same values
- Fix invalid MKL function override after a refactoring
- Add more information and context to several error messages
v1.1.0 (2019-10-18)
- New Docker images:
latest-ubuntu16-gpu,latest-ubuntu18,latest-ubuntu18-gpu - Support OpenNMT-tf Transformer models with shared embeddings
- Update to TensorRT 6
- Make OpenMP runtime configurable
- Reduce the size of models with shared weights on disk and in memory
- Shared words vocabulary is no longer duplicated on disk and in memory
- Improve performance of translation with a vocabulary map on GPU
- Statically link against Intel MKL
- Remove some implementation details from public headers
v1.0.1 (2019-10-08)
- Fix loading of newer OpenNMT-py models
- Promote FP16 to FP32 in model converter scripts
- Improve INT8 performance on CPU and GPU
- Improve performance on GPU by fusing the layer normalization operation
x * gamma + beta - Enable INT8 and INT16 computation on all platforms with Intel MKL 2019.5 and above
v1.0.0 (2019-09-23)
First stable release.