Skip to content

Marian v1.5.0
Compare
Choose a tag to compare
@emjotde emjotde released this 17 Jun 22:12
· 4932 commits to master since this release
1.5.0
f429d4a

[1.5.0] - 2018-06-17

Added

  • Average Attention Networks for Transformer model
  • 16-bit matrix multiplication on CPU
  • Memoization for constant nodes for decoding
  • Autotuning for decoding

Fixed

  • GPU decoding optimizations, about 2x faster decoding of transformer models
Assets 2
Loading