Skip to content

Marian v1.12.0

Latest
Compare
Choose a tag to compare
@emjotde emjotde released this 21 Feb 18:20
· 67 commits to master since this release
65bf82f

[1.12.0] - 2023-02-20

Added

  • Fused inplace-dropout in FFN layer in Transformer
  • --force-decode option for marian-decoder
  • --output-sampling now works with ensembles (requires proper normalization via e.g --weights 0.5 0.5)
  • --valid-reset-all option

Fixed

  • Make concat factors not break old vector implementation
  • Use allocator in hashing
  • Read/restore checkpoints from main process only when training with MPI
  • Multi-loss casts type to first loss-type before accumulation (aborted before due to missing cast)
  • Throw ShapeSizeException if total expanded shape size exceeds numeric capacity of the maximum int value (2^31-1)
  • During mini-batch-fitting, catch ShapeSizeException and use another sizing hint. Aborts outside mini-batch-fitting.
  • Fix incorrect/missing gradient accumulation with delay > 1 or large effective batch size of biases of affine operations.
  • Fixed case augmentation with multi-threaded reading.
  • Scripts using PyYAML now use safe_load; see https://msg.pyyaml.org/load
  • Fixed check for fortran_ordering in cnpy
  • Fixed fp16 training/inference with factors-combine concat method
  • Fixed clang 13.0.1 compatibility
  • Fixed potential vulnerabilities from lxml<4.9.1 or mistune<2.0.31
  • Fixed the --best-deep RNN alias not setting the s2s model type

Changed

  • Parameter synchronization in local sharding model now executes hash checksum before syncing
  • Make guided-alignment faster via sparse memory layout, add alignment points for EOS, remove losses other than ce
  • Negative --workspace -N value allocates workspace as total available GPU memory minus N megabytes.
  • Set default parameters for cost-scaling to 8.f 10000 1.f 8.f, i.e. when scaling scale by 8 and do not try to automatically scale up or down. This seems most stable.
  • Make guided-alignment faster via sparse memory layout, add alignment points for EOS, remove losses other than ce.
  • Changed minimal C++ standard to C++-17
  • Faster LSH top-k search on CPU
  • Updated intgemm to the latest upstream version
  • Parameters in npz files are no longer implicitly assumed to be row-ordered. Non row-ordered parameters will result in an abort
  • Updated Catch2 header from 2.10.1 to 2.13.9